WEBVTT 00:00:00.080 --> 00:00:02.397 The following is a conversation with Yan 00:00:02.480 --> 00:00:05.517 Lun, his third time on this podcast. He 00:00:05.600 --> 00:00:07.917 is the chief AI scientist at Meta, 00:00:08.000 --> 00:00:11.197 professor at NYU, touring award winner, 00:00:11.280 --> 00:00:13.517 and one of the seinal figures in the 00:00:13.600 --> 00:00:16.237 history of artificial intelligence. He 00:00:16.320 --> 00:00:19.757 and Meta AI have been big proponents of 00:00:19.840 --> 00:00:22.317 open sourcing AI development and have 00:00:22.400 --> 00:00:24.717 been walking the walk by open sourcing 00:00:24.800 --> 00:00:26.797 many of their biggest models including 00:00:26.880 --> 00:00:31.517 Llama 2 and eventually Llama 3. Also, 00:00:31.600 --> 00:00:34.477 Yan has been an outspoken critic of 00:00:34.560 --> 00:00:37.117 those people in the AI community who 00:00:37.200 --> 00:00:39.197 warned about the looming danger and 00:00:39.280 --> 00:00:42.797 existential threat of AGI. He believes 00:00:42.880 --> 00:00:46.797 the AGI will be created one day, but it 00:00:46.880 --> 00:00:49.037 will be good. It will not escape human 00:00:49.120 --> 00:00:52.717 control, nor will it dominate and kill 00:00:52.800 --> 00:00:55.837 all humans. At this moment of rapid AI 00:00:55.920 --> 00:00:59.197 development, this happens to be somewhat 00:00:59.280 --> 00:01:01.757 a controversial position. And so, it's 00:01:01.840 --> 00:01:06.557 been fun seeing Yan get into a lot of intense and fascinating discussions 00:01:06.640 --> 00:01:10.797 online, as we do in this very conversation. And 00:01:10.880 --> 00:01:14.397 now a quick few second mention of each sponsor. Check them out in the 00:01:14.480 --> 00:01:16.237 description. It's the best way to 00:01:16.320 --> 00:01:18.637 support this podcast. We got Hidden 00:01:18.720 --> 00:01:21.997 Layer for securing your AI models, 00:01:22.080 --> 00:01:25.197 Element for electrolytes, Shopify for 00:01:25.280 --> 00:01:29.437 shopping for stuff online, and AG1 for 00:01:29.520 --> 00:01:31.597 delicious health. Choose wisely, my 00:01:31.680 --> 00:01:34.317 friends. Also, if you want to get in 00:01:34.400 --> 00:01:37.357 touch with me for whatever reason, maybe 00:01:37.440 --> 00:01:40.317 to uh work with our amazing team, go to 00:01:40.400 --> 00:01:42.397 lexfreman.com/cont. 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 00:01:47.520 --> 00:01:49.277 these interesting. I don't know why I'm 00:01:49.360 --> 00:01:52.477 talking like this, but I am. There's a 00:01:52.560 --> 00:01:56.717 staccato nature to it. Speaking of staccato, I've been playing a bit of 00:01:56.800 --> 00:02:01.437 piano. Anyway, if you skip these ads, please still check out the sponsors. We 00:02:01.520 --> 00:02:03.517 love them. I love them. I enjoy their 00:02:03.600 --> 00:02:07.037 stuff. Maybe you will too. 00:02:07.120 --> 00:02:11.357 This episode is brought to you by a on 00:02:11.440 --> 00:02:15.037 theme in context. See what I did there? 00:02:15.120 --> 00:02:18.557 Uh sponsor since this is Yan Lun 00:02:18.640 --> 00:02:21.917 artificial intelligence machine learning 00:02:22.000 --> 00:02:24.477 one of the seminal figures in the field. 00:02:24.560 --> 00:02:27.517 So of course you got to have a sponsor that's related to artificial 00:02:27.600 --> 00:02:30.557 intelligence. hidden layer. They provide 00:02:30.640 --> 00:02:33.517 a platform that keeps your machine 00:02:33.600 --> 00:02:36.077 learning models secure. 