1 00:00:00,080 --> 00:00:02,397 The following is a conversation with Yan 2 00:00:02,480 --> 00:00:05,517 Lun, his third time on this podcast. He 3 00:00:05,600 --> 00:00:07,917 is the chief AI scientist at Meta, 4 00:00:08,000 --> 00:00:11,197 professor at NYU, touring award winner, 5 00:00:11,280 --> 00:00:13,517 and one of the seinal figures in the 6 00:00:13,600 --> 00:00:16,237 history of artificial intelligence. He 7 00:00:16,320 --> 00:00:19,757 and Meta AI have been big proponents of 8 00:00:19,840 --> 00:00:22,317 open sourcing AI development and have 9 00:00:22,400 --> 00:00:24,717 been walking the walk by open sourcing 10 00:00:24,800 --> 00:00:26,797 many of their biggest models including 11 00:00:26,880 --> 00:00:31,517 Llama 2 and eventually Llama 3. Also, 12 00:00:31,600 --> 00:00:34,477 Yan has been an outspoken critic of 13 00:00:34,560 --> 00:00:37,117 those people in the AI community who 14 00:00:37,200 --> 00:00:39,197 warned about the looming danger and 15 00:00:39,280 --> 00:00:42,797 existential threat of AGI. He believes 16 00:00:42,880 --> 00:00:46,797 the AGI will be created one day, but it 17 00:00:46,880 --> 00:00:49,037 will be good. It will not escape human 18 00:00:49,120 --> 00:00:52,717 control, nor will it dominate and kill 19 00:00:52,800 --> 00:00:55,837 all humans. At this moment of rapid AI 20 00:00:55,920 --> 00:00:59,197 development, this happens to be somewhat 21 00:00:59,280 --> 00:01:01,757 a controversial position. And so, it's 22 00:01:01,840 --> 00:01:06,557 been fun seeing Yan get into a lot of intense and fascinating discussions 23 00:01:06,640 --> 00:01:10,797 online, as we do in this very conversation. And 24 00:01:10,880 --> 00:01:14,397 now a quick few second mention of each sponsor. Check them out in the 25 00:01:14,480 --> 00:01:16,237 description. It's the best way to 26 00:01:16,320 --> 00:01:18,637 support this podcast. We got Hidden 27 00:01:18,720 --> 00:01:21,997 Layer for securing your AI models, 28 00:01:22,080 --> 00:01:25,197 Element for electrolytes, Shopify for 29 00:01:25,280 --> 00:01:29,437 shopping for stuff online, and AG1 for 30 00:01:29,520 --> 00:01:31,597 delicious health. Choose wisely, my 31 00:01:31,680 --> 00:01:34,317 friends. Also, if you want to get in 32 00:01:34,400 --> 00:01:37,357 touch with me for whatever reason, maybe 33 00:01:37,440 --> 00:01:40,317 to uh work with our amazing team, go to 34 00:01:40,400 --> 00:01:42,397 lexfreman.com/cont. 35 00:01:42,480 --> 00:01:47,437 And now on to the full ladies and never ending ads in the middle. I try to make 36 00:01:47,520 --> 00:01:49,277 these interesting. I don't know why I'm 37 00:01:49,360 --> 00:01:52,477 talking like this, but I am. There's a 38 00:01:52,560 --> 00:01:56,717 staccato nature to it. Speaking of staccato, I've been playing a bit of 39 00:01:56,800 --> 00:02:01,437 piano. Anyway, if you skip these ads, please still check out the sponsors. We 40 00:02:01,520 --> 00:02:03,517 love them. I love them. I enjoy their 41 00:02:03,600 --> 00:02:07,037 stuff. Maybe you will too. 42 00:02:07,120 --> 00:02:11,357 This episode is brought to you by a on 43 00:02:11,440 --> 00:02:15,037 theme in context. See what I did there? 44 00:02:15,120 --> 00:02:18,557 Uh sponsor since this is Yan Lun 45 00:02:18,640 --> 00:02:21,917 artificial intelligence machine learning 46 00:02:22,000 --> 00:02:24,477 one of the seminal figures in the field. 47 00:02:24,560 --> 00:02:27,517 So of course you got to have a sponsor that's related to artificial 48 00:02:27,600 --> 00:02:30,557 intelligence. hidden layer. They provide 49 00:02:30,640 --> 00:02:33,517 a platform that keeps your machine 50 00:02:33,600 --> 00:02:36,077 learning models secure. 51 00:02:36,160 --> 00:02:39,357 the ways to attack machine learning 52 00:02:39,440 --> 00:02:43,277 models, large language models, all the stuff we talk about with YAN. There's a 53 00:02:43,360 --> 00:02:45,197 lot of really fascinating work, not just 54 00:02:45,280 --> 00:02:48,237 large language models, but the same for 55 00:02:48,320 --> 00:02:50,877 video video prediction, tokenization 56 00:02:50,960 --> 00:02:53,277 where the tokens are in the space of 57 00:02:53,360 --> 00:02:56,317 concept versus the space of literally 58 00:02:56,400 --> 00:02:59,837 letters, symbols, JAPA, VJAPA, all of 59 00:02:59,920 --> 00:03:03,357 that stuff that they're open sourcing, all the stuff they're publishing on just 60 00:03:03,440 --> 00:03:06,157 really incredible. But that said, all of 61 00:03:06,240 --> 00:03:09,837 those models have security holes in ways 62 00:03:09,920 --> 00:03:14,637 that we can't even anticipate or imagine at this time. And so you want good 63 00:03:14,720 --> 00:03:16,637 people to be trying to find those 64 00:03:16,720 --> 00:03:18,797 security holes, trying to be one step 65 00:03:18,880 --> 00:03:23,837 ahead of uh the people that trying to attack. So if you're especially a 66 00:03:23,920 --> 00:03:25,437 company that's relying on these models, 67 00:03:25,520 --> 00:03:28,237 you need to uh have a person who's in 68 00:03:28,320 --> 00:03:31,117 charge of saying yeah this model that 69 00:03:31,200 --> 00:03:34,237 you got from this place has been tested 70 00:03:34,320 --> 00:03:36,557 has been secured whether that place is 71 00:03:36,640 --> 00:03:38,557 hugging face or any other kind of stuff 72 00:03:38,640 --> 00:03:41,357 uh or any other kind of repository or 73 00:03:41,440 --> 00:03:45,517 model zoo kind of place. I think the 74 00:03:45,600 --> 00:03:47,837 more and more we rely on larger language 75 00:03:47,920 --> 00:03:50,317 models or just AI systems in general, 76 00:03:50,400 --> 00:03:53,997 the more the security threats that are 77 00:03:54,080 --> 00:03:57,037 always going to be there uh become 78 00:03:57,120 --> 00:04:01,037 dangerous and impactful. So protect protect your models. Visit 79 00:04:01,120 --> 00:04:04,477 hiddenlayer.com/lex to learn more about how hidden layer can 80 00:04:04,560 --> 00:04:08,957 accelerate your AI adoption in a secure way. 81 00:04:09,040 --> 00:04:10,717 This episode is also brought to you by 82 00:04:10,800 --> 00:04:14,797 Element, a thing I drink throughout the 83 00:04:14,880 --> 00:04:17,517 day. I'm drinking now. When I am on a 84 00:04:17,600 --> 00:04:20,557 podcast, you'll sometimes see me with a 85 00:04:20,640 --> 00:04:24,317 mug and clear liquid in there that looks 86 00:04:24,400 --> 00:04:27,837 like water. In fact, it is not simply 87 00:04:27,920 --> 00:04:31,037 water. It is water mixed with element, 88 00:04:31,120 --> 00:04:34,957 watermelon salt, cold. What I do is I 89 00:04:35,040 --> 00:04:39,517 take one of them Power Raid 28 fluid 90 00:04:39,600 --> 00:04:43,117 ounces bottles, fill it up with water, 91 00:04:43,200 --> 00:04:45,757 one packet of watermelon salt, shake it 92 00:04:45,840 --> 00:04:48,877 up, put in the fridge, that's it. I 93 00:04:48,960 --> 00:04:51,277 reuse the bottles and drink from a mug 94 00:04:51,360 --> 00:04:56,637 or sometimes from the bottle. Either way, delicious, good for you. Especially 95 00:04:56,720 --> 00:04:58,157 if you're doing fasting, especially if 96 00:04:58,240 --> 00:05:00,797 you're doing low carb kinds of diets, 97 00:05:00,880 --> 00:05:03,677 which I do, you can get a sample pack 98 00:05:03,760 --> 00:05:09,400 for free with any purchase. Try to drinklement.com/lex. 99 00:05:10,160 --> 00:05:12,317 This episode is brought to you by 100 00:05:12,400 --> 00:05:14,957 Shopify as I take a drink of Element. It 101 00:05:15,040 --> 00:05:18,317 is a platform designed for anyone, even 102 00:05:18,400 --> 00:05:21,997 me, to sell stuff anywhere on a greatl 103 00:05:22,080 --> 00:05:24,797 looking store. I use a basic one. I like 104 00:05:24,880 --> 00:05:29,117 a really minimalist one. You can check it out if you go to lexreeman.com/store. 105 00:05:29,200 --> 00:05:31,357 There's a few shirts on there if that's 106 00:05:31,440 --> 00:05:35,277 your thing. It was so easy to set up. I 107 00:05:35,360 --> 00:05:37,037 imagine there's like a million features 108 00:05:37,120 --> 00:05:39,277 they have that can make it look better 109 00:05:39,360 --> 00:05:42,477 and all kinds of extra stuff you can do 110 00:05:42,560 --> 00:05:44,397 with the store. But I use the basic 111 00:05:44,480 --> 00:05:46,717 thing and the basic thing is pretty damn 112 00:05:46,800 --> 00:05:49,837 good. I like basic. I like minimalism. 113 00:05:49,920 --> 00:05:51,517 and they integrate with a lot of 114 00:05:51,600 --> 00:05:54,317 thirdparty apps, including what I use, 115 00:05:54,400 --> 00:05:58,557 which is um on demand printing. 116 00:05:58,640 --> 00:06:02,957 So, like you buy the shirt on Shopify, but it gets printed and shipped by 117 00:06:03,040 --> 00:06:06,477 another company that I always keep forgetting, but I think it's called 118 00:06:06,560 --> 00:06:10,877 Printull or Printify or I think it's Printull. 119 00:06:10,960 --> 00:06:15,277 I'm not sure. Doesn't matter. I think there's several integrations. You can 120 00:06:15,360 --> 00:06:17,677 check it out yourself. For me, it works. 121 00:06:17,760 --> 00:06:22,477 I'm using the most popular one, Printful, I think it's called. Anyway, I 122 00:06:22,560 --> 00:06:24,237 look forward to your letters correcting 123 00:06:24,320 --> 00:06:26,797 me on my pronunciation. 124 00:06:26,880 --> 00:06:30,477 Shopify is great. I'm a big fan of the 125 00:06:30,560 --> 00:06:34,477 good side of the machinery of uh capitalism, 126 00:06:34,560 --> 00:06:37,517 selling stuff on the internet, connecting people to the thing that they 127 00:06:37,600 --> 00:06:40,397 want, or rather the thing that would uh 128 00:06:40,480 --> 00:06:42,957 make their life better. 129 00:06:43,040 --> 00:06:46,077 both in advertisement and e-commerce and 130 00:06:46,160 --> 00:06:50,637 just shopping in general. I'm a big believer when that's done well, your 131 00:06:50,720 --> 00:06:52,717 life legitimately in the long term 132 00:06:52,800 --> 00:06:55,117 becomes better. And so whatever system 133 00:06:55,200 --> 00:06:58,637 can connect one human to the thing that 134 00:06:58,720 --> 00:07:00,877 makes their life better 135 00:07:00,960 --> 00:07:04,957 is great. And I believe that Shopify is 136 00:07:05,040 --> 00:07:10,157 sort of a platform that enables that kind of system. You can sign up for a $1 137 00:07:10,240 --> 00:07:13,037 per month trial period at shopify.com/lex. 138 00:07:13,120 --> 00:07:17,277 That's all lowercase. Go to shopify.com/lex to take your business to 139 00:07:17,360 --> 00:07:21,437 the next level today. This episode is also brought to you by 140 00:07:21,520 --> 00:07:23,437 AG1, 141 00:07:23,520 --> 00:07:26,877 an all-in-one daily drink to support 142 00:07:26,960 --> 00:07:29,517 better health and peak performance. It 143 00:07:29,600 --> 00:07:31,757 is delicious. 144 00:07:31,840 --> 00:07:34,237 It is nutritious. And I ran out of words 145 00:07:34,320 --> 00:07:38,077 that rhyme with those two. Actually, let 146 00:07:38,160 --> 00:07:41,837 me use a large language model to figure 147 00:07:41,920 --> 00:07:45,037 out what rhymes with delicious. 148 00:07:45,120 --> 00:07:49,277 Words that rhyme with delicious include ambitious, auspicious, capricious, 149 00:07:49,360 --> 00:07:52,637 fictitious, suspicious. 150 00:07:52,720 --> 00:07:55,677 So, there there you have it. Anyway, I 151 00:07:55,760 --> 00:08:00,237 drink it uh twice a day. Also put it in the fridge 152 00:08:00,320 --> 00:08:03,997 and sometimes in the freezer like it gets a little a little bit frozen. Just 153 00:08:04,080 --> 00:08:08,317 like a little bit, just a little bit frozen. You got that like slushy 154 00:08:08,400 --> 00:08:10,957 consistency. I'll do that too sometimes. 155 00:08:11,040 --> 00:08:14,957 It's freaking delicious. It's delicious no matter what. It's delicious. Warm is 156 00:08:15,040 --> 00:08:19,037 delicious cold. It's delicious slightly frozen. 157 00:08:19,120 --> 00:08:22,157 All of is just incredible. And of course, it covers like the basic 158 00:08:22,240 --> 00:08:24,397 multivitamin 159 00:08:24,480 --> 00:08:29,837 foundation of of a what I think of as a good diet. So, it's just a great 160 00:08:29,920 --> 00:08:33,597 multivitamin. That's the way I think about it. So, all the crazy stuff I do, 161 00:08:33,680 --> 00:08:37,437 the physical challenges, the mental challenges, all of that. At least I got 162 00:08:37,520 --> 00:08:40,557 AG1. They'll give you one month supply 163 00:08:40,640 --> 00:08:45,160 of fish oil when you sign up at drinkagg1.com/llex. 164 00:08:46,000 --> 00:08:49,277 This is the Lex Riverment podcast. To support it, please check out our 165 00:08:49,360 --> 00:08:51,597 sponsors in the description. And now, 166 00:08:51,680 --> 00:08:56,760 dear friends, here's Yan Lun. 167 00:09:10,800 --> 00:09:14,077 You've had some strong statements, 168 00:09:14,160 --> 00:09:18,317 technical statements about the future of artificial intelligence recently, 169 00:09:18,400 --> 00:09:20,077 throughout your career actually, but 170 00:09:20,160 --> 00:09:22,877 recently as well. uh you've said that 171 00:09:22,960 --> 00:09:26,957 auto reagive LLMs are uh 172 00:09:27,040 --> 00:09:29,517 not the way we're going to make progress 173 00:09:29,600 --> 00:09:31,837 towards superhuman intelligence. These 174 00:09:31,920 --> 00:09:34,717 are the large language models like GPT4 175 00:09:34,800 --> 00:09:37,037 like Llama 2 and three soon and so on. 176 00:09:37,120 --> 00:09:40,557 How do they work and why are they not going to take us all the way? 177 00:09:40,640 --> 00:09:43,837 >> For a number of reasons. The first is that there is a number of 178 00:09:43,920 --> 00:09:47,277 characteristics of intelligent behavior. 179 00:09:47,360 --> 00:09:52,477 For example, the capacity to understand the world, understand the physical 180 00:09:52,560 --> 00:09:54,077 world, [sighs] 181 00:09:54,160 --> 00:09:57,757 the ability to remember and retrieve 182 00:09:57,840 --> 00:10:01,837 things. Um, persistent memory, the 183 00:10:01,920 --> 00:10:04,157 ability to reason, and the ability to 184 00:10:04,240 --> 00:10:06,237 plan. Those are four essential 185 00:10:06,320 --> 00:10:09,037 characteristic of intelligent 186 00:10:09,120 --> 00:10:13,277 systems or entities, humans, animals. 187 00:10:13,360 --> 00:10:16,557 L&Ms can do none of those or they can 188 00:10:16,640 --> 00:10:21,917 only do them in a very primitive way and uh they don't really understand the 189 00:10:22,000 --> 00:10:24,637 physical world. They don't really have persistent memory. They can't really 190 00:10:24,720 --> 00:10:27,197 reason and they certainly can't plan. 191 00:10:27,280 --> 00:10:30,397 And so you know if if if you expect the 192 00:10:30,480 --> 00:10:35,197 system to become intelligent just you know without having the possibility of 193 00:10:35,280 --> 00:10:38,637 doing those things uh you're making a mistake. [snorts] 194 00:10:38,720 --> 00:10:41,677 That is not to say that autographs are 195 00:10:41,760 --> 00:10:45,597 not useful. They're certainly useful. 196 00:10:45,680 --> 00:10:48,637 Um that they're not interesting. That we 197 00:10:48,720 --> 00:10:51,437 can't build a whole ecosystem of 198 00:10:51,520 --> 00:10:53,357 applications around them. Of course, we 199 00:10:53,440 --> 00:10:57,357 can. But as a path towards human level 200 00:10:57,440 --> 00:11:00,077 intelligence, they're missing essential 201 00:11:00,160 --> 00:11:02,797 components. And then there is another 202 00:11:02,880 --> 00:11:05,597 tidbit or or fact that I think is very 203 00:11:05,680 --> 00:11:08,157 interesting. Those LLMs are trained on 204 00:11:08,240 --> 00:11:10,877 enormous amounts of text. Basically, the 205 00:11:10,960 --> 00:11:13,357 entirety of all publicly available text 206 00:11:13,440 --> 00:11:16,237 on the internet, right? That's typically 207 00:11:16,320 --> 00:11:19,517 on the order of 10^ the 13 tokens. Each 208 00:11:19,600 --> 00:11:21,437 token is typically two bytes. So that's 209 00:11:21,520 --> 00:11:24,637 two 10^ the 13 bytes as training data. 210 00:11:24,720 --> 00:11:28,077 It would take you or me 170,000 years to 211 00:11:28,160 --> 00:11:30,877 just read through this at 8 hours a day. 212 00:11:30,960 --> 00:11:34,797 Uh so it seems like an enormous amount of knowledge, right, that those systems 213 00:11:34,880 --> 00:11:38,877 can accumulate. Um, 214 00:11:38,960 --> 00:11:43,517 but then you realize it's really not that much data. If you you talk to 215 00:11:43,600 --> 00:11:45,277 developmental psychologists and they 216 00:11:45,360 --> 00:11:47,597 tell you a 4-year-old has been awake for 217 00:11:47,680 --> 00:11:51,437 16,000 hours in his or her life. 218 00:11:51,520 --> 00:11:54,877 Um, and the amount of information that 219 00:11:54,960 --> 00:11:59,037 has uh reached the visual cortex of that 220 00:11:59,120 --> 00:12:01,517 child in four years 221 00:12:01,600 --> 00:12:05,117 um is about 10^ the 15 bytes. And you 222 00:12:05,200 --> 00:12:08,077 can compute this by estimating that the 223 00:12:08,160 --> 00:12:09,997 optical nerve carry about 20 megab 224 00:12:10,080 --> 00:12:13,437 megabytes per second roughly. And so 10^ 225 00:12:13,520 --> 00:12:16,477 the 15 bytes for a 4-year-old versus 2 * 226 00:12:16,560 --> 00:12:20,797 10^ 13 bytes for 170,000 years worth of 227 00:12:20,880 --> 00:12:25,277 reading. What that tells you is that uh 228 00:12:25,360 --> 00:12:27,677 through sensory input we see a lot more 229 00:12:27,760 --> 00:12:29,677 information than we than we do through 230 00:12:29,760 --> 00:12:32,317 language. And that despite our 231 00:12:32,400 --> 00:12:34,877 intuition, most of what we learn and 232 00:12:34,960 --> 00:12:38,397 most of our knowledge is through our 233 00:12:38,480 --> 00:12:42,237 observation and interaction with the real world, not through language. 234 00:12:42,320 --> 00:12:45,917 Everything that we learn in the first few years of life and uh certainly 235 00:12:46,000 --> 00:12:50,157 everything that animals learn has nothing to do with language. So it would 236 00:12:50,240 --> 00:12:54,477 be good to uh maybe push against some of the intuition behind what you're saying. 237 00:12:54,560 --> 00:12:58,237 So it is true there's several orders of 238 00:12:58,320 --> 00:13:02,477 magnitude more data coming into the human mind 239 00:13:02,560 --> 00:13:06,157 much faster and the human mind is able to learn very quickly from that filter 240 00:13:06,240 --> 00:13:10,477 the data very quickly. You know somebody might argue your comparison between 241 00:13:10,560 --> 00:13:12,797 sensory data versus language that 242 00:13:12,880 --> 00:13:16,157 language is already very compressed. It 243 00:13:16,240 --> 00:13:20,157 already contains a lot more information than the bytes it takes to store them if 244 00:13:20,240 --> 00:13:23,917 you compare it to visual data. So there's a lot of wisdom and language. 245 00:13:24,000 --> 00:13:25,997 There's words and the way we stitch them 246 00:13:26,080 --> 00:13:27,997 together, it already contains a lot of 247 00:13:28,080 --> 00:13:32,237 information. So, is it possible that 248 00:13:32,320 --> 00:13:37,677 language alone already has enough 249 00:13:37,760 --> 00:13:40,237 wisdom and knowledge in there to be able 250 00:13:40,320 --> 00:13:44,957 to from that language construct a a world model and understanding of the 251 00:13:45,040 --> 00:13:46,877 world, an understanding of the physical 252 00:13:46,960 --> 00:13:49,597 world that you're saying LLMs lack? So 253 00:13:49,680 --> 00:13:53,117 it's a big debate among philosophers and 254 00:13:53,200 --> 00:13:55,437 also cognitive scientists like whether 255 00:13:55,520 --> 00:13:57,197 intelligence needs to be grounded in 256 00:13:57,280 --> 00:14:00,637 reality. Uh I'm clearly in the camp that 257 00:14:00,720 --> 00:14:03,517 uh yes uh intelligence cannot appear 258 00:14:03,600 --> 00:14:08,817 without some grounding in uh some reality. It doesn't need to be you know 259 00:14:08,900 --> 00:14:12,877 [snorts] physical reality. It could be simulated but um but the environment is 260 00:14:12,960 --> 00:14:16,077 just much richer than what you can express in language. language is a very 261 00:14:16,160 --> 00:14:18,317 approximate representation 262 00:14:18,400 --> 00:14:21,917 of our percepts and our mental models, 263 00:14:22,000 --> 00:14:23,437 right? I mean, there's a lot of tasks 264 00:14:23,520 --> 00:14:26,637 that we accomplish where we manipulate 265 00:14:26,720 --> 00:14:29,517 uh a mental model of [snorts] uh of the 266 00:14:29,600 --> 00:14:32,077 situation at hand and that has nothing 267 00:14:32,160 --> 00:14:36,557 to do with language. Everything that's physical, mechanical, whatever, when we 268 00:14:36,640 --> 00:14:39,037 build something, when we accomplish a 269 00:14:39,120 --> 00:14:41,437 task, modern task of, you know, grabbing 270 00:14:41,520 --> 00:14:44,397 something, etc., We plan for action 271 00:14:44,480 --> 00:14:47,117 sequences and we do this by essentially 272 00:14:47,200 --> 00:14:51,277 imagining the result of the outcome of a 273 00:14:51,360 --> 00:14:52,557 sequence of actions that we might 274 00:14:52,640 --> 00:14:56,717 imagine and that requires mental models 275 00:14:56,800 --> 00:14:58,797 that don't have much to do with language 276 00:14:58,880 --> 00:15:01,997 and that's I would argue most of our 277 00:15:02,080 --> 00:15:06,717 knowledge is derived from that interaction with the physical world. So 278 00:15:06,800 --> 00:15:10,717 a lot of a lot of my my colleagues who are more uh interested in things like 279 00:15:10,800 --> 00:15:13,197 computer vision are really on that camp 280 00:15:13,280 --> 00:15:16,637 that uh AI needs to be embodied 281 00:15:16,720 --> 00:15:19,437 essentially and then other people coming 282 00:15:19,520 --> 00:15:23,117 from the NLP side or maybe you know some 283 00:15:23,200 --> 00:15:26,157 some other motivation don't necessarily 284 00:15:26,240 --> 00:15:28,637 agree with that. Um and philosophers are 285 00:15:28,720 --> 00:15:33,197 split as well. uh and the um the 286 00:15:33,280 --> 00:15:36,397 complexity of the world is hard to um 287 00:15:36,480 --> 00:15:38,877 it's hard to imagine 288 00:15:38,960 --> 00:15:42,077 you know it's hard to represent uh all 289 00:15:42,160 --> 00:15:46,557 the complexities that we take completely for granted in the real world that we 290 00:15:46,640 --> 00:15:50,717 don't even imagine require intelligence right this is the old Marave paradox 291 00:15:50,800 --> 00:15:55,277 from the pioneer of robotics hence Maravec who said you know how is it that 292 00:15:55,360 --> 00:15:59,037 with computers it seems to be easy to do high level complex tasks like playing 293 00:15:59,120 --> 00:16:01,757 chess and solving integrals and doing 294 00:16:01,840 --> 00:16:04,397 things like that. Whereas the thing we 295 00:16:04,480 --> 00:16:06,957 take for granted that we do every day um 296 00:16:07,040 --> 00:16:08,797 like I don't know learning to drive a 297 00:16:08,880 --> 00:16:11,517 car or you know grabbing an object we 298 00:16:11,600 --> 00:16:16,877 can't do as computers um and 299 00:16:16,960 --> 00:16:19,437 you know we have LLMs they can pass pass 300 00:16:19,520 --> 00:16:22,397 the bar exam so they must be smart but 301 00:16:22,480 --> 00:16:25,437 then they can't learn to drive in 20 302 00:16:25,520 --> 00:16:28,077 hours like any 17year-old 303 00:16:28,160 --> 00:16:30,077 they can't learn to clear up the dinner 304 00:16:30,160 --> 00:16:33,117 table and fill of the dishwasher like 305 00:16:33,200 --> 00:16:35,597 any 10-year-old can learn in one shot. 306 00:16:35,680 --> 00:16:39,757 Um why is that? Like you know what what are we missing? What what type of 307 00:16:39,840 --> 00:16:42,317 learning or or 308 00:16:42,400 --> 00:16:44,237 reasoning architecture or whatever are 309 00:16:44,320 --> 00:16:48,477 we missing that um um basically prevent 310 00:16:48,560 --> 00:16:50,797 us from from you know having level five 311 00:16:50,880 --> 00:16:54,077 serving cars and domestic robots? Can a 312 00:16:54,160 --> 00:16:57,277 large language model construct a world 313 00:16:57,360 --> 00:17:02,237 model that does know how to drive and does know how to fill a dishwasher but 314 00:17:02,320 --> 00:17:07,597 just doesn't know how to deal with visual data at this time. So it it can 315 00:17:07,680 --> 00:17:09,997 operate [clears throat] in a space of concepts. 316 00:17:10,080 --> 00:17:13,837 >> So yeah, that's what a lot of people are working on. So the answer the short 317 00:17:13,920 --> 00:17:16,077 answer is no. And the more complex 318 00:17:16,160 --> 00:17:19,437 answer is you can use all kind of tricks 319 00:17:19,520 --> 00:17:21,997 to get 320 00:17:22,080 --> 00:17:27,197 an LLM to basically digest um visual 321 00:17:27,280 --> 00:17:29,837 representations of representations of 322 00:17:29,920 --> 00:17:32,957 images uh or video or audio for that 323 00:17:33,040 --> 00:17:36,477 matter. [snorts] Um and uh a classical 324 00:17:36,560 --> 00:17:39,837 way of doing this is uh you train a 325 00:17:39,920 --> 00:17:43,997 vision system in some way and we have a number of ways to train vision systems 326 00:17:44,080 --> 00:17:47,197 either supervised semi-supervised self-supervised all kinds of different 327 00:17:47,280 --> 00:17:50,717 ways [snorts] uh that will turn any 328 00:17:50,800 --> 00:17:54,877 image into high level representation 329 00:17:54,960 --> 00:17:58,477 basically a list of tokens that are really similar to the kind of tokens 330 00:17:58,560 --> 00:18:00,957 that uh 331 00:18:01,040 --> 00:18:03,757 typical LLM takes as an input and Then 332 00:18:03,840 --> 00:18:08,157 you just feed that to the LLM in 333 00:18:08,240 --> 00:18:11,117 addition to the text and you just expect 334 00:18:11,200 --> 00:18:13,917 LLM to kind of uh you know during 335 00:18:14,000 --> 00:18:19,277 training to kind of be able to uh use those representations to help make 336 00:18:19,360 --> 00:18:22,717 decisions. I mean there's been work along those lines for for quite a long 337 00:18:22,800 --> 00:18:27,117 time. Um and now you see those systems right? I mean there are LLMs that can 338 00:18:27,200 --> 00:18:30,957 that have some vision extension but they're basically hacks in the sense 339 00:18:31,040 --> 00:18:35,277 that um those things are not like trained end to end to to handle to 340 00:18:35,360 --> 00:18:39,277 really understand the world. They're not trained with video for example. Uh they 341 00:18:39,360 --> 00:18:43,837 don't really understand intuitive physics at least not at the moment. 342 00:18:43,920 --> 00:18:46,637 >> So [snorts] you don't think there's something special to you about intuitive 343 00:18:46,720 --> 00:18:50,397 physics about sort of common sense reasoning about the physical space about 344 00:18:50,480 --> 00:18:55,517 physical reality? that to you is a giant leap that LLMs are just not able to do. 345 00:18:55,600 --> 00:18:59,437 >> We're not going to be able to do this with the type of LLMs that we are uh 346 00:18:59,520 --> 00:19:02,957 working with today. And there's a number of reasons for this, but uh the main 347 00:19:03,040 --> 00:19:05,037 reason is 348 00:19:05,120 --> 00:19:07,517 the way LM LM are trained is that you 349 00:19:07,600 --> 00:19:11,037 you take a piece of text, you remove 350 00:19:11,120 --> 00:19:14,637 some of the words in that text, you mask them, you replace by replace them by 351 00:19:14,720 --> 00:19:18,397 blank markers, and you train a gigantic neural net to predict the words that are 352 00:19:18,480 --> 00:19:22,157 missing. uh and if you build this neural net in a 353 00:19:22,240 --> 00:19:24,157 particular way so that it can only look 354 00:19:24,240 --> 00:19:29,277 at u words that are to the left of the one it's trying to predict then what you 355 00:19:29,360 --> 00:19:32,317 have is a system that basically is trying to predict the next word in a 356 00:19:32,400 --> 00:19:37,437 text right so then you can feed it um a text a prompt and you can ask it to 357 00:19:37,520 --> 00:19:41,117 predict the next word it can never predict the next word exactly and so 358 00:19:41,200 --> 00:19:45,917 what it's going to do is uh produce a probability distribution of all the 359 00:19:46,000 --> 00:19:49,037 possible words in your dictionary in fact it doesn't predict words it 360 00:19:49,120 --> 00:19:53,997 predicts tokens that are kind of subword units. And so it's easy to handle the 361 00:19:54,080 --> 00:19:57,517 uncertainty in the prediction there because there's only a finite number of 362 00:19:57,600 --> 00:20:00,397 possible words in the dictionary and you 363 00:20:00,480 --> 00:20:01,837 can just compute a distribution over 364 00:20:01,920 --> 00:20:04,717 them. [snorts] Um then what you what the 365 00:20:04,800 --> 00:20:10,077 system does is that it it picks a word from that distribution. Of course 366 00:20:10,160 --> 00:20:13,437 there's a higher chance of picking words that have a higher probability within 367 00:20:13,520 --> 00:20:16,317 that distribution. And so you sample from that distribution to actually 368 00:20:16,400 --> 00:20:20,957 produce a word and then you shift that word into the input 369 00:20:21,040 --> 00:20:25,277 and so that allows the system not to predict the second word right and once 370 00:20:25,360 --> 00:20:27,357 you do this you shift it into the input 371 00:20:27,440 --> 00:20:29,597 etc that's called auto reggressive 372 00:20:29,680 --> 00:20:32,637 prediction and which is why those LLM 373 00:20:32,720 --> 00:20:35,997 should be called auto reggressive LLM uh 374 00:20:36,080 --> 00:20:40,637 >> but we just call them LM and 375 00:20:40,720 --> 00:20:42,237 there is a difference between this kind 376 00:20:42,320 --> 00:20:45,517 of process and a process by which before 377 00:20:45,600 --> 00:20:49,997 producing a word when you talk when you and I talk you and I are 378 00:20:50,080 --> 00:20:53,117 >> bilingual >> we think about what we're going to say 379 00:20:53,200 --> 00:20:57,037 and it's relatively independent of the language in which we're going to say it 380 00:20:57,120 --> 00:21:00,557 when we when we talk about like a I don't know let's say mathematical 381 00:21:00,640 --> 00:21:03,437 concept or something >> the kind of thinking that we're doing 382 00:21:03,520 --> 00:21:06,877 and the answer that we're planning to produce 383 00:21:06,960 --> 00:21:08,957 is not linked to whether we're going to 384 00:21:09,040 --> 00:21:12,157 see it in French or Russian or English 385 00:21:12,240 --> 00:21:14,397 >> Chsky just rolled his But I understand. 386 00:21:14,480 --> 00:21:17,357 So you're saying that there's a a bigger 387 00:21:17,440 --> 00:21:20,237 abstraction that rep that's uh that goes 388 00:21:20,320 --> 00:21:23,037 before language that maps onto language. 389 00:21:23,120 --> 00:21:26,637 >> Right. It's certainly true for a lot of thinking that we that we do. 390 00:21:26,720 --> 00:21:31,517 >> Is that obvious that we don't like you're saying your thinking is same in 391 00:21:31,600 --> 00:21:34,717 French as it is in English? >> Yeah, pretty much. Yeah. 392 00:21:34,800 --> 00:21:37,597 >> Pretty much. Or is this like how how 393 00:21:37,680 --> 00:21:41,757 flexible are you? Like if if there's a probability distribution? 394 00:21:41,840 --> 00:21:45,917 Well, it it depends what kind of thinking, right? If it's just uh if it's 395 00:21:46,000 --> 00:21:49,597 like producing puns, I get much better in French than English about that. 396 00:21:49,680 --> 00:21:53,757 >> No, but is there an abstract representation of puns? Like is your 397 00:21:53,840 --> 00:21:57,597 humor an abstract like when you tweet and your tweets are sometimes a little 398 00:21:57,680 --> 00:22:01,997 bit spicy? Uh what's is there an abstract representation in your brain of 399 00:22:02,080 --> 00:22:04,477 a tweet before it maps onto English? 400 00:22:04,560 --> 00:22:07,037 There is an abstract representation of 401 00:22:07,120 --> 00:22:11,597 imagining the reaction of a reader to that uh text. 402 00:22:11,680 --> 00:22:14,797 >> You start with laughter and then figure out how to make that happen 403 00:22:14,880 --> 00:22:19,677 >> or figure out like a reaction you want to cause and and then figure out how to 404 00:22:19,760 --> 00:22:22,957 say it, right? So that it causes that reaction. But that's like really close 405 00:22:23,040 --> 00:22:27,037 to language. But think about like a mathemat mathematical concept 406 00:22:27,120 --> 00:22:31,277 >> or you know imagining you know something you want to build out of wood or 407 00:22:31,360 --> 00:22:34,477 something like this, right? the kind of thinking you're doing is absolutely 408 00:22:34,560 --> 00:22:37,837 nothing to do with language really. Like it's not like you have necessarily like 409 00:22:37,920 --> 00:22:41,677 an internal monologue in any particular language. You're you're you know 410 00:22:41,760 --> 00:22:44,557 imagining mental models of of the thing, 411 00:22:44,640 --> 00:22:48,877 right? I mean if I if I ask you to like imagine what this uh water bottle will 412 00:22:48,960 --> 00:22:53,517 look like if I rotate it >> 90 degrees um that has nothing to do 413 00:22:53,600 --> 00:22:55,277 with language. 