00:02:36.160 --> 00:02:39.357 the ways to attack machine learning 00:02:39.440 --> 00:02:43.277 models, large language models, all the stuff we talk about with YAN. There's a 00:02:43.360 --> 00:02:45.197 lot of really fascinating work, not just 00:02:45.280 --> 00:02:48.237 large language models, but the same for 00:02:48.320 --> 00:02:50.877 video video prediction, tokenization 00:02:50.960 --> 00:02:53.277 where the tokens are in the space of 00:02:53.360 --> 00:02:56.317 concept versus the space of literally 00:02:56.400 --> 00:02:59.837 letters, symbols, JAPA, VJAPA, all of 00:02:59.920 --> 00:03:03.357 that stuff that they're open sourcing, all the stuff they're publishing on just 00:03:03.440 --> 00:03:06.157 really incredible. But that said, all of 00:03:06.240 --> 00:03:09.837 those models have security holes in ways 00:03:09.920 --> 00:03:14.637 that we can't even anticipate or imagine at this time. And so you want good 00:03:14.720 --> 00:03:16.637 people to be trying to find those 00:03:16.720 --> 00:03:18.797 security holes, trying to be one step 00:03:18.880 --> 00:03:23.837 ahead of uh the people that trying to attack. So if you're especially a 00:03:23.920 --> 00:03:25.437 company that's relying on these models, 00:03:25.520 --> 00:03:28.237 you need to uh have a person who's in 00:03:28.320 --> 00:03:31.117 charge of saying yeah this model that 00:03:31.200 --> 00:03:34.237 you got from this place has been tested 00:03:34.320 --> 00:03:36.557 has been secured whether that place is 00:03:36.640 --> 00:03:38.557 hugging face or any other kind of stuff 00:03:38.640 --> 00:03:41.357 uh or any other kind of repository or 00:03:41.440 --> 00:03:45.517 model zoo kind of place. I think the 00:03:45.600 --> 00:03:47.837 more and more we rely on larger language 00:03:47.920 --> 00:03:50.317 models or just AI systems in general, 00:03:50.400 --> 00:03:53.997 the more the security threats that are 00:03:54.080 --> 00:03:57.037 always going to be there uh become 00:03:57.120 --> 00:04:01.037 dangerous and impactful. So protect protect your models. Visit 00:04:01.120 --> 00:04:04.477 hiddenlayer.com/lex to learn more about how hidden layer can 00:04:04.560 --> 00:04:08.957 accelerate your AI adoption in a secure way. 00:04:09.040 --> 00:04:10.717 This episode is also brought to you by 00:04:10.800 --> 00:04:14.797 Element, a thing I drink throughout the 00:04:14.880 --> 00:04:17.517 day. I'm drinking now. When I am on a 00:04:17.600 --> 00:04:20.557 podcast, you'll sometimes see me with a 00:04:20.640 --> 00:04:24.317 mug and clear liquid in there that looks 00:04:24.400 --> 00:04:27.837 like water. In fact, it is not simply 00:04:27.920 --> 00:04:31.037 water. It is water mixed with element, 00:04:31.120 --> 00:04:34.957 watermelon salt, cold. What I do is I 00:04:35.040 --> 00:04:39.517 take one of them Power Raid 28 fluid 00:04:39.600 --> 00:04:43.117 ounces bottles, fill it up with water, 00:04:43.200 --> 00:04:45.757 one packet of watermelon salt, shake it 00:04:45.840 --> 00:04:48.877 up, put in the fridge, that's it. I 00:04:48.960 --> 00:04:51.277 reuse the bottles and drink from a mug 00:04:51.360 --> 00:04:56.637 or sometimes from the bottle. Either way, delicious, good for you. Especially 00:04:56.720 --> 00:04:58.157 if you're doing fasting, especially if 00:04:58.240 --> 00:05:00.797 you're doing low carb kinds of diets, 00:05:00.880 --> 00:05:03.677 which I do, you can get a sample pack 00:05:03.760 --> 00:05:09.400 for free with any purchase. Try to drinklement.com/lex. 00:05:10.160 --> 00:05:12.317 This episode is brought to you by 00:05:12.400 --> 00:05:14.957 Shopify as I take a drink of Element. It 00:05:15.040 --> 00:05:18.317 is a platform designed for anyone, even 00:05:18.400 --> 00:05:21.997 me, to sell stuff anywhere on a greatl 00:05:22.080 --> 00:05:24.797 looking store. I use a basic one. I like 00:05:24.880 --> 00:05:29.117 a really minimalist one. You can check it out if you go to lexreeman.com/store. 00:05:29.200 --> 00:05:31.357 There's a few shirts on there if that's 00:05:31.440 --> 00:05:35.277 your thing. It was so easy to set up. I 00:05:35.360 --> 00:05:37.037 imagine there's like a million features 00:05:37.120 --> 00:05:39.277 they have that can make it look better 00:05:39.360 --> 00:05:42.477 and all kinds of extra stuff you can do 00:05:42.560 --> 00:05:44.397 with the store. But I use the basic 00:05:44.480 --> 00:05:46.717 thing and the basic thing is pretty damn 00:05:46.800 --> 00:05:49.837 good. I like basic. I like minimalism. 00:05:49.920 --> 00:05:51.517 and they integrate with a lot of 00:05:51.600 --> 00:05:54.317 thirdparty apps, including what I use, 00:05:54.400 --> 00:05:58.557 which is um on demand printing. 00:05:58.640 --> 00:06:02.957 So, like you buy the shirt on Shopify, but it gets printed and shipped by 00:06:03.040 --> 00:06:06.477 another company that I always keep forgetting, but I think it's called 00:06:06.560 --> 00:06:10.877 Printull or Printify or I think it's Printull. 00:06:10.960 --> 00:06:15.277 I'm not sure. Doesn't matter. I think there's several integrations. You can 00:06:15.360 --> 00:06:17.677 check it out yourself. For me, it works. 00:06:17.760 --> 00:06:22.477 I'm using the most popular one, Printful, I think it's called. Anyway, I 00:06:22.560 --> 00:06:24.237 look forward to your letters correcting 00:06:24.320 --> 00:06:26.797 me on my pronunciation. 00:06:26.880 --> 00:06:30.477 Shopify is great. I'm a big fan of the 00:06:30.560 --> 00:06:34.477 good side of the machinery of uh capitalism, 00:06:34.560 --> 00:06:37.517 selling stuff on the internet, connecting people to the thing that they 00:06:37.600 --> 00:06:40.397 want, or rather the thing that would uh 00:06:40.480 --> 00:06:42.957 make their life better. 00:06:43.040 --> 00:06:46.077 both in advertisement and e-commerce and 00:06:46.160 --> 00:06:50.637 just shopping in general. I'm a big believer when that's done well, your 00:06:50.720 --> 00:06:52.717 life legitimately in the long term 00:06:52.800 --> 00:06:55.117 becomes better. And so whatever system 00:06:55.200 --> 00:06:58.637 can connect one human to the thing that 00:06:58.720 --> 00:07:00.877 makes their life better 00:07:00.960 --> 00:07:04.957 is great. And I believe that Shopify is 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 00:07:10.240 --> 00:07:13.037 per month trial period at shopify.com/lex. 00:07:13.120 --> 00:07:17.277 That's all lowercase. Go to shopify.com/lex to take your business to 00:07:17.360 --> 00:07:21.437 the next level today. This episode is also brought to you by 00:07:21.520 --> 00:07:23.437 AG1, 00:07:23.520 --> 00:07:26.877 an all-in-one daily drink to support 00:07:26.960 --> 00:07:29.517 better health and peak performance. It 00:07:29.600 --> 00:07:31.757 is delicious. 00:07:31.840 --> 00:07:34.237 It is nutritious. And I ran out of words 00:07:34.320 --> 00:07:38.077 that rhyme with those two. Actually, let 00:07:38.160 --> 00:07:41.837 me use a large language model to figure 00:07:41.920 --> 00:07:45.037 out what rhymes with delicious. 