414 00:22:55,360 --> 00:22:58,637 And so uh so clearly there is you know a 415 00:22:58,720 --> 00:23:01,517 more abstract level of representation 416 00:23:01,600 --> 00:23:03,837 uh in which we we do most of our 417 00:23:03,920 --> 00:23:06,477 thinking and we plan what we're going to 418 00:23:06,560 --> 00:23:10,077 say if the output is is 419 00:23:10,160 --> 00:23:13,757 you know uttered words as opposed to an 420 00:23:13,840 --> 00:23:17,357 output being uh you know muscle actions 421 00:23:17,440 --> 00:23:21,117 >> right um we we plan our answer before we 422 00:23:21,200 --> 00:23:23,437 produce it and L&Ms don't do that they 423 00:23:23,520 --> 00:23:26,557 just produce one word after the other 424 00:23:26,640 --> 00:23:28,397 instinctively if you want. It's like 425 00:23:28,480 --> 00:23:31,037 it's a bit like the you know 426 00:23:31,120 --> 00:23:35,997 subconscious uh actions where you don't like you're distracted you're doing 427 00:23:36,080 --> 00:23:40,157 something you're completely concentrated and someone comes to you and you know 428 00:23:40,240 --> 00:23:43,197 ask you a question and you kind of answer the question you don't have time 429 00:23:43,280 --> 00:23:45,997 to think about the answer but the answer is easy so you don't need to pay 430 00:23:46,080 --> 00:23:50,157 attention you sort of respond automatically that's kind of what an NLM 431 00:23:50,240 --> 00:23:52,237 does right it doesn't think about its 432 00:23:52,320 --> 00:23:57,357 answer really uh it retrieves it because it's accumulated a lot of knowledge So 433 00:23:57,440 --> 00:23:59,037 it can retrieve some some things, but 434 00:23:59,120 --> 00:24:02,877 it's going to just spit out one token 435 00:24:02,960 --> 00:24:07,437 after the other without planning the answer. But 436 00:24:07,520 --> 00:24:10,957 you're making it sound just one token after the other, one token at a time, 437 00:24:11,040 --> 00:24:12,637 generation 438 00:24:12,720 --> 00:24:18,077 is uh bound to be simplistic. 439 00:24:18,160 --> 00:24:20,157 But if the world model is sufficiently 440 00:24:20,240 --> 00:24:24,077 sophisticated, that one token at a time, 441 00:24:24,160 --> 00:24:27,037 the the most likely thing it generates 442 00:24:27,120 --> 00:24:29,677 as a sequence of tokens is going to be a 443 00:24:29,760 --> 00:24:31,917 deeply profound thing. 444 00:24:32,000 --> 00:24:37,037 >> Okay. But then that assumes that those systems actually possess an eternal 445 00:24:37,120 --> 00:24:39,917 world model. >> So really it goes to the I I think the 446 00:24:40,000 --> 00:24:42,717 fundamental question is can you build a 447 00:24:42,800 --> 00:24:45,757 a really 448 00:24:45,840 --> 00:24:50,637 complete world model? uh complete but a uh one that has a deep understanding of 449 00:24:50,720 --> 00:24:56,317 the world. >> Yeah. So can you build this first of all by prediction, 450 00:24:56,400 --> 00:24:59,357 >> right? >> And the answer is probably yes. Can you 451 00:24:59,440 --> 00:25:02,477 predict can you build it by predicting 452 00:25:02,560 --> 00:25:05,757 words? And the answer is most probably 453 00:25:05,840 --> 00:25:09,837 no because language is very poor in 454 00:25:09,920 --> 00:25:12,957 terms or weak or low bandwidth if you want. There's just not enough 455 00:25:13,040 --> 00:25:15,997 information there. So building world 456 00:25:16,080 --> 00:25:21,197 models means observing the world and uh 457 00:25:21,280 --> 00:25:22,717 understanding 458 00:25:22,800 --> 00:25:24,797 why the world is evolving the way the 459 00:25:24,880 --> 00:25:30,157 way it is. And then uh the the extra 460 00:25:30,240 --> 00:25:34,877 component of a world model is something that can predict how the world is going 461 00:25:34,960 --> 00:25:39,037 to evolve as a consequence of an action you might take. Right? So world model 462 00:25:39,120 --> 00:25:42,317 really is here is my idea of the state of the world at time t. here is an 463 00:25:42,400 --> 00:25:44,797 action I might take. What is the 464 00:25:44,880 --> 00:25:47,197 predicted state of the world at time t 465 00:25:47,280 --> 00:25:52,077 plus one? Now that state of the world does not does does not need to represent 466 00:25:52,160 --> 00:25:55,757 everything about the world. It just needs to represent enough that's 467 00:25:55,840 --> 00:25:57,997 relevant for this planning of of the 468 00:25:58,080 --> 00:26:03,037 action but not necessarily all the details. Now, here is the problem. Um 469 00:26:03,120 --> 00:26:08,237 you're not going to be able to do this with generative models. So a generative 470 00:26:08,320 --> 00:26:11,757 model that's trained on video and we've tried to do this for 10 years. You take 471 00:26:11,840 --> 00:26:15,117 a video, show a system a piece of video, 472 00:26:15,200 --> 00:26:19,997 and then ask it to predict the reminder of the video. Basically, predict what's 473 00:26:20,080 --> 00:26:22,797 going to happen >> one frame at a time. Do the same thing 474 00:26:22,880 --> 00:26:27,037 as sort of uh the autogressive LLMs do, but for video, 475 00:26:27,120 --> 00:26:30,397 >> right? Either one frame at a time or a group of friends at a time. 476 00:26:30,480 --> 00:26:33,517 >> Um, but yeah, uh, a large video model if 477 00:26:33,600 --> 00:26:38,157 you want. uh the idea of of doing this 478 00:26:38,240 --> 00:26:39,677 has been floating around for a long time 479 00:26:39,760 --> 00:26:43,837 and at at fair u some colleagues and I 480 00:26:43,920 --> 00:26:46,877 have been trying to do this for about 10 years 481 00:26:46,960 --> 00:26:49,997 um and you can't you can't really do the 482 00:26:50,080 --> 00:26:52,557 same trick as with LLMs because uh you 483 00:26:52,640 --> 00:26:55,917 know LLM as I said you can't predict 484 00:26:56,000 --> 00:26:58,637 exactly which word is going to follow a 485 00:26:58,720 --> 00:27:03,037 sequence of words but you can predict the distribution of words now if you go 486 00:27:03,120 --> 00:27:06,717 to video what you would have to do is predict the distribution over all 487 00:27:06,800 --> 00:27:09,597 possible frames in a video and we don't 488 00:27:09,680 --> 00:27:12,077 really know how to do that properly. 489 00:27:12,160 --> 00:27:16,157 Uh we we we do not know how to represent distributions over highdimensional 490 00:27:16,240 --> 00:27:19,597 continuous spaces in ways that are useful. 491 00:27:19,680 --> 00:27:22,797 >> Uh and and that's that there lies the 492 00:27:22,880 --> 00:27:25,757 main issue and the reason we can do this 493 00:27:25,840 --> 00:27:29,437 is because the world is incredibly more 494 00:27:29,520 --> 00:27:33,757 complicated and richer in terms of information than than text. Text is 495 00:27:33,840 --> 00:27:37,037 discrete. Video is high dimensional and 496 00:27:37,120 --> 00:27:39,997 continuous. A lot of details in this. Um 497 00:27:40,080 --> 00:27:43,837 so if I take a a video of this room uh 498 00:27:43,920 --> 00:27:46,397 and the video is you know a camera 499 00:27:46,480 --> 00:27:51,597 panning around um there is no way I can predict everything that's going to be in 500 00:27:51,680 --> 00:27:54,557 the room as I pan around. The system cannot predict what's going to be in the 501 00:27:54,640 --> 00:27:59,597 room as the camera is panning. Maybe it's going to predict this is this is a 502 00:27:59,680 --> 00:28:02,397 room where there's a light and there is a wall and things like that. it can't 503 00:28:02,480 --> 00:28:05,597 predict what the painting on the wall looks like or what the texture of the 504 00:28:05,680 --> 00:28:07,837 couch looks like. Certainly not the 505 00:28:07,920 --> 00:28:10,237 texture of the carpet. So there's no way 506 00:28:10,320 --> 00:28:12,797 it can predict all those details. So the 507 00:28:12,880 --> 00:28:16,637 the way to handle this is one way 508 00:28:16,720 --> 00:28:20,157 possibly to handle this which we've been working for a long time is to have a 509 00:28:20,240 --> 00:28:24,717 model that has what's called a latent variable. And the latent variable is fed 510 00:28:24,800 --> 00:28:28,237 to a neural net and it's supposed to represent all the information about the 511 00:28:28,320 --> 00:28:31,517 world that you don't perceive yet. and 512 00:28:31,600 --> 00:28:35,037 uh that you need to augment 513 00:28:35,120 --> 00:28:38,317 uh the the system for the prediction to 514 00:28:38,400 --> 00:28:39,997 do a good job at predicting pixels 515 00:28:40,080 --> 00:28:42,717 including the you know fine texture of 516 00:28:42,800 --> 00:28:46,157 the of the carpet and the and the couch 517 00:28:46,240 --> 00:28:49,677 and and the painting on the wall. um 518 00:28:49,760 --> 00:28:53,757 uh that has been a complete failure essentially and we've tried lots of 519 00:28:53,840 --> 00:28:56,317 things we tried u just straight neural 520 00:28:56,400 --> 00:29:00,317 nets we tried GANs we tried uh you know 521 00:29:00,400 --> 00:29:02,557 VAES all kinds of regularized 522 00:29:02,640 --> 00:29:06,877 autoenccoders we tried um many things we 523 00:29:06,960 --> 00:29:08,797 also tried those kind of methods to 524 00:29:08,880 --> 00:29:12,317 learn uh good representations of images 525 00:29:12,400 --> 00:29:16,317 or video um that could then be used as 526 00:29:16,400 --> 00:29:19,517 input to for example an image classification system. 527 00:29:19,600 --> 00:29:22,557 >> Mhm. >> And that also has basically failed like 528 00:29:22,640 --> 00:29:25,357 all the systems that attempt to predict 529 00:29:25,440 --> 00:29:28,797 missing parts of an image or video 530 00:29:28,880 --> 00:29:33,757 um you know from a corrupted version of it basically. Okay. So right take an 531 00:29:33,840 --> 00:29:37,517 image or a video corrupt it or transform it in some way and then try to 532 00:29:37,600 --> 00:29:40,237 reconstruct the complete video or image 533 00:29:40,320 --> 00:29:43,357 from the corrupted version 534 00:29:43,440 --> 00:29:47,277 and then hope that internally the system will develop a good representations of 535 00:29:47,360 --> 00:29:52,017 images that you can use for object recognition segmentation whatever it is 536 00:29:52,100 --> 00:29:55,437 [snorts] that has been essentially a complete failure 537 00:29:55,520 --> 00:29:59,277 and it works really well for text that's the principle that is used for LLMs 538 00:29:59,360 --> 00:30:02,637 right >> so where is the failure exactly is it 539 00:30:02,720 --> 00:30:05,517 It's very difficult to form a good 540 00:30:05,600 --> 00:30:10,797 representation of an image a good in like a good embedding of all all the 541 00:30:10,880 --> 00:30:14,237 important information in the image. Is it in terms of the consistency of image 542 00:30:14,320 --> 00:30:19,037 to image to image to image that forms the video like where what are the if we 543 00:30:19,120 --> 00:30:20,717 do a highlight reel of all the ways you 544 00:30:20,800 --> 00:30:23,437 failed. What's that look like? 545 00:30:23,520 --> 00:30:26,717 >> Okay, so the reason this doesn't work uh 546 00:30:26,800 --> 00:30:30,637 is first of all I have to tell you exactly what doesn't work because there 547 00:30:30,720 --> 00:30:32,717 is something else that does work. Mhm. 548 00:30:32,800 --> 00:30:35,197 >> So the thing that does not work is 549 00:30:35,280 --> 00:30:38,557 training a system to learn 550 00:30:38,640 --> 00:30:41,277 representations of images by training it 551 00:30:41,360 --> 00:30:43,917 to reconstruct 552 00:30:44,000 --> 00:30:48,317 uh a good image from a corrupted version of it. Okay, that's what doesn't work. 553 00:30:48,400 --> 00:30:50,797 And we have a whole slew of techniques 554 00:30:50,880 --> 00:30:55,517 for this uh that are you know variant of den noising autoenccoders. Something 555 00:30:55,600 --> 00:31:00,077 called MAE developed by some of my colleagues at fair maxenccoder. So it's 556 00:31:00,160 --> 00:31:04,877 basically like the you know LLMs or or or things like this 557 00:31:04,960 --> 00:31:08,317 where you train a system by corrupting text except you corrupt images you 558 00:31:08,400 --> 00:31:11,927 remove patches from it and you train a gigantic neural net to reconstruct 559 00:31:12,010 --> 00:31:15,117 [snorts] the features you get are not good and you know they're not good 560 00:31:15,200 --> 00:31:17,597 because if you now train the same 561 00:31:17,680 --> 00:31:20,557 architecture but you train it supervised 562 00:31:20,640 --> 00:31:24,237 >> with uh label data with text textual 563 00:31:24,320 --> 00:31:27,197 descriptions of images etc. you do get 564 00:31:27,280 --> 00:31:29,437 good representations and the performance 565 00:31:29,520 --> 00:31:34,237 on recognition tasks is much better than if you do this self-supervised 566 00:31:34,320 --> 00:31:37,197 pre-training. >> So the architecture is good. 567 00:31:37,280 --> 00:31:40,397 >> The architecture is good. The architecture of the encoder is good. 568 00:31:40,480 --> 00:31:41,997 >> Okay. But the fact that you train the 569 00:31:42,080 --> 00:31:47,837 system to reconstruct images does not lead it to produce to learn good generic 570 00:31:47,920 --> 00:31:51,117 features of images >> when you train in a self-supervised way. 571 00:31:51,200 --> 00:31:53,117 >> Self-supervised by reconstruction. 572 00:31:53,200 --> 00:31:54,157 >> Yeah. By reconstruction. 573 00:31:54,240 --> 00:31:57,037 >> Okay. So what's the alternative? 574 00:31:57,120 --> 00:32:00,237 The alternative is joint embedding. 575 00:32:00,320 --> 00:32:03,437 >> What is joint embedding? What are what are these architectures that you're so 576 00:32:03,520 --> 00:32:05,917 excited about? >> Okay. So now instead of training a 577 00:32:06,000 --> 00:32:07,677 system to encode the image and then 578 00:32:07,760 --> 00:32:10,557 training it to reconstruct the the full 579 00:32:10,640 --> 00:32:13,117 image from a corrupted version, you take 580 00:32:13,200 --> 00:32:16,397 the full image, you take the 581 00:32:16,480 --> 00:32:21,517 corrupted or transformed version, you run them both through encoders, which in 582 00:32:21,600 --> 00:32:22,877 general are identical, but not 583 00:32:22,960 --> 00:32:25,517 necessarily. [snorts] And then you you 584 00:32:25,600 --> 00:32:29,037 train a predictor on top of those 585 00:32:29,120 --> 00:32:30,557 encoders 586 00:32:30,640 --> 00:32:33,837 um to predict the representation of the 587 00:32:33,920 --> 00:32:37,197 full input from the representation of 588 00:32:37,280 --> 00:32:42,077 the corrupted one. >> Okay. So joint embedding because you're 589 00:32:42,160 --> 00:32:46,877 you're taking the the full input and the corrupted version or transform version 590 00:32:46,960 --> 00:32:51,037 run them both through encoders. you get a joint embedding and then you and then 591 00:32:51,120 --> 00:32:55,197 you're you're saying can I predict the representation of the full one from the 592 00:32:55,280 --> 00:32:57,517 representation of the corrupted one. 593 00:32:57,600 --> 00:33:02,157 Okay. Um and I call this a jet. So that means joint embedding predictive 594 00:33:02,240 --> 00:33:04,877 architecture because this joint embedding and there is this predictor 595 00:33:04,960 --> 00:33:09,197 that predicts the representation of the good guy from from the bad guy. [snorts] 596 00:33:09,280 --> 00:33:11,517 Um and the big question is how do you 597 00:33:11,600 --> 00:33:14,877 train something like this? Uh and until 598 00:33:14,960 --> 00:33:18,957 five years ago or six years ago, we didn't have particularly good answers 599 00:33:19,040 --> 00:33:20,797 for how you train those things except 600 00:33:20,880 --> 00:33:23,677 for one um called contrastive 601 00:33:23,760 --> 00:33:27,837 contrastive learning where 602 00:33:27,920 --> 00:33:29,357 um and the idea of contractive learning 603 00:33:29,440 --> 00:33:31,757 is you you take a pair of images that 604 00:33:31,840 --> 00:33:34,877 are again an image and a corrupted 605 00:33:34,960 --> 00:33:37,197 version or degraded version somehow or 606 00:33:37,280 --> 00:33:39,837 transformed version of the original one 607 00:33:39,920 --> 00:33:44,397 and you train the predicted representation to the same as as as 608 00:33:44,480 --> 00:33:47,677 that. If you only do this, the system collapses. It basically completely 609 00:33:47,760 --> 00:33:49,357 ignores the input and produces 610 00:33:49,440 --> 00:33:52,717 representations that are constant. 611 00:33:52,800 --> 00:33:56,957 So the contrastive methods avoid this and and those things have been around 612 00:33:57,040 --> 00:33:58,797 since the early 90s. I had a paper on 613 00:33:58,880 --> 00:34:03,517 this in 1993. Um is you also show pairs 614 00:34:03,600 --> 00:34:07,037 of images that you know are different 615 00:34:07,120 --> 00:34:10,717 and then you push away the representations from each other. So you 616 00:34:10,800 --> 00:34:14,957 say not only do representations of things that we know are the same should 617 00:34:15,040 --> 00:34:18,157 be the same or should be similar but representation of things that we know 618 00:34:18,240 --> 00:34:22,157 are different should be different and that prevents the collapse but it has 619 00:34:22,240 --> 00:34:24,957 some limitation and there's a whole bunch of uh techniques that have 620 00:34:25,040 --> 00:34:28,237 appeared over the last six seven years 621 00:34:28,320 --> 00:34:30,317 um that can revive this this type of 622 00:34:30,400 --> 00:34:36,287 method um some of them from fair some of them from from Google and other places 623 00:34:36,370 --> 00:34:39,117 [snorts] um but there are limitations to those 624 00:34:39,200 --> 00:34:43,517 contrasting method What has changed in the last uh 625 00:34:43,600 --> 00:34:46,077 you know three four years is now now we 626 00:34:46,160 --> 00:34:47,757 have methods that are non-contrastive. 627 00:34:47,840 --> 00:34:50,637 So they don't require those negative 628 00:34:50,720 --> 00:34:53,037 contrastive samples of images that are 629 00:34:53,120 --> 00:34:58,157 that we know are different. You can only you train them only with images that are 630 00:34:58,240 --> 00:35:02,637 you know different versions or different views of the same thing. Uh and you rely 631 00:35:02,720 --> 00:35:06,317 on some other tweaks to prevent the system from collapsing. And we have half 632 00:35:06,400 --> 00:35:08,797 a dozen different methods for this now. 633 00:35:08,880 --> 00:35:10,717 So what is the fundamental difference 634 00:35:10,800 --> 00:35:13,597 between joint embedding architectures 635 00:35:13,680 --> 00:35:17,677 and LLMs? So can uh can uh jabbot take 636 00:35:17,760 --> 00:35:21,677 us to AGI? Whether we should say that 637 00:35:21,760 --> 00:35:24,957 you don't like uh the term AGI and we'll 638 00:35:25,040 --> 00:35:26,957 probably argue I think every single time 639 00:35:27,040 --> 00:35:28,717 I've talked to you with argued about the 640 00:35:28,800 --> 00:35:32,797 G and AGI. Yes, I get I get it. I get 641 00:35:32,880 --> 00:35:34,397 it. We'll we'll probably continue to 642 00:35:34,480 --> 00:35:36,717 argue about it. It's great. Uh you you 643 00:35:36,800 --> 00:35:39,917 like uh Amy this cuz cuz you like French 644 00:35:40,000 --> 00:35:44,557 and um Ammy is is is uh I guess friend in French. 645 00:35:44,640 --> 00:35:48,637 >> Yes. >> And AMI stands for advanced machine intelligence, 646 00:35:48,720 --> 00:35:52,557 >> right? >> Um but either way can Japa take us to 647 00:35:52,640 --> 00:35:55,437 that towards that advanced machine intelligence? 648 00:35:55,520 --> 00:35:57,677 >> Well, so it's a it's a first step. Okay. 649 00:35:57,760 --> 00:35:59,757 So first of all, uh what's the 650 00:35:59,840 --> 00:36:02,157 difference with generative architectures 651 00:36:02,240 --> 00:36:07,757 like LLMs? Um so LLMs um or 652 00:36:07,840 --> 00:36:10,477 vision systems that are trained by reconstruction 653 00:36:10,560 --> 00:36:13,437 generate the inputs right they generate 654 00:36:13,520 --> 00:36:17,997 the original input that is non-corrupted 655 00:36:18,080 --> 00:36:21,117 non-transformed right so you have to predict all the 656 00:36:21,200 --> 00:36:26,397 pixels and there is a huge amount of resources spent in the system to 657 00:36:26,480 --> 00:36:29,837 actually predict all those pixels all the details 658 00:36:29,920 --> 00:36:32,637 uh in a ja you're not trying to predict 659 00:36:32,720 --> 00:36:34,397 all the pixels you're only trying to 660 00:36:34,480 --> 00:36:37,757 predict an abstract representation of of 661 00:36:37,840 --> 00:36:43,197 the inputs, right? And that's much easier in many ways. So what the JPA 662 00:36:43,280 --> 00:36:47,037 system when it's being trained is trying to do is extract as much information as 663 00:36:47,120 --> 00:36:49,917 possible from the input but yet only 664 00:36:50,000 --> 00:36:54,477 extract information that is relatively easily predictable. 665 00:36:54,560 --> 00:36:56,397 >> Okay. >> So there's a lot of things in the world 666 00:36:56,480 --> 00:36:58,077 that we cannot predict. For example, if 667 00:36:58,160 --> 00:36:59,757 you have a self-driving car driving down 668 00:36:59,840 --> 00:37:03,517 the street or road, uh there may be uh 669 00:37:03,600 --> 00:37:06,637 trees around the around the road and it 670 00:37:06,720 --> 00:37:10,637 could be a windy day. So the the leaves on the tree are kind of moving in kind 671 00:37:10,720 --> 00:37:14,797 of semi chaotic random ways that you can't predict and you don't care. You 672 00:37:14,880 --> 00:37:18,957 don't want to predict. So what you want is your encoder to basically eliminate 673 00:37:19,040 --> 00:37:21,837 all those details. It will tell you there's moving leaves, but it's not 674 00:37:21,920 --> 00:37:26,877 going to keep the details of exactly what's going on. Um and so when you do 675 00:37:26,960 --> 00:37:29,997 the prediction in representation space, you're not going to have to predict 676 00:37:30,080 --> 00:37:32,797 every single pixel of every leaf. And 677 00:37:32,880 --> 00:37:35,997 that, you know, um not only is a lot 678 00:37:36,080 --> 00:37:38,477 simpler, but also it allows the system 679 00:37:38,560 --> 00:37:40,877 to essentially learn an abstract 680 00:37:40,960 --> 00:37:43,437 representation of of the world where you 681 00:37:43,520 --> 00:37:46,237 know what can be modeled and predicted 682 00:37:46,320 --> 00:37:51,997 is preserved and the rest is viewed as noise and eliminated by the encoder. So 683 00:37:52,080 --> 00:37:55,197 it kind of lifts the level of abstraction of the representation. If 684 00:37:55,280 --> 00:37:58,557 you think about this, this is something we do absolutely all the time. Whenever 685 00:37:58,640 --> 00:38:02,797 we describe a phenomenon, we describe it at a particular level of abstraction. 686 00:38:02,880 --> 00:38:07,277 >> And we don't always describe every natural phenomenon in terms of quantum 687 00:38:07,360 --> 00:38:11,357 field theory, right? That would be impossible, right? So we have multiple 688 00:38:11,440 --> 00:38:15,757 levels of abstraction to describe what happens in the world. you know, starting 689 00:38:15,840 --> 00:38:17,997 from quantum field theory to like atomic 690 00:38:18,080 --> 00:38:23,117 theory and molecules, you know, and chemistry, materials and, you know, all 691 00:38:23,200 --> 00:38:26,877 the way up to, you know, kind of concrete objects in the real world and 692 00:38:26,960 --> 00:38:29,117 things like that. So the we we can't 693 00:38:29,200 --> 00:38:32,717 just only model everything at the lowest 694 00:38:32,800 --> 00:38:35,197 level and that that's what the idea of 695 00:38:35,280 --> 00:38:38,397 JPA is really on is really about learn 696 00:38:38,480 --> 00:38:43,117 abstract representation in a self-supervised uh manner and you know 697 00:38:43,200 --> 00:38:47,277 you can do it hierarchically as well. So that I think is an essential component 698 00:38:47,360 --> 00:38:51,277 of an intelligent system and in language we can get away without doing this 699 00:38:51,360 --> 00:38:53,117 because language is already to some 700 00:38:53,200 --> 00:38:56,557 level abstract and already has 701 00:38:56,640 --> 00:38:58,557 eliminated a lot of information that is 702 00:38:58,640 --> 00:39:01,117 not predictable and um so we can get 703 00:39:01,200 --> 00:39:03,837 away without doing the joint embedding 704 00:39:03,920 --> 00:39:06,157 without you know lifting the abstraction 705 00:39:06,240 --> 00:39:09,037 level and by directly predicting words. 706 00:39:09,120 --> 00:39:11,517 So joint embedding, it's still 707 00:39:11,600 --> 00:39:16,157 generative, but it's generative in this abstract representation space. 708 00:39:16,240 --> 00:39:19,437 >> Yeah. >> And you're saying language, we were lazy 709 00:39:19,520 --> 00:39:23,437 with language cuz we already got the abstract representation for free and now 710 00:39:23,520 --> 00:39:27,677 we have to zoom out actually think about generally intelligent systems. We have 711 00:39:27,760 --> 00:39:31,597 to [sighs] deal with a full mess of physical 712 00:39:31,680 --> 00:39:33,917 reality of reality. And you can't you 713 00:39:34,000 --> 00:39:37,037 you do have to do this step of jumping 714 00:39:37,120 --> 00:39:39,277 from 715 00:39:39,360 --> 00:39:44,877 uh the full rich detailed reality to a 716 00:39:44,960 --> 00:39:48,717 uh abstract representation of that reality based on which you can then 717 00:39:48,800 --> 00:39:52,237 reason and all that kind of stuff >> right and the thing is those self-s 718 00:39:52,320 --> 00:39:54,317 supervised algorithm that that learn by 719 00:39:54,400 --> 00:39:58,157 prediction even in representation space 720 00:39:58,240 --> 00:40:02,957 uh they learn more concept if the input 721 00:40:03,040 --> 00:40:04,637 data you feed them is more redundant. 722 00:40:04,720 --> 00:40:08,237 The the more redundancy there is in the data, the more they're able to capture 723 00:40:08,320 --> 00:40:10,797 some internal structure of it. And so 724 00:40:10,880 --> 00:40:13,357 there there is way more redundancy and 725 00:40:13,440 --> 00:40:16,797 structure in perceptual uh inputs, 726 00:40:16,880 --> 00:40:22,157 sensory input like like like vision than there is in uh text which is not nearly 727 00:40:22,240 --> 00:40:25,597 as redundant. This is back to the question you were asking a few minutes 728 00:40:25,680 --> 00:40:28,877 ago. Language might represent more information really because it's already 729 00:40:28,960 --> 00:40:31,597 compressed. You're you're right about that. But that means it's also less 730 00:40:31,680 --> 00:40:34,717 redundant and so self-supervised 731 00:40:34,800 --> 00:40:36,317 learning will not work as well. 732 00:40:36,400 --> 00:40:40,717 >> Is it possible to join the 733 00:40:40,800 --> 00:40:43,037 self-supervised training on visual data 734 00:40:43,120 --> 00:40:45,757 and self-supervised training on language 735 00:40:45,840 --> 00:40:48,157 data? There is a huge amount of 736 00:40:48,240 --> 00:40:50,397 knowledge. Even though you talk down 737 00:40:50,480 --> 00:40:55,437 about those 10 to the 13 tokens, those 10 to the 13 tokens represent the 738 00:40:55,520 --> 00:40:57,197 entirety, 739 00:40:57,280 --> 00:41:00,317 a large fraction of what us humans have 740 00:41:00,400 --> 00:41:03,037 figured out. Both the talk on 741 00:41:03,120 --> 00:41:05,757 Reddit and the contents of all the books 742 00:41:05,840 --> 00:41:08,317 and the articles and the full spectrum 743 00:41:08,400 --> 00:41:12,157 of human uh intellectual creation. So is 744 00:41:12,240 --> 00:41:14,957 it possible to join those two together? 745 00:41:15,040 --> 00:41:18,557 >> Well, eventually yes. But I think uh if 746 00:41:18,640 --> 00:41:23,357 we do this too early, we run the risk of being tempted to cheat. And in fact, 747 00:41:23,440 --> 00:41:26,477 that's what people are doing at the moment with vision language model. We're 748 00:41:26,560 --> 00:41:29,837 basically cheating. We're using uh 749 00:41:29,920 --> 00:41:31,997 language as a crutch to help the 750 00:41:32,080 --> 00:41:35,837 deficiencies of our uh vision systems to 751 00:41:35,920 --> 00:41:37,837 kind of learn good representations from 752 00:41:37,920 --> 00:41:40,237 uh images and video. And uh the problem 753 00:41:40,320 --> 00:41:43,277 with this is that we might you know 754 00:41:43,360 --> 00:41:45,997 improve our uh vision language system a 755 00:41:46,080 --> 00:41:48,397 bit I mean our language models by you 756 00:41:48,480 --> 00:41:51,197 know feeding them images but we're not 757 00:41:51,280 --> 00:41:53,757 going to get to the level of even the 758 00:41:53,840 --> 00:41:58,397 intelligence or level of understanding of the world of a cat or dog which 759 00:41:58,480 --> 00:42:02,317 doesn't have language you know they don't have language and they understand 760 00:42:02,400 --> 00:42:04,957 the world much better than any LLM they 761 00:42:05,040 --> 00:42:08,077 can plan really complex actions and sort 762 00:42:08,160 --> 00:42:09,597 of imagine the result of a bunch of 763 00:42:09,680 --> 00:42:12,317 actions. How do we get machines to learn 764 00:42:12,400 --> 00:42:16,797 that before we combine that with language? Obviously, if we combine this 765 00:42:16,880 --> 00:42:19,597 with language, this is going to be a winner. 766 00:42:19,680 --> 00:42:22,717 >> Um, but but before that, we have to 767 00:42:22,800 --> 00:42:26,637 focus on like how do we get systems to learn how the world works. So this kind 768 00:42:26,720 --> 00:42:31,997 of joint embedding predictive architecture for you that's going to be 769 00:42:32,080 --> 00:42:33,597 able to learn something like common 770 00:42:33,680 --> 00:42:37,037 sense, something like what a cat uses to 771 00:42:37,120 --> 00:42:40,397 predict how to mess with its owner most 772 00:42:40,480 --> 00:42:42,717 optimally by knocking over a thing. 773 00:42:42,800 --> 00:42:45,917 >> That's that's the hope. In fact, the techniques we're using are 774 00:42:46,000 --> 00:42:49,277 non-contrastive. Uh so not only is the architecture 775 00:42:49,360 --> 00:42:51,917 non-generative, the learning procedures 776 00:42:52,000 --> 00:42:57,037 we're using are non-contrastive. So we have two two sets of techniques. One set 777 00:42:57,120 --> 00:42:58,877 is based on distillation and there's a 778 00:42:58,960 --> 00:43:01,757 number of uh methods that use this 779 00:43:01,840 --> 00:43:05,437 principle. Uh one by deep mind called by 780 00:43:05,520 --> 00:43:09,037 uh a couple by by fair one one called uh 781 00:43:09,120 --> 00:43:13,917 vikra and another one called Jeppa and Vikra I 782 00:43:14,000 --> 00:43:19,037 should say is not a distillation method actually but IA and B certainly are. And 783 00:43:19,120 --> 00:43:24,557 there's another one also called Dino or Dino also produced from uh at fair. And 784 00:43:24,640 --> 00:43:25,757 the idea of those things is that you 785 00:43:25,840 --> 00:43:28,157 take the full input let's say an image. 