00:07:45.120 --> 00:07:49.277 Words that rhyme with delicious include ambitious, auspicious, capricious, 00:07:49.360 --> 00:07:52.637 fictitious, suspicious. 00:07:52.720 --> 00:07:55.677 So, there there you have it. Anyway, I 00:07:55.760 --> 00:08:00.237 drink it uh twice a day. Also put it in the fridge 00:08:00.320 --> 00:08:03.997 and sometimes in the freezer like it gets a little a little bit frozen. Just 00:08:04.080 --> 00:08:08.317 like a little bit, just a little bit frozen. You got that like slushy 00:08:08.400 --> 00:08:10.957 consistency. I'll do that too sometimes. 00:08:11.040 --> 00:08:14.957 It's freaking delicious. It's delicious no matter what. It's delicious. Warm is 00:08:15.040 --> 00:08:19.037 delicious cold. It's delicious slightly frozen. 00:08:19.120 --> 00:08:22.157 All of is just incredible. And of course, it covers like the basic 00:08:22.240 --> 00:08:24.397 multivitamin 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 00:08:29.920 --> 00:08:33.597 multivitamin. That's the way I think about it. So, all the crazy stuff I do, 00:08:33.680 --> 00:08:37.437 the physical challenges, the mental challenges, all of that. At least I got 00:08:37.520 --> 00:08:40.557 AG1. They'll give you one month supply 00:08:40.640 --> 00:08:45.160 of fish oil when you sign up at drinkagg1.com/llex. 00:08:46.000 --> 00:08:49.277 This is the Lex Riverment podcast. To support it, please check out our 00:08:49.360 --> 00:08:51.597 sponsors in the description. And now, 00:08:51.680 --> 00:08:56.760 dear friends, here's Yan Lun. 00:09:10.800 --> 00:09:14.077 You've had some strong statements, 00:09:14.160 --> 00:09:18.317 technical statements about the future of artificial intelligence recently, 00:09:18.400 --> 00:09:20.077 throughout your career actually, but 00:09:20.160 --> 00:09:22.877 recently as well. uh you've said that 00:09:22.960 --> 00:09:26.957 auto reagive LLMs are uh 00:09:27.040 --> 00:09:29.517 not the way we're going to make progress 00:09:29.600 --> 00:09:31.837 towards superhuman intelligence. These 00:09:31.920 --> 00:09:34.717 are the large language models like GPT4 00:09:34.800 --> 00:09:37.037 like Llama 2 and three soon and so on. 00:09:37.120 --> 00:09:40.557 How do they work and why are they not going to take us all the way? 00:09:40.640 --> 00:09:43.837 >> For a number of reasons. The first is that there is a number of 00:09:43.920 --> 00:09:47.277 characteristics of intelligent behavior. 00:09:47.360 --> 00:09:52.477 For example, the capacity to understand the world, understand the physical 00:09:52.560 --> 00:09:54.077 world, [sighs] 00:09:54.160 --> 00:09:57.757 the ability to remember and retrieve 00:09:57.840 --> 00:10:01.837 things. Um, persistent memory, the 00:10:01.920 --> 00:10:04.157 ability to reason, and the ability to 00:10:04.240 --> 00:10:06.237 plan. Those are four essential 00:10:06.320 --> 00:10:09.037 characteristic of intelligent 00:10:09.120 --> 00:10:13.277 systems or entities, humans, animals. 00:10:13.360 --> 00:10:16.557 L&Ms can do none of those or they can 00:10:16.640 --> 00:10:21.917 only do them in a very primitive way and uh they don't really understand the 00:10:22.000 --> 00:10:24.637 physical world. They don't really have persistent memory. They can't really 00:10:24.720 --> 00:10:27.197 reason and they certainly can't plan. 00:10:27.280 --> 00:10:30.397 And so you know if if if you expect the 00:10:30.480 --> 00:10:35.197 system to become intelligent just you know without having the possibility of 00:10:35.280 --> 00:10:38.637 doing those things uh you're making a mistake. [snorts] 00:10:38.720 --> 00:10:41.677 That is not to say that autographs are 00:10:41.760 --> 00:10:45.597 not useful. They're certainly useful. 