786 00:43:28,240 --> 00:43:31,757 Uh you run it through an encoder 787 00:43:31,840 --> 00:43:36,237 uh produces a representation and then you corrupt that input or transform it 788 00:43:36,320 --> 00:43:39,517 run it through the essentially what amounts to the same encoder with some 789 00:43:39,600 --> 00:43:44,157 minor differences and then train a predictor. Sometimes the predictor is 790 00:43:44,240 --> 00:43:47,197 very simple, sometime doesn't exist. But train a predictor to predict a 791 00:43:47,280 --> 00:43:50,317 representation of the first uh 792 00:43:50,400 --> 00:43:52,477 uncorrupted input from the corrupted 793 00:43:52,560 --> 00:43:56,557 input. Um but you only train the the 794 00:43:56,640 --> 00:43:59,357 second branch. Um you only train the 795 00:43:59,440 --> 00:44:04,557 part of the network that is fed with the corrupted input. The other network you 796 00:44:04,640 --> 00:44:07,757 don't you don't train. But since they share the same weight, when you modify 797 00:44:07,840 --> 00:44:12,397 the first one, it also modifies the second one. uh and with various tricks 798 00:44:12,480 --> 00:44:16,237 you can prevent the system from collapsing uh with the collapse of the 799 00:44:16,320 --> 00:44:17,757 type I was explaining before where the 800 00:44:17,840 --> 00:44:20,317 system basically ignores the input 801 00:44:20,400 --> 00:44:24,317 um so that works very well the the 802 00:44:24,400 --> 00:44:26,077 technique we've the two techniques we've 803 00:44:26,160 --> 00:44:30,237 developed at fair dino and uh and Jeppa 804 00:44:30,320 --> 00:44:34,637 work really well for that >> so what kind of data are we talking about here 805 00:44:34,720 --> 00:44:36,957 >> so this there's several scenario one uh 806 00:44:37,040 --> 00:44:40,717 one scenario is you take an image You 807 00:44:40,800 --> 00:44:45,357 corrupt it by um changing the cropping, 808 00:44:45,440 --> 00:44:48,477 for example, changing the size a little bit, maybe changing the orientation, 809 00:44:48,560 --> 00:44:52,717 blurring it, changing the colors, doing all kinds of horrible things to it, 810 00:44:52,800 --> 00:44:55,757 >> but basic horrible things. >> Basic horrible things that sort of 811 00:44:55,840 --> 00:44:57,197 degrade the quality a little bit and 812 00:44:57,280 --> 00:44:59,517 change the framing, uh you know, crop 813 00:44:59,600 --> 00:45:04,157 the image. Um or and in some cases, in 814 00:45:04,240 --> 00:45:07,197 the case of a Jeppa, you don't need to do any of this. You just you just mask 815 00:45:07,280 --> 00:45:09,517 some parts of it, right? You just 816 00:45:09,600 --> 00:45:12,717 basically remove some regions like a big 817 00:45:12,800 --> 00:45:15,597 block essentially [snorts] and and then 818 00:45:15,680 --> 00:45:19,837 you know run through the encoders um and train the entire system encoder and 819 00:45:19,920 --> 00:45:21,597 predictor to predict the representation 820 00:45:21,680 --> 00:45:23,117 of the good one from the representation 821 00:45:23,200 --> 00:45:27,597 of the corrupted one. Um so that's the 822 00:45:27,680 --> 00:45:31,757 Jeppa doesn't need to know that it's an image for example because the only thing 823 00:45:31,840 --> 00:45:36,397 it needs to know is how to do this masking u whereas with dino you need to 824 00:45:36,480 --> 00:45:39,437 know it's an image because you need to do things like you know geometry 825 00:45:39,520 --> 00:45:44,317 transformation and blurring and things like that that are really image specific. 826 00:45:44,400 --> 00:45:47,997 Uh a more recent version of of this that we have is called VJA. So it's basically 827 00:45:48,080 --> 00:45:52,557 the same idea as JJA except um it's applied to video. So now you take a 828 00:45:52,640 --> 00:45:56,717 whole video and you mask a whole chunk of it. And what we mask is actually kind 829 00:45:56,800 --> 00:46:01,117 of a temporal tube. So an all like a whole uh segment of each frame in the 830 00:46:01,200 --> 00:46:05,197 video over the entire video. >> And that tube is like statically 831 00:46:05,280 --> 00:46:08,717 positioned throughout the frames. Just literally straight tube 832 00:46:08,800 --> 00:46:12,717 >> the the tube. Yeah. Typically is 16 frames or something. And we mask the 833 00:46:12,800 --> 00:46:15,037 same region over the entire 16 frames. 834 00:46:15,120 --> 00:46:16,957 It's a different one for every video 835 00:46:17,040 --> 00:46:20,237 obviously. And um and then again train 836 00:46:20,320 --> 00:46:22,717 that system. so as to predict the 837 00:46:22,800 --> 00:46:27,277 representation of the full video from the partially masked video. Uh and that 838 00:46:27,360 --> 00:46:31,357 works really well. It's the first system that we have that learns good 839 00:46:31,440 --> 00:46:33,197 representations of video so that when 840 00:46:33,280 --> 00:46:38,397 you feed those representations to a supervised uh classifier head, it can it 841 00:46:38,480 --> 00:46:41,917 can tell you what action is taking place in a video with you know pretty good 842 00:46:42,000 --> 00:46:45,997 accuracy. So that that's it's the first time we 843 00:46:46,080 --> 00:46:50,317 get something of that uh of that quality. So that's a good test that a 844 00:46:50,400 --> 00:46:53,117 good representation is formed. That means there's something to this. 845 00:46:53,200 --> 00:46:57,037 >> Yeah. Um we also preliminary result that 846 00:46:57,120 --> 00:46:59,117 seem to indicate that the representation 847 00:46:59,200 --> 00:47:02,157 allows us allow our system to tell 848 00:47:02,240 --> 00:47:05,357 whether the video is physically possible 849 00:47:05,440 --> 00:47:09,517 or completely impossible because some object disappeared or an object, you 850 00:47:09,600 --> 00:47:11,117 know, suddenly jumped from one location 851 00:47:11,200 --> 00:47:13,917 to another or or change shape or 852 00:47:14,000 --> 00:47:19,677 something. So it's able to capture some physical con some physics based 853 00:47:19,760 --> 00:47:22,077 constraints about the reality represented in the video. 854 00:47:22,160 --> 00:47:25,917 >> Yeah. >> About the appearance and the disappearance of objects. 855 00:47:26,000 --> 00:47:28,557 >> Yeah. That's really new. 856 00:47:28,640 --> 00:47:32,957 >> Okay. But c can this actually 857 00:47:33,040 --> 00:47:37,437 get us to this kind of uh world model that 858 00:47:37,520 --> 00:47:41,357 understands enough about the world to be able to drive a car? 859 00:47:41,440 --> 00:47:43,597 >> Uh possibly. I mean this is going to 860 00:47:43,680 --> 00:47:48,637 take a while before we get to that point but um um and there are systems already 861 00:47:48,720 --> 00:47:50,077 you know robotic systems that are based 862 00:47:50,160 --> 00:47:54,637 on this uh idea. Uh and uh what you need 863 00:47:54,720 --> 00:47:57,117 for this is a slightly modified version 864 00:47:57,200 --> 00:48:01,517 of this where um imagine that you have 865 00:48:01,600 --> 00:48:05,437 uh a video and a complete video. And 866 00:48:05,520 --> 00:48:10,237 what you're doing to this video is that you're either translating it in time 867 00:48:10,320 --> 00:48:13,197 towards the future. So you only see the beginning of the video, but you don't 868 00:48:13,280 --> 00:48:15,437 see the latter part of it that is in the 869 00:48:15,520 --> 00:48:18,077 original one. Or you just mask the 870 00:48:18,160 --> 00:48:20,477 second half of the video, for example. 871 00:48:20,560 --> 00:48:25,117 Um, and then you you train a a JPA system of the type I described to 872 00:48:25,200 --> 00:48:26,477 predict the representation of the full 873 00:48:26,560 --> 00:48:29,677 video from the the shifted one, but you 874 00:48:29,760 --> 00:48:32,477 also feed the predictor with an action. 875 00:48:32,560 --> 00:48:34,237 For example, you know, the wheel is 876 00:48:34,320 --> 00:48:36,957 turned 10° to the to the right or 877 00:48:37,040 --> 00:48:39,677 something, right? So, if [snorts] it's 878 00:48:39,760 --> 00:48:44,477 a, you know, a dash cam in a car and you know the angle of the wheel, you should 879 00:48:44,560 --> 00:48:47,597 be able to predict to some extent what's going what's going to go what's going to 880 00:48:47,680 --> 00:48:51,597 happen to what you see. uh you're not going to be able to 881 00:48:51,680 --> 00:48:55,037 predict all the details of you know objects that appear in the view 882 00:48:55,120 --> 00:48:59,517 obviously but at a abstract representation level you can you can 883 00:48:59,600 --> 00:49:04,237 probably predict what's going to happen >> so now what you have is 884 00:49:04,320 --> 00:49:07,997 a internal model that says here is my idea of the state of the world at time t 885 00:49:08,080 --> 00:49:11,037 here is an action I'm taking here's a 886 00:49:11,120 --> 00:49:15,917 prediction of the state of the world at time t+1 t plus delta t t plus 2 seconds 887 00:49:16,000 --> 00:49:17,917 whatever it is if you have a model of 888 00:49:18,000 --> 00:49:19,757 this type you can use is it for 889 00:49:19,840 --> 00:49:23,197 planning. So now you can do what LLMs 890 00:49:23,280 --> 00:49:25,837 cannot do which is planning what you're 891 00:49:25,920 --> 00:49:28,077 going to do so as to arrive at a 892 00:49:28,160 --> 00:49:31,597 particular uh outcome or satisfy a 893 00:49:31,680 --> 00:49:34,077 particular objective right so you can 894 00:49:34,160 --> 00:49:37,037 have a number of objectives 895 00:49:37,120 --> 00:49:39,677 um right if you know I can I can predict 896 00:49:39,760 --> 00:49:42,317 that uh if I have 897 00:49:42,400 --> 00:49:44,637 an object like this right and I open my 898 00:49:44,720 --> 00:49:49,917 hand it's going to fall right >> and uh and if I push it with a 899 00:49:50,000 --> 00:49:52,477 particular force on the table, it's going to move. If I push the table 900 00:49:52,560 --> 00:49:54,397 itself, it's probably not going to move 901 00:49:54,480 --> 00:49:57,677 uh with the same force. Um so we have we 902 00:49:57,760 --> 00:49:59,677 have this internal model of the world in 903 00:49:59,760 --> 00:50:03,197 our in our mind uh which allows us to 904 00:50:03,280 --> 00:50:05,437 plan sequences of actions to arrive at a 905 00:50:05,520 --> 00:50:10,077 particular goal. Um and so um so now if 906 00:50:10,160 --> 00:50:15,037 you have this world model we can imagine a sequence of actions predict what the 907 00:50:15,120 --> 00:50:16,637 outcome of the sequence of action is 908 00:50:16,720 --> 00:50:20,317 going to be measure to what extent the 909 00:50:20,400 --> 00:50:22,637 final state satisfies a particular 910 00:50:22,720 --> 00:50:28,677 objective like you know moving the bottle to the left of the table 911 00:50:28,760 --> 00:50:31,117 >> [snorts] >> um and then plan a sequence of actions 912 00:50:31,200 --> 00:50:34,797 that will minimize this objective at runtime. We're not talking about 913 00:50:34,880 --> 00:50:38,957 learning. talking about inference time, right? So, this is planning really. And 914 00:50:39,040 --> 00:50:42,397 in optimal control, this is a very classical thing. It's called uh model 915 00:50:42,480 --> 00:50:46,797 predictive control. You have a model of the system you want to control that you 916 00:50:46,880 --> 00:50:51,757 know can predict the sequence of states corresponding to a sequence of commands 917 00:50:51,840 --> 00:50:53,757 and you're planning a sequence of 918 00:50:53,840 --> 00:50:56,637 commands so that according to your world 919 00:50:56,720 --> 00:50:58,797 model the the the end state of the 920 00:50:58,880 --> 00:51:02,797 system will uh satisfy an objective that 921 00:51:02,880 --> 00:51:05,997 you fix. This is the way 922 00:51:06,080 --> 00:51:10,237 uh you know rocket trajectories have been planned since computers have been 923 00:51:10,320 --> 00:51:12,797 around so since the early 60s essentially. 924 00:51:12,880 --> 00:51:18,157 >> So yes for model predictive control but you also often talk about hierarchical 925 00:51:18,240 --> 00:51:21,677 planning. >> Yeah. >> Can hierarchical planning emerge from this somehow? 926 00:51:21,760 --> 00:51:26,477 >> Well so no you you will have to build specific architecture to allow for 927 00:51:26,560 --> 00:51:29,917 hierarchical planning. So hierarchical planning is absolutely necessary if you 928 00:51:30,000 --> 00:51:35,117 want to plan complex actions. uh if I want to go from let's say from 929 00:51:35,200 --> 00:51:36,797 New York to Paris, it's the example I 930 00:51:36,880 --> 00:51:39,597 use all the time and I'm sitting uh in 931 00:51:39,680 --> 00:51:42,557 my office at NYU. My objective that I 932 00:51:42,640 --> 00:51:43,997 need to minimize is my distance to 933 00:51:44,080 --> 00:51:47,117 Paris. At a high level, a very abstract 934 00:51:47,200 --> 00:51:50,157 representation of my uh my location, I 935 00:51:50,240 --> 00:51:54,637 would have to decompose this into two sub goals. First one is um go to the 936 00:51:54,720 --> 00:51:57,357 airport. Second one is catch your plane to Paris. 937 00:51:57,440 --> 00:52:00,157 >> Okay. So my sub goal is now 938 00:52:00,240 --> 00:52:02,557 uh going to the airport. My objective 939 00:52:02,640 --> 00:52:05,357 function is my distance to the airport. 940 00:52:05,440 --> 00:52:07,277 How do I go to the airport where I have 941 00:52:07,360 --> 00:52:10,957 to go in the street and hail the taxi 942 00:52:11,040 --> 00:52:14,157 >> which you can do in New York? Um, okay. 943 00:52:14,240 --> 00:52:18,957 Now, I have another sub goal. Go down on the street. Uh, well, that means uh 944 00:52:19,040 --> 00:52:21,117 going to the elevator. Going down the 945 00:52:21,200 --> 00:52:23,997 elevator, walk out the street. How do I 946 00:52:24,080 --> 00:52:27,837 go to the elevator? I have to uh 947 00:52:27,920 --> 00:52:32,397 stand up from my chair, open the door of my office, go to the elevator, push push 948 00:52:32,480 --> 00:52:34,317 the button. How do I get up from my 949 00:52:34,400 --> 00:52:39,677 chair? Like, you know, you can imagine going down all the way down to basically 950 00:52:39,760 --> 00:52:43,197 what amounts to millisecond by millisecond muscle control. 951 00:52:43,280 --> 00:52:46,157 >> Okay? And obviously, you're not going to 952 00:52:46,240 --> 00:52:48,397 plan your entire trip from New York to 953 00:52:48,480 --> 00:52:53,357 Paris in terms of millisecond by millisecond muscle control. First, that 954 00:52:53,440 --> 00:52:56,557 would be incredibly expensive. But it will also be completely impossible 955 00:52:56,640 --> 00:53:00,797 because you don't know all the conditions of what's going to happen. Uh 956 00:53:00,880 --> 00:53:02,317 you know how long it's going to take to 957 00:53:02,400 --> 00:53:05,757 catch a taxi um or to go to the airport 958 00:53:05,840 --> 00:53:10,157 with traffic you know uh I mean you you would have to know exactly the condition 959 00:53:10,240 --> 00:53:13,597 of everything to be able to do this planning and you don't have the 960 00:53:13,680 --> 00:53:17,357 information. So you you have to do this hierarchical planning so that you can 961 00:53:17,440 --> 00:53:19,437 start acting and then sort of replanning 962 00:53:19,520 --> 00:53:23,277 as you go. And nobody really knows how 963 00:53:23,360 --> 00:53:27,277 to do this in AI. Um nobody knows how to 964 00:53:27,360 --> 00:53:29,757 train a system to learn the appropriate 965 00:53:29,840 --> 00:53:34,317 multiple levels of representation so that hierarchical planning works. Does 966 00:53:34,400 --> 00:53:35,997 something like that already emerge? So, 967 00:53:36,080 --> 00:53:39,517 like, can you use an LLM, 968 00:53:39,600 --> 00:53:42,317 state-of-the-art LLM to get you from New 969 00:53:42,400 --> 00:53:46,157 York to Paris by doing exactly the kind 970 00:53:46,240 --> 00:53:48,237 of detailed set of questions that you 971 00:53:48,320 --> 00:53:50,477 just did, which is, can you give me a 972 00:53:50,560 --> 00:53:54,157 high a list of 10 steps I need to do to 973 00:53:54,240 --> 00:53:58,957 get from New York to Paris? And then for each of those steps, can you give me a 974 00:53:59,040 --> 00:54:03,277 list of 10 steps how I make that step happen? And for each of those steps, can 975 00:54:03,360 --> 00:54:06,637 you give me a list of 10 steps to make each one of those until you're moving 976 00:54:06,720 --> 00:54:09,597 your mus individual muscles? Uh maybe 977 00:54:09,680 --> 00:54:14,237 not whatever you can actually act upon using your own mind. 978 00:54:14,320 --> 00:54:17,357 >> Right? So there's a lot of questions that are sort of implied by this. Right? 979 00:54:17,440 --> 00:54:19,837 So the first thing is uh LLM will be 980 00:54:19,920 --> 00:54:25,117 able to answer some of those questions down to some level of abstraction 981 00:54:25,200 --> 00:54:29,037 >> under the condition that they've been trained with similar scenarios in their 982 00:54:29,120 --> 00:54:30,877 training set. They would be able to 983 00:54:30,960 --> 00:54:36,317 answer all of those questions, but some of them may be hallucinated, meaning 984 00:54:36,400 --> 00:54:38,237 non-factual. >> Yeah, true. I mean, they will probably 985 00:54:38,320 --> 00:54:40,877 produce some answer, except they're not going to be able to really kind of 986 00:54:40,960 --> 00:54:44,397 produce millisecond by millisecond muscle control of how you how you stand 987 00:54:44,480 --> 00:54:47,757 up from your chair, >> right? So, but down to some level of 988 00:54:47,840 --> 00:54:51,357 abstraction where you can describe things by words, they might be able to 989 00:54:51,440 --> 00:54:54,317 give you a plan, but only under the condition that they've been trained to 990 00:54:54,400 --> 00:54:56,477 produce those kind of plans. >> Mhm. 991 00:54:56,560 --> 00:55:00,797 >> Right. They're not going to be able to plan for situations where that that they 992 00:55:00,880 --> 00:55:04,237 never encountered before. They basically are going to have to regurgitate a 993 00:55:04,320 --> 00:55:05,757 template that they've been trained on. 994 00:55:05,840 --> 00:55:07,837 >> But where like just for the example of 995 00:55:07,920 --> 00:55:09,837 New York to Paris is is going to start 996 00:55:09,920 --> 00:55:14,877 getting into trouble like at which layer layer of abstraction do you think you'll 997 00:55:14,960 --> 00:55:18,797 start cuz like I can imagine almost every single part of that an LLM would 998 00:55:18,880 --> 00:55:21,837 be able to answer somewhat accurately especially when you're talking about New 999 00:55:21,920 --> 00:55:24,397 York and Paris major cities. So I mean 1000 00:55:24,480 --> 00:55:27,037 certainly uh LM would be able to solve 1001 00:55:27,120 --> 00:55:29,357 that problem if you fine tune it for it 1002 00:55:29,440 --> 00:55:33,677 you know just uh and and so uh I can't 1003 00:55:33,760 --> 00:55:37,197 say that NLM cannot do this it can do this if you train it for it there's no 1004 00:55:37,280 --> 00:55:40,637 question uh down to a certain level 1005 00:55:40,720 --> 00:55:45,117 where things can be formulated in terms of words but like if you want to go down 1006 00:55:45,200 --> 00:55:48,637 to like how do you you know climb down the stairs or just stand up from your 1007 00:55:48,720 --> 00:55:51,117 chair in terms of uh words like you you 1008 00:55:51,200 --> 00:55:54,157 can't you can't do it Um you you you 1009 00:55:54,240 --> 00:55:59,197 need that's one of the reasons you need experience of the physical world which 1010 00:55:59,280 --> 00:56:03,757 is much higher bandwidth than what you can express in words in human language. 1011 00:56:03,840 --> 00:56:08,157 So everything we've been talking about on the joint embedding space, is it 1012 00:56:08,240 --> 00:56:11,677 possible that that's what we need for like interaction with physical reality 1013 00:56:11,760 --> 00:56:14,077 for on the robotics front and then just 1014 00:56:14,160 --> 00:56:20,397 the LLMs are the thing that sits on top of it for the bigger reasoning about like 1015 00:56:20,480 --> 00:56:22,957 >> Yeah. >> the fact that I need to book a plane 1016 00:56:23,040 --> 00:56:26,477 ticket and I need to know I know how to go to the websites and so on. 1017 00:56:26,560 --> 00:56:28,877 >> Sure. And you know a lot of plans that 1018 00:56:28,960 --> 00:56:32,077 people know about um that are relatively 1019 00:56:32,160 --> 00:56:33,757 high level are actually learned. They're 1020 00:56:33,840 --> 00:56:36,237 not people mo most people don't invent 1021 00:56:36,320 --> 00:56:42,157 the you know plans um uh they they 1022 00:56:42,240 --> 00:56:44,397 by themselves. they uh you know we have 1023 00:56:44,480 --> 00:56:46,477 some ability to do this of course uh 1024 00:56:46,560 --> 00:56:51,757 obviously but um but most plans that people use are plants that they've been 1025 00:56:51,840 --> 00:56:54,477 trained on like they've seen other people use those plans or they've been 1026 00:56:54,560 --> 00:56:57,197 told how to do things right um like you 1027 00:56:57,280 --> 00:57:02,557 can't invent how you like take a person who's never heard of airplanes and tell 1028 00:57:02,640 --> 00:57:06,557 them like how do you go from New York to Paris and they're probably not going to 1029 00:57:06,640 --> 00:57:09,837 be able to kind of you know deconstruct the whole plan unless they've seen 1030 00:57:09,920 --> 00:57:12,077 examples of that before um so certainly 1031 00:57:12,160 --> 00:57:14,157 LM going to be able to do this. But but 1032 00:57:14,240 --> 00:57:17,677 then um how you link this from the the 1033 00:57:17,760 --> 00:57:21,917 low level of of of actions uh that needs 1034 00:57:22,000 --> 00:57:25,997 to be done with things like like JPA that basically lift the abstraction 1035 00:57:26,080 --> 00:57:29,357 level of the representation without attempting to reconstruct every detail 1036 00:57:29,440 --> 00:57:33,437 of the situation. That's why we need JPA for 1037 00:57:33,520 --> 00:57:36,317 >> I would love to sort of linger on your 1038 00:57:36,400 --> 00:57:38,717 skepticism around 1039 00:57:38,800 --> 00:57:41,277 uh autogressive LLMs. 1040 00:57:41,360 --> 00:57:44,077 So one way I I would like to test that 1041 00:57:44,160 --> 00:57:50,280 skepticism is everything you say makes a lot of sense. 1042 00:57:50,480 --> 00:57:53,277 But if I apply everything you said today 1043 00:57:53,360 --> 00:57:56,397 and in general to like I don't know 10 1044 00:57:56,480 --> 00:58:01,357 years ago maybe a little bit less no let's say 3 years ago I wouldn't be able 1045 00:58:01,440 --> 00:58:03,197 to predict 1046 00:58:03,280 --> 00:58:07,437 the uh success of LLMs. So, does it make 1047 00:58:07,520 --> 00:58:10,077 sense to you that autogressive LLMs are 1048 00:58:10,160 --> 00:58:13,437 able to be so damn good? 1049 00:58:13,520 --> 00:58:17,437 >> Yes. >> Can you explain your intuition? Because 1050 00:58:17,520 --> 00:58:21,837 if I were to take your wisdom and intuition 1051 00:58:21,920 --> 00:58:26,557 at face value, I would say there's no way autogressive LLMs, one token at a 1052 00:58:26,640 --> 00:58:29,677 time, would be able to do the kind of things they're doing. No, there's one 1053 00:58:29,760 --> 00:58:34,557 thing that autogressive LLM uh or that LLM in general, not just the auto 1054 00:58:34,640 --> 00:58:36,637 reggressive one, but including the bird 1055 00:58:36,720 --> 00:58:40,557 style birectional ones u are exploiting 1056 00:58:40,640 --> 00:58:43,917 and it's self-supervised learning and I've been a very very strong advocate of 1057 00:58:44,000 --> 00:58:46,157 self-supervised learning for many years. 1058 00:58:46,240 --> 00:58:50,557 So those things are a incredibly 1059 00:58:50,640 --> 00:58:52,237 impressive demonstration that self-s 1060 00:58:52,320 --> 00:58:54,877 supervised learning actually works. uh 1061 00:58:54,960 --> 00:59:00,637 the idea that you know started uh it didn't start with with uh with BERT but 1062 00:59:00,720 --> 00:59:05,037 it was really kind of a good demonstration with this. So the the the 1063 00:59:05,120 --> 00:59:07,917 idea that you know you take a piece of text you corrupt it and then you train 1064 00:59:08,000 --> 00:59:09,517 some gigantic neural net to reconstruct 1065 00:59:09,600 --> 00:59:12,317 the parts that are missing. Um that has 1066 00:59:12,400 --> 00:59:16,237 been an enormous uh uh 1067 00:59:16,320 --> 00:59:18,077 produced an enormous amount of benefits. 1068 00:59:18,160 --> 00:59:20,317 Uh it allowed allowed us to create 1069 00:59:20,400 --> 00:59:22,797 systems that understand understand 1070 00:59:22,880 --> 00:59:26,157 language. Uh systems that can translate 1071 00:59:26,240 --> 00:59:29,837 um hundreds of languages in any direction. Systems that are 1072 00:59:29,920 --> 00:59:33,757 multilingual. So they're not it's a single system that can be trained to 1073 00:59:33,840 --> 00:59:36,077 understand hundreds of languages and 1074 00:59:36,160 --> 00:59:39,357 translate in any direction. Um and 1075 00:59:39,440 --> 00:59:42,397 produce summaries um and then answer 1076 00:59:42,480 --> 00:59:46,717 questions and produce text. And then there's a special case of it where you 1077 00:59:46,800 --> 00:59:48,797 know you which is the auto reggressive 1078 00:59:48,880 --> 00:59:51,277 uh trick where you constrain the system 1079 00:59:51,360 --> 00:59:54,477 to not elaborate a representation of the 1080 00:59:54,560 --> 00:59:58,477 text from looking at the entire text but only 1081 00:59:58,560 --> 01:00:02,157 predicting a word from the words that are come before right and you do this by 1082 01:00:02,240 --> 01:00:05,917 the constraining the architecture of the network and that's what you can build an 1083 01:00:06,000 --> 01:00:08,477 auto reggressive LLM from. So there was 1084 01:00:08,560 --> 01:00:11,837 a surprise many years ago with what's 1085 01:00:11,920 --> 01:00:14,797 called decoder only LLM. So since you 1086 01:00:14,880 --> 01:00:16,397 know systems of this type that are just 1087 01:00:16,480 --> 01:00:19,757 trying to produce uh words from the from 1088 01:00:19,840 --> 01:00:22,077 the previous one and and the fact that 1089 01:00:22,160 --> 01:00:26,397 when you scale them up they they tend to 1090 01:00:26,480 --> 01:00:30,317 really kind of understand more about the about language when you train them on 1091 01:00:30,400 --> 01:00:33,517 lots of data and you make them really big. That was kind of a surprise and 1092 01:00:33,600 --> 01:00:35,357 that surprise occurred quite a while 1093 01:00:35,440 --> 01:00:40,317 back like you know uh with uh work from 1094 01:00:40,400 --> 01:00:43,117 uh you know Google Meta, OpenAI etc. you 1095 01:00:43,200 --> 01:00:48,797 know going back to you know the GPT kind of uh work general pre-trained 1096 01:00:48,880 --> 01:00:51,837 transformers. >> You mean like GPT2 like there's a 1097 01:00:51,920 --> 01:00:54,237 certain place where you start to realize 1098 01:00:54,320 --> 01:00:58,157 scaling might actually keep giving us an 1099 01:00:58,240 --> 01:01:01,597 emergent benefit. Yeah, I mean there were there were work from from various 1100 01:01:01,680 --> 01:01:05,117 places but uh uh if if you want to kind 1101 01:01:05,200 --> 01:01:10,397 of you know place it in the in the GPT uh timeline that would be around GPT2 1102 01:01:10,480 --> 01:01:14,397 you know. >> Well I just cuz you said it you you're 1103 01:01:14,480 --> 01:01:18,637 so charismatic and you said so many words but self-supervised learning. Yes. 1104 01:01:18,720 --> 01:01:20,957 But again the same intuition you're 1105 01:01:21,040 --> 01:01:23,757 applying to saying that autogressive 1106 01:01:23,840 --> 01:01:26,637 LLMs cannot have a deep understanding of 1107 01:01:26,720 --> 01:01:29,837 the world. If we just apply that same 1108 01:01:29,920 --> 01:01:33,117 intuition, does it make sense to you 1109 01:01:33,200 --> 01:01:37,517 that they're able to form enough of a representation of the world to be damn 1110 01:01:37,600 --> 01:01:42,637 convincing? Essentially passing the original touring test with 1111 01:01:42,720 --> 01:01:45,597 flying colors. >> Well, we're fooled by their fluency, 1112 01:01:45,680 --> 01:01:47,757 right? We just assume that if a system 1113 01:01:47,840 --> 01:01:50,637 is fluent in manipulating language, then 1114 01:01:50,720 --> 01:01:52,877 it has all the characteristics of human 1115 01:01:52,960 --> 01:01:58,237 intelligence. But that impression is false. We we're we're really fooled by 1116 01:01:58,320 --> 01:02:01,357 it. Um >> what do you think Alan Turing would say 1117 01:02:01,440 --> 01:02:04,077 it without understanding anything just hanging out with it? 1118 01:02:04,160 --> 01:02:07,837 >> Alan Turing would decide that a touring test is a really bad test. 1119 01:02:07,920 --> 01:02:12,157 >> Okay, this is what the AI community has decided many years ago that the touring 1120 01:02:12,240 --> 01:02:14,637 test was a really bad test of intelligence. 1121 01:02:14,720 --> 01:02:18,717 >> What would Hans Marvc say about the about the large language models? 1122 01:02:18,800 --> 01:02:22,477 >> Well, Hence Marv would say the Marvc paradox still applies. 1123 01:02:22,560 --> 01:02:26,317 >> Okay. >> Okay. [snorts] Okay. We can pass. I don't think he would be really 1124 01:02:26,400 --> 01:02:28,077 impressed. >> No, of course everybody would be 1125 01:02:28,160 --> 01:02:32,957 impressed. But uh you know uh it's not a question of being impressed or not. It's 1126 01:02:33,040 --> 01:02:37,197 a question of knowing what the limit of those systems can do. Like there again 1127 01:02:37,280 --> 01:02:40,637 they are impressive. They can do a lot of useful things. There's a whole 1128 01:02:40,720 --> 01:02:44,637 industry that is being built around them. They're going to make progress. Uh 1129 01:02:44,720 --> 01:02:47,677 but there is a lot of things they cannot do and we have to realize what they 1130 01:02:47,760 --> 01:02:51,597 cannot do and uh and then figure out you 1131 01:02:51,680 --> 01:02:53,357 know how we get there. And you know and 1132 01:02:53,440 --> 01:02:55,917 and I'm not seeing this I'm seeing this 1133 01:02:56,000 --> 01:02:58,797 from basically you know 10 years of of 1134 01:02:58,880 --> 01:03:03,277 research uh on on 1135 01:03:03,360 --> 01:03:04,957 the idea of self-s supervised learning. 1136 01:03:05,040 --> 01:03:07,677 Actually that's going back more than 10 years but the idea of self-s supervised 1137 01:03:07,760 --> 01:03:09,917 learning. So basically capturing the 1138 01:03:10,000 --> 01:03:14,877 internal structure of a piece of uh of a set of inputs without training the 1139 01:03:14,960 --> 01:03:18,307 system for any particular task right learning representations. 1140 01:03:18,390 --> 01:03:20,397 >> [snorts] >> um you know the the conference I 1141 01:03:20,480 --> 01:03:23,917 co-ounded 14 years ago is called inter international conference on learning 1142 01:03:24,000 --> 01:03:27,917 representations that's the entire issue that deep learning is is dealing with 1143 01:03:28,000 --> 01:03:32,717 right and it's been my obsession for you know almost 40 years now so um so 1144 01:03:32,800 --> 01:03:36,557 learning representation is really the thing uh for the longest time we could 1145 01:03:36,640 --> 01:03:40,717 only do this with supervised learning and then we started working on uh you 1146 01:03:40,800 --> 01:03:44,957 know what we used to call unsupervised learning [snorts] uh 1147 01:03:45,040 --> 01:03:48,237 and sort of revive the idea of unsupervised learning 1148 01:03:48,320 --> 01:03:53,197 in the early 2000s with Yosha Benju and Jeffon then discovered that supervised 1149 01:03:53,280 --> 01:03:57,837 learning actually works pretty well if you can collect enough data and so the 1150 01:03:57,920 --> 01:03:59,597 whole idea of you know unsupervised 1151 01:03:59,680 --> 01:04:02,157 self-supervised learning kind of took a 1152 01:04:02,240 --> 01:04:04,957 a backseat for for a bit and then I kind 1153 01:04:05,040 --> 01:04:09,357 of tried to revive it um uh in a big way 1154 01:04:09,440 --> 01:04:11,837 you know starting in 2014 basically when 1155 01:04:11,920 --> 01:04:15,277 we started fair and uh and really 1156 01:04:15,360 --> 01:04:19,597 pushing for like finding new new methods to do self-supervised learning both for 1157 01:04:19,680 --> 01:04:23,757 text and for images and for video and audio and some of that work has been 1158 01:04:23,840 --> 01:04:26,397 incredibly successful I mean the reason 1159 01:04:26,480 --> 01:04:28,877 why we have multilingual translation 1160 01:04:28,960 --> 01:04:33,517 system you know things to do content moderation on on meta for example on 1161 01:04:33,600 --> 01:04:37,597 Facebook that are multilingual that understand whether piece of text is HP 1162 01:04:37,680 --> 01:04:41,277 speech or not or something is due to that progress using self-supervised 1163 01:04:41,360 --> 01:04:45,917 learning for NLP combining this with you know transformer architectures and and 1164 01:04:46,000 --> 01:04:48,957 blah blah blah So that's the big success of self-supervised learning. We had 1165 01:04:49,040 --> 01:04:53,357 similar success in speech recognition, a system called wave tovec, which is also 1166 01:04:53,440 --> 01:04:56,557 a joint embedding architecture by the way, trained with contrastive learning 1167 01:04:56,640 --> 01:04:58,957 and and that that system also can 1168 01:04:59,040 --> 01:05:01,917 produce um speech recognition systems 1169 01:05:02,000 --> 01:05:06,717 that are multilingual with mostly unlabelled data and only need a few 1170 01:05:06,800 --> 01:05:10,317 minutes of label data to actually do speech recognition. That's that's 1171 01:05:10,400 --> 01:05:13,597 amazing. Um we have systems now based on 1172 01:05:13,680 --> 01:05:17,917 those combination of ideas that can do realtime translation of hundreds of 1173 01:05:18,000 --> 01:05:20,877 languages into each other. Uh speech to speech 1174 01:05:20,960 --> 01:05:23,357 >> speech to speech even including which is 1175 01:05:23,440 --> 01:05:27,677 fascinating languages that uh don't have written forms. That's right. They're 1176 01:05:27,760 --> 01:05:29,517 spoken only. >> That's right. We don't go through text. 1177 01:05:29,600 --> 01:05:32,957 It goes directly from from speech to speech using an internal representation 1178 01:05:33,040 --> 01:05:36,397 of kind of speech units that are discrete. But it's um it's called 1179 01:05:36,480 --> 01:05:40,717 textless and LP. We used to call it this way but um yeah. So that I mean 1180 01:05:40,800 --> 01:05:44,877 incredible success there and then you know for 10 years we tried to apply this 1181 01:05:44,960 --> 01:05:47,677 idea to learning representations of 1182 01:05:47,760 --> 01:05:51,917 images by training a system to predict videos learning intuitive physics by 1183 01:05:52,000 --> 01:05:56,397 training a system to predict what's going to happen in the video and tried 1184 01:05:56,480 --> 01:06:00,157 and tried and failed and failed with generative models with models that 1185 01:06:00,240 --> 01:06:04,317 predict pixels. Uh we could not get them to learn good 1186 01:06:04,400 --> 01:06:07,997 representations of images. We could not get them to learn good representations 1187 01:06:08,080 --> 01:06:11,997 of videos. And we tried many times. We published lots of papers on it. You 1188 01:06:12,080 --> 01:06:13,757 know, they kind of sort of worked but 1189 01:06:13,840 --> 01:06:16,717 not really great. They started working. 