00:10:45.680 --> 00:10:48.637 Um that they're not interesting. That we 00:10:48.720 --> 00:10:51.437 can't build a whole ecosystem of 00:10:51.520 --> 00:10:53.357 applications around them. Of course, we 00:10:53.440 --> 00:10:57.357 can. But as a path towards human level 00:10:57.440 --> 00:11:00.077 intelligence, they're missing essential 00:11:00.160 --> 00:11:02.797 components. And then there is another 00:11:02.880 --> 00:11:05.597 tidbit or or fact that I think is very 00:11:05.680 --> 00:11:08.157 interesting. Those LLMs are trained on 00:11:08.240 --> 00:11:10.877 enormous amounts of text. Basically, the 00:11:10.960 --> 00:11:13.357 entirety of all publicly available text 00:11:13.440 --> 00:11:16.237 on the internet, right? That's typically 00:11:16.320 --> 00:11:19.517 on the order of 10^ the 13 tokens. Each 00:11:19.600 --> 00:11:21.437 token is typically two bytes. So that's 00:11:21.520 --> 00:11:24.637 two 10^ the 13 bytes as training data. 00:11:24.720 --> 00:11:28.077 It would take you or me 170,000 years to 00:11:28.160 --> 00:11:30.877 just read through this at 8 hours a day. 00:11:30.960 --> 00:11:34.797 Uh so it seems like an enormous amount of knowledge, right, that those systems 00:11:34.880 --> 00:11:38.877 can accumulate. Um, 00:11:38.960 --> 00:11:43.517 but then you realize it's really not that much data. If you you talk to 00:11:43.600 --> 00:11:45.277 developmental psychologists and they 00:11:45.360 --> 00:11:47.597 tell you a 4-year-old has been awake for 00:11:47.680 --> 00:11:51.437 16,000 hours in his or her life. 00:11:51.520 --> 00:11:54.877 Um, and the amount of information that 00:11:54.960 --> 00:11:59.037 has uh reached the visual cortex of that 00:11:59.120 --> 00:12:01.517 child in four years 00:12:01.600 --> 00:12:05.117 um is about 10^ the 15 bytes. And you 00:12:05.200 --> 00:12:08.077 can compute this by estimating that the 00:12:08.160 --> 00:12:09.997 optical nerve carry about 20 megab 00:12:10.080 --> 00:12:13.437 megabytes per second roughly. And so 10^ 00:12:13.520 --> 00:12:16.477 the 15 bytes for a 4-year-old versus 2 * 00:12:16.560 --> 00:12:20.797 10^ 13 bytes for 170,000 years worth of 00:12:20.880 --> 00:12:25.277 reading. What that tells you is that uh 00:12:25.360 --> 00:12:27.677 through sensory input we see a lot more 00:12:27.760 --> 00:12:29.677 information than we than we do through 00:12:29.760 --> 00:12:32.317 language. And that despite our 00:12:32.400 --> 00:12:34.877 intuition, most of what we learn and 00:12:34.960 --> 00:12:38.397 most of our knowledge is through our 00:12:38.480 --> 00:12:42.237 observation and interaction with the real world, not through language. 00:12:42.320 --> 00:12:45.917 Everything that we learn in the first few years of life and uh certainly 00:12:46.000 --> 00:12:50.157 everything that animals learn has nothing to do with language. So it would 00:12:50.240 --> 00:12:54.477 be good to uh maybe push against some of the intuition behind what you're saying. 00:12:54.560 --> 00:12:58.237 So it is true there's several orders of 00:12:58.320 --> 00:13:02.477 magnitude more data coming into the human mind 00:13:02.560 --> 00:13:06.157 much faster and the human mind is able to learn very quickly from that filter 00:13:06.240 --> 00:13:10.477 the data very quickly. You know somebody might argue your comparison between 00:13:10.560 --> 00:13:12.797 sensory data versus language that 00:13:12.880 --> 00:13:16.157 language is already very compressed. It 00:13:16.240 --> 00:13:20.157 already contains a lot more information than the bytes it takes to store them if 00:13:20.240 --> 00:13:23.917 you compare it to visual data. So there's a lot of wisdom and language. 