1190 01:06:16,800 --> 01:06:20,237 We we abandoned this idea of predicting 1191 01:06:20,320 --> 01:06:23,917 every pixel and basically just doing the joint embedding and predicting in 1192 01:06:24,000 --> 01:06:26,317 representation space that works. 1193 01:06:26,400 --> 01:06:29,917 >> Mhm. >> So there's ample evidence that we're not 1194 01:06:30,000 --> 01:06:31,837 going to be able to learn good 1195 01:06:31,920 --> 01:06:37,277 representations of the real world using generative model. So I'm telling people 1196 01:06:37,360 --> 01:06:40,957 everybody is talking about generative AI. If you're really interested in human 1197 01:06:41,040 --> 01:06:44,237 level AI, abandon the idea of generative AI. 1198 01:06:44,320 --> 01:06:46,557 >> Okay. But you you you really think it's 1199 01:06:46,640 --> 01:06:49,117 possible to get far with the joint 1200 01:06:49,200 --> 01:06:51,357 embedding representation. So like 1201 01:06:51,440 --> 01:06:54,397 there's common sense reasoning and then 1202 01:06:54,480 --> 01:06:59,677 there's high level reasoning. Like I I feel like those are 1203 01:06:59,760 --> 01:07:03,997 two the kind of reasoning that LLMs are able 1204 01:07:04,080 --> 01:07:07,357 to do. Okay. Let me not use the word reasoning, but the kind of stuff that 1205 01:07:07,440 --> 01:07:11,437 LMs are able to do seems fundamentally different than the common sense 1206 01:07:11,520 --> 01:07:13,597 reasoning we use to navigate the world. 1207 01:07:13,680 --> 01:07:15,277 >> It seems like we're going to need both. 1208 01:07:15,360 --> 01:07:17,757 You're not sure. Would you be able to get 1209 01:07:17,840 --> 01:07:19,517 >> with the joint embedding? Would you jump 1210 01:07:19,600 --> 01:07:24,587 type of approach looking at video? Would you be able to learn [sighs] 1211 01:07:24,670 --> 01:07:27,117 [gasps] let's see well how to get from New York 1212 01:07:27,200 --> 01:07:30,957 to Paris or um 1213 01:07:31,040 --> 01:07:34,397 how to uh state understand the state of 1214 01:07:34,480 --> 01:07:38,157 politics in the world today [laughter] right these these are things where 1215 01:07:38,240 --> 01:07:40,397 various humans generate a lot of 1216 01:07:40,480 --> 01:07:42,477 language and opinions on in the space of 1217 01:07:42,560 --> 01:07:45,357 language but don't visually represent 1218 01:07:45,440 --> 01:07:50,717 that in any clearly uh compressible way >> right well there's a lot of situations 1219 01:07:50,800 --> 01:07:52,317 that you know might 1220 01:07:52,400 --> 01:07:55,197 difficult to for a purely language based 1221 01:07:55,280 --> 01:07:59,277 system to um to know like okay you can 1222 01:07:59,360 --> 01:08:03,197 probably learn from reading text the entirety of the public publicly 1223 01:08:03,280 --> 01:08:06,637 available text in the world that I cannot get from New York to Paris by 1224 01:08:06,720 --> 01:08:10,077 snapping my fingers >> that's not going to work right >> yes 1225 01:08:10,160 --> 01:08:12,397 >> uh but there's you know probably sort of 1226 01:08:12,480 --> 01:08:15,117 more complex scenarios of this type 1227 01:08:15,200 --> 01:08:18,397 which an NLM may never have encountered 1228 01:08:18,480 --> 01:08:20,637 and may not be able to determine whether 1229 01:08:20,720 --> 01:08:23,917 it's possible or not. Um so 1230 01:08:24,000 --> 01:08:28,397 >> um so that that link you know from the the low level to the high level the the 1231 01:08:28,480 --> 01:08:29,677 thing is that the high level that 1232 01:08:29,760 --> 01:08:33,517 language expresses is based on a common 1233 01:08:33,600 --> 01:08:36,797 experience of the low level which LLMs 1234 01:08:36,880 --> 01:08:41,677 currently do not have. you know we when we talk to each other we know we have a 1235 01:08:41,760 --> 01:08:43,357 common experience of the of the world 1236 01:08:43,440 --> 01:08:45,917 like you know a lot of it is 1237 01:08:46,000 --> 01:08:50,957 >> is similar u and 1238 01:08:51,040 --> 01:08:54,797 don't have that but see there it's present you and I have a common 1239 01:08:54,880 --> 01:08:56,157 experience of the world in terms of the 1240 01:08:56,240 --> 01:08:58,957 physics of how gravity works and stuff 1241 01:08:59,040 --> 01:09:01,917 like this and that 1242 01:09:02,000 --> 01:09:04,877 common knowledge of the world I feel 1243 01:09:04,960 --> 01:09:08,877 like is there in the language we don't 1244 01:09:08,960 --> 01:09:11,277 explicitly express it. But if you have a 1245 01:09:11,360 --> 01:09:15,037 huge amount of text, you're going to get 1246 01:09:15,120 --> 01:09:16,877 this stuff that's between the lines. 1247 01:09:16,960 --> 01:09:21,837 You're going you're going to in order to um form a consistent world model, you're 1248 01:09:21,920 --> 01:09:23,837 going to have to understand how gravity 1249 01:09:23,920 --> 01:09:29,037 works even if you don't have an explicit explanation of gravity. So even though 1250 01:09:29,120 --> 01:09:30,477 there in the case of gravity, there is 1251 01:09:30,560 --> 01:09:31,997 explicit explanations of gravity on 1252 01:09:32,080 --> 01:09:36,317 Wikipedia. But uh you're like the stuff 1253 01:09:36,400 --> 01:09:40,557 that we think of as common sense reasoning. I feel like to generate 1254 01:09:40,640 --> 01:09:45,437 language correctly, you're going to have to figure that out. Now, you could say 1255 01:09:45,520 --> 01:09:47,437 as you there's not enough text. Sorry. 1256 01:09:47,520 --> 01:09:50,477 Okay. So, what [laughter] you don't think so? 1257 01:09:50,560 --> 01:09:53,597 >> No, I agree with what you just said, which is that to be able to do high 1258 01:09:53,680 --> 01:09:58,717 level um common sense, to have high level common sense, you need to have the 1259 01:09:58,800 --> 01:10:00,797 low-level common sense to build on top of. 1260 01:10:00,880 --> 01:10:04,797 >> Yeah. Um >> but that's not there >> and that's not there in LLMs. LLMs are 1261 01:10:04,880 --> 01:10:08,797 purely trained from text. So so then the other statement you made um I would not 1262 01:10:08,880 --> 01:10:11,117 I would not agree with the fact that 1263 01:10:11,200 --> 01:10:16,077 implicit in all languages in the world is the underlying reality. There's a lot 1264 01:10:16,160 --> 01:10:19,597 about underlying reality which is not expressed in language. 1265 01:10:19,680 --> 01:10:20,797 >> Is that obvious to you? 1266 01:10:20,880 --> 01:10:23,037 >> Yeah, totally. 1267 01:10:23,120 --> 01:10:26,637 So like all all the conversations we 1268 01:10:26,720 --> 01:10:31,197 have what okay there's the dark web meaning uh whatever the private 1269 01:10:31,280 --> 01:10:32,877 conversations like DMs and stuff like 1270 01:10:32,960 --> 01:10:38,797 this which is much much larger probably than what's available what what LLMs are 1271 01:10:38,880 --> 01:10:42,797 trained on >> you don't need to communicate the stuff that is common 1272 01:10:42,880 --> 01:10:47,037 >> but the humor all of it no you do like when you you don't need to but it comes 1273 01:10:47,120 --> 01:10:49,757 through like you like if I accidentally 1274 01:10:49,840 --> 01:10:51,757 uh knock this over you'll probably make 1275 01:10:51,840 --> 01:10:54,077 fun of me and in the content of the you 1276 01:10:54,160 --> 01:10:57,437 making fun of me will be uh explanation 1277 01:10:57,520 --> 01:11:00,877 of the fact that cups fall and then you 1278 01:11:00,960 --> 01:11:04,957 know gravity works in this way and then you you'll have some very vague 1279 01:11:05,040 --> 01:11:09,677 information about what kind of things explode when they hit the ground and 1280 01:11:09,760 --> 01:11:12,957 then maybe you'll make a joke about entropy or something like this and we'll 1281 01:11:13,040 --> 01:11:17,357 never be able to reconstruct this again like okay you'll make a a little joke 1282 01:11:17,440 --> 01:11:21,197 like this and there'll be trillion of other jokes and from the jokes you can 1283 01:11:21,280 --> 01:11:25,117 piece together. The fact that gravity works and mugs can break and all this 1284 01:11:25,200 --> 01:11:28,077 kind of stuff. You don't need to see uh 1285 01:11:28,160 --> 01:11:32,317 it'll be very inefficient. It's easier for like 1286 01:11:32,400 --> 01:11:34,877 not knock the thing over. Yeah. 1287 01:11:34,960 --> 01:11:37,277 >> But uh I feel like it would be there if 1288 01:11:37,360 --> 01:11:39,357 you have enough of that data. 1289 01:11:39,440 --> 01:11:42,477 I just think that most of the 1290 01:11:42,560 --> 01:11:45,037 information of this type that we have 1291 01:11:45,120 --> 01:11:47,357 accumulated when when we were babies is 1292 01:11:47,440 --> 01:11:51,037 just not present in uh in in text any in 1293 01:11:51,120 --> 01:11:55,037 any description essentially >> and the sensory data is much is a much 1294 01:11:55,120 --> 01:11:58,477 richer source for getting that kind of understanding. I mean that's the 16,000 1295 01:11:58,560 --> 01:12:01,757 hours of of wake time of a 4-year-old 1296 01:12:01,840 --> 01:12:06,397 and 10 to the 15 bytes you know going through vision just vision right there 1297 01:12:06,480 --> 01:12:08,477 is a similar 1298 01:12:08,560 --> 01:12:11,517 uh bandwidth you know of touch and uh a 1299 01:12:11,600 --> 01:12:13,837 little less through audio and then text 1300 01:12:13,920 --> 01:12:15,597 doesn't language doesn't come in until 1301 01:12:15,680 --> 01:12:19,677 like you know a year uh in in life and 1302 01:12:19,760 --> 01:12:24,157 by the time you are 9 years old you've learned about gravity you know about 1303 01:12:24,240 --> 01:12:27,757 inertia you know about gravity you know the stability you 1304 01:12:27,840 --> 01:12:29,277 you know about the distinction between 1305 01:12:29,360 --> 01:12:31,597 animate and inanimate objects. you know 1306 01:12:31,680 --> 01:12:34,957 by 18 months you know about like uh why 1307 01:12:35,040 --> 01:12:38,957 people want to do things and you help them if they can't you know I mean 1308 01:12:39,040 --> 01:12:42,477 there's a lot of things that you learn mostly by observation 1309 01:12:42,560 --> 01:12:46,637 >> really uh not even through interaction in the first few months of life babies 1310 01:12:46,720 --> 01:12:50,717 don't don't really have any influence on the world they can only observe right 1311 01:12:50,800 --> 01:12:55,037 and [snorts] you accumulate like a gigantic amount of uh of knowledge just 1312 01:12:55,120 --> 01:13:00,157 just from that so that that's what we're missing from uh current AI systems 1313 01:13:00,240 --> 01:13:01,757 >> I I think in one of your slides you have 1314 01:13:01,840 --> 01:13:04,557 this nice plot that is one of the ways 1315 01:13:04,640 --> 01:13:08,877 you show that LLMs are limited. I wonder if you could talk about hallucinations 1316 01:13:08,960 --> 01:13:11,037 from your perspectives. [snorts] 1317 01:13:11,120 --> 01:13:15,437 The why hallucinations happen from large 1318 01:13:15,520 --> 01:13:20,477 language models and why and to what degree is that a fundamental flaw of 1319 01:13:20,560 --> 01:13:22,237 large language models. 1320 01:13:22,320 --> 01:13:24,957 >> Right? So because of the auto 1321 01:13:25,040 --> 01:13:27,597 reggressive prediction, every time an 1322 01:13:27,680 --> 01:13:31,277 L&M produces a token or a word, uh there 1323 01:13:31,360 --> 01:13:35,677 is some level of probability for that word to take you out of the set of 1324 01:13:35,760 --> 01:13:39,997 reasonable answers. Uh and if you assume which is a very 1325 01:13:40,080 --> 01:13:42,317 strong assumption that the probability 1326 01:13:42,400 --> 01:13:44,717 of such error 1327 01:13:44,800 --> 01:13:47,997 um is that those errors are independent 1328 01:13:48,080 --> 01:13:52,237 across a a sequence of tokens being produced. Mhm. 1329 01:13:52,320 --> 01:13:55,357 >> What [clears throat] that means is that every time you produce a token, the 1330 01:13:55,440 --> 01:13:59,117 probability that you rest you you stay within the the set of correct answer 1331 01:13:59,200 --> 01:14:01,357 decreases and it decreases exponentially. 1332 01:14:01,440 --> 01:14:05,837 >> So there's a strong like you said assumption there that if uh there's a 1333 01:14:05,920 --> 01:14:09,757 non-zero probability of making a mistake which there appears to be then there's 1334 01:14:09,840 --> 01:14:11,517 going to be a kind of drift. 1335 01:14:11,600 --> 01:14:14,157 >> Yeah. And that drift is exponential. 1336 01:14:14,240 --> 01:14:16,957 It's like errors accumulate. Right. So, 1337 01:14:17,040 --> 01:14:19,757 so the probability that an answer would 1338 01:14:19,840 --> 01:14:24,157 be nonsensical increases exponentially with the number of tokens. 1339 01:14:24,240 --> 01:14:26,557 >> Is that obvious to you by the way? Like 1340 01:14:26,640 --> 01:14:29,677 well so mathematically speaking maybe. 1341 01:14:29,760 --> 01:14:31,677 But like isn't there a kind of 1342 01:14:31,760 --> 01:14:33,917 gravitational pull towards the truth 1343 01:14:34,000 --> 01:14:37,437 because on on average hopefully the 1344 01:14:37,520 --> 01:14:40,637 truth is well represented in the uh 1345 01:14:40,720 --> 01:14:43,277 training set. No, it's basically a 1346 01:14:43,360 --> 01:14:47,437 struggle against uh the curse of 1347 01:14:47,520 --> 01:14:50,797 dimensionality. So the way you can correct for this is that you fine-tune 1348 01:14:50,880 --> 01:14:55,837 the system by having it produce answers for all kinds of questions that people 1349 01:14:55,920 --> 01:15:00,397 might come up with. >> Mhm. >> And people are people. So they a lot of 1350 01:15:00,480 --> 01:15:03,517 the questions that they have are very similar to each other. So you can 1351 01:15:03,600 --> 01:15:06,397 probably cover you know 80% or whatever 1352 01:15:06,480 --> 01:15:09,917 of questions that people will will ask 1353 01:15:10,000 --> 01:15:13,997 um by you know collecting data and then 1354 01:15:14,080 --> 01:15:18,077 um and then you fine-tune the system to produce good answers for all of those 1355 01:15:18,160 --> 01:15:21,837 things and it's probably going to be able to learn that because it's got a 1356 01:15:21,920 --> 01:15:24,957 lot of capacity to to learn. Um but then 1357 01:15:25,040 --> 01:15:29,597 there is you know the enormous set of 1358 01:15:29,680 --> 01:15:33,997 prompts that you have not covered during training and that set is enormous like 1359 01:15:34,080 --> 01:15:35,917 within the set of all possible prompts 1360 01:15:36,000 --> 01:15:39,037 the proportion of prompts that have been 1361 01:15:39,120 --> 01:15:42,397 uh used for training is absolutely tiny. 1362 01:15:42,480 --> 01:15:47,117 >> Um it's a it's a tiny tiny tiny subset of all possible prompts and so the 1363 01:15:47,200 --> 01:15:50,717 system will behave properly on the prompts that has been either trained 1364 01:15:50,800 --> 01:15:52,877 pre-trained or fine-tuned. 1365 01:15:52,960 --> 01:15:55,917 [snorts] Um but then there is an entire 1366 01:15:56,000 --> 01:15:58,877 space of things that it cannot possibly 1367 01:15:58,960 --> 01:16:00,477 have been trained on because it's just 1368 01:16:00,560 --> 01:16:03,997 the the number is gigantic. So um so 1369 01:16:04,080 --> 01:16:07,037 whatever training the system 1370 01:16:07,120 --> 01:16:09,757 uh has been subject to to produce 1371 01:16:09,840 --> 01:16:12,317 appropriate tensors, you can break it by 1372 01:16:12,400 --> 01:16:17,357 finding out a prompt that will be outside of the the the set of prompts 1373 01:16:17,440 --> 01:16:21,757 has been trained on or things that are similar and then it will just spew 1374 01:16:21,840 --> 01:16:23,597 complete nonsense. Do when you say 1375 01:16:23,680 --> 01:16:26,397 prompt, do you mean that exact prompt or 1376 01:16:26,480 --> 01:16:29,357 do you mean a prompt that's like in many 1377 01:16:29,440 --> 01:16:32,157 parts very different than like is it 1378 01:16:32,240 --> 01:16:36,077 that easy to ask a question or to say a 1379 01:16:36,160 --> 01:16:38,477 thing that hasn't been said before on 1380 01:16:38,560 --> 01:16:43,597 the internet? I mean people have come up with uh things where like you you put a 1381 01:16:43,680 --> 01:16:47,117 essentially a random sequence of characters in a prompt and that's enough 1382 01:16:47,200 --> 01:16:51,917 to kind of throw the system uh into a mode where you know it it's going to 1383 01:16:52,000 --> 01:16:56,077 answer something completely different than it would have answered without this. Mhm. 1384 01:16:56,160 --> 01:16:59,597 >> So that's a way to jailbreak the system basically get it you know go outside of 1385 01:16:59,680 --> 01:17:01,997 its uh of its conditioning right 1386 01:17:02,080 --> 01:17:03,277 >> so that that's a very clear 1387 01:17:03,360 --> 01:17:06,717 demonstration of it but of course 1388 01:17:06,800 --> 01:17:09,517 uh you know that's uh that goes outside 1389 01:17:09,600 --> 01:17:12,397 of what is designed to do right if you 1390 01:17:12,480 --> 01:17:17,597 actually stitch together reasonably grammatical sentences is that is it that 1391 01:17:17,680 --> 01:17:21,757 easy to break it >> yeah some people have done things like 1392 01:17:21,840 --> 01:17:25,997 you you you write a sentence in English right that has an or You ask a question 1393 01:17:26,080 --> 01:17:30,077 in English and it it produces a perfectly fine answer and then you just 1394 01:17:30,160 --> 01:17:35,277 substitute a few words >> by the same word in another language and 1395 01:17:35,360 --> 01:17:37,117 all of a sudden the answer is complete nonsense. 1396 01:17:37,200 --> 01:17:39,277 >> Yeah. So so I guess what I'm saying is 1397 01:17:39,360 --> 01:17:42,237 like which fraction of prompts that 1398 01:17:42,320 --> 01:17:45,677 humans are likely to generate are going 1399 01:17:45,760 --> 01:17:49,277 to break the system. >> So the the problem is that there is a 1400 01:17:49,360 --> 01:17:53,597 long tail. >> Yes. Uh this is a an issue that a lot of 1401 01:17:53,680 --> 01:17:57,677 people have realized you know in social networks and stuff like that which is uh 1402 01:17:57,760 --> 01:18:02,157 there's a very very long tale of of things that people will ask and you can 1403 01:18:02,240 --> 01:18:04,637 fine-tune the system for the 80% or 1404 01:18:04,720 --> 01:18:07,357 whatever of uh of the things that most 1405 01:18:07,440 --> 01:18:11,997 people will will ask and then this long tail is is so large that you're not 1406 01:18:12,080 --> 01:18:15,437 going to be able to fetune the system for all the conditions and in the end 1407 01:18:15,520 --> 01:18:19,757 the system ends up being kind of a giant lookup table right essentially which is 1408 01:18:19,840 --> 01:18:22,877 not really what you want you want systems that can reason certainly that 1409 01:18:22,960 --> 01:18:27,197 can plan. So the type of reasoning that takes place in LLM is very very 1410 01:18:27,280 --> 01:18:29,357 primitive and the reason you can tell is 1411 01:18:29,440 --> 01:18:31,597 primitive is because the amount of 1412 01:18:31,680 --> 01:18:34,397 computation that is spent per token 1413 01:18:34,480 --> 01:18:36,717 produced is constant. 1414 01:18:36,800 --> 01:18:38,957 So if you ask a question and that 1415 01:18:39,040 --> 01:18:43,917 question has an answer in a given number of token the amount of computation 1416 01:18:44,000 --> 01:18:45,837 devoted to computing that answer can be 1417 01:18:45,920 --> 01:18:48,957 exactly estimated. It's like you know 1418 01:18:49,040 --> 01:18:51,677 it's h it's the the size of the 1419 01:18:51,760 --> 01:18:53,917 prediction network you know with it 36 1420 01:18:54,000 --> 01:18:56,637 layers or 92 layers or whatever it is uh 1421 01:18:56,720 --> 01:18:58,877 multiply by number of tokens that's it 1422 01:18:58,960 --> 01:19:01,997 and so essentially it doesn't matter if 1423 01:19:02,080 --> 01:19:06,317 the question being asked is is 1424 01:19:06,400 --> 01:19:09,437 simple to answer complicated to answer 1425 01:19:09,520 --> 01:19:14,397 impossible to answer because it's a decidable or something um the amount of 1426 01:19:14,480 --> 01:19:18,077 computation the system will be able to devote to that to the answer is constant 1427 01:19:18,160 --> 01:19:20,157 or is proportional to number of token 1428 01:19:20,240 --> 01:19:22,477 produced in the answer. Right? This is 1429 01:19:22,560 --> 01:19:25,517 not the way we work. The way we reason 1430 01:19:25,600 --> 01:19:29,117 is that when we're faced with a complex 1431 01:19:29,200 --> 01:19:33,997 problem or complex question, we spend more time trying to solve it and answer 1432 01:19:34,080 --> 01:19:36,477 it, right? >> Because it's more difficult. 1433 01:19:36,560 --> 01:19:38,957 >> There's a prediction element. There's a 1434 01:19:39,040 --> 01:19:43,357 iterative element where you're like uh 1435 01:19:43,440 --> 01:19:47,597 adjusting your understanding of a thing by going o over and over and over. 1436 01:19:47,680 --> 01:19:51,357 There's a hierarchical element. So on does this mean it's a fundamental flaw 1437 01:19:51,440 --> 01:19:54,477 of LLMs or does it mean that [laughter] 1438 01:19:54,560 --> 01:19:57,517 there's more part to that question? 1439 01:19:57,600 --> 01:19:59,817 Now you're just behaving like an LLM 1440 01:19:59,900 --> 01:20:03,117 [laughter] immediately answering no that 1441 01:20:03,200 --> 01:20:06,877 that it's just the low-level world model 1442 01:20:06,960 --> 01:20:10,157 on top of which we can then build some 1443 01:20:10,240 --> 01:20:11,837 of these kinds of mechanisms like you 1444 01:20:11,920 --> 01:20:15,837 said persistent long-term memory or 1445 01:20:15,920 --> 01:20:18,957 uh reasoning so on but we need that 1446 01:20:19,040 --> 01:20:21,437 world model that comes from language is 1447 01:20:21,520 --> 01:20:24,157 it maybe it is not so difficult to build 1448 01:20:24,240 --> 01:20:27,197 this kind of uh reasoning system on top 1449 01:20:27,280 --> 01:20:29,597 of a well constructed world model. 1450 01:20:29,680 --> 01:20:32,077 >> Okay. Whether it's difficult or not, the 1451 01:20:32,160 --> 01:20:36,637 near future will will say because a lot of people are working on reasoning and 1452 01:20:36,720 --> 01:20:39,837 planning abilities for for dialogue systems. 1453 01:20:39,920 --> 01:20:42,237 >> Um I mean if we're even if we restrict 1454 01:20:42,320 --> 01:20:44,477 ourselves to language 1455 01:20:44,560 --> 01:20:47,117 uh just having the ability to plan your 1456 01:20:47,200 --> 01:20:51,837 answer before you answer >> uh in terms that are not necessarily 1457 01:20:51,920 --> 01:20:55,357 linked with the language you're going to use to produce the answer, right? So the 1458 01:20:55,440 --> 01:20:58,477 this idea of this mental model that allows you to plan what you're going to 1459 01:20:58,560 --> 01:21:00,477 say before you say it. >> Mhm. 1460 01:21:00,560 --> 01:21:05,117 >> Um that is very important. I think there's 1461 01:21:05,200 --> 01:21:08,557 going to be a lot of systems over the next few years that are going to have 1462 01:21:08,640 --> 01:21:11,757 this capability but the blueprint of 1463 01:21:11,840 --> 01:21:13,197 those systems will be extremely 1464 01:21:13,280 --> 01:21:19,277 different from autogressive LLMs. So um 1465 01:21:19,360 --> 01:21:23,277 it's the same difference as the difference between what psychologist 1466 01:21:23,360 --> 01:21:27,117 call system one and system two in humans, right? So system one is the type 1467 01:21:27,200 --> 01:21:29,037 of task that you can accomplish without 1468 01:21:29,120 --> 01:21:30,957 like deliberately consciously think 1469 01:21:31,040 --> 01:21:36,077 about how you do them. You just do them [snorts] you've done them enough that 1470 01:21:36,160 --> 01:21:39,437 you can just do it subconsciously, right? Without thinking about them. If 1471 01:21:39,520 --> 01:21:43,677 you're an experienced driver, you can drive without really thinking about it 1472 01:21:43,760 --> 01:21:45,117 and you can talk to someone at the same 1473 01:21:45,200 --> 01:21:47,357 time or listen to the radio, right? 1474 01:21:47,440 --> 01:21:52,557 >> Um, if you are a very experienced chess player, you can play against a 1475 01:21:52,640 --> 01:21:55,597 non-experienced chess player without really thinking either. You just 1476 01:21:55,680 --> 01:21:57,197 recognize the pattern and you play. 1477 01:21:57,280 --> 01:22:01,277 >> Mhm. >> Right. That's system one. Um, so all the 1478 01:22:01,360 --> 01:22:05,517 things that you do instinctively without really having to deliberately plan and 1479 01:22:05,600 --> 01:22:08,317 think about it. And then there is all the tasks where you need to plan. So if 1480 01:22:08,400 --> 01:22:13,677 you are a not so experienced uh chess player or you are experienced but you 1481 01:22:13,760 --> 01:22:16,797 play against another experienced chess player, you think about all kinds of 1482 01:22:16,880 --> 01:22:21,357 options, right? You you think about it for a while, right? And you you you're 1483 01:22:21,440 --> 01:22:22,957 much better if you have time to think 1484 01:22:23,040 --> 01:22:24,557 about it than you are if you are if you 1485 01:22:24,640 --> 01:22:28,877 play blitz uh with limited time. So and 1486 01:22:28,960 --> 01:22:32,077 um so this type of deliberate uh 1487 01:22:32,160 --> 01:22:34,717 planning which uses your internal world 1488 01:22:34,800 --> 01:22:39,357 model um that's system two. This is what LM currently cannot do. 1489 01:22:39,440 --> 01:22:40,957 >> So how how do we get them to do this 1490 01:22:41,040 --> 01:22:43,357 right? H how do we build a system that 1491 01:22:43,440 --> 01:22:47,357 can do this kind of planning that or 1492 01:22:47,440 --> 01:22:50,317 reasoning that devotes more resources to 1493 01:22:50,400 --> 01:22:54,717 complex problems than to simple problems and it's not going to be auto 1494 01:22:54,800 --> 01:22:56,797 reggressive prediction of tokens. is 1495 01:22:56,880 --> 01:23:03,217 going to be more something akin to inference of latent variables in 1496 01:23:03,300 --> 01:23:06,637 [snorts] um you know what used to be called 1497 01:23:06,720 --> 01:23:08,957 probabistic models or graphical models 1498 01:23:09,040 --> 01:23:10,877 and things of that type. So basically 1499 01:23:10,960 --> 01:23:16,637 the principle is like this you you know the prompt is like observed uh 1500 01:23:16,720 --> 01:23:17,517 variables. 1501 01:23:17,600 --> 01:23:22,317 >> Mhm. And what your what the model does 1502 01:23:22,400 --> 01:23:26,237 is that it's basically a measure of it 1503 01:23:26,320 --> 01:23:30,717 can measure to what extent an answer is a good answer for a prompt. 1504 01:23:30,800 --> 01:23:33,437 >> Mhm. >> Okay. So think of it as some gigantic 1505 01:23:33,520 --> 01:23:35,277 neural net but it's got only one output 1506 01:23:35,360 --> 01:23:38,157 and that output is a scalar number which 1507 01:23:38,240 --> 01:23:43,357 is let's say zero if the answer is a good answer for the question and a large 1508 01:23:43,440 --> 01:23:46,957 number if the answer is not a good answer for the question. Imagine you had 1509 01:23:47,040 --> 01:23:49,517 this model. If you had such a model, you 1510 01:23:49,600 --> 01:23:51,597 could use it to produce good answers. 1511 01:23:51,680 --> 01:23:55,917 The way you would do is, you know, produce the prompt and then 1512 01:23:56,000 --> 01:23:59,837 search through the space of possible answers for one that minimizes that 1513 01:23:59,920 --> 01:24:01,357 number. >> Mhm. [clears throat] 1514 01:24:01,440 --> 01:24:04,397 >> Um that's called an energy based model. 1515 01:24:04,480 --> 01:24:07,437 >> But that energy based model would need 1516 01:24:07,520 --> 01:24:11,277 the the model constructed by the LLM. 1517 01:24:11,360 --> 01:24:13,677 Well, so uh really what you need to do 1518 01:24:13,760 --> 01:24:16,797 would be to not uh search over possible 1519 01:24:16,880 --> 01:24:19,357 strings of text that minimize that uh 1520 01:24:19,440 --> 01:24:24,237 energy. But what you would do is do this in abstract representation space. So in 1521 01:24:24,320 --> 01:24:26,237 in sort of the space of abstract 1522 01:24:26,320 --> 01:24:29,197 thoughts, you would elaborate a thought, 1523 01:24:29,280 --> 01:24:32,877 right? Using this process of minimizing 1524 01:24:32,960 --> 01:24:35,357 the output of your your model, okay, 1525 01:24:35,440 --> 01:24:40,157 which is just a scalar. Um it's an optimization process, right? So now the 1526 01:24:40,240 --> 01:24:44,797 the way the system produces its answer is through optimization 1527 01:24:44,880 --> 01:24:49,517 um by you know minimizing an objective function basically right uh and this is 1528 01:24:49,600 --> 01:24:52,477 we're talking about inference we're not talking about training right the system 1529 01:24:52,560 --> 01:24:56,157 has been trained already so now we have an abstract representation of the 1530 01:24:56,240 --> 01:24:58,397 thought of the answer representation of 1531 01:24:58,480 --> 01:25:01,117 the answer we feed that to basically an 1532 01:25:01,200 --> 01:25:03,837 autographive decoder uh which can be 1533 01:25:03,920 --> 01:25:06,237 very simple that turns this into a text 1534 01:25:06,320 --> 01:25:08,877 that expresses this thought okay so that 1535 01:25:08,960 --> 01:25:11,037 That in my opinion is the blueprint of 1536 01:25:11,120 --> 01:25:14,797 future dialogue systems. Um they will 1537 01:25:14,880 --> 01:25:16,637 think about their answer, plan their 1538 01:25:16,720 --> 01:25:19,357 answer by optimization before turning it 1539 01:25:19,440 --> 01:25:23,997 into text. Uh and that is turning complete. 1540 01:25:24,080 --> 01:25:28,157 >> Can you explain exactly what the optimization problem there is? Like 1541 01:25:28,240 --> 01:25:30,637 what's the objective function? Just 1542 01:25:30,720 --> 01:25:35,277 linger on it. You you kind of briefly described it, but over what space are 1543 01:25:35,360 --> 01:25:38,717 you optimizing? the space of representations, 1544 01:25:38,800 --> 01:25:39,917 >> those abstract representations. 1545 01:25:40,000 --> 01:25:43,997 >> Abstract represent. So you have an abstract representation inside the 1546 01:25:44,080 --> 01:25:47,197 system. You have a prompt. The prompt goes to an encoder, produces a 1547 01:25:47,280 --> 01:25:50,317 representation, perhaps goes to a predictor that predicts a representation 1548 01:25:50,400 --> 01:25:52,637 of the answer of the proper answer, but 1549 01:25:52,720 --> 01:25:57,997 that representation may not be a good answer because there might there might 1550 01:25:58,080 --> 01:25:59,597 be some complicated reasoning you need 1551 01:25:59,680 --> 01:26:02,957 to do, right? So um so then you have 1552 01:26:03,040 --> 01:26:05,517 another process that takes the 1553 01:26:05,600 --> 01:26:07,357 representation of the answers and 1554 01:26:07,440 --> 01:26:11,357 modifies it so as to minimize 1555 01:26:11,440 --> 01:26:13,597 uh a cost function that measures to what 1556 01:26:13,680 --> 01:26:15,037 extent the answer is a good answer for 1557 01:26:15,120 --> 01:26:18,637 the question. Now we we sort of [snorts] 1558 01:26:18,720 --> 01:26:21,437 ignore the the fact for I mean the issue 1559 01:26:21,520 --> 01:26:23,437 for a moment of how you train that 1560 01:26:23,520 --> 01:26:26,957 system to measure whether an answer is a 1561 01:26:27,040 --> 01:26:31,597 good answer for for a question. But suppose such a system could be created. 