00:13:24.000 --> 00:13:25.997 There's words and the way we stitch them 00:13:26.080 --> 00:13:27.997 together, it already contains a lot of 00:13:28.080 --> 00:13:32.237 information. So, is it possible that 00:13:32.320 --> 00:13:37.677 language alone already has enough 00:13:37.760 --> 00:13:40.237 wisdom and knowledge in there to be able 00:13:40.320 --> 00:13:44.957 to from that language construct a a world model and understanding of the 00:13:45.040 --> 00:13:46.877 world, an understanding of the physical 00:13:46.960 --> 00:13:49.597 world that you're saying LLMs lack? So 00:13:49.680 --> 00:13:53.117 it's a big debate among philosophers and 00:13:53.200 --> 00:13:55.437 also cognitive scientists like whether 00:13:55.520 --> 00:13:57.197 intelligence needs to be grounded in 00:13:57.280 --> 00:14:00.637 reality. Uh I'm clearly in the camp that 00:14:00.720 --> 00:14:03.517 uh yes uh intelligence cannot appear 00:14:03.600 --> 00:14:08.817 without some grounding in uh some reality. It doesn't need to be you know 00:14:08.900 --> 00:14:12.877 [snorts] physical reality. It could be simulated but um but the environment is 00:14:12.960 --> 00:14:16.077 just much richer than what you can express in language. language is a very 00:14:16.160 --> 00:14:18.317 approximate representation 00:14:18.400 --> 00:14:21.917 of our percepts and our mental models, 00:14:22.000 --> 00:14:23.437 right? I mean, there's a lot of tasks 00:14:23.520 --> 00:14:26.637 that we accomplish where we manipulate 00:14:26.720 --> 00:14:29.517 uh a mental model of [snorts] uh of the 00:14:29.600 --> 00:14:32.077 situation at hand and that has nothing 00:14:32.160 --> 00:14:36.557 to do with language. Everything that's physical, mechanical, whatever, when we 00:14:36.640 --> 00:14:39.037 build something, when we accomplish a 00:14:39.120 --> 00:14:41.437 task, modern task of, you know, grabbing 00:14:41.520 --> 00:14:44.397 something, etc., We plan for action 00:14:44.480 --> 00:14:47.117 sequences and we do this by essentially 00:14:47.200 --> 00:14:51.277 imagining the result of the outcome of a 00:14:51.360 --> 00:14:52.557 sequence of actions that we might 00:14:52.640 --> 00:14:56.717 imagine and that requires mental models 00:14:56.800 --> 00:14:58.797 that don't have much to do with language 00:14:58.880 --> 00:15:01.997 and that's I would argue most of our 00:15:02.080 --> 00:15:06.717 knowledge is derived from that interaction with the physical world. So 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 00:15:10.800 --> 00:15:13.197 computer vision are really on that camp 00:15:13.280 --> 00:15:16.637 that uh AI needs to be embodied 00:15:16.720 --> 00:15:19.437 essentially and then other people coming 00:15:19.520 --> 00:15:23.117 from the NLP side or maybe you know some 00:15:23.200 --> 00:15:26.157 some other motivation don't necessarily 00:15:26.240 --> 00:15:28.637 agree with that. Um and philosophers are 00:15:28.720 --> 00:15:33.197 split as well. uh and the um the 00:15:33.280 --> 00:15:36.397 complexity of the world is hard to um 00:15:36.480 --> 00:15:38.877 it's hard to imagine 00:15:38.960 --> 00:15:42.077 you know it's hard to represent uh all 00:15:42.160 --> 00:15:46.557 the complexities that