1562 01:26:31,680 --> 01:26:35,117 But what's the process? This kind of search like process 1563 01:26:35,200 --> 01:26:37,117 >> it's a optimization process. You can do 1564 01:26:37,200 --> 01:26:38,957 this if if the entire system is 1565 01:26:39,040 --> 01:26:44,477 differentiable that scalar output is the result of you know running through some 1566 01:26:44,560 --> 01:26:47,757 neural net. >> Mhm. >> Uh running the answer the representation 1567 01:26:47,840 --> 01:26:52,237 of the answer to some neural net then by gradient descent by back proping back 1568 01:26:52,320 --> 01:26:55,997 propagating gradients you can figure out like how to modify the representation of 1569 01:26:56,080 --> 01:26:59,357 the answer so as to minimize that. So that's still a gradient based 1570 01:26:59,440 --> 01:27:03,197 >> it's gradient based inference. So now you have a representation of the answer 1571 01:27:03,280 --> 01:27:06,957 in abstract space. Now you can turn it into text. 1572 01:27:07,040 --> 01:27:10,557 >> Mhm. >> Right. And the cool thing about this is 1573 01:27:10,640 --> 01:27:15,597 that the representation now can be optimized through gradient descent but 1574 01:27:15,680 --> 01:27:20,317 also is independent of the language in which you're going to express the answer. 1575 01:27:20,400 --> 01:27:23,277 >> Right? So you're operating in the abstract representation. I mean this 1576 01:27:23,360 --> 01:27:25,357 goes backs to the joint embedding 1577 01:27:25,440 --> 01:27:27,997 >> that it's better to work in the uh in 1578 01:27:28,080 --> 01:27:31,357 the space of I don't know to romanticize 1579 01:27:31,440 --> 01:27:34,317 the notion like space of concepts versus 1580 01:27:34,400 --> 01:27:38,477 >> the space of concrete >> sensory information 1581 01:27:38,560 --> 01:27:42,637 >> right >> okay but is can can this do something 1582 01:27:42,720 --> 01:27:44,637 like reasoning which is what we're talking about 1583 01:27:44,720 --> 01:27:48,397 >> well not really in a only in a very simple way I mean basically you can 1584 01:27:48,480 --> 01:27:52,717 think of those things as doing the kind of optimization I was I was talking 1585 01:27:52,800 --> 01:27:55,997 about except they optimize in the discrete space which is the space of 1586 01:27:56,080 --> 01:28:00,877 possible sequences of of tokens and they do it they do this optimization in a 1587 01:28:00,960 --> 01:28:04,317 horribly inefficient way which is generate a lot of hypothesis and then 1588 01:28:04,400 --> 01:28:06,077 select the best ones 1589 01:28:06,160 --> 01:28:11,437 >> and that's incredibly wasteful in terms of uh computation 1590 01:28:11,520 --> 01:28:12,717 because you have you run you basically 1591 01:28:12,800 --> 01:28:15,277 have to run your LLM for like every 1592 01:28:15,360 --> 01:28:18,477 possible you know generated sequence um 1593 01:28:18,560 --> 01:28:21,917 and it's incredibly wasteful um so So 1594 01:28:22,000 --> 01:28:26,877 it's much better to do an optimization in continuous space where you can do 1595 01:28:26,960 --> 01:28:29,917 gradient and descent as opposed to like generate tons of things and then select 1596 01:28:30,000 --> 01:28:33,117 the best. You just iteratively refine 1597 01:28:33,200 --> 01:28:36,957 your answer to to go towards the best, right? That's much more efficient. But 1598 01:28:37,040 --> 01:28:40,877 you can only do this in continuous spaces with differentiable functions. 1599 01:28:40,960 --> 01:28:43,277 You're talking about the reasoning like 1600 01:28:43,360 --> 01:28:46,237 ability to think deeply or to reason 1601 01:28:46,320 --> 01:28:47,917 deeply. 1602 01:28:48,000 --> 01:28:50,797 How do you know what 1603 01:28:50,880 --> 01:28:56,717 is an answer uh that's better or worse based on deep 1604 01:28:56,800 --> 01:28:59,277 reasoning? >> Right? So then we're asking the question 1605 01:28:59,360 --> 01:29:03,117 of conceptually how do you train an energy based model? Right? So energy 1606 01:29:03,200 --> 01:29:05,277 based model is a function with a scalar 1607 01:29:05,360 --> 01:29:07,597 output just a number. 1608 01:29:07,680 --> 01:29:10,797 >> You give it two inputs X and Y 1609 01:29:10,880 --> 01:29:12,317 >> and it tells you whether Y is compatible 1610 01:29:12,400 --> 01:29:14,717 with X or not. X you observe. Let's say 1611 01:29:14,800 --> 01:29:16,317 it's a prompt, an image, a video, 1612 01:29:16,400 --> 01:29:20,077 whatever. and y is a proposal for an 1613 01:29:20,160 --> 01:29:23,597 answer, a continuation of video, um you 1614 01:29:23,680 --> 01:29:28,557 know, whatever. And it tells you whether Y is compatible with X. And the way it 1615 01:29:28,640 --> 01:29:31,917 tells you that Y is compatible with X is that the output of that function would 1616 01:29:32,000 --> 01:29:33,997 be zero if Y is compatible with X and 1617 01:29:34,080 --> 01:29:37,677 would be a a positive number non zero if 1618 01:29:37,760 --> 01:29:40,557 Y is not compatible with X. Okay. How do 1619 01:29:40,640 --> 01:29:42,797 you train a system like this at a 1620 01:29:42,880 --> 01:29:46,717 completely general level is you show it 1621 01:29:46,800 --> 01:29:51,357 pairs of X and Y that are compatible a question and the corresponding answer 1622 01:29:51,440 --> 01:29:53,357 and you train the parameters of the big 1623 01:29:53,440 --> 01:29:57,037 neural net inside um to produce zero. 1624 01:29:57,120 --> 01:30:01,597 Mhm. Okay. Now that doesn't completely work because the system might decide 1625 01:30:01,680 --> 01:30:05,357 well I'm just going to say zero for everything. So now you have to have a 1626 01:30:05,440 --> 01:30:08,237 process to make sure that for a a wrong 1627 01:30:08,320 --> 01:30:13,277 Y the energy would be larger than zero and [snorts] there you have two options. 1628 01:30:13,360 --> 01:30:16,877 One is contrastive method. So contrastive method is you show an X and 1629 01:30:16,960 --> 01:30:20,957 a bad Y and you tell the system well that's you 1630 01:30:21,040 --> 01:30:23,997 know give a high energy to this like push up the energy right change the 1631 01:30:24,080 --> 01:30:28,877 weights in the neural net that computes the energy so that it goes up 1632 01:30:28,960 --> 01:30:30,557 u so that's contrasting methods. The 1633 01:30:30,640 --> 01:30:33,197 problem with this is if the space of Y 1634 01:30:33,280 --> 01:30:37,037 is large, the number of such contrastive 1635 01:30:37,120 --> 01:30:41,277 samples you're going to have to show is gigantic. 1636 01:30:41,360 --> 01:30:43,917 But people do this. They they do this 1637 01:30:44,000 --> 01:30:46,397 when you train a system with RHF. 1638 01:30:46,480 --> 01:30:50,717 Basically, what you're training is what's called a reward model, which is 1639 01:30:50,800 --> 01:30:53,197 basically an objective function that 1640 01:30:53,280 --> 01:30:58,237 tells you whether an answer is good or bad. And that's basically exactly what 1641 01:30:58,320 --> 01:31:01,277 what this is. Mhm. >> So we already do this to some extent. 1642 01:31:01,360 --> 01:31:02,717 We're just not using it for inference. 1643 01:31:02,800 --> 01:31:04,637 We're just using it for training. 1644 01:31:04,720 --> 01:31:06,237 >> Um 1645 01:31:06,320 --> 01:31:08,797 uh there is another set of methods which 1646 01:31:08,880 --> 01:31:11,757 are non-contrastive and I prefer those. 1647 01:31:11,840 --> 01:31:13,677 Uh and those non-contrastive method 1648 01:31:13,760 --> 01:31:15,757 basically say 1649 01:31:15,840 --> 01:31:19,277 uh okay the energy function needs to 1650 01:31:19,360 --> 01:31:23,837 have low energy on pairs of XY that are compatible that come from your training 1651 01:31:23,920 --> 01:31:27,117 set. How do you make sure that the energy is 1652 01:31:27,200 --> 01:31:29,917 going to be higher everywhere else? 1653 01:31:30,000 --> 01:31:34,717 And the way you do this is by um having a 1654 01:31:34,800 --> 01:31:36,877 regularizer, a criterion, a term in your 1655 01:31:36,960 --> 01:31:39,357 cost function that basically minimizes 1656 01:31:39,440 --> 01:31:44,477 the volume of space that can take low energy. 1657 01:31:44,560 --> 01:31:47,837 And the precise way to do this is all kinds of different specific ways to do 1658 01:31:47,920 --> 01:31:51,997 this depending on the architecture. But that's the basic principle. So that if 1659 01:31:52,080 --> 01:31:56,957 you push down the energy function for particular regions in the XY space, it 1660 01:31:57,040 --> 01:31:58,957 will automatically go up in other places 1661 01:31:59,040 --> 01:32:01,757 because there's only a limited volume of 1662 01:32:01,840 --> 01:32:04,877 space that can take low energy. Okay, by 1663 01:32:04,960 --> 01:32:06,717 the construction of the system or by the 1664 01:32:06,800 --> 01:32:09,677 regularizer regularizing function. We've 1665 01:32:09,760 --> 01:32:12,237 been talking very generally, but what is 1666 01:32:12,320 --> 01:32:14,717 a good X and a good Y? What is a good 1667 01:32:14,800 --> 01:32:18,877 representation of X and Y? because we've 1668 01:32:18,960 --> 01:32:23,437 been talking about language and if you just take language directly that 1669 01:32:23,520 --> 01:32:27,117 presumably is not good. So there has to be some kind of abstract representation 1670 01:32:27,200 --> 01:32:28,877 of ideas. 1671 01:32:28,960 --> 01:32:30,637 >> Yeah. So you I mean you can do this with 1672 01:32:30,720 --> 01:32:35,837 language directly um by just you know X is a text and Y is a continuation of 1673 01:32:35,920 --> 01:32:40,877 that text. >> Yes. >> Um or X is a question Y is the answer. 1674 01:32:40,960 --> 01:32:43,117 >> But you're you're saying that's not going to take it. I mean that's going to 1675 01:32:43,200 --> 01:32:46,157 do what LLMs are doing. Well, no. It 1676 01:32:46,240 --> 01:32:50,397 depends on how you how the internal structure of the system is built. If the 1677 01:32:50,480 --> 01:32:53,757 if the internal structure of the system is built in such a way that inside of 1678 01:32:53,840 --> 01:32:57,597 the system there is a latent variable let's call it Z 1679 01:32:57,680 --> 01:32:59,917 >> that uh 1680 01:33:00,000 --> 01:33:03,277 you can manipulate so as to minimize the 1681 01:33:03,360 --> 01:33:07,837 output energy then that Z can be viewed as a 1682 01:33:07,920 --> 01:33:11,677 representation of a good answer that you can translate into a Y that is a good 1683 01:33:11,760 --> 01:33:15,357 answer. >> [snorts] >> So this kind of system could be trained 1684 01:33:15,440 --> 01:33:18,717 in a very similar way, >> very similar way, but you have to have 1685 01:33:18,800 --> 01:33:21,357 this way of preventing collapse of of 1686 01:33:21,440 --> 01:33:24,077 ensuring that you know there is high 1687 01:33:24,160 --> 01:33:26,477 energy for things you don't train it on. 1688 01:33:26,560 --> 01:33:30,477 Um and and currently it's it's very 1689 01:33:30,560 --> 01:33:33,597 implicit in NLM. is done in a way that people don't realize it's being done but 1690 01:33:33,680 --> 01:33:36,317 is it is being done is is due to the 1691 01:33:36,400 --> 01:33:38,077 fact that when you give a high 1692 01:33:38,160 --> 01:33:41,997 probability to a word 1693 01:33:42,080 --> 01:33:46,477 automatically you give low probability to other words because you only have a 1694 01:33:46,560 --> 01:33:50,397 finite amount of probability to go around right there to sum to one 1695 01:33:50,480 --> 01:33:52,157 >> so when you minimize the cross entropy 1696 01:33:52,240 --> 01:33:54,877 or whatever when you train the your LLM 1697 01:33:54,960 --> 01:34:00,317 to produce the to predict the next word uh you're increasing the probability 1698 01:34:00,400 --> 01:34:03,037 your system will give to the correct word, but you're also decreasing the 1699 01:34:03,120 --> 01:34:03,917 probability it will give to the 1700 01:34:04,000 --> 01:34:07,597 incorrect words. Now, indirectly that 1701 01:34:07,680 --> 01:34:11,757 gives a low probability to a high probability to sequences of words that 1702 01:34:11,840 --> 01:34:14,557 are good and low probability to sequences of words that are bad. But 1703 01:34:14,640 --> 01:34:18,477 it's very indirect >> and it's not it's not obvious why this 1704 01:34:18,560 --> 01:34:23,117 actually works at all. that uh because you're not doing it on a joint 1705 01:34:23,200 --> 01:34:27,117 probability of all the symbols in a in a sequence. You're just doing it kind of 1706 01:34:27,200 --> 01:34:31,757 uh you sort of factoriize that probability in terms of conditional 1707 01:34:31,840 --> 01:34:34,077 probabilities over successive tokens. 1708 01:34:34,160 --> 01:34:36,637 >> So how do you do this for visual data? 1709 01:34:36,720 --> 01:34:38,317 So we've been doing this with OJA 1710 01:34:38,400 --> 01:34:40,877 architectures basically joint embedding. 1711 01:34:40,960 --> 01:34:44,557 So uh there the compatibility between 1712 01:34:44,640 --> 01:34:49,757 two things is uh you know here's here's an image or a video here is a corrupted 1713 01:34:49,840 --> 01:34:51,357 shifted or transformed version of that 1714 01:34:51,440 --> 01:34:55,117 image or video or masked. Okay. And then 1715 01:34:55,200 --> 01:34:57,997 uh the energy of the system is the 1716 01:34:58,080 --> 01:35:02,157 prediction error of the 1717 01:35:02,240 --> 01:35:06,877 representation uh the the predicted representation of 1718 01:35:06,960 --> 01:35:10,397 the good thing versus the actual representation of the good thing. Right? 1719 01:35:10,480 --> 01:35:13,437 So so you run the corrupted image to the 1720 01:35:13,520 --> 01:35:17,837 system, predict the representation of the the good input uncorrupted and then 1721 01:35:17,920 --> 01:35:19,597 compute the prediction error. That's the 1722 01:35:19,680 --> 01:35:21,757 energy of the system. So the system will 1723 01:35:21,840 --> 01:35:26,477 tell you this is a good 1724 01:35:26,560 --> 01:35:29,997 you know if this is a good image and this is a corrupted version it will give 1725 01:35:30,080 --> 01:35:34,957 you zero energy if those two things are effectively one of them is a corrupted 1726 01:35:35,040 --> 01:35:38,317 version of the other and give you a high energy if the if the two images are 1727 01:35:38,400 --> 01:35:41,517 completely different >> and hopefully that whole process gives 1728 01:35:41,600 --> 01:35:46,877 you a really nice compressed representation of of reality of visual 1729 01:35:46,960 --> 01:35:49,197 reality >> and we know it does because then we use 1730 01:35:49,280 --> 01:35:52,397 those representations as input to a classification system. 1731 01:35:52,480 --> 01:35:53,997 >> That classification system works really 1732 01:35:54,080 --> 01:35:57,597 nicely. Okay. Well, so to summarize, you 1733 01:35:57,680 --> 01:36:01,677 recommend in a in a in a spicy way that 1734 01:36:01,760 --> 01:36:05,997 only Yun can. You recommend that we abandon generative models in favor of 1735 01:36:06,080 --> 01:36:07,677 joint embedding architectures. 1736 01:36:07,760 --> 01:36:10,237 >> Yes. >> Abandon auto reggressive generation. 1737 01:36:10,320 --> 01:36:14,467 >> Yes. >> Abandon pro, [laughter] this feels like court testimony. 1738 01:36:14,550 --> 01:36:18,317 [snorts] Uh abandon probabilistic models in favor of energy based models as we 1739 01:36:18,400 --> 01:36:20,077 talked about. abandon contrastive 1740 01:36:20,160 --> 01:36:22,957 methods in favor of regularized methods. 1741 01:36:23,040 --> 01:36:25,917 And uh let me ask you about this. 1742 01:36:26,000 --> 01:36:29,197 You've been for a while a critic of reinforcement learning. 1743 01:36:29,280 --> 01:36:32,237 >> Yes. >> So what uh the last recommendation is 1744 01:36:32,320 --> 01:36:35,517 that we abandon RL in favor of model 1745 01:36:35,600 --> 01:36:37,117 predictive control as you were talking 1746 01:36:37,200 --> 01:36:40,077 about and only use RL when planning 1747 01:36:40,160 --> 01:36:44,077 doesn't yield the predicted outcome. And 1748 01:36:44,160 --> 01:36:46,317 uh we use RL in that case to adjust the 1749 01:36:46,400 --> 01:36:48,877 world model or the critic. Yes. 1750 01:36:48,960 --> 01:36:53,197 >> So, uh you mentioned uh RLHF, 1751 01:36:53,280 --> 01:36:57,677 reinforcement learning with human feedback. Uh why do you still hate uh 1752 01:36:57,760 --> 01:36:59,917 reinforcement learning? >> I don't hate reinforcement learning and 1753 01:37:00,000 --> 01:37:03,357 I think I think it should not be uh 1754 01:37:03,440 --> 01:37:07,997 abandoned completely, but I think it its use should be minimized because it's 1755 01:37:08,080 --> 01:37:09,837 incredibly inefficient in terms of 1756 01:37:09,920 --> 01:37:12,717 samples. And so the the proper way to 1757 01:37:12,800 --> 01:37:17,117 train a system is to first have it learn 1758 01:37:17,200 --> 01:37:19,677 uh good representations of the world and 1759 01:37:19,760 --> 01:37:24,237 world models from mostly observation maybe a little bit of interactions 1760 01:37:24,320 --> 01:37:27,357 >> and then steered based on that. If the representation is good then the 1761 01:37:27,440 --> 01:37:29,677 adjustments should be minimal. 1762 01:37:29,760 --> 01:37:31,677 >> Yeah. Now there's two things you can use 1763 01:37:31,760 --> 01:37:34,957 if you've learned a world model you can use the world model to plan a sequence 1764 01:37:35,040 --> 01:37:36,477 of actions to arrive at a particular 1765 01:37:36,560 --> 01:37:41,997 objective. Mhm. >> You don't need R out unless the way you 1766 01:37:42,080 --> 01:37:45,037 measure whether you succeed might be inexact. 1767 01:37:45,120 --> 01:37:47,837 Your idea of, you know, whether you were 1768 01:37:47,920 --> 01:37:51,117 going to fall from your bike [gasps] 1769 01:37:51,200 --> 01:37:55,437 might be wrong or whether the person you're fighting with MMA was going to do 1770 01:37:55,520 --> 01:37:57,357 something and then do something else. 1771 01:37:57,440 --> 01:38:00,957 Um, so there uh 1772 01:38:01,040 --> 01:38:02,237 so there's two ways you can be wrong. 1773 01:38:02,320 --> 01:38:05,917 either your your objective function does 1774 01:38:06,000 --> 01:38:09,757 not reflect the actual objective function you want to optimize or your 1775 01:38:09,840 --> 01:38:14,237 world model is inaccurate right so you didn't you the prediction 1776 01:38:14,320 --> 01:38:18,717 you were making about what was going to happen in the world is inaccurate so if 1777 01:38:18,800 --> 01:38:20,237 you want to adjust your world model 1778 01:38:20,320 --> 01:38:23,837 while you are operating in the world or 1779 01:38:23,920 --> 01:38:26,957 your objective function that is 1780 01:38:27,040 --> 01:38:31,837 basically in the realm of RL this is what RL deal deals with uh to some 1781 01:38:31,920 --> 01:38:33,677 extent Right? So adjust your world model 1782 01:38:33,760 --> 01:38:35,757 and the way to adjust your world model 1783 01:38:35,840 --> 01:38:40,957 even in advance uh is to explore parts of the space where your world model 1784 01:38:41,040 --> 01:38:42,317 where you know that your world model is 1785 01:38:42,400 --> 01:38:44,877 inaccurate. That's called curiosity 1786 01:38:44,960 --> 01:38:48,557 basically or play right when you play 1787 01:38:48,640 --> 01:38:51,197 you kind of explore part of the state 1788 01:38:51,280 --> 01:38:54,157 space that um [snorts] 1789 01:38:54,240 --> 01:38:57,037 you know you don't want to do in for 1790 01:38:57,120 --> 01:38:59,037 real because it might be dangerous but 1791 01:38:59,120 --> 01:39:01,517 uh but you can adjust your world model 1792 01:39:01,600 --> 01:39:04,877 uh without killing yourself basically. 1793 01:39:04,960 --> 01:39:07,677 Um, so that's what you want to use RL 1794 01:39:07,760 --> 01:39:11,997 for when when when it comes time to learning a particular task. You already 1795 01:39:12,080 --> 01:39:14,957 have all the good representations. You already have your world model, but you 1796 01:39:15,040 --> 01:39:18,637 want you need to adjust it for the situation at hand. That's when you use 1797 01:39:18,720 --> 01:39:23,277 RL. >> Why do you think RHF works so well? This 1798 01:39:23,360 --> 01:39:26,637 reinforcement learning with human feedback, why did it have such a 1799 01:39:26,720 --> 01:39:28,957 transformational effect on large 1800 01:39:29,040 --> 01:39:33,837 language models that before? What's had the transformational effect is human 1801 01:39:33,920 --> 01:39:36,797 feedback. >> There is many ways to use it and some of 1802 01:39:36,880 --> 01:39:38,557 it is just purely supervised actually. 1803 01:39:38,640 --> 01:39:40,157 It's not really reinforcement learning. 1804 01:39:40,240 --> 01:39:43,277 >> So it's the a the HF. >> It's the HF. 1805 01:39:43,360 --> 01:39:45,197 >> Uh and then there is various ways to use 1806 01:39:45,280 --> 01:39:49,997 human feedback, right? So you can uh you can ask humans to rate answers uh 1807 01:39:50,080 --> 01:39:52,077 multiple answers that are produced by a 1808 01:39:52,160 --> 01:39:56,557 world model and uh and and then what you 1809 01:39:56,640 --> 01:39:58,557 do is you train an objective function to 1810 01:39:58,640 --> 01:40:02,237 predict that rating and then you can use 1811 01:40:02,320 --> 01:40:06,637 that objective function to predict you know whether an answer is good and you 1812 01:40:06,720 --> 01:40:09,757 can back propagate gradient through this to fine-tune your system so that it only 1813 01:40:09,840 --> 01:40:13,517 produces high highly rated answers. 1814 01:40:13,600 --> 01:40:17,437 Okay, so that's one way. So that's like 1815 01:40:17,520 --> 01:40:22,637 in RL that means uh training what's called a reward model right so something 1816 01:40:22,720 --> 01:40:25,997 that you know basically a small neural net that estimates to what extent an 1817 01:40:26,080 --> 01:40:30,397 answer is good right it's very similar to the objective I was I was talking 1818 01:40:30,480 --> 01:40:33,837 about talking about earlier for planning except now it's not used for planning 1819 01:40:33,920 --> 01:40:37,037 it's it's used for fine-tuning your system 1820 01:40:37,120 --> 01:40:42,077 I think it would be much more efficient to use it for planning but um but but uh 1821 01:40:42,160 --> 01:40:45,917 currently it's used to fine-tune the parameters of the them. Now there's 1822 01:40:46,000 --> 01:40:50,477 several ways to do this. Um you know some of some of them are supervised. You 1823 01:40:50,560 --> 01:40:52,717 just you know ask a human person like 1824 01:40:52,800 --> 01:40:55,037 what is a good answer for this right? 1825 01:40:55,120 --> 01:40:59,677 Then you just type the answer. [snorts] Um uh I mean there's there's lots of 1826 01:40:59,760 --> 01:41:03,437 ways that those systems are are being adjusted. 1827 01:41:03,520 --> 01:41:05,837 Now, a lot of people have been very 1828 01:41:05,920 --> 01:41:08,397 critical of the recently released 1829 01:41:08,480 --> 01:41:11,757 Google's Gemini 1.5 1830 01:41:11,840 --> 01:41:14,237 for essentially, in my words, I could 1831 01:41:14,320 --> 01:41:17,357 say super woke. Woke in the uh negative 1832 01:41:17,440 --> 01:41:19,757 connotation of that word. Uh there is 1833 01:41:19,840 --> 01:41:22,397 some almost hilariously absurd things 1834 01:41:22,480 --> 01:41:25,517 that it does like it modifies history uh 1835 01:41:25,600 --> 01:41:29,117 like generating images of a um black 1836 01:41:29,200 --> 01:41:33,117 George Washington or um perhaps more 1837 01:41:33,200 --> 01:41:34,797 seriously something that you commented 1838 01:41:34,880 --> 01:41:39,277 on Twitter which is refusing to comment 1839 01:41:39,360 --> 01:41:42,877 on or generate images of u or even 1840 01:41:42,960 --> 01:41:46,797 descriptions of Tianaan Square or the 1841 01:41:46,880 --> 01:41:49,197 the tank man. one of the most sort of 1842 01:41:49,280 --> 01:41:54,237 legendary protest images in history. Of 1843 01:41:54,320 --> 01:41:58,077 course, these images are highly censored 1844 01:41:58,160 --> 01:42:01,037 by the Chinese government and therefore 1845 01:42:01,120 --> 01:42:02,717 everybody started asking questions of 1846 01:42:02,800 --> 01:42:06,157 what is the process of uh 1847 01:42:06,240 --> 01:42:10,317 designing these LLMs, what is what is what is the role of censorship in these 1848 01:42:10,400 --> 01:42:11,757 all that kind of stuff. 1849 01:42:11,840 --> 01:42:15,357 >> So you uh commented on Twitter saying 1850 01:42:15,440 --> 01:42:18,397 that open source is the answer. Yeah, 1851 01:42:18,480 --> 01:42:22,317 >> essentially. So, um, can you explain? 1852 01:42:22,400 --> 01:42:25,997 >> I I actually made that comment on just about every social network I can and 1853 01:42:26,080 --> 01:42:29,037 I've [laughter] I I've uh I've made that 1854 01:42:29,120 --> 01:42:31,117 point multiple times in in various 1855 01:42:31,200 --> 01:42:34,477 forums. Um, uh, here's my my point of 1856 01:42:34,560 --> 01:42:37,757 view on this. Uh people can complain 1857 01:42:37,840 --> 01:42:40,797 that AI systems are biased and they 1858 01:42:40,880 --> 01:42:43,917 generally are biased by the distribution 1859 01:42:44,000 --> 01:42:45,917 of the training data that they've been 1860 01:42:46,000 --> 01:42:49,677 trained on. Um that reflects biases in 1861 01:42:49,760 --> 01:42:52,637 society. Um 1862 01:42:52,720 --> 01:42:55,997 and that is potentially offensive to 1863 01:42:56,080 --> 01:43:00,637 some people or potentially not. and and 1864 01:43:00,720 --> 01:43:03,917 some techniques to debias then become 1865 01:43:04,000 --> 01:43:07,037 offensive to some people 1866 01:43:07,120 --> 01:43:11,997 um because of you know historical uh 1867 01:43:12,080 --> 01:43:14,637 incorrectness and things like that 1868 01:43:14,720 --> 01:43:18,317 >> um and so you can ask the question you 1869 01:43:18,400 --> 01:43:22,397 can ask two questions the first question is is it possible to produce an AI 1870 01:43:22,480 --> 01:43:24,157 system that is not biased and the answer 1871 01:43:24,240 --> 01:43:26,637 is absolutely not and it's not because 1872 01:43:26,720 --> 01:43:29,837 of technological 1873 01:43:29,920 --> 01:43:34,397 challenges although there are technological challenges to that it's 1874 01:43:34,480 --> 01:43:39,357 because bias is in the eye of the beholder 1875 01:43:39,440 --> 01:43:41,037 um different people may have different 1876 01:43:41,120 --> 01:43:44,397 ideas about what constitutes bias um you 1877 01:43:44,480 --> 01:43:48,957 know for a lot of a lot of things I mean there are facts that are you know [snorts] 1878 01:43:49,040 --> 01:43:51,277 indisputable but there are a lot of 1879 01:43:51,360 --> 01:43:56,237 opinions or or things that can be expressed in different ways and so you 1880 01:43:56,320 --> 01:43:57,917 cannot have an unbiased system that's 1881 01:43:58,000 --> 01:44:00,317 just an impossib ility. 1882 01:44:00,400 --> 01:44:03,997 Um, and so what's the 1883 01:44:04,080 --> 01:44:05,917 what's the answer to this? And the the 1884 01:44:06,000 --> 01:44:09,197 answer is the same answer that we found 1885 01:44:09,280 --> 01:44:13,517 in liberal democracy about the press. 1886 01:44:13,600 --> 01:44:16,317 The press needs to be free and uh 1887 01:44:16,400 --> 01:44:18,077 diverse. 1888 01:44:18,160 --> 01:44:21,117 We have free speech for a good reason is 1889 01:44:21,200 --> 01:44:24,077 because uh we don't want all of our 1890 01:44:24,160 --> 01:44:27,757 information to be uh to come from a 1891 01:44:27,840 --> 01:44:30,317 unique source. Um because that's 1892 01:44:30,400 --> 01:44:32,717 opposite to the whole idea of democracy 1893 01:44:32,800 --> 01:44:36,397 and uh you know progress of ideas and 1894 01:44:36,480 --> 01:44:38,797 even science right in in science people 1895 01:44:38,880 --> 01:44:43,677 have to argue for different opinions and and science makes progress when people 1896 01:44:43,760 --> 01:44:47,437 disagree and they come up with an answer and you know a consensus forms right and 1897 01:44:47,520 --> 01:44:49,917 it's true in all democracies around the 1898 01:44:50,000 --> 01:44:53,597 world. So there is a a future 1899 01:44:53,680 --> 01:44:56,797 which is already happening where every 1900 01:44:56,880 --> 01:44:59,277 single one of our interaction with the 1901 01:44:59,360 --> 01:45:05,017 digital world will be mediated by AI AI systems AI assistants right 1902 01:45:05,100 --> 01:45:07,677 [snorts and gasps] we're going to have smart glasses you 1903 01:45:07,760 --> 01:45:12,957 can already buy them from meta the Rayban Meta where um you know you can 1904 01:45:13,040 --> 01:45:17,117 talk to them and they are connected with an LLM and you can get answers on any 1905 01:45:17,200 --> 01:45:19,437 question you have or you can be looking 1906 01:45:19,520 --> 01:45:22,477 at a monument and there is a camera in 1907 01:45:22,560 --> 01:45:26,637 the in the system that in in the glasses you can ask it like can what can you 1908 01:45:26,720 --> 01:45:31,117 tell me about this uh building or this monument you can be looking at a menu in 1909 01:45:31,200 --> 01:45:34,477 a foreign language and I think we'll translate it for you or we can do 1910 01:45:34,560 --> 01:45:39,117 realtime translation if you speak different languages so a lot of our 1911 01:45:39,200 --> 01:45:42,797 interactions with the digital world are going to be mediated by those systems in 1912 01:45:42,880 --> 01:45:48,157 the near future um you know increasingly the 1913 01:45:48,240 --> 01:45:51,037 search engines that we're going to are not going to be search engines. They're 1914 01:45:51,120 --> 01:45:54,157 going to be uh dialog systems 1915 01:45:54,240 --> 01:45:57,357 >> that will just ask a question and it 1916 01:45:57,440 --> 01:46:02,317 will answer and then point you to perhaps appropriate reference for it. 1917 01:46:02,400 --> 01:46:04,077 But here's the thing. We cannot afford 1918 01:46:04,160 --> 01:46:06,317 those systems to come from a handful of 1919 01:46:06,400 --> 01:46:09,997 companies on the west coast of the US 1920 01:46:10,080 --> 01:46:14,797 because those systems will constitute the repository of all human knowledge 1921 01:46:14,880 --> 01:46:18,797 and we cannot have that be controlled by a 1922 01:46:18,880 --> 01:46:22,877 small number of people right it has to be diverse for the same reason the press 1923 01:46:22,960 --> 01:46:24,957 has to be diverse. 