we take completely for granted in the real world that we 00:15:46.640 --> 00:15:50.717 don't even imagine require intelligence right this is the old Marave paradox 00:15:50.800 --> 00:15:55.277 from the pioneer of robotics hence Maravec who said you know how is it that 00:15:55.360 --> 00:15:59.037 with computers it seems to be easy to do high level complex tasks like playing 00:15:59.120 --> 00:16:01.757 chess and solving integrals and doing 00:16:01.840 --> 00:16:04.397 things like that. Whereas the thing we 00:16:04.480 --> 00:16:06.957 take for granted that we do every day um 00:16:07.040 --> 00:16:08.797 like I don't know learning to drive a 00:16:08.880 --> 00:16:11.517 car or you know grabbing an object we 00:16:11.600 --> 00:16:16.877 can't do as computers um and 00:16:16.960 --> 00:16:19.437 you know we have LLMs they can pass pass 00:16:19.520 --> 00:16:22.397 the bar exam so they must be smart but 00:16:22.480 --> 00:16:25.437 then they can't learn to drive in 20 00:16:25.520 --> 00:16:28.077 hours like any 17year-old 00:16:28.160 --> 00:16:30.077 they can't learn to clear up the dinner 00:16:30.160 --> 00:16:33.117 table and fill of the dishwasher like 00:16:33.200 --> 00:16:35.597 any 10-year-old can learn in one shot. 00:16:35.680 --> 00:16:39.757 Um why is that? Like you know what what are we missing? What what type of 00:16:39.840 --> 00:16:42.317 learning or or 00:16:42.400 --> 00:16:44.237 reasoning architecture or whatever are 00:16:44.320 --> 00:16:48.477 we missing that um um basically prevent 00:16:48.560 --> 00:16:50.797 us from from you know having level five 00:16:50.880 --> 00:16:54.077 serving cars and domestic robots? Can a 00:16:54.160 --> 00:16:57.277 large language model construct a world 00:16:57.360 --> 00:17:02.237 model that does know how to drive and does know how to fill a dishwasher but 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 00:17:07.680 --> 00:17:09.997 operate [clears throat] in a space of concepts. 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 00:17:13.920 --> 00:17:16.077 answer is no. And the more complex 00:17:16.160 --> 00:17:19.437 answer is you can use all kind of tricks 00:17:19.520 --> 00:17:21.997 to get 00:17:22.080 --> 00:17:27.197 an LLM to basically digest um visual 00:17:27.280 --> 00:17:29.837 representations of representations of 00:17:29.920 --> 00:17:32.957 images uh or video or audio for that 00:17:33.040 --> 00:17:36.477 matter. [snorts] Um and uh a classical 00:17:36.560 --> 00:17:39.837 way of doing this is uh you train a 00:17:39.920 --> 00:17:43.997 vision system in some way and we have a number of ways to train vision systems 00:17:44.080 --> 00:17:47.197 either supervised semi-supervised self-supervised all kinds of different 00:17:47.280 --> 00:17:50.717 ways [snorts] uh that will turn any 00:17:50.800 --> 00:17:54.877 image into high level representation 00:17:54.960 --> 00:17:58.477 basically a list of tokens that are really similar to the kind of tokens 00:17:58.560 --> 00:18:00.957 that uh 00:18:01.040 --> 00:18:03.757 typical LLM takes as an input and Then 00:18:03.840 --> 00:18:08.157 you just feed that to the LLM in 00:18:08.240 --> 00:18:11.117 addition to the text and you just expect 00:18:11.200 --> 00:18:13.917 LLM to kind of uh you know during 00:18:14.000 --> 00:18:19.277 training to kind of be able to uh use those representations to help make 00:18:19.360 --> 00:18:22.717 decisions. I mean there's been work along those lines for for quite a long 00:18:22.800 --> 00:18:27.117 time. Um and now you see those systems right? I mean there are LLMs that can 00:18:27.200 --> 00:18:30.957 that have some vision extension but they're basically hacks in the sense 00:18:31.040 --> 00:18:35.277 that um those things are not like trained end to end to to handle to 00:18:35.360 --> 00:18:39.277 really understand the world. They're not trained with video for example. Uh they 00:18:39.360 --> 00:18:43.837 don't really understand intuitive physics at least not at the moment. 