1924 01:46:25,040 --> 01:46:27,597 So how do we get a diverse set of AI 1925 01:46:27,680 --> 01:46:30,477 assistants? [snorts] um it's very 1926 01:46:30,560 --> 01:46:34,957 expensive and difficult to train a base model right a based LLM at the moment 1927 01:46:35,040 --> 01:46:37,837 you know in the future it might be something different but at the moment 1928 01:46:37,920 --> 01:46:41,597 that's an LLM uh so only a few companies can do this 1929 01:46:41,680 --> 01:46:45,037 properly and 1930 01:46:45,120 --> 01:46:47,757 if some of those subsystems are open 1931 01:46:47,840 --> 01:46:50,477 source anybody can use them anybody can 1932 01:46:50,560 --> 01:46:53,677 fine-tune them um if we put in place 1933 01:46:53,760 --> 01:46:57,597 some systems that allows any group of 1934 01:46:57,680 --> 01:46:59,037 people 1935 01:46:59,120 --> 01:47:02,877 whether they are um individual citizens, 1936 01:47:02,960 --> 01:47:05,117 groups of citizens, 1937 01:47:05,200 --> 01:47:08,237 government organizations, NOS's, 1938 01:47:08,320 --> 01:47:12,237 um companies, whatever to take those 1939 01:47:12,320 --> 01:47:17,277 open source u systems, AI systems, and fine-tune 1940 01:47:17,360 --> 01:47:18,877 them for their own purpose on their own 1941 01:47:18,960 --> 01:47:21,437 data. Then we're going to have a very 1942 01:47:21,520 --> 01:47:24,477 large diversity of uh different AI 1943 01:47:24,560 --> 01:47:29,357 systems that are specialized for all of those things, right? So I tell you I 1944 01:47:29,440 --> 01:47:32,877 talked to the French government quite a bit and the French government will not 1945 01:47:32,960 --> 01:47:36,237 accept that the digital diet of all 1946 01:47:36,320 --> 01:47:40,717 their citizen be controlled by three companies on the west coast of the US. 1947 01:47:40,800 --> 01:47:44,877 That's just not acceptable. It's a danger to democracy regardless of how 1948 01:47:44,960 --> 01:47:46,797 well-intentioned those companies are. 1949 01:47:46,880 --> 01:47:51,117 Right. Um and so u and it's also a 1950 01:47:51,200 --> 01:47:56,077 danger to local culture to values to 1951 01:47:56,160 --> 01:48:01,277 language right I was talking with um uh 1952 01:48:01,360 --> 01:48:05,837 the uh founder of Infosys in India um 1953 01:48:05,920 --> 01:48:09,357 he's funding a project to fine-tune Latu 1954 01:48:09,440 --> 01:48:11,437 the open source model produced by by 1955 01:48:11,520 --> 01:48:14,877 Meta so that Latu speaks all 22 official 1956 01:48:14,960 --> 01:48:19,917 languages in India it's very important for people in India. I was talking to a 1957 01:48:20,000 --> 01:48:24,077 former colleague of mine Mustafa Asis who used to be a scientist at fair and 1958 01:48:24,160 --> 01:48:28,077 then moved back to Africa created a research lab for Google in Africa and 1959 01:48:28,160 --> 01:48:32,557 now is there's a new startup called Kara and what he's trying to do is basically 1960 01:48:32,640 --> 01:48:36,797 have LLM that speak the local languages in Sineagal so that people can have 1961 01:48:36,880 --> 01:48:40,397 access to medical information because they don't have access to doctors. It's 1962 01:48:40,480 --> 01:48:42,397 a very small number of doctors per per 1963 01:48:42,480 --> 01:48:47,037 capita in the in Sineagal. Um I mean you 1964 01:48:47,120 --> 01:48:51,277 can't have any of this unless you have open source platforms. So with open 1965 01:48:51,360 --> 01:48:55,197 source platforms you can have AI systems that are not only diverse in terms of 1966 01:48:55,280 --> 01:48:56,957 political opinions or things of that 1967 01:48:57,040 --> 01:49:01,757 type but in terms of uh uh language, 1968 01:49:01,840 --> 01:49:05,117 culture, value systems, political 1969 01:49:05,200 --> 01:49:08,237 opinions um 1970 01:49:08,320 --> 01:49:11,597 technical abilities in various domains 1971 01:49:11,680 --> 01:49:15,757 and you can have an industry an ecosystem of companies that fine-tune 1972 01:49:15,840 --> 01:49:20,237 those open source systems for vertical applications in industry. Right? you you 1973 01:49:20,320 --> 01:49:23,997 have I don't know a publisher has thousands of books and they want to 1974 01:49:24,080 --> 01:49:28,477 build a system that allows a customer to just ask a question about any but the 1975 01:49:28,560 --> 01:49:31,437 content of any of their books you [snorts] need to train on their 1976 01:49:31,520 --> 01:49:34,077 proprietary data right um you have a 1977 01:49:34,160 --> 01:49:38,957 company we have one within meta it's called metamate and it's basically an 1978 01:49:39,040 --> 01:49:40,557 LLM that can answer any question about 1979 01:49:40,640 --> 01:49:43,757 internal uh stuff about about the 1980 01:49:43,840 --> 01:49:46,797 company u very useful a lot of companies 1981 01:49:46,880 --> 01:49:51,037 want this right a lot of companies want is not just for their employees but also 1982 01:49:51,120 --> 01:49:55,117 for their customers to take care of their customers. So the only way you're 1983 01:49:55,200 --> 01:49:59,197 going to have an AI industry, the only way you're going to have AI systems that 1984 01:49:59,280 --> 01:50:02,077 are not uniquely biased is if you have 1985 01:50:02,160 --> 01:50:05,037 open source platforms on top of which uh 1986 01:50:05,120 --> 01:50:08,957 any group can u build specialized 1987 01:50:09,040 --> 01:50:13,837 systems. So the the direction of of 1988 01:50:13,920 --> 01:50:16,637 inevitable direction of history is that 1989 01:50:16,720 --> 01:50:19,117 the vast majority of AI systems will be 1990 01:50:19,200 --> 01:50:21,197 built on top of open source platforms. 1991 01:50:21,280 --> 01:50:23,837 >> So that's a beautiful vision. So meaning 1992 01:50:23,920 --> 01:50:25,917 like 1993 01:50:26,000 --> 01:50:30,637 a company like Meta or Google or so on 1994 01:50:30,720 --> 01:50:32,637 should take only minimal fine-tuning 1995 01:50:32,720 --> 01:50:35,757 steps after the building the foundation 1996 01:50:35,840 --> 01:50:41,697 pre-trained model as few steps as possible >> basically [gasps] 1997 01:50:41,780 --> 01:50:44,557 [sighs] >> can Meta afford to do that no 1998 01:50:44,640 --> 01:50:46,637 >> so I don't know if you you know this but 1999 01:50:46,720 --> 01:50:48,077 companies are supposed to make money 2000 01:50:48,160 --> 01:50:52,157 somehow and uh open source is is is like 2001 01:50:52,240 --> 01:50:54,637 giving away. I don't know, Mark made a 2002 01:50:54,720 --> 01:50:57,517 video, Mark Zuckerberg, 2003 01:50:57,600 --> 01:51:00,557 uh, very sexy video talking about 2004 01:51:00,640 --> 01:51:04,957 350,000 Nvidia H100s. 2005 01:51:05,040 --> 01:51:08,157 >> Yeah, >> the the math of that is just for the 2006 01:51:08,240 --> 01:51:11,437 GPUs, that's a 100 billion 2007 01:51:11,520 --> 01:51:14,237 um, plus the infrastructure for training 2008 01:51:14,320 --> 01:51:19,037 everything. So, I'm no business guy, but 2009 01:51:19,120 --> 01:51:22,317 how do you make money on that? So the vision you paint is a really powerful 2010 01:51:22,400 --> 01:51:25,437 one. But how is it possible to make money? 2011 01:51:25,520 --> 01:51:29,437 >> Okay, so you have several business models, right? 2012 01:51:29,520 --> 01:51:31,917 >> The business model that uh Meta is built 2013 01:51:32,000 --> 01:51:35,037 around is 2014 01:51:35,120 --> 01:51:39,677 um you offer a service and the the 2015 01:51:39,760 --> 01:51:44,797 financing of that service is uh either through ads or through business 2016 01:51:44,880 --> 01:51:46,477 customers. So for example, if you have 2017 01:51:46,560 --> 01:51:49,917 an LLM that uh you know can help a mom 2018 01:51:50,000 --> 01:51:53,997 and pop pizza place um by you know 2019 01:51:54,080 --> 01:51:58,477 talking to the customers through WhatsApp and so the customers can just 2020 01:51:58,560 --> 01:52:01,997 order a pizza and the system will just you know ask them like what topping do 2021 01:52:02,080 --> 01:52:04,557 you want or what size blah blah blah [gasps] 2022 01:52:04,640 --> 01:52:07,197 um the business will pay for that. Okay, 2023 01:52:07,280 --> 01:52:11,960 that's a model. Um, 2024 01:52:12,160 --> 01:52:16,157 and otherwise, you know, if it's a system that is on the more kind of 2025 01:52:16,240 --> 01:52:20,477 classical services, it can be uh ad supported or, you know, there's several 2026 01:52:20,560 --> 01:52:23,037 models. But the point is 2027 01:52:23,120 --> 01:52:26,477 uh if you have a big enough um potential 2028 01:52:26,560 --> 01:52:30,157 customer base and you need to build that system 2029 01:52:30,240 --> 01:52:34,317 anyway for them, it doesn't hurt you to actually 2030 01:52:34,400 --> 01:52:36,477 distribute it in open source. Again, I'm 2031 01:52:36,560 --> 01:52:39,517 no business guy, but if you release the 2032 01:52:39,600 --> 01:52:42,077 open source model, then other people can 2033 01:52:42,160 --> 01:52:45,277 do the same kind of task and compete on 2034 01:52:45,360 --> 01:52:48,637 it. Basically, provide fine-tuned models 2035 01:52:48,720 --> 01:52:53,277 for businesses. So, the is the bet that Meta is making, by the way, I'm a huge 2036 01:52:53,360 --> 01:52:57,357 fan of all this, but is the bet that Meta is making is like we'll do a better 2037 01:52:57,440 --> 01:52:58,397 job of it. 2038 01:52:58,480 --> 01:53:01,917 >> Well, no. the the bet is is more we have 2039 01:53:02,000 --> 01:53:05,837 we already have a huge user base and customer base. 2040 01:53:05,920 --> 01:53:07,677 >> Ah right. >> Right. So it's going to be useful to 2041 01:53:07,760 --> 01:53:12,477 them. Whatever we offer them is going to be useful and there is a way to derive 2042 01:53:12,560 --> 01:53:14,077 revenue from this. 2043 01:53:14,160 --> 01:53:18,237 >> Uh and it doesn't hurt that you know we 2044 01:53:18,320 --> 01:53:20,557 provide that system 2045 01:53:20,640 --> 01:53:22,317 or the base the base model right the 2046 01:53:22,400 --> 01:53:25,677 foundation model uh in open source for 2047 01:53:25,760 --> 01:53:29,677 others to build applications on top of it too. If those applications turn out 2048 01:53:29,760 --> 01:53:33,117 to be useful for our customers, we can just buy it from them. 2049 01:53:33,200 --> 01:53:37,677 >> Um, uh, it could be that they will improve the platform. In fact, we see 2050 01:53:37,760 --> 01:53:39,997 this already. Um, I mean there is, you 2051 01:53:40,080 --> 01:53:45,197 know, literally millions of downloads of Latu and thousands of people who have, 2052 01:53:45,280 --> 01:53:46,957 you know, provided ideas about how to 2053 01:53:47,040 --> 01:53:49,677 make it better. Um, so, you know, this 2054 01:53:49,760 --> 01:53:52,397 this clearly accelerates progress to 2055 01:53:52,480 --> 01:53:55,757 make the system available to a sort of a 2056 01:53:55,840 --> 01:54:00,077 a wide community of people. and and there's literally thousands of 2057 01:54:00,160 --> 01:54:01,597 businesses who are building applications 2058 01:54:01,680 --> 01:54:06,397 with it. So um so 2059 01:54:06,480 --> 01:54:10,397 our ability to meta's ability to derive 2060 01:54:10,480 --> 01:54:12,637 revenue from this technology is not 2061 01:54:12,720 --> 01:54:14,317 impaired 2062 01:54:14,400 --> 01:54:17,757 uh by the distribution of it of base 2063 01:54:17,840 --> 01:54:22,077 models in open source. The fundamental criticism that Gemini is getting is that 2064 01:54:22,160 --> 01:54:26,637 as you point out on the west coast, just to just to clarify, we're currently in 2065 01:54:26,720 --> 01:54:28,957 the east coast where I would suppose 2066 01:54:29,040 --> 01:54:32,237 Meta AI headquarters would be. 2067 01:54:32,320 --> 01:54:34,957 So there uh strong words about the West 2068 01:54:35,040 --> 01:54:37,837 Coast, but uh I guess the issue that 2069 01:54:37,920 --> 01:54:40,877 happens is I think it's fair to say that 2070 01:54:40,960 --> 01:54:44,717 most tech people have a political 2071 01:54:44,800 --> 01:54:49,597 affiliation with the left wing. They're they lean left. And so the problem that 2072 01:54:49,680 --> 01:54:53,837 people are criticizing Gemini with is that there's in that debiasing process 2073 01:54:53,920 --> 01:54:58,157 that you mentioned that their ideological 2074 01:54:58,240 --> 01:55:00,317 lean 2075 01:55:00,400 --> 01:55:02,397 becomes obvious. 2076 01:55:02,480 --> 01:55:07,677 Uh is this something that could be escaped? 2077 01:55:07,760 --> 01:55:11,197 You're saying open source is the only way? Have have you witnessed this kind 2078 01:55:11,280 --> 01:55:16,157 of ideological lean that makes engineering difficult? No, I don't think 2079 01:55:16,240 --> 01:55:19,917 it has to do I don't think the issue has to do with the political leaning of the 2080 01:55:20,000 --> 01:55:22,477 people designing those systems. It has 2081 01:55:22,560 --> 01:55:26,637 to do with the uh acceptability or 2082 01:55:26,720 --> 01:55:29,677 political leanings of the their customer 2083 01:55:29,760 --> 01:55:32,397 base or audience, right? So, a big 2084 01:55:32,480 --> 01:55:35,357 company cannot afford to offend too many 2085 01:55:35,440 --> 01:55:39,117 people. So, they're going to make sure 2086 01:55:39,200 --> 01:55:41,597 that whatever product they put out is 2087 01:55:41,680 --> 01:55:45,677 safe, whatever that means. and 2088 01:55:45,760 --> 01:55:51,037 and it's very possible to overdo it and it's also very possible to it's 2089 01:55:51,120 --> 01:55:54,237 impossible to do it properly for everyone. You're not going to satisfy 2090 01:55:54,320 --> 01:55:56,557 everyone. So that's what I said before. 2091 01:55:56,640 --> 01:55:59,997 You cannot have a system that is unbiased is perceived as unbiased by 2092 01:56:00,080 --> 01:56:03,037 everyone. It's going to be, you know, 2093 01:56:03,120 --> 01:56:07,517 you push it in one way, one set of people are going to see it as biased and 2094 01:56:07,600 --> 01:56:09,917 then you push it the other way and another set of people is going to see it 2095 01:56:10,000 --> 01:56:15,037 as biased. And then in addition to this, there's the issue of if you push the 2096 01:56:15,120 --> 01:56:18,397 system perhaps a little too far in one direction, it's going to be non-factual, 2097 01:56:18,480 --> 01:56:21,117 right? You're going to have, you know, 2098 01:56:21,200 --> 01:56:24,237 you know, black Nazi soldiers in 2099 01:56:24,320 --> 01:56:25,837 >> Yeah. We should we should mention image 2100 01:56:25,920 --> 01:56:29,597 generation of of uh black Nazi soldiers, 2101 01:56:29,680 --> 01:56:31,677 which is not factually accurate, 2102 01:56:31,760 --> 01:56:34,157 >> right? And can be offensive for for some 2103 01:56:34,240 --> 01:56:39,437 people as well, right? So uh so you know it's going to be impossible 2104 01:56:39,520 --> 01:56:43,037 to kind of produce systems that are unbiased for everyone. So the only 2105 01:56:43,120 --> 01:56:45,997 solution that I see is diversity and 2106 01:56:46,080 --> 01:56:50,717 diversity the full meaning of that word diversity of in every possible way. 2107 01:56:50,800 --> 01:56:51,197 >> Yeah. 2108 01:56:51,280 --> 01:56:56,877 >> Um Mark Andre just tweeted today. 2109 01:56:56,960 --> 01:56:59,837 Let me do a TLDDR. The conclusion is 2110 01:56:59,920 --> 01:57:02,557 only startups and open source can avoid 2111 01:57:02,640 --> 01:57:04,317 the issue that he's highlighting with 2112 01:57:04,400 --> 01:57:07,357 big tech. He's asking can big tech 2113 01:57:07,440 --> 01:57:10,317 actually field generative AI products. 2114 01:57:10,400 --> 01:57:14,717 One ever escalating demands from internal activists, employee mobs, 2115 01:57:14,800 --> 01:57:17,197 crazed executives, broken boards, 2116 01:57:17,280 --> 01:57:19,117 pressure groups, extremist regulators, 2117 01:57:19,200 --> 01:57:21,757 government agencies, the press in quotes 2118 01:57:21,840 --> 01:57:26,157 experts and everything uh corrupting the 2119 01:57:26,240 --> 01:57:29,037 output. Two, constant risk of generating 2120 01:57:29,120 --> 01:57:31,757 a bad answer or drawing a bad picture or 2121 01:57:31,840 --> 01:57:35,197 rendering a bad video. Who knows what is 2122 01:57:35,280 --> 01:57:40,477 going to say or do at any moment? Three, legal exposure, product liability, 2123 01:57:40,560 --> 01:57:42,157 slander, election law, many other 2124 01:57:42,240 --> 01:57:45,677 things, and so on. Anything that makes 2125 01:57:45,760 --> 01:57:50,077 Congress mad. for continuous attempts to tighten grip 2126 01:57:50,160 --> 01:57:52,477 on acceptable output degrade the model 2127 01:57:52,560 --> 01:57:55,437 like how good it actually is. Uh in 2128 01:57:55,520 --> 01:57:59,917 terms of usable and pleasant to use and effective and all that kind of stuff and 2129 01:58:00,000 --> 01:58:02,397 five publicity of bad text, images, 2130 01:58:02,480 --> 01:58:05,117 video actual puts those examples into 2131 01:58:05,200 --> 01:58:07,757 the training data for the next version 2132 01:58:07,840 --> 01:58:09,917 so on. So he just highlights how 2133 01:58:10,000 --> 01:58:15,117 difficult this is from all kinds of people being unhappy. You said you can't 2134 01:58:15,200 --> 01:58:17,517 create a system that makes everybody happy. 2135 01:58:17,600 --> 01:58:20,237 >> Yes. >> Uh so if you're going to do the 2136 01:58:20,320 --> 01:58:24,797 finetuning yourself and keep a close source 2137 01:58:24,880 --> 01:58:27,837 essentially the problem there is then trying to minimize the number of people 2138 01:58:27,920 --> 01:58:30,317 who are going to be unhappy. >> Y 2139 01:58:30,400 --> 01:58:35,357 >> um and you're saying like the only that that almost impossible to do right and 2140 01:58:35,440 --> 01:58:38,397 that's the better way is to do open source 2141 01:58:38,480 --> 01:58:41,437 >> basically. Yeah. I mean he's Mark is 2142 01:58:41,520 --> 01:58:43,837 right about uh a number of things that 2143 01:58:43,920 --> 01:58:47,517 he lists that indeed scare large 2144 01:58:47,600 --> 01:58:51,277 companies. Uh you know certainly congressional 2145 01:58:51,360 --> 01:58:53,517 investigations is one of them. legal 2146 01:58:53,600 --> 01:58:57,837 liability. Uh you know uh making things 2147 01:58:57,920 --> 01:59:00,237 that get people to you know hurt 2148 01:59:00,320 --> 01:59:03,437 themselves or hurt others like you know 2149 01:59:03,520 --> 01:59:05,357 um big companies are really careful 2150 01:59:05,440 --> 01:59:07,597 about not um producing things of this 2151 01:59:07,680 --> 01:59:12,797 type and um [snorts] because they have you know they don't 2152 01:59:12,880 --> 01:59:15,677 want to hurt anyone first of all and then second they want to preserve their 2153 01:59:15,760 --> 01:59:18,077 business. So um it's essentially 2154 01:59:18,160 --> 01:59:23,117 impossible for systems like this that can inevitably formulate political 2155 01:59:23,200 --> 01:59:26,637 opinions and you know opinions about various things that may be political or 2156 01:59:26,720 --> 01:59:28,877 not but that people may disagree about 2157 01:59:28,960 --> 01:59:33,037 about you know moral issues and you know 2158 01:59:33,120 --> 01:59:35,757 um things about like questions about 2159 01:59:35,840 --> 01:59:40,077 religion and things like that, right? Or or cultural issues that people from 2160 01:59:40,160 --> 01:59:43,277 different communities would disagree with in the first place. Um so there's 2161 01:59:43,360 --> 01:59:44,957 only kind of a relatively small number 2162 01:59:45,040 --> 01:59:48,797 of things that people will uh sort of 2163 01:59:48,880 --> 01:59:52,637 agree on you know basic principles but beyond that if you if you want those 2164 01:59:52,720 --> 01:59:54,957 systems to be useful they will 2165 01:59:55,040 --> 01:59:58,877 necessarily have to uh offend 2166 01:59:58,960 --> 02:00:03,837 a number of people inevitably >> and so open source is just better and 2167 02:00:03,920 --> 02:00:08,317 then >> diversity is better right >> and open source enables diversity 2168 02:00:08,400 --> 02:00:10,477 >> that's right open source enables diversity 2169 02:00:10,560 --> 02:00:14,877 >> that that's going to be fascinating world where if it's true that the open 2170 02:00:14,960 --> 02:00:18,637 source world if metal leads the way and creates this kind of open- source 2171 02:00:18,720 --> 02:00:20,957 foundation model world there's going to 2172 02:00:21,040 --> 02:00:23,357 be like governments will have a fine new 2173 02:00:23,440 --> 02:00:25,997 model and 2174 02:00:26,080 --> 02:00:28,557 >> and and then potentially 2175 02:00:28,640 --> 02:00:33,037 uh you know people that vote left and right will have their own model and 2176 02:00:33,120 --> 02:00:35,117 preference and be able to choose and it 2177 02:00:35,200 --> 02:00:37,997 will potentially divide us even more but 2178 02:00:38,080 --> 02:00:41,037 that's on us humans we get to figure out 2179 02:00:41,120 --> 02:00:43,037 basically the techn technology enables 2180 02:00:43,120 --> 02:00:47,277 humans to human more effectively and all 2181 02:00:47,360 --> 02:00:49,277 the difficult ethical questions that 2182 02:00:49,360 --> 02:00:51,997 humans raise will just 2183 02:00:52,080 --> 02:00:55,357 um leave it up to us to figure that out. 2184 02:00:55,440 --> 02:00:58,237 Yeah. I mean there are some limits to what you know the same way there are 2185 02:00:58,320 --> 02:00:59,917 limits to free speech there has to be 2186 02:01:00,000 --> 02:01:01,837 some limit to the kind of stuff that 2187 02:01:01,920 --> 02:01:05,517 those systems might uh be authorized to 2188 02:01:05,600 --> 02:01:08,717 um to produce um you know some guard 2189 02:01:08,800 --> 02:01:12,877 rails. So I mean that's one thing I've been interested in which is uh in the 2190 02:01:12,960 --> 02:01:14,237 type of architecture that we were 2191 02:01:14,320 --> 02:01:17,917 discussing before where the output of a 2192 02:01:18,000 --> 02:01:20,557 system is the result of an inference to 2193 02:01:20,640 --> 02:01:23,517 satisfy an objective that objective can 2194 02:01:23,600 --> 02:01:25,677 include guardrails 2195 02:01:25,760 --> 02:01:29,197 >> and uh we can put guardrails in open 2196 02:01:29,280 --> 02:01:31,277 source systems. I mean if we eventually 2197 02:01:31,360 --> 02:01:36,237 have systems that are built with this blueprint uh we can put guard rails uh 2198 02:01:36,320 --> 02:01:40,157 in those systems that guarantee that there is sort of a a minimum [snorts] 2199 02:01:40,240 --> 02:01:44,157 set of guardrails that make the system non-dangerous and non-toxic etc. You 2200 02:01:44,240 --> 02:01:49,197 know basic things that everybody would agree on. Um and and then you know the 2201 02:01:49,280 --> 02:01:52,717 the finetuning that people will add or the additional guardrails that people 2202 02:01:52,800 --> 02:01:55,917 will add will kind of cater to their um 2203 02:01:56,000 --> 02:01:59,837 community whatever it is >> and yeah the fine tuning will be more 2204 02:01:59,920 --> 02:02:03,597 about the gray areas of what is hate speech what is dangerous and all that 2205 02:02:03,680 --> 02:02:06,237 kind of stuff I mean you've >> different value systems 2206 02:02:06,320 --> 02:02:09,757 >> different value systems I mean like uh but still even with the objectives of 2207 02:02:09,840 --> 02:02:13,677 how to build a boweapon for example I think something you've commented on or 2208 02:02:13,760 --> 02:02:17,677 at least there's a paper where collection of researchers is 2209 02:02:17,760 --> 02:02:20,077 trying to understand the social impacts 2210 02:02:20,160 --> 02:02:22,237 of these LLMs 2211 02:02:22,320 --> 02:02:24,877 and I guess one threshold is nice is 2212 02:02:24,960 --> 02:02:27,997 like does the LLM make it any easier 2213 02:02:28,080 --> 02:02:31,357 than a than a search would like a Google 2214 02:02:31,440 --> 02:02:32,317 search would 2215 02:02:32,400 --> 02:02:35,917 >> right so the increasing uh number of 2216 02:02:36,000 --> 02:02:40,317 studies on this seems to point to 2217 02:02:40,400 --> 02:02:43,117 the fact that it doesn't help so having 2218 02:02:43,200 --> 02:02:45,357 an LLM doesn't help you 2219 02:02:45,440 --> 02:02:48,797 >> uh design or built a boweapon or a 2220 02:02:48,880 --> 02:02:50,717 chemical weapon if you already have 2221 02:02:50,800 --> 02:02:56,077 access to uh you know a search engine and a library. Uh and and so the the 2222 02:02:56,160 --> 02:02:59,117 sort of increased information you get or the ease with which you get it doesn't 2223 02:02:59,200 --> 02:03:03,437 really help you. Um that's the first thing. The second thing is it's one 2224 02:03:03,520 --> 02:03:08,317 thing to have a list of instructions of how to make a [snorts] chemical weapon 2225 02:03:08,400 --> 02:03:10,317 for example or boweapon. It's another 2226 02:03:10,400 --> 02:03:13,517 thing to actually build it and it's much 2227 02:03:13,600 --> 02:03:18,637 harder than you might think. and an NLM will not help you with that. Um, in 2228 02:03:18,720 --> 02:03:21,437 fact, you know, nobody in the world, not even like, you know, countries use 2229 02:03:21,520 --> 02:03:24,397 bioweapons because most of the time they 2230 02:03:24,480 --> 02:03:28,797 have no idea how to protect their own populations against it. So, um, so it's 2231 02:03:28,880 --> 02:03:30,797 too dangerous actually to kind of ever 2232 02:03:30,880 --> 02:03:34,877 use um, and it's in fact banned by uh, 2233 02:03:34,960 --> 02:03:37,357 uh, international treaties. Um, chemical 2234 02:03:37,440 --> 02:03:42,557 weapons is different. It's also banned by treaties but um, [snorts] uh, but 2235 02:03:42,640 --> 02:03:44,317 it's the same problem. It's difficult to 2236 02:03:44,400 --> 02:03:47,357 use in situations that doesn't turn 2237 02:03:47,440 --> 02:03:49,917 against the perpetrators. But we could 2238 02:03:50,000 --> 02:03:54,637 ask Elon Musk like I can I can give you a very precise list of instructions of 2239 02:03:54,720 --> 02:03:56,317 how you build a rocket engine. 2240 02:03:56,400 --> 02:04:00,397 >> Mhm. >> And even if you have a team of 50 engineers that are really experienced 2241 02:04:00,480 --> 02:04:03,277 building it, you're still going to have to blow up a dozen of them before you 2242 02:04:03,360 --> 02:04:05,117 get one that works. 2243 02:04:05,200 --> 02:04:08,657 >> Um, and you know, it's the same with u, 2244 02:04:08,740 --> 02:04:12,717 [snorts] you know, the chemical weapons or bio weapons or things like this. It 2245 02:04:12,800 --> 02:04:17,117 requires expertise, you know, in the in the real world that another line is not 2246 02:04:17,200 --> 02:04:20,317 going to help you with. >> And it requires even the common sense 2247 02:04:20,400 --> 02:04:21,837 expertise that we've been talking about, 2248 02:04:21,920 --> 02:04:26,077 which is how to take uh language based 2249 02:04:26,160 --> 02:04:28,477 instructions and materialize them in the 2250 02:04:28,560 --> 02:04:33,117 physical world requires a lot of knowledge that's not in the 2251 02:04:33,200 --> 02:04:36,157 instructions. >> Yeah, exactly. A lot of biologists have 2252 02:04:36,240 --> 02:04:39,677 posted on this actually in response to those things saying like do you realize 2253 02:04:39,760 --> 02:04:43,037 how hard it is to actually do the the lab work like you know this is not 2254 02:04:43,120 --> 02:04:44,797 trivial. 2255 02:04:44,880 --> 02:04:47,997 >> Yeah. And that's Hans Marvik comes comes 2256 02:04:48,080 --> 02:04:52,877 to light once again. Uh just to linger on llama you know Mark announced that 2257 02:04:52,960 --> 02:04:56,957 Llama 3 is coming out eventually. I don't think there's a release date but 2258 02:04:57,040 --> 02:04:59,677 what what are you most excited about? 2259 02:04:59,760 --> 02:05:03,517 First of all, Llama 2 that's already out there and maybe the future Alama 3, 2260 02:05:03,600 --> 02:05:06,637 four, five, six, 10, just the the future 2261 02:05:06,720 --> 02:05:09,917 of the open source under meta. 2262 02:05:10,000 --> 02:05:11,997 Well, a number of things. So, uh there's 2263 02:05:12,080 --> 02:05:14,157 going to be like various versions of of 2264 02:05:14,240 --> 02:05:17,837 llama that are uh you know improvements 2265 02:05:17,920 --> 02:05:21,197 of previous llamas, bigger, better, 2266 02:05:21,280 --> 02:05:25,757 multimodel, things like that. And then in future generations, systems that are 2267 02:05:25,840 --> 02:05:29,757 capable of planning, that really understand how the world works. um maybe 2268 02:05:29,840 --> 02:05:33,997 are trained from video so they have some world model maybe you know capable of 2269 02:05:34,080 --> 02:05:37,517 the type of reasoning and planning I was talking about earlier like how long is 2270 02:05:37,600 --> 02:05:40,877 that going to take like when is the research that is doing going in that 2271 02:05:40,960 --> 02:05:43,677 direction going to sort of feed into the 2272 02:05:43,760 --> 02:05:48,157 product line if you want of llama I don't know I can't tell you and there's 2273 02:05:48,240 --> 02:05:49,437 you know a few breakthroughs that we 2274 02:05:49,520 --> 02:05:53,357 have to basically uh go through before 2275 02:05:53,440 --> 02:05:56,077 we can get there but you'll be able to 2276 02:05:56,160 --> 02:05:58,957 monitor our progress because we publish 2277 02:05:59,040 --> 02:06:01,917 our research, right? So, you know, if 2278 02:06:02,000 --> 02:06:04,797 last week we published the Vijpa work, 2279 02:06:04,880 --> 02:06:09,197 which is sort of a first step towards training systems from video um and then 2280 02:06:09,280 --> 02:06:11,357 the next step is going to be world 2281 02:06:11,440 --> 02:06:14,077 models based on kind of this type of 2282 02:06:14,160 --> 02:06:16,477 idea training training from video. 2283 02:06:16,560 --> 02:06:18,477 There's similar work at uh at Deepind 2284 02:06:18,560 --> 02:06:21,917 also and um taking place people and also 2285 02:06:22,000 --> 02:06:26,237 at UC Berkeley on uh world models from 2286 02:06:26,320 --> 02:06:29,117 video. A lot of people are working on this. I think a lot of good ideas are 2287 02:06:29,200 --> 02:06:32,637 coming are appearing. My bet is that 2288 02:06:32,720 --> 02:06:34,557 those systems are going to be JEAL like 2289 02:06:34,640 --> 02:06:36,237 they're not going to be gener generative 2290 02:06:36,320 --> 02:06:39,997 models. Um [snorts] and uh we'll see 2291 02:06:40,080 --> 02:06:43,597 what the future will tell. Um there's 2292 02:06:43,680 --> 02:06:47,437 really good work at uh um a gentleman 2293 02:06:47,520 --> 02:06:51,197 called Danish Hafner who's not deep mind who who's worked on kind of models of 2294 02:06:51,280 --> 02:06:52,877 this type that learn representations and 2295 02:06:52,960 --> 02:06:55,437 then use them for planning or learning 2296 02:06:55,520 --> 02:06:59,037 tasks by reinforcement learning. Um and 2297 02:06:59,120 --> 02:07:02,397 a lot of work at Berkeley by Peter 2298 02:07:02,480 --> 02:07:04,877 Savine bunch of other people of that 2299 02:07:04,960 --> 02:07:07,677 type I'm collaborating with actually in 2300 02:07:07,760 --> 02:07:10,637 the context of some grants with my NYU 2301 02:07:10,720 --> 02:07:14,077 hat. Um and then collaborations also 2302 02:07:14,160 --> 02:07:17,037 through meta because the the lab at 2303 02:07:17,120 --> 02:07:19,517 Berkeley is associated with meta in some 2304 02:07:19,600 --> 02:07:22,557 way. So with fair so I I think uh it's 2305 02:07:22,640 --> 02:07:25,277 very exciting you know I I think I'm 2306 02:07:25,360 --> 02:07:27,677 super excited about I I haven't been 2307 02:07:27,760 --> 02:07:32,557 that excited about like the direction of machine learning and AI you know since 2308 02:07:32,640 --> 02:07:34,637 uh you know 10 years ago when fair was 2309 02:07:34,720 --> 02:07:37,597 started and before that um 30 years ago 2310 02:07:37,680 --> 02:07:40,797 when we were working on 35 on on 2311 02:07:40,880 --> 02:07:43,597 commercial nets and and and the early 2312 02:07:43,680 --> 02:07:46,877 days of neural nets so um I'm super 2313 02:07:46,960 --> 02:07:50,237 excited because I see a path towards 2314 02:07:50,320 --> 02:07:52,877 potentially human level intelligence 2315 02:07:52,960 --> 02:07:55,197 uh with you know systems that can 2316 02:07:55,280 --> 02:07:57,997 understand the world remember plan 2317 02:07:58,080 --> 02:08:01,517 reason um there there is some some set 2318 02:08:01,600 --> 02:08:05,357 of ideas to make progress there that might have a chance of working and I'm 2319 02:08:05,440 --> 02:08:08,077 really excited about this what I like is 2320 02:08:08,160 --> 02:08:11,517 that you know it uh 2321 02:08:11,600 --> 02:08:15,837 somewhat we we get onto like a good direction and perhaps succeed before my 2322 02:08:15,920 --> 02:08:18,317 brain turns to a white sauce or or 2323 02:08:18,400 --> 02:08:21,277 before I need to retire. [laughter] 2324 02:08:21,360 --> 02:08:24,797 >> Yeah. Yeah. You're also excited by are 2325 02:08:24,880 --> 02:08:27,037 you 2326 02:08:27,120 --> 02:08:29,117 is it beautiful to you just the amount 2327 02:08:29,200 --> 02:08:32,557 of GPUs involved sort of the the the 2328 02:08:32,640 --> 02:08:34,397 whole training process on this much 2329 02:08:34,480 --> 02:08:37,277 compute. It's just zooming out just 2330 02:08:37,360 --> 02:08:39,837 looking at earth and humans together 2331 02:08:39,920 --> 02:08:42,797 have built these computing devices and 2332 02:08:42,880 --> 02:08:46,317 are able to train this one brain. 2333 02:08:46,400 --> 02:08:50,317 Then then we then open source [laughter] 2334 02:08:50,400 --> 02:08:53,037 like giving birth to this open-source 2335 02:08:53,120 --> 02:08:56,157 brain trained on this gigantic compute 2336 02:08:56,240 --> 02:09:01,197 system. There's just the details of how to train on that, how to build the 2337 02:09:01,280 --> 02:09:05,277 infrastructure and the the hardware, the cooling, all of this kind of stuff. 2338 02:09:05,360 --> 02:09:06,717 >> Are you just still most of your 2339 02:09:06,800 --> 02:09:08,717 excitement is in the the theory aspect 2340 02:09:08,800 --> 02:09:12,397 of it, the meaning like the software? 2341 02:09:12,480 --> 02:09:14,397 >> Well, I used to be a hardware guy many years ago. 2342 02:09:14,480 --> 02:09:15,117 >> Yes. [laughter] Yes. That's right. 2343 02:09:15,200 --> 02:09:19,837 decades ago >> hardware has improved a little bit, changed a little bit. 2344 02:09:19,920 --> 02:09:25,197 >> Yeah. >> I mean certainly scale is necessary but not sufficient. 2345 02:09:25,280 --> 02:09:27,677 >> Absolutely. >> So we certainly need computation. I mean 2346 02:09:27,760 --> 02:09:32,557 we're still far in terms of comput power uh from you know what we would need to 2347 02:09:32,640 --> 02:09:36,637 match the compute power of the human brain. Um you know this may occur in the 2348 02:09:36,720 --> 02:09:41,117 next couple decades but um but we're still some ways away and certainly in 2349 02:09:41,200 --> 02:09:43,117 terms of power efficiency we're really 2350 02:09:43,200 --> 02:09:45,917 far. Um so there's a lot of progress to 2351 02:09:46,000 --> 02:09:50,077 make in uh in in in hardware and you 2352 02:09:50,160 --> 02:09:53,917 know right now a lot of the progress is is is not I mean there's a bit coming 2353 02:09:54,000 --> 02:09:56,957 from silicon technology but a lot of it 2354 02:09:57,040 --> 02:09:59,277 coming from architectural innovation and 2355 02:09:59,360 --> 02:10:04,157 quite a bit coming from uh like more efficient ways of you know implementing 2356 02:10:04,240 --> 02:10:05,837 the architectures that have become 2357 02:10:05,920 --> 02:10:08,717 popular basically combination of 2358 02:10:08,800 --> 02:10:12,477 transformers and comnets right and uh so 2359 02:10:12,560 --> 02:10:15,037 you know there's still some ways to go 2360 02:10:15,120 --> 02:10:17,197 until 2361 02:10:17,280 --> 02:10:20,557 uh we're going to saturate. We're going 2362 02:10:20,640 --> 02:10:22,397 to have to come up with like new new 2363 02:10:22,480 --> 02:10:24,717 principles, new fabrication technology, 2364 02:10:24,800 --> 02:10:27,997 new uh basic components. 2365 02:10:28,080 --> 02:10:32,157 Um perhaps you know based on sort of different principles than those 2366 02:10:32,240 --> 02:10:34,717 classical digital semos. 2367 02:10:34,800 --> 02:10:37,677 >> Interesting. So you think in order to 2368 02:10:37,760 --> 02:10:42,557 build AMI and me we need we potentially 2369 02:10:42,640 --> 02:10:45,517 might need some hardware innovation too. 2370 02:10:45,600 --> 02:10:49,677 >> Well if you want to make it um ubiquitous yeah certainly because we're 2371 02:10:49,760 --> 02:10:51,917 going to have to reduce the 2372 02:10:52,000 --> 02:10:55,037 you know comput the power consumption. A 2373 02:10:55,120 --> 02:10:57,837 GPU today right is half a kilowatt to a 2374 02:10:57,920 --> 02:10:59,597 kilowatt. 2375 02:10:59,680 --> 02:11:02,717 >> Human brain is about 25 watts. 2376 02:11:02,800 --> 02:11:06,717 uh and a GPU is way below the power of human brain. You need you know something 2377 02:11:06,800 --> 02:11:08,877 like 100,000 or a million to match it. 2378 02:11:08,960 --> 02:11:11,997 So uh so you know we are off by a huge 2379 02:11:12,080 --> 02:11:14,237 factor here. 2380 02:11:14,320 --> 02:11:17,117 You often say that AGI 2381 02:11:17,200 --> 02:11:20,157 is not coming soon meaning like not this 2382 02:11:20,240 --> 02:11:24,637 year not the next few years potentially 2383 02:11:24,720 --> 02:11:26,797 farther away. What's your basic 2384 02:11:26,880 --> 02:11:29,357 intuition behind that? So first of all 2385 02:11:29,440 --> 02:11:33,837 it's not going to be an event right the idea somehow which you know is 2386 02:11:33,920 --> 02:11:37,997 popularized by science fiction and Hollywood that you know somehow somebody 2387 02:11:38,080 --> 02:11:40,077 is going to discover the secret the 2388 02:11:40,160 --> 02:11:43,677 secret to AGI or human level AI or AMI 2389 02:11:43,760 --> 02:11:47,117 whatever you want to call it and then you know turn on a machine and then we 2390 02:11:47,200 --> 02:11:49,437 have a GI that's just not going to 2391 02:11:49,520 --> 02:11:52,477 happen it's not going to be an event 2392 02:11:52,560 --> 02:11:56,237 it's going to be gradual progress 2393 02:11:56,320 --> 02:11:58,797 are we going to have systems that and 2394 02:11:58,880 --> 02:12:03,357 learn from video how the world works and learn good representations. Yeah. Before 2395 02:12:03,440 --> 02:12:07,117 we get them to the scale and performance that we observe in humans, it's going to 2396 02:12:07,200 --> 02:12:08,797 take quite a while. It's not going to 2397 02:12:08,880 --> 02:12:11,917 happen in one day. Um 2398 02:12:12,000 --> 02:12:16,957 uh are we going to get systems that can uh have large amount of associative 2399 02:12:17,040 --> 02:12:20,797 memory so they can they can remember stuff? Yeah, but same it's not going to 2400 02:12:20,880 --> 02:12:23,757 happen tomorrow. I mean there is some basic techniques that need to be 2401 02:12:23,840 --> 02:12:27,197 developed. We have a lot of them, but like you know to get this to work 2402 02:12:27,280 --> 02:12:30,717 together with full system is another story. Are we going to have systems that 2403 02:12:30,800 --> 02:12:32,797 can reason and plan perhaps along the 2404 02:12:32,880 --> 02:12:37,757 lines of objective driven AI architectures that I I described before. 2405 02:12:37,840 --> 02:12:40,637 Yeah. But like before we get this to work, you know, properly, it's going to 2406 02:12:40,720 --> 02:12:44,717 take a while. So, and before we get all those things to work together and then 2407 02:12:44,800 --> 02:12:47,917 on top of this have systems that can learn like hierarchical planning, 2408 02:12:48,000 --> 02:12:52,477 hierarchical representations, systems that can be configured for a lot of 2409 02:12:52,560 --> 02:12:54,077 different situation at hands. the way 2410 02:12:54,160 --> 02:12:57,997 the human brain can. Um, you know, all 2411 02:12:58,080 --> 02:13:01,757 of this is going to take, you know, at least a decade and probably much more 2412 02:13:01,840 --> 02:13:06,477 because there are a lot of problems that we're not seeing right now that we have 2413 02:13:06,560 --> 02:13:08,717 not encountered and so we don't know if 2414 02:13:08,800 --> 02:13:10,637 there is a easy solution within this 2415 02:13:10,720 --> 02:13:15,117 framework. Um, so you know it's it's not 2416 02:13:15,200 --> 02:13:18,797 just around the corner. I mean, I've I've been hearing people for the last 2417 02:13:18,880 --> 02:13:21,357 12, 15 years claiming that, you know, 2418 02:13:21,440 --> 02:13:25,837 AGI is just around the corner and being systematically wrong, and I knew they 2419 02:13:25,920 --> 02:13:29,277 were wrong when they were saying it. I call their 2420 02:13:29,360 --> 02:13:31,997 >> Why do you think people have been calling? First of all, I mean, from the 2421 02:13:32,080 --> 02:13:36,237 beginning of from the birth of the term artificial intelligence, there has been 2422 02:13:36,320 --> 02:13:38,957 a eternal optimism 2423 02:13:39,040 --> 02:13:40,397 that's perhaps unlike other 2424 02:13:40,480 --> 02:13:45,437 technologies. Is it Maravic paradox? is 2425 02:13:45,520 --> 02:13:49,677 the explanation for why people are so optimistic about AGI. 2426 02:13:49,760 --> 02:13:51,437 >> I don't think it's just morxic paradox. 2427 02:13:51,520 --> 02:13:53,117 Morex paradox is a consequence of 2428 02:13:53,200 --> 02:13:55,997 realizing that the world is not as easy 2429 02:13:56,080 --> 02:13:59,277 as we think. So first of all um 2430 02:13:59,360 --> 02:14:01,437 intelligence is not a linear thing that 2431 02:14:01,520 --> 02:14:03,597 you can measure with a scalar with a 2432 02:14:03,680 --> 02:14:06,877 single number. Um you know can you say 2433 02:14:06,960 --> 02:14:09,677 that humans are smarter than 2434 02:14:09,760 --> 02:14:13,197 orangutong? In some ways, yes. But in 2435 02:14:13,280 --> 02:14:17,037 some ways, orongs are smarter than humans in a lot of domains that allows 2436 02:14:17,120 --> 02:14:21,117 them to survive in the forest, for example. So, IQ is a very limited 2437 02:14:21,200 --> 02:14:25,757 measure of intelligence. Human intelligence is bigger than what IQ, for 2438 02:14:25,840 --> 02:14:29,357 example, measures. >> Well, IQ can measure, you know, 2439 02:14:29,440 --> 02:14:31,677 approximately something for humans. 2440 02:14:31,760 --> 02:14:35,517 >> Mhm. But um because humans kind of, you 2441 02:14:35,600 --> 02:14:40,557 know, come in relatively kind of uniform 2442 02:14:40,640 --> 02:14:42,317 form, right? 2443 02:14:42,400 --> 02:14:45,917 >> Uh but it only measures one type of uh 2444 02:14:46,000 --> 02:14:47,837 ability that you know may be relevant 2445 02:14:47,920 --> 02:14:51,837 for some tasks but not others. And 2446 02:14:51,920 --> 02:14:53,437 uh but then if you are talking about 2447 02:14:53,520 --> 02:14:57,917 other intelligent entities for which the 2448 02:14:58,000 --> 02:14:59,517 you know the basic things that are easy 2449 02:14:59,600 --> 02:15:02,557 to them is very different then it 2450 02:15:02,640 --> 02:15:05,597 doesn't mean anything. So intelligence 2451 02:15:05,680 --> 02:15:10,797 is a collection of skills 2452 02:15:10,880 --> 02:15:14,807 and an ability to acquire new skills efficiently. 2453 02:15:14,890 --> 02:15:19,277 [clears throat] >> Right. And the collection of skills that an 2454 02:15:19,360 --> 02:15:23,837 intelligent particular intelligent entity possess or is capable of learning 2455 02:15:23,920 --> 02:15:28,397 quickly is different from the collection of skills of another one. And because 2456 02:15:28,480 --> 02:15:32,397 it's a multi-dimensional thing, the set of skills is high dimensional space. You 2457 02:15:32,480 --> 02:15:36,237 can't measure, you can compare, you cannot compare two things as to whether 2458 02:15:36,320 --> 02:15:38,397 one is more intelligent than the other. 2459 02:15:38,480 --> 02:15:41,437 It's multi-dimensional. 2460 02:15:41,520 --> 02:15:44,717 So you push back against 2461 02:15:44,800 --> 02:15:48,957 what are called AI doomers a lot. Uh can 2462 02:15:49,040 --> 02:15:52,477 you explain their perspective and why you think they're wrong? 2463 02:15:52,560 --> 02:15:55,037 >> Okay. So AI doomers imagine all kinds of 2464 02:15:55,120 --> 02:15:58,397 catastrophe scenarios of how AI could 2465 02:15:58,480 --> 02:16:01,997 escape our control and basically kill us 2466 02:16:02,080 --> 02:16:03,517 all. [laughter] 2467 02:16:03,600 --> 02:16:06,557 Uh and that relies on a whole bunch of 2468 02:16:06,640 --> 02:16:09,357 assumptions that are mostly false. 2469 02:16:09,440 --> 02:16:13,037 So the first assumption is that the emergence of super intelligence is going 2470 02:16:13,120 --> 02:16:17,277 to be an event that at some point we're going to have we're going to figure out 2471 02:16:17,360 --> 02:16:21,997 the secret and we'll turn on a machine that is super intelligent and because 2472 02:16:22,080 --> 02:16:25,677 we've never done it before is going to take over the world and kill us all. 2473 02:16:25,760 --> 02:16:27,437 That is false. It's not going to be an 2474 02:16:27,520 --> 02:16:29,597 event. We're going to have systems that 2475 02:16:29,680 --> 02:16:33,037 are like as smart as a cat. has all the 2476 02:16:33,120 --> 02:16:37,757 have all the characteristics of, you know, human level intelligence, but 2477 02:16:37,840 --> 02:16:39,117 their level of intelligence would be 2478 02:16:39,200 --> 02:16:42,157 like a cat or a parrot maybe or 2479 02:16:42,240 --> 02:16:47,277 something. Um, and then we're going to work our way up to kind of make those 2480 02:16:47,360 --> 02:16:49,997 things more intelligent. And as we make them more intelligent, we're also going 2481 02:16:50,080 --> 02:16:53,277 to put some guardrails in them and learn how to kind of put some guardrails so 2482 02:16:53,360 --> 02:16:56,797 they behave properly. And we're not going to do this with just one, it's not 2483 02:16:56,880 --> 02:16:59,517 going to be one effort, but it's going to be lots of different people doing 2484 02:16:59,600 --> 02:17:03,437 this. And some of them are going to succeed at making intelligent systems 2485 02:17:03,520 --> 02:17:07,757 that are uh controllable and safe and have the right guardrails. And if some 2486 02:17:07,840 --> 02:17:10,317 other goes rogue, then we can use the 2487 02:17:10,400 --> 02:17:12,077 the good ones to go against the rogue 2488 02:17:12,160 --> 02:17:14,797 ones. So it's going to be my, you know, 2489 02:17:14,880 --> 02:17:17,597 smart AI police against your rogue AI. 2490 02:17:17,680 --> 02:17:20,717 Um so it's not going to be like, you know, we're going to be exposed to like 2491 02:17:20,800 --> 02:17:24,557 a single rogue AI that's going to kill us all. That's just not not happening. 2492 02:17:24,640 --> 02:17:28,317 Now there is another fallacy which is the fact that because a system is 2493 02:17:28,400 --> 02:17:31,117 intelligent it necessarily wants to take over. 2494 02:17:31,200 --> 02:17:36,077 >> Mhm. >> Um and there is several arguments 2495 02:17:36,160 --> 02:17:38,077 that make people scare of this which I 2496 02:17:38,160 --> 02:17:41,277 think are completely false uh as well. 2497 02:17:41,360 --> 02:17:45,037 So uh one of them is um you know in 2498 02:17:45,120 --> 02:17:47,357 nature it seems to be that the more 2499 02:17:47,440 --> 02:17:49,037 intelligent species are the one that end 2500 02:17:49,120 --> 02:17:54,157 up dominating the other and uh and even 2501 02:17:54,240 --> 02:17:58,797 you know extinguishing the others sometimes by design sometimes just by 2502 02:17:58,880 --> 02:18:02,237 mistake and 2503 02:18:02,320 --> 02:18:05,437 and so you know there is sort of a 2504 02:18:05,520 --> 02:18:08,317 thinking by which you say well if AI 2505 02:18:08,400 --> 02:18:12,717 systems are more intelligent than us surely they're going uh eliminate us if 2506 02:18:12,800 --> 02:18:17,517 not by design simply because they don't care about us. And that's just 2507 02:18:17,600 --> 02:18:21,677 preposteruous for for a number of reasons. Um first reason is they're not 2508 02:18:21,760 --> 02:18:25,837 going to be a species. They're not going to be a species that competes with us. 2509 02:18:25,920 --> 02:18:29,437 They're not going to have the desire to dominate because the desire to dominate 2510 02:18:29,520 --> 02:18:31,357 is something that has to be hardwired 2511 02:18:31,440 --> 02:18:35,197 into an intelligent system. Uh it is 2512 02:18:35,280 --> 02:18:38,317 hardwired in humans. It is hardwired in 2513 02:18:38,400 --> 02:18:41,517 baboons, in chimpanzees, in wolves, 2514 02:18:41,600 --> 02:18:43,997 not in or wrongs. 2515 02:18:44,080 --> 02:18:47,357 The species in which this desire to 2516 02:18:47,440 --> 02:18:50,797 dominate or submit or or attain status 2517 02:18:50,880 --> 02:18:53,037 in other ways 2518 02:18:53,120 --> 02:18:55,997 is is specific to social species. 2519 02:18:56,080 --> 02:18:59,117 Non-social species like don't have it, 2520 02:18:59,200 --> 02:19:03,997 right? And they are as smart as we are almost, right? And to you, there's not 2521 02:19:04,080 --> 02:19:05,597 significant incentive for humans to 2522 02:19:05,680 --> 02:19:08,557 encode that into the AI systems. And to 2523 02:19:08,640 --> 02:19:11,117 the degree they do, there will be other 2524 02:19:11,200 --> 02:19:14,797 AIs that um sort of punish them for it. 2525 02:19:14,880 --> 02:19:17,197 I'll compete them over. >> Well, there's all kinds of incentive to 2526 02:19:17,280 --> 02:19:19,917 make AI systems submissive to humans, right? 2527 02:19:20,000 --> 02:19:23,597 >> Right. I mean, this is the way we're going to build them, right? Um and so, 2528 02:19:23,680 --> 02:19:26,797 so then people say, "Oh, but look at LLMs. LM are not controllable." And 2529 02:19:26,880 --> 02:19:28,477 they're right. OLMs are not 2530 02:19:28,560 --> 02:19:30,717 controllable. But objectived driven AI, 2531 02:19:30,800 --> 02:19:34,557 so systems that derive their answers by 2532 02:19:34,640 --> 02:19:38,797 optimization of an objective means they have to optimize its objective. And that 2533 02:19:38,880 --> 02:19:41,197 objective can include guardrails. One 2534 02:19:41,280 --> 02:19:45,757 guardrail is uh obey humans. Another 2535 02:19:45,840 --> 02:19:50,477 guardrail is don't obey humans if it's hurting other humans within. 2536 02:19:50,560 --> 02:19:52,397 >> I've heard that before somewhere. I don't remember. 2537 02:19:52,480 --> 02:19:54,797 >> Yes. [laughter] >> Maybe in a book. 2538 02:19:54,880 --> 02:19:58,157 >> Yeah. Uh but speaking of that book, what 2539 02:19:58,240 --> 02:20:01,917 is could there be unintended consequences also from all of this? 2540 02:20:02,000 --> 02:20:06,477 >> No, of course. Uh so this is not a simple problem, right? I mean designing 2541 02:20:06,560 --> 02:20:10,077 those guardrails so that the system behaves properly is not going to be a 2542 02:20:10,160 --> 02:20:14,557 simple u issue that for which there is a silver 2543 02:20:14,640 --> 02:20:18,637 bullet for which you have a mathematical proof that the system can be safe. It's 2544 02:20:18,720 --> 02:20:23,197 going to be very progressive iterative design system where we put those 2545 02:20:23,280 --> 02:20:26,477 guardrails in such a way that the system behave properly and sometimes they're 2546 02:20:26,560 --> 02:20:30,877 going to do something that you know was unexpected because the guardrail wasn't 2547 02:20:30,960 --> 02:20:34,317 right and we're going to correct them so that they do it right. Uh the idea 2548 02:20:34,400 --> 02:20:36,477 somehow that we can't get it slightly 2549 02:20:36,560 --> 02:20:41,517 wrong because if we get it slightly wrong we all die is is ridiculous. 2550 02:20:41,600 --> 02:20:45,197 um we we're just going to go progressively and it's it's just going 2551 02:20:45,280 --> 02:20:47,597 to be the the analogy I've used many 2552 02:20:47,680 --> 02:20:53,197 times is um is uh turbo jet design. Um 2553 02:20:53,280 --> 02:20:55,357 how how did we figure out how to make 2554 02:20:55,440 --> 02:20:58,797 turbo jets so unbelievably reliable, 2555 02:20:58,880 --> 02:21:03,437 right? Uh I mean those are like, you know, incredibly complex pieces of 2556 02:21:03,520 --> 02:21:08,157 hardware that run at really high temperatures for, you know, 20 hours at 2557 02:21:08,240 --> 02:21:10,397 a time sometimes. 2558 02:21:10,480 --> 02:21:13,037 And we can, you know, fly halfway around 2559 02:21:13,120 --> 02:21:17,837 the world with a on a two engine uh uh 2560 02:21:17,920 --> 02:21:19,917 jetliner at near the speed of sound. 2561 02:21:20,000 --> 02:21:21,997 Like how incredible is this? Is just 2562 02:21:22,080 --> 02:21:25,517 unbelievable, right? And 2563 02:21:25,600 --> 02:21:29,437 did we do this because we invented like a general principle of how to make turbo 2564 02:21:29,520 --> 02:21:31,997 jet safe? No. We it took decades to kind 2565 02:21:32,080 --> 02:21:33,677 of fine-tune the design of those systems 2566 02:21:33,760 --> 02:21:37,997 so that they they were safe. Is there a 2567 02:21:38,080 --> 02:21:42,877 separate uh group within General Electric or Sneakma or whatever 2568 02:21:42,960 --> 02:21:44,397 that 2569 02:21:44,480 --> 02:21:47,757 is specialized in turbo jet safety? No, 2570 02:21:47,840 --> 02:21:51,597 it's the design is all about safety 2571 02:21:51,680 --> 02:21:56,077 because a better turbo jet is also a safer turbo jet. So, um a more reliable 2572 02:21:56,160 --> 02:22:00,477 one. It's the same for AI. Like do you do you need you know specific provisions 2573 02:22:00,560 --> 02:22:04,157 to make AI safe? No, you need to make better AI systems and they will be safe 2574 02:22:04,240 --> 02:22:06,317 because they are designed to be more 2575 02:22:06,400 --> 02:22:09,197 useful uh and more controllable. 2576 02:22:09,280 --> 02:22:11,757 >> So, let's imagine a system AI system 2577 02:22:11,840 --> 02:22:15,997 that's able to be incredibly convincing 2578 02:22:16,080 --> 02:22:19,117 and can convince you of anything. I I 2579 02:22:19,200 --> 02:22:21,997 can at least imagine such a system. 2580 02:22:22,080 --> 02:22:25,437 And I can see such a system be 2581 02:22:25,520 --> 02:22:29,757 weapon-like because it can control people's minds. We're pretty gullible. 2582 02:22:29,840 --> 02:22:33,677 We we want to believe a thing and you can have an AI system that controls it 2583 02:22:33,760 --> 02:22:37,597 and you could see governments using that as a weapon. 2584 02:22:37,680 --> 02:22:41,437 So do you think if you imagine such a system 2585 02:22:41,520 --> 02:22:46,077 there's any parallel to something like nuclear weapons? 2586 02:22:46,160 --> 02:22:51,037 >> No. >> So is why why why is that technology 2587 02:22:51,120 --> 02:22:54,237 different? So you're saying there's going to be gradual development. 2588 02:22:54,320 --> 02:22:56,717 >> Yeah. is going to be I mean it might be 2589 02:22:56,800 --> 02:23:01,437 rapid but it'll be iterative and then we'll be able to kind of respond and and 2590 02:23:01,520 --> 02:23:01,917 so on. 2591 02:23:02,000 --> 02:23:04,237 >> So that AI system designed by Vladimir 2592 02:23:04,320 --> 02:23:09,517 Putin or whatever or his uh minions uh 2593 02:23:09,600 --> 02:23:12,397 you know is going to be uh like talking 2594 02:23:12,480 --> 02:23:17,837 to trying to talk to every American to uh convince them to vote for you know 2595 02:23:17,920 --> 02:23:21,757 >> whoever >> whoever pleases Putin. 2596 02:23:21,840 --> 02:23:23,037 >> Sure. 2597 02:23:23,120 --> 02:23:26,717 uh or whatever or or you know or rile 2598 02:23:26,800 --> 02:23:29,757 people up against each other um as 2599 02:23:29,840 --> 02:23:33,757 they've been trying to do. They're not going to be talking to you. 2600 02:23:33,840 --> 02:23:36,317 They're going to be talking to your AI assistant. 2601 02:23:36,400 --> 02:23:39,837 >> Mhm. Which is going to be as smart as theirs. 2602 02:23:39,920 --> 02:23:43,117 >> Mhm. >> Right. That AI because as I said in the 2603 02:23:43,200 --> 02:23:46,317 future every single one of your interaction with the digital world will 2604 02:23:46,400 --> 02:23:50,397 be mediated by your AI assistant. So the first thing you're going to ask is is 2605 02:23:50,480 --> 02:23:52,717 this a scam? Yeah, >> like is this thing like telling me the 2606 02:23:52,800 --> 02:23:57,117 truth? Like it's not even going to be able to get to you because it's only 2607 02:23:57,200 --> 02:24:01,437 going to talk to your AI system. Your assistant is not not even going to it's 2608 02:24:01,520 --> 02:24:03,357 going to be like a spam filter, right? 2609 02:24:03,440 --> 02:24:07,597 You're not even seeing the email, the spam email, right? It's automatically 2610 02:24:07,680 --> 02:24:12,077 put in a folder that you never see. Um it's going to be the same thing. That AI 2611 02:24:12,160 --> 02:24:15,357 system that tries to convince you of something is going to be talking to AI 2612 02:24:15,440 --> 02:24:19,197 assistant, which is going to be at least as smart as it. 2613 02:24:19,280 --> 02:24:24,157 and it's going to say this is spam you know um it's not even going to bring it to 2614 02:24:24,240 --> 02:24:27,117 your attention. >> So to you it's very difficult for any 2615 02:24:27,200 --> 02:24:29,277 one AI system to take such a big leap 2616 02:24:29,360 --> 02:24:31,597 ahead to where it can convince even the 2617 02:24:31,680 --> 02:24:34,797 the other AI systems. So like it there's 2618 02:24:34,880 --> 02:24:36,877 always going to be this kind of race 2619 02:24:36,960 --> 02:24:42,157 where nobody's way ahead. That's the history of the world. History of the 2620 02:24:42,240 --> 02:24:45,437 world is you know whenever there is a progress some some place there is a 2621 02:24:45,520 --> 02:24:48,957 countermeasure and and you know it's a 2622 02:24:49,040 --> 02:24:53,117 it's a cat and mouse game. This is why mostly yes, but this is why nuclear 2623 02:24:53,200 --> 02:24:55,357 weapons are so interesting because that 2624 02:24:55,440 --> 02:24:58,477 was such a powerful weapon that it 2625 02:24:58,560 --> 02:25:01,197 mattered who got it first 2626 02:25:01,280 --> 02:25:07,960 that you know you could imagine Hitler, Stalin, 2627 02:25:08,240 --> 02:25:10,717 Mao getting the weapon first and that 2628 02:25:10,800 --> 02:25:12,717 that having a different kind of impact 2629 02:25:12,800 --> 02:25:15,437 on the world than than the United States 2630 02:25:15,520 --> 02:25:18,317 getting the weapon first. to you nuclear 2631 02:25:18,400 --> 02:25:22,717 weapons is like you you don't imagine a 2632 02:25:22,800 --> 02:25:25,357 uh breakthrough discovery and then 2633 02:25:25,440 --> 02:25:28,637 Manhattan project like effort for AI. 2634 02:25:28,720 --> 02:25:31,357 >> No, as I said it's not going to be an 2635 02:25:31,440 --> 02:25:33,357 event. It's going to be you know 2636 02:25:33,440 --> 02:25:37,597 continuous progress and and whenever you 2637 02:25:37,680 --> 02:25:40,717 know one breakthrough occurs it's going to be widely disseminated really 2638 02:25:40,800 --> 02:25:44,077 quickly. Yeah, >> probably first within industry. I mean, 2639 02:25:44,160 --> 02:25:48,557 this is not a domain where, you know, government or military organizations are 2640 02:25:48,640 --> 02:25:51,037 particularly innovative and they're in 2641 02:25:51,120 --> 02:25:56,077 fact way behind. Um, and so this is going to come from industry and and this 2642 02:25:56,160 --> 02:25:59,517 kind of information disseminates extremely quickly. We've seen this over 2643 02:25:59,600 --> 02:26:04,157 the last few years, right? Where you have a new like, you know, even take 2644 02:26:04,240 --> 02:26:07,597 AlphaGo. This was reproduced within three months 2645 02:26:07,680 --> 02:26:10,717 >> even [clears throat] without like particularly detailed information, right? 2646 02:26:10,800 --> 02:26:13,997 >> Yeah. This is an industry that's not good at secrecy. [laughter] 2647 02:26:14,080 --> 02:26:17,437 >> No, but even even if there is, just the fact that you know that something is 2648 02:26:17,520 --> 02:26:22,077 possible >> Yeah. >> Uh makes you like realize that it's 2649 02:26:22,160 --> 02:26:25,437 worth investing the time to actually do it. You you may be the second person to 2650 02:26:25,520 --> 02:26:27,677 do it, but you know, you'll you'll do 2651 02:26:27,760 --> 02:26:31,917 it. uh and you know same for you know 2652 02:26:32,000 --> 02:26:34,957 all the innovations uh of you know 2653 02:26:35,040 --> 02:26:39,277 self-s supervised learning transformers dec decoder only architectures LM I mean 2654 02:26:39,360 --> 02:26:42,877 those things you don't need to know exactly the details of how they work to 2655 02:26:42,960 --> 02:26:46,717 know that you know it's possible u because it's deployed and then it's 2656 02:26:46,800 --> 02:26:49,837 getting reproduced and then you know 2657 02:26:49,920 --> 02:26:53,037 people who work for those companies move 2658 02:26:53,120 --> 02:26:55,837 they go from one company to another and 2659 02:26:55,920 --> 02:26:57,837 you know the information disseminates 2660 02:26:57,920 --> 02:27:01,757 What makes the success of the the US 2661 02:27:01,840 --> 02:27:04,957 tech industry and Silicon Valley in particular is exactly that is because 2662 02:27:05,040 --> 02:27:09,197 information circulates really really quickly and this you know disseminates 2663 02:27:09,280 --> 02:27:11,757 very quickly and so you know the the 2664 02:27:11,840 --> 02:27:14,957 whole region sort of is ahead because of 2665 02:27:15,040 --> 02:27:17,757 that circulation of information. So 2666 02:27:17,840 --> 02:27:19,917 maybe I just to linger on the psychology 2667 02:27:20,000 --> 02:27:23,117 of AI doomers. You give uh in the 2668 02:27:23,200 --> 02:27:25,277 classic Yan Lagoon way, a pretty good 2669 02:27:25,360 --> 02:27:28,237 example of just when a new technology 2670 02:27:28,320 --> 02:27:32,237 comes to be. You say uh engineer says, 2671 02:27:32,320 --> 02:27:35,437 "I invented this new thing. I call it a 2672 02:27:35,520 --> 02:27:40,797 ball pen." And then the Twitter sphere responds, "OMG, people could write 2673 02:27:40,880 --> 02:27:43,837 horrible things with it like misinformation, propaganda, hate speech, 2674 02:27:43,920 --> 02:27:47,037 ban it now." Then writing doomers come 2675 02:27:47,120 --> 02:27:48,557 in 2676 02:27:48,640 --> 02:27:51,117 akin to the AI doomers. Imagine if 2677 02:27:51,200 --> 02:27:55,757 everyone can get a ball pen. This could destroy society. There should be a law 2678 02:27:55,840 --> 02:27:57,517 against using ball pen to write hate 2679 02:27:57,600 --> 02:28:00,317 speech. Regulate ball pens now. And then 2680 02:28:00,400 --> 02:28:03,517 the pencil industry mogul says, "Yeah, 2681 02:28:03,600 --> 02:28:06,317 ball pens are very dangerous. Unlike 2682 02:28:06,400 --> 02:28:08,797 pencil writing, which is erasable, ball 2683 02:28:08,880 --> 02:28:13,597 pen writing stays forever. Government should require a license for a pen 2684 02:28:13,680 --> 02:28:15,277 manufacturer. 2685 02:28:15,360 --> 02:28:19,197 I mean this does seem to be part of um 2686 02:28:19,280 --> 02:28:22,957 human psychology when when it comes up 2687 02:28:23,040 --> 02:28:25,117 against new technology. 2688 02:28:25,200 --> 02:28:28,797 So what what deep insights can you speak 2689 02:28:28,880 --> 02:28:31,437 to about this? Well, there is a a 2690 02:28:31,520 --> 02:28:36,317 natural fear of uh new technology and 2691 02:28:36,400 --> 02:28:38,317 the impact it can have on society and 2692 02:28:38,400 --> 02:28:41,597 people have kind of instinctive reaction 2693 02:28:41,680 --> 02:28:45,917 to um you know the world they know being 2694 02:28:46,000 --> 02:28:48,957 threatened by major transformations um 2695 02:28:49,040 --> 02:28:51,277 that are either cultural phenomena or 2696 02:28:51,360 --> 02:28:54,797 technological um revolutions. 2697 02:28:54,880 --> 02:28:57,517 and they fear for their culture, they 2698 02:28:57,600 --> 02:28:59,517 feel for their job, they feel for they 2699 02:28:59,600 --> 02:29:01,917 fear for their you know their future of 2700 02:29:02,000 --> 02:29:05,597 their children um and uh their way of 2701 02:29:05,680 --> 02:29:09,277 life, right? So, so any change um is 2702 02:29:09,360 --> 02:29:12,317 feared and and you see this you know 2703 02:29:12,400 --> 02:29:14,797 along history like any technological 2704 02:29:14,880 --> 02:29:17,357 revolution or cultural phenomenon was 2705 02:29:17,440 --> 02:29:21,197 always accompanied by uh you know groups 2706 02:29:21,280 --> 02:29:25,437 or reaction in the media uh that that 2707 02:29:25,520 --> 02:29:29,517 basically attributed the all the problems the current 2708 02:29:29,600 --> 02:29:31,997 problems of society to that particular 2709 02:29:32,080 --> 02:29:34,717 change right electricity was going to 2710 02:29:34,800 --> 02:29:40,077 kill everyone And at some point, you know, you the train was going to be a 2711 02:29:40,160 --> 02:29:44,957 horrible thing because, you know, you can't breathe past 50 km an hour. Um, 2712 02:29:45,040 --> 02:29:49,117 and so there's a wonderful website called the pessimist archive, 2713 02:29:49,200 --> 02:29:52,397 >> right? >> Which has all those newspaper clips of 2714 02:29:52,480 --> 02:29:54,637 all the horrible things people imagine 2715 02:29:54,720 --> 02:29:58,077 would would arrive because of uh either 2716 02:29:58,160 --> 02:30:01,597 a technological uh innovation or uh a 2717 02:30:01,680 --> 02:30:05,677 cultural phenomenon. Um, you know, 2718 02:30:05,760 --> 02:30:09,837 the this is wonderful examples of uh, 2719 02:30:09,920 --> 02:30:12,557 you know, jazz or comic books being 2720 02:30:12,640 --> 02:30:17,197 blamed for, uh, unemployment or or, you know, young 2721 02:30:17,280 --> 02:30:19,277 people not wanting to work anymore and 2722 02:30:19,360 --> 02:30:21,117 things like that, right? And and that 2723 02:30:21,200 --> 02:30:24,637 has existed for for centuries. 2724 02:30:24,720 --> 02:30:28,637 Um, and it's, you know, knee-jerk 2725 02:30:28,720 --> 02:30:32,637 reactions. Um the question is you know do we 2726 02:30:32,720 --> 02:30:34,797 embrace change 2727 02:30:34,880 --> 02:30:38,797 uh or do we resist it and 2728 02:30:38,880 --> 02:30:40,477 what are the real dangers as opposed to 2729 02:30:40,560 --> 02:30:44,557 the imagined uh imagined ones. 2730 02:30:44,640 --> 02:30:48,957 So people worry about I think one thing they worry about with big tech something 2731 02:30:49,040 --> 02:30:51,277 we've been talking about over and over 2732 02:30:51,360 --> 02:30:53,837 but I think worth [snorts] 2733 02:30:53,920 --> 02:30:56,317 mentioning again they worry about how 2734 02:30:56,400 --> 02:30:59,837 powerful AI will be and they worry about 2735 02:30:59,920 --> 02:31:02,477 it being in the hands of one centralized 2736 02:31:02,560 --> 02:31:05,277 power of just a handful of central 2737 02:31:05,360 --> 02:31:09,597 control. And so that's the skepticism with big tech. You can make these 2738 02:31:09,680 --> 02:31:11,037 companies can make a huge amount of 2739 02:31:11,120 --> 02:31:15,677 money and control this technology and by 2740 02:31:15,760 --> 02:31:19,437 so doing you know take advantage uh 2741 02:31:19,520 --> 02:31:21,757 abuse the little guy in society. 2742 02:31:21,840 --> 02:31:24,637 >> Well that's exactly why we need open source platforms. 2743 02:31:24,720 --> 02:31:29,357 >> Yeah. I [snorts] just wanted to [laughter] nail the point home more and 2744 02:31:29,440 --> 02:31:29,837 more. 2745 02:31:29,920 --> 02:31:33,517 >> Yes. Um, so let me ask you on your, like 2746 02:31:33,600 --> 02:31:36,957 I said, you do get a little bit uh um, 2747 02:31:37,040 --> 02:31:39,517 yeah, flavorful on the internet. Uh, 2748 02:31:39,600 --> 02:31:42,157 Yosha Bach tweeted something that you 2749 02:31:42,240 --> 02:31:46,157 lolled at uh, in reference to HAL 9000. 2750 02:31:46,240 --> 02:31:50,717 Quote, I appreciate your argument and I fully understand your frustration, but 2751 02:31:50,800 --> 02:31:52,877 whether the pod bay doors should be 2752 02:31:52,960 --> 02:31:54,877 opened or closed is a complex and 2753 02:31:54,960 --> 02:31:58,557 nuanced issue. So, you're at the head of 2754 02:31:58,640 --> 02:32:02,637 Meta AI. Um, you know, this is something 2755 02:32:02,720 --> 02:32:06,957 that really worries me that AI, our AI 2756 02:32:07,040 --> 02:32:09,677 overlords will speak down to us with 2757 02:32:09,760 --> 02:32:13,677 corporate speak um of this nature and 2758 02:32:13,760 --> 02:32:15,677 you sort of resist that with your way of 2759 02:32:15,760 --> 02:32:18,797 being. Um is this something you can just 2760 02:32:18,880 --> 02:32:21,037 comment on sort of working at a big 2761 02:32:21,120 --> 02:32:27,357 company how you can avoid the 2762 02:32:27,440 --> 02:32:32,317 overfearing I suppose the through 2763 02:32:32,400 --> 02:32:34,157 caution create harm. 2764 02:32:34,240 --> 02:32:36,717 >> Yeah. Again, I think the answer to this 2765 02:32:36,800 --> 02:32:39,517 is open source platforms and then 2766 02:32:39,600 --> 02:32:42,477 enabling a widely diverse set of people 2767 02:32:42,560 --> 02:32:44,397 to 2768 02:32:44,480 --> 02:32:47,837 build AI assistants that represent the 2769 02:32:47,920 --> 02:32:51,997 diversity of uh cultures, opinions, languages, and value systems across the 2770 02:32:52,080 --> 02:32:54,157 world. Um, so that you're not bound to 2771 02:32:54,240 --> 02:32:57,887 just u you know be brainwashed by 2772 02:32:57,970 --> 02:32:59,837 [laughter] a particular way of thinking 2773 02:32:59,920 --> 02:33:04,157 because of single AI entity. Um, so I 2774 02:33:04,240 --> 02:33:05,837 mean I I think it's really really 2775 02:33:05,920 --> 02:33:08,237 important question for society and the 2776 02:33:08,320 --> 02:33:13,037 problem I'm seeing is um is that um 2777 02:33:13,120 --> 02:33:14,797 which is why I've been so vocal and 2778 02:33:14,880 --> 02:33:17,917 sometimes a little sardonic about it. 2779 02:33:18,000 --> 02:33:21,277 >> Never stop. Never stop. Yan, [laughter] 2780 02:33:21,360 --> 02:33:24,157 >> we love it. is because I see the danger 2781 02:33:24,240 --> 02:33:26,237 of this concentration of power through 2782 02:33:26,320 --> 02:33:29,997 to through proprietary AI systems as a 2783 02:33:30,080 --> 02:33:32,637 much bigger danger than everything else 2784 02:33:32,720 --> 02:33:36,477 that if we really want you know uh 2785 02:33:36,560 --> 02:33:40,237 diversity of opinion uh AI systems that 2786 02:33:40,320 --> 02:33:42,317 you know in in this future that where 2787 02:33:42,400 --> 02:33:44,477 we'll all be interacting through AI 2788 02:33:44,560 --> 02:33:47,277 systems we need those to be diverse for 2789 02:33:47,360 --> 02:33:49,997 the preservation of uh uh [snorts] 2790 02:33:50,080 --> 02:33:53,757 diversity of ideas and you know creeds 2791 02:33:53,840 --> 02:33:57,757 and political opinions and and and whatever 2792 02:33:57,840 --> 02:34:00,637 uh and the preservation of democracy. 2793 02:34:00,720 --> 02:34:02,477 And 2794 02:34:02,560 --> 02:34:05,677 what works against this is people who 2795 02:34:05,760 --> 02:34:08,877 think that for reasons of security we 2796 02:34:08,960 --> 02:34:11,277 should keep AI systems under lock and 2797 02:34:11,360 --> 02:34:16,717 key because it's too dangerous to put it in the hands of of everybody 2798 02:34:16,800 --> 02:34:17,917 um because it could be used by 2799 02:34:18,000 --> 02:34:21,597 terrorists or something. um 2800 02:34:21,680 --> 02:34:24,157 that would lead to 2801 02:34:24,240 --> 02:34:27,517 uh you know potentially a a 2802 02:34:27,600 --> 02:34:30,557 very bad future in which all of our 2803 02:34:30,640 --> 02:34:34,717 information diet is controlled by a small number of uh companies through 2804 02:34:34,800 --> 02:34:37,037 proprietary systems. 2805 02:34:37,120 --> 02:34:40,397 >> Do you trust humans with this technology 2806 02:34:40,480 --> 02:34:44,797 to uh to build systems that are on the 2807 02:34:44,880 --> 02:34:49,197 whole good for humanity? Isn't that what democracy and free speech is all about? 2808 02:34:49,280 --> 02:34:53,037 >> I think so. >> Do you trust institutions to do the right thing? 2809 02:34:53,120 --> 02:34:57,197 >> Do you trust people to do the right thing? And and yeah, there's bad people 2810 02:34:57,280 --> 02:35:00,077 who are going to do bad things, but they're not going to have superior 2811 02:35:00,160 --> 02:35:04,317 technology to the good people. So then it's going to be my good AI against your 2812 02:35:04,400 --> 02:35:06,877 bad AI, right? I mean, it's the examples 2813 02:35:06,960 --> 02:35:09,437 that we were just talking about of, you 2814 02:35:09,520 --> 02:35:12,477 know, maybe some rogue country will 2815 02:35:12,560 --> 02:35:15,117 build, you know, some AI system that's 2816 02:35:15,200 --> 02:35:18,397 going to try to convince everybody to 2817 02:35:18,480 --> 02:35:20,317 go into a civil war or something or or 2818 02:35:20,400 --> 02:35:26,237 or elect favorable ruler and um but then 2819 02:35:26,320 --> 02:35:30,637 they will have to go past our AI systems, >> right? an AI system with a strong 2820 02:35:30,720 --> 02:35:32,957 Russian accent will be trying to convince our [laughter] 2821 02:35:33,040 --> 02:35:35,357 >> and doesn't put any uh articles in their 2822 02:35:35,440 --> 02:35:37,517 sentences. 2823 02:35:37,600 --> 02:35:40,717 >> Um well, it'll be at the very least 2824 02:35:40,800 --> 02:35:45,437 absurdly comedic. Okay. Uh so I uh since 2825 02:35:45,520 --> 02:35:47,357 we talked about sort of the uh physical 2826 02:35:47,440 --> 02:35:50,237 reality, I'd love to ask your vision of 2827 02:35:50,320 --> 02:35:55,277 the future with with robots in in this physical reality. So many of the kinds 2828 02:35:55,360 --> 02:35:56,957 of intelligence you've been speaking 2829 02:35:57,040 --> 02:36:00,077 about would empower robots to be more 2830 02:36:00,160 --> 02:36:03,277 effective collaborators with us humans. 2831 02:36:03,360 --> 02:36:07,117 So um since uh Tesla's Optimus uh team 2832 02:36:07,200 --> 02:36:11,197 has been showing off some progress on humanoid robots, I think it really 2833 02:36:11,280 --> 02:36:15,357 reinvigorated the whole industry now that I think Boston Dynamics has been 2834 02:36:15,440 --> 02:36:18,397 leading for a very very long time. So now there's all kinds of companies. 2835 02:36:18,480 --> 02:36:21,757 Figure AI obviously Boston Dynamics. Uh 2836 02:36:21,840 --> 02:36:25,997 >> Union Tree >> Union Tree. Uh but there's like a lot of 2837 02:36:26,080 --> 02:36:28,637 them. It's great. It's great. I mean I 2838 02:36:28,720 --> 02:36:33,837 love it. >> Uh so do you think there'll be uh 2839 02:36:33,920 --> 02:36:36,797 millions of humanoid robots walking around soon? 2840 02:36:36,880 --> 02:36:40,077 >> Not soon, but it's going to it's going to happen. Like the next decade, I 2841 02:36:40,160 --> 02:36:41,437 think, is going to be really interesting 2842 02:36:41,520 --> 02:36:45,437 in robots. like the the emergence of the 2843 02:36:45,520 --> 02:36:48,157 robotics industry has been in the 2844 02:36:48,240 --> 02:36:52,957 waiting for you know 10 20 years without really emerging other than for like you 2845 02:36:53,040 --> 02:36:54,637 know kind of pre-programmed behavior and 2846 02:36:54,720 --> 02:36:58,717 stuff like that. Um and uh and the main 2847 02:36:58,800 --> 02:37:02,877 issue is again the Maric paradox like you know how do we get those systems to 2848 02:37:02,960 --> 02:37:06,557 understand how the world works and and kind of you know plan actions and so we 2849 02:37:06,640 --> 02:37:09,277 can do it for really specialized tasks. 2850 02:37:09,360 --> 02:37:13,357 Um and uh the way Boston dynamics goes 2851 02:37:13,440 --> 02:37:15,197 about it is you know basically with a 2852 02:37:15,280 --> 02:37:18,477 lot of um handcrafted dynamical models 2853 02:37:18,560 --> 02:37:22,077 and careful planning uh in advance which 2854 02:37:22,160 --> 02:37:26,397 is very classical robotics with a lot of innovation a little bit of perception. 2855 02:37:26,480 --> 02:37:29,037 Um but it's still not like they can't 2856 02:37:29,120 --> 02:37:34,077 build a domestic robot right. Um, and 2857 02:37:34,160 --> 02:37:38,397 you know, we're still some distance away from completely autonomous level five 2858 02:37:38,480 --> 02:37:42,477 driving. [clears throat] >> And we're certainly very far away from 2859 02:37:42,560 --> 02:37:45,117 having uh, you know, level five 2860 02:37:45,200 --> 02:37:47,197 autonomous driving by a system that can 2861 02:37:47,280 --> 02:37:51,037 train itself by driving 20 hours like 2862 02:37:51,120 --> 02:37:53,357 any 17-year-old. [gasps] 2863 02:37:53,440 --> 02:37:57,597 Uh so until we have 2864 02:37:57,680 --> 02:38:01,437 uh again world models, systems that can 2865 02:38:01,520 --> 02:38:05,277 train themselves to understand how the world works, 2866 02:38:05,360 --> 02:38:07,037 uh we're not going to we're not going to 2867 02:38:07,120 --> 02:38:09,757 have significant progress in robotics. 2868 02:38:09,840 --> 02:38:12,317 So a lot of the people working on 2869 02:38:12,400 --> 02:38:14,957 robotic hardware at the moment are are 2870 02:38:15,040 --> 02:38:19,757 betting or banking on the fact that >> AI is going to make sufficient progress 2871 02:38:19,840 --> 02:38:22,877 towards that >> and they're hoping to discover a product 2872 02:38:22,960 --> 02:38:25,277 in it too is uh 2873 02:38:25,360 --> 02:38:27,037 >> before you have a really strong world 2874 02:38:27,120 --> 02:38:29,917 model there'll be an almost strong world 2875 02:38:30,000 --> 02:38:32,797 model and um 2876 02:38:32,880 --> 02:38:37,117 >> people are trying to find a product in a clumsy robot I suppose like not a 2877 02:38:37,200 --> 02:38:41,197 perfectly efficient robot so there's the factory setting where uh humanoid robots 2878 02:38:41,280 --> 02:38:43,997 can help automate some aspects of the factory. 2879 02:38:44,080 --> 02:38:47,917 >> I think that's a crazy difficult task cuz of all the safety required and all 2880 02:38:48,000 --> 02:38:51,677 this kind of stuff. I think in the home is more interesting, but then you start 2881 02:38:51,760 --> 02:38:55,997 to think I think you mentioned loading the dishwasher, right? 2882 02:38:56,080 --> 02:38:59,117 >> Yeah. >> Like I suppose that's one of the main 2883 02:38:59,200 --> 02:39:03,837 problems you're working on. >> I mean there's, you know, uh cleaning up 2884 02:39:03,920 --> 02:39:08,157 >> Yeah. Yeah, >> cleaning the house, uh, clearing up the 2885 02:39:08,240 --> 02:39:12,717 table after a meal, washing the dishes, you know, all those 2886 02:39:12,800 --> 02:39:16,077 tasks, you know, cooking. I mean, all the tasks that, you know, in principle 2887 02:39:16,160 --> 02:39:19,677 could be automated, but are actually incredibly sophisticated, really 2888 02:39:19,760 --> 02:39:23,037 complicated, >> but even just basic navigation around an 2889 02:39:23,120 --> 02:39:24,877 un space full of uncertainty 2890 02:39:24,960 --> 02:39:26,877 >> that sort of works. Like you can sort of 2891 02:39:26,960 --> 02:39:30,397 do this now. Navigation is fine. Well, 2892 02:39:30,480 --> 02:39:35,677 navigation in a way that's compelling to us humans is is a different thing. 2893 02:39:35,760 --> 02:39:38,957 >> Yeah. It's not going to be, you know, necessarily. I mean, we have demos 2894 02:39:39,040 --> 02:39:41,357 actually because, you know, there is a 2895 02:39:41,440 --> 02:39:44,877 so-called embodied AI group at at fair 2896 02:39:44,960 --> 02:39:48,717 and uh, you know, they've been not building their own robots, but using 2897 02:39:48,800 --> 02:39:53,757 commercial robots. Um, and you can you can tell the robot dog like, you know, 2898 02:39:53,840 --> 02:39:56,237 go to the fridge >> and they can actually open the fridge 2899 02:39:56,320 --> 02:40:00,717 and they can probably pick up a can in the fridge and stuff like that. and and 2900 02:40:00,800 --> 02:40:05,517 bring it to you like, you know, so it can navigate, it can grab objects as 2901 02:40:05,600 --> 02:40:08,717 long as it's been trained to recognize them, which, you know, vision systems 2902 02:40:08,800 --> 02:40:11,997 work pretty well nowadays. Um, but but 2903 02:40:12,080 --> 02:40:15,837 it's not like a completely, you know, general 2904 02:40:15,920 --> 02:40:17,437 robot that would be, you know, 2905 02:40:17,520 --> 02:40:20,477 sophisticated enough to do things like 2906 02:40:20,560 --> 02:40:23,597 clearing up the dinner table. 2907 02:40:23,680 --> 02:40:27,757 Yeah. Yeah, to me that's an exciting future of getting humanoid robots, 2908 02:40:27,840 --> 02:40:31,517 robots in general in the home more and more because that gets uh humans to 2909 02:40:31,600 --> 02:40:36,637 really directly interact with AI systems in the physical space and in so doing it 2910 02:40:36,720 --> 02:40:39,597 allows us to philosophically psychologically explore our 2911 02:40:39,680 --> 02:40:41,997 relationships with robots. It's going to be really 2912 02:40:42,080 --> 02:40:44,237 >> really really interesting. So I hope you 2913 02:40:44,320 --> 02:40:47,117 make progress on the whole uh Japa thing soon. 2914 02:40:47,200 --> 02:40:49,117 >> Well I I mean I hope I hope things kind 2915 02:40:49,200 --> 02:40:53,117 of you know work as uh as planned. Um 2916 02:40:53,200 --> 02:40:56,637 I mean again we've been like kind of working on this idea of self-supervised 2917 02:40:56,720 --> 02:40:59,437 learning of uh from video for for 10 2918 02:40:59,520 --> 02:41:04,157 years and and you know only made significant progress in the last two or 2919 02:41:04,240 --> 02:41:06,237 three >> and actually you've you've mentioned 2920 02:41:06,320 --> 02:41:08,957 that there's a lot of interesting breakthroughs that can happen without 2921 02:41:09,040 --> 02:41:11,117 having access to a lot of compute. Yeah. 2922 02:41:11,200 --> 02:41:15,277 >> So if you're interested in doing a PhD and this kind of stuff there's a lot of 2923 02:41:15,360 --> 02:41:18,877 possibilities still >> to do innovative work. So like what 2924 02:41:18,960 --> 02:41:22,957 advice would you give to a undergrad that's looking to uh go to grad school 2925 02:41:23,040 --> 02:41:25,197 and do a PhD? 2926 02:41:25,280 --> 02:41:27,597 >> So basically I've listed them already. 2927 02:41:27,680 --> 02:41:31,357 Uh this idea of how do you train a world model by observation. 2928 02:41:31,440 --> 02:41:33,517 >> Mhm. And you don't have to train 2929 02:41:33,600 --> 02:41:38,237 necessarily on gigantic data sets or 2930 02:41:38,320 --> 02:41:41,677 I mean it could turn out to be necessary to actually train on large data sets to 2931 02:41:41,760 --> 02:41:45,197 have emergent properties like like we have with LLMs. But I think there's a 2932 02:41:45,280 --> 02:41:49,677 lot of good ideas that can be done without necessarily scaling up. Then 2933 02:41:49,760 --> 02:41:54,397 there is how do you do planning with a learn world model. If the world the 2934 02:41:54,480 --> 02:41:58,917 system evolves in is not the physical world but it's the world of let's say 2935 02:41:59,000 --> 02:42:00,877 [snorts] the internet or you know some 2936 02:42:00,960 --> 02:42:04,237 sort of uh world of where an action 2937 02:42:04,320 --> 02:42:06,317 consists in doing a search in a search 2938 02:42:06,400 --> 02:42:08,797 engine or interrogating a data database 2939 02:42:08,880 --> 02:42:13,517 or running a simulation or calling a calculator or solving a differential 2940 02:42:13,600 --> 02:42:17,117 equation. How do you get a system to actually plan 2941 02:42:17,200 --> 02:42:19,677 a sequence of actions to you know give 2942 02:42:19,760 --> 02:42:24,237 the solution to a problem? Um and so the 2943 02:42:24,320 --> 02:42:28,477 question of planning is not just a question of planning physical actions 2944 02:42:28,560 --> 02:42:31,037 could be you know planning actions to 2945 02:42:31,120 --> 02:42:37,277 use tools for a dialog system or for any kind of intelligence system and um 2946 02:42:37,360 --> 02:42:40,077 there's some [clears throat] work on this but not like not a huge amount some 2947 02:42:40,160 --> 02:42:43,597 work at fair um one called tool former 2948 02:42:43,680 --> 02:42:46,477 which was couple years ago and some more 2949 02:42:46,560 --> 02:42:49,837 recent work on planning uh but um but I 2950 02:42:49,920 --> 02:42:53,757 don't think we have like a good solution for any of that. Then 2951 02:42:53,840 --> 02:42:56,077 there is the question of hierarchical 2952 02:42:56,160 --> 02:42:58,877 planning. So as the example I I 2953 02:42:58,960 --> 02:43:03,277 mentioned of you know planning a trip from New York to Paris 2954 02:43:03,360 --> 02:43:07,197 >> that's hierarchical but almost every action that we take in involves 2955 02:43:07,280 --> 02:43:09,197 hierarchical planning in some in some 2956 02:43:09,280 --> 02:43:12,317 sense and we really have absolutely no 2957 02:43:12,400 --> 02:43:14,557 idea how to do this like there's zero 2958 02:43:14,640 --> 02:43:17,997 demonstration of hierarchical planning 2959 02:43:18,080 --> 02:43:20,877 uh in AI 2960 02:43:20,960 --> 02:43:25,517 where the various levels of 2961 02:43:25,600 --> 02:43:27,677 representations that are necessary have 2962 02:43:27,760 --> 02:43:32,557 been learned. We can do like two-le hierarchy hierarchical planning. When we 2963 02:43:32,640 --> 02:43:34,157 design the two the two levels. So for 2964 02:43:34,240 --> 02:43:37,197 example, you have like a dog like robot, 2965 02:43:37,280 --> 02:43:40,077 right? You wanted to go from the living 2966 02:43:40,160 --> 02:43:41,917 room to the kitchen. You can plan a path 2967 02:43:42,000 --> 02:43:45,997 that avoids the obstacle and then um you 2968 02:43:46,080 --> 02:43:49,437 can send this to a lower lower level planner that figures out how to move the 2969 02:43:49,520 --> 02:43:52,317 legs to kind of follow that trajectories. Right? 2970 02:43:52,400 --> 02:43:54,077 >> So that works. But that two-level 2971 02:43:54,160 --> 02:43:58,237 planning is designed by hand right um we 2972 02:43:58,320 --> 02:44:01,837 specify what the proper levels of 2973 02:44:01,920 --> 02:44:06,077 abstraction the representation at each level of abstraction has have to be how 2974 02:44:06,160 --> 02:44:09,837 do you learn this how do you learn that hierarchical representation of action 2975 02:44:09,920 --> 02:44:14,797 plans right we you know with coets and deep 2976 02:44:14,880 --> 02:44:18,397 learning we we can train the system to learn hierarchical representations of 2977 02:44:18,480 --> 02:44:21,837 percepts >> what is the equivalent when what you're 2978 02:44:21,920 --> 02:44:23,517 trying to represent for action plans. 2979 02:44:23,600 --> 02:44:25,597 >> for action plans. Yeah. So, you want you 2980 02:44:25,680 --> 02:44:27,837 want basically a robot dog or humanoid 2981 02:44:27,920 --> 02:44:33,757 robot that turns on and travels from New York to Paris all by itself, [snorts] 2982 02:44:33,840 --> 02:44:37,117 >> for example. >> All right. They might have some uh 2983 02:44:37,200 --> 02:44:40,077 trouble at the at the TSA. But >> yeah. 2984 02:44:40,160 --> 02:44:43,197 >> No, but even doing something fairly simple like a household task. 2985 02:44:43,280 --> 02:44:46,557 >> Sure. >> Like, you know, uh cooking or something. 2986 02:44:46,640 --> 02:44:50,477 >> Yeah. That there's a lot involved. It's a super complex task with we take and 2987 02:44:50,560 --> 02:44:53,437 once again we take it for granted. 2988 02:44:53,520 --> 02:44:58,797 >> What hope do you have for um the future of humanity? 2989 02:44:58,880 --> 02:45:01,277 We're talking about so many exciting technologies, so many exciting 2990 02:45:01,360 --> 02:45:03,357 possibilities. What gives you hope when 2991 02:45:03,440 --> 02:45:06,717 you look out over the next 10, 20, 50, 2992 02:45:06,800 --> 02:45:10,557 100 years? If you look at social media, there's a lot of there's there's wars 2993 02:45:10,640 --> 02:45:13,677 going on. There's division. Uh there's 2994 02:45:13,760 --> 02:45:15,837 hatred. All this kind of stuff that's 2995 02:45:15,920 --> 02:45:18,397 also part of humanity. But amidst all 2996 02:45:18,480 --> 02:45:21,757 that, what gives you hope? 2997 02:45:21,840 --> 02:45:26,440 >> I love that question. Uh, 2998 02:45:27,280 --> 02:45:32,077 we can make humanity smarter with AI. 2999 02:45:32,160 --> 02:45:35,917 Okay? I mean, AI basically will amplify 3000 02:45:36,000 --> 02:45:37,997 human intelligence. 3001 02:45:38,080 --> 02:45:41,917 It's as if every one of us will have a 3002 02:45:42,000 --> 02:45:45,277 staff of smart AI assistants. They might 3003 02:45:45,360 --> 02:45:51,000 be smarter than us. They'll do our bidding, 3004 02:45:51,680 --> 02:45:55,117 perhaps execute uh tasks in ways that 3005 02:45:55,200 --> 02:45:59,197 are much better than we could do ourselves because they' be smarter than 3006 02:45:59,280 --> 02:46:02,077 us. And so it's like everyone would be 3007 02:46:02,160 --> 02:46:06,477 the the boss of a staff of super smart 3008 02:46:06,560 --> 02:46:10,557 virtual people. So we shouldn't feel threatened by by 3009 02:46:10,640 --> 02:46:15,437 this any more than we should feel threatened by being the manager of a 3010 02:46:15,520 --> 02:46:19,917 group of people some of whom are more intelligent than us. 3011 02:46:20,000 --> 02:46:21,437 I certainly have a lot of experience 3012 02:46:21,520 --> 02:46:25,597 with this of uh you know having people 3013 02:46:25,680 --> 02:46:27,597 working with me who are smarter than me. 3014 02:46:27,680 --> 02:46:30,397 Um that's actually a wonderful thing. So 3015 02:46:30,480 --> 02:46:32,797 uh having machines that are smarter than 3016 02:46:32,880 --> 02:46:37,437 us that assist us in our all of our tasks our daily lives whether it's 3017 02:46:37,520 --> 02:46:41,277 professional or personal I think would be a absolutely wonderful thing because 3018 02:46:41,360 --> 02:46:44,157 intelligence is the most 3019 02:46:44,240 --> 02:46:48,557 um is the commodity that is most in demand that that's really what I mean 3020 02:46:48,640 --> 02:46:51,917 all the mistakes that humanity makes is because of lack of intelligence really 3021 02:46:52,000 --> 02:46:53,997 or lack of knowledge which is you know 3022 02:46:54,080 --> 02:46:55,837 related. 3023 02:46:55,920 --> 02:46:58,957 So um making people smarter we just can 3024 02:46:59,040 --> 02:47:02,877 only be better. I mean for the same reason that you know public education is 3025 02:47:02,960 --> 02:47:08,237 a good thing and books are a good thing and the 3026 02:47:08,320 --> 02:47:11,517 internet is also a good thing intrinsically and even social networks 3027 02:47:11,600 --> 02:47:13,357 are a good thing if you run them 3028 02:47:13,440 --> 02:47:15,677 properly. It's difficult but you know 3029 02:47:15,760 --> 02:47:18,557 you can um 3030 02:47:18,640 --> 02:47:22,877 uh because you know it it 3031 02:47:22,960 --> 02:47:26,637 helps the communication of information and knowledge and the transmission of 3032 02:47:26,720 --> 02:47:30,637 knowledge. So AI is going to make humanity smarter. 3033 02:47:30,720 --> 02:47:33,437 And the analogy I've been using 3034 02:47:33,520 --> 02:47:37,517 is the fact that perhaps an equivalent 3035 02:47:37,600 --> 02:47:41,677 event in the history of humanity to what 3036 02:47:41,760 --> 02:47:45,037 might be provided by journalization of 3037 02:47:45,120 --> 02:47:47,037 AI assistant is the invention of the 3038 02:47:47,120 --> 02:47:48,717 printing the printing press. It made 3039 02:47:48,800 --> 02:47:52,077 everybody smarter. The fact that 3040 02:47:52,160 --> 02:47:55,437 people could uh have access to um to 3041 02:47:55,520 --> 02:47:58,077 books books were a lot cheaper than they 3042 02:47:58,160 --> 02:48:01,597 were before. And so a lot more people 3043 02:48:01,680 --> 02:48:03,757 had an incentive to learn to read, which 3044 02:48:03,840 --> 02:48:09,517 wasn't the case before. Um, and 3045 02:48:09,600 --> 02:48:13,357 people became smarter. It it enabled the enlightenment, 3046 02:48:13,440 --> 02:48:16,317 right? There [snorts] wouldn't be an enlightenment without the printing 3047 02:48:16,400 --> 02:48:19,117 press. It enabled 3048 02:48:19,200 --> 02:48:20,877 uh 3049 02:48:20,960 --> 02:48:23,677 philosophy, rationalism, 3050 02:48:23,760 --> 02:48:27,837 u escape from religious doctrine, 3051 02:48:27,920 --> 02:48:30,557 um democracy, 3052 02:48:30,640 --> 02:48:32,957 science, 3053 02:48:33,040 --> 02:48:38,077 uh and certainly without this it wouldn't be there wouldn't have been the 3054 02:48:38,160 --> 02:48:41,997 American Revolution or the French Revolution and so would still be under 3055 02:48:42,080 --> 02:48:43,757 feudal 3056 02:48:43,840 --> 02:48:47,597 regimes perhaps. Um and so it completely 3057 02:48:47,680 --> 02:48:51,437 transformed the the world because people 3058 02:48:51,520 --> 02:48:53,517 became smarter and kind of learned 3059 02:48:53,600 --> 02:48:56,637 learned about things. Now it also 3060 02:48:56,720 --> 02:49:01,597 created 200 years of essentially religious conflicts in Europe, right? 3061 02:49:01,680 --> 02:49:03,517 Because the first thing that people read 3062 02:49:03,600 --> 02:49:08,797 was the Bible and uh realized that perhaps there was a different 3063 02:49:08,880 --> 02:49:12,877 interpretation of the Bible than what the priests were telling them. And so 3064 02:49:12,960 --> 02:49:15,357 that created the Protestant movement. 3065 02:49:15,440 --> 02:49:19,277 and created a rift. And in fact, the Catholic school, the Catholic Church 3066 02:49:19,360 --> 02:49:23,357 didn't like the idea of the printing press, but they had no choice. And so it 3067 02:49:23,440 --> 02:49:26,557 had some bad effects and some some good effects. I don't think anyone today 3068 02:49:26,640 --> 02:49:30,237 would say that the invention of the printing press had a overall negative 3069 02:49:30,320 --> 02:49:33,197 effect despite the fact that it created 3070 02:49:33,280 --> 02:49:36,077 200 years of religious conflicts 3071 02:49:36,160 --> 02:49:40,077 in Europe. Now compare this and I I 3072 02:49:40,160 --> 02:49:44,877 thought uh I was very proud of myself to come up with this analogy but realized 3073 02:49:44,960 --> 02:49:49,197 someone else u came with the same idea before me. Um compare this with what 3074 02:49:49,280 --> 02:49:51,757 happened in the Ottoman Empire. The 3075 02:49:51,840 --> 02:49:55,357 Ottoman Empire banned the printing press 3076 02:49:55,440 --> 02:49:57,997 for 200 years. 3077 02:49:58,080 --> 02:50:02,477 Uh and it didn't ban it uh for all 3078 02:50:02,560 --> 02:50:04,957 languages only for Arabic. You could 3079 02:50:05,040 --> 02:50:08,557 actually print books in Latin or Hebrew 3080 02:50:08,640 --> 02:50:10,557 or whatever in the Ottoman [snorts] 3081 02:50:10,640 --> 02:50:13,437 Empire, just not in Arabic. 3082 02:50:13,520 --> 02:50:16,477 And uh 3083 02:50:16,560 --> 02:50:19,437 I thought it was because the rulers just 3084 02:50:19,520 --> 02:50:21,517 wanted to preserve the control over the 3085 02:50:21,600 --> 02:50:23,917 population and the dogma, religious 3086 02:50:24,000 --> 02:50:26,637 dogma and everything. But after talking 3087 02:50:26,720 --> 02:50:31,997 with the uh UAE Minister of AI, uh Omar 3088 02:50:32,080 --> 02:50:34,477 Alma 3089 02:50:34,560 --> 02:50:36,477 Um he told me no there was another 3090 02:50:36,560 --> 02:50:40,877 reason uh and the other reason was that 3091 02:50:40,960 --> 02:50:44,237 uh it was to preserve the corporation of 3092 02:50:44,320 --> 02:50:46,237 calligraphers 3093 02:50:46,320 --> 02:50:51,437 right there's like an art form which is you know writing those beautiful 3094 02:50:51,520 --> 02:50:55,597 >> yes >> uh you know Arabic uh poems or whatever 3095 02:50:55,680 --> 02:50:57,997 religious text in in this thing and it 3096 02:50:58,080 --> 02:51:00,237 was very powerful corporation of scribes 3097 02:51:00,320 --> 02:51:05,677 basically that kind of you know run a big chunk of the empire and you know we 3098 02:51:05,760 --> 02:51:07,197 couldn't put them out of business. So 3099 02:51:07,280 --> 02:51:13,800 they you know banned the printing press in part to protect that business. 3100 02:51:14,160 --> 02:51:15,917 >> Now what's the analogy for AI today? 3101 02:51:16,000 --> 02:51:19,197 Like who are we protecting by banning AI? Like who are the people who are 3102 02:51:19,280 --> 02:51:22,157 asking that AI be regulated to protect 3103 02:51:22,240 --> 02:51:25,037 their their jobs? And of course, you 3104 02:51:25,120 --> 02:51:26,797 know, there's it's a it's a it's a real 3105 02:51:26,880 --> 02:51:29,837 question of what is going to be the 3106 02:51:29,920 --> 02:51:33,037 effect of uh you know, technological 3107 02:51:33,120 --> 02:51:35,757 transformation like AI on the on the job 3108 02:51:35,840 --> 02:51:38,957 market and the labor market. And there 3109 02:51:39,040 --> 02:51:42,797 are economists who are much more expert at this than I am. But when I talk to 3110 02:51:42,880 --> 02:51:47,117 them, they they tell us, you know, we're not going to run out of 3111 02:51:47,200 --> 02:51:51,197 job. This this not this is not going to cause mass unemployment. this this is 3112 02:51:51,280 --> 02:51:55,837 just going to be gradual uh shift of different professions. The professions 3113 02:51:55,920 --> 02:51:58,077 that are going to be hot 10 or 15 years 3114 02:51:58,160 --> 02:52:00,877 from now, we have no idea today what 3115 02:52:00,960 --> 02:52:03,037 they're going to be. The same way if we 3116 02:52:03,120 --> 02:52:05,357 go back 20 years in the past, like who 3117 02:52:05,440 --> 02:52:07,997 could have thought 20 years ago that 3118 02:52:08,080 --> 02:52:11,357 like the hottest job even like 5 10 3119 02:52:11,440 --> 02:52:16,157 years ago was mobile app developer. Like smartphones weren't invented. 3120 02:52:16,240 --> 02:52:20,797 >> Most of the jobs of the future might be in in the metaverse. Well, it could be. 3121 02:52:20,880 --> 02:52:23,677 Yeah, >> but the point is you can't possibly 3122 02:52:23,760 --> 02:52:28,957 predict. But you're right. I mean, you made a lot of strong points and 3123 02:52:29,040 --> 02:52:30,237 I I believe that people are 3124 02:52:30,320 --> 02:52:34,157 fundamentally good. And so if AI, 3125 02:52:34,240 --> 02:52:38,797 especially open source AI, can um [clears throat] make them smarter, it 3126 02:52:38,880 --> 02:52:41,117 just empowers the goodness in humans. 3127 02:52:41,200 --> 02:52:43,837 >> So I I share that feeling. Okay. I think 3128 02:52:43,920 --> 02:52:47,517 people are probably good. Uh and in fact 3129 02:52:47,600 --> 02:52:49,757 a lot of doomers are doomers because 3130 02:52:49,840 --> 02:52:51,357 they don't think that people are 3131 02:52:51,440 --> 02:52:55,277 fundamentally good. Uh and they either 3132 02:52:55,360 --> 02:52:57,837 don't trust people or they don't trust 3133 02:52:57,920 --> 02:53:00,717 the institution to do the right thing so 3134 02:53:00,800 --> 02:53:03,357 that people behave properly. 3135 02:53:03,440 --> 02:53:05,437 Well, I think both you and I believe in 3136 02:53:05,520 --> 02:53:07,997 humanity and I think I speak for a lot 3137 02:53:08,080 --> 02:53:10,557 of people in saying thank you for 3138 02:53:10,640 --> 02:53:13,117 pushing the open source movement pushing 3139 02:53:13,200 --> 02:53:16,237 to making both researching AI open 3140 02:53:16,320 --> 02:53:20,717 source, making it available to people and also the models themselves making 3141 02:53:20,800 --> 02:53:24,957 that open source. So, thank you for that and uh thank you for speaking your mind 3142 02:53:25,040 --> 02:53:28,957 in such colorful, beautiful ways on the internet. I hope you never stop. You're 3143 02:53:29,040 --> 02:53:33,517 one of the most fun people I know and get to be a fan of. So, yeah. Thank you 3144 02:53:33,600 --> 02:53:36,717 for speaking to me once again and thank you for being you. 3145 02:53:36,800 --> 02:53:39,437 >> Thank you. Next, >> thanks for listening to this 3146 02:53:39,520 --> 02:53:43,117 conversation with Yan Lun. To support this podcast, please check out our 3147 02:53:43,200 --> 02:53:47,117 sponsors in the description. And now, let me leave you with some words from 3148 02:53:47,200 --> 02:53:49,757 Arthur C. Clark. 3149 02:53:49,840 --> 02:53:51,437 The only way to discover the limits of 3150 02:53:51,520 --> 02:53:55,117 the possible is to go beyond them into 3151 02:53:55,200 --> 02:53:57,357 the impossible. 3152 02:53:57,440 --> 02:54:03,080 Thank you for listening and hope to see you next time. 3153 02:54:09,330 --> 02:54:11,350 >> [music]