00:18:43.920 --> 00:18:46.637 >> So [snorts] you don't think there's something special to you about intuitive 00:18:46.720 --> 00:18:50.397 physics about sort of common sense reasoning about the physical space about 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. 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 00:18:59.520 --> 00:19:02.957 working with today. And there's a number of reasons for this, but uh the main 00:19:03.040 --> 00:19:05.037 reason is 00:19:05.120 --> 00:19:07.517 the way LM LM are trained is that you 00:19:07.600 --> 00:19:11.037 you take a piece of text, you remove 00:19:11.120 --> 00:19:14.637 some of the words in that text, you mask them, you replace by replace them by 00:19:14.720 --> 00:19:18.397 blank markers, and you train a gigantic neural net to predict the words that are 00:19:18.480 --> 00:19:22.157 missing. uh and if you build this neural net in a 00:19:22.240 --> 00:19:24.157 particular way so that it can only look 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 00:19:29.360 --> 00:19:32.317 have is a system that basically is trying to predict the next word in a 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 00:19:37.520 --> 00:19:41.117 predict the next word it can never predict the next word exactly and so 00:19:41.200 --> 00:19:45.917 what it's going to do is uh produce a probability distribution of all the 00:19:46.000 --> 00:19:49.037 possible words in your dictionary in fact it doesn't predict words it 00:19:49.120 --> 00:19:53.997 predicts tokens that are kind of subword units. And so it's easy to handle the 00:19:54.080 --> 00:19:57.517 uncertainty in the prediction there because there's only a finite number of 00:19:57.600 --> 00:20:00.397 possible words in the dictionary and you 00:20:00.480 --> 00:20:01.837 can just compute a distribution over 00:20:01.920 --> 00:20:04.717 them. [snorts] Um then what you what the 00:20:04.800 --> 00:20:10.077 system does is that it it picks a word from that distribution. Of course 00:20:10.160 --> 00:20:13.437 there's a higher chance of picking words that have a higher probability within 00:20:13.520 --> 00:20:16.317 that distribution. And so you sample from that distribution to actually 00:20:16.400 --> 00:20:20.957 produce a word and then you shift that word into the input 00:20:21.040 --> 00:20:25.277 and so that allows the system not to predict the second word right and once 00:20:25.360 --> 00:20:27.357 you do this you shift it into the input 00:20:27.440 --> 00:20:29.597 etc that's called auto reggressive 00:20:29.680 --> 00:20:32.637 prediction and which is why those LLM 00:20:32.720 --> 00:20:35.997 should be called auto reggressive LLM uh 00:20:36.080 --> 00:20:40.637 >> but we just call them LM and 00:20:40.720 --> 00:20:42.237 there is a difference between this kind 00:20:42.320 --> 00:20:45.517 of process and a process by which before 00:20:45.600 --> 00:20:49.997 producing a word when you talk when you and I talk you and I are 00:20:50.080 --> 00:20:53.117 >> bilingual >> we think about what we're going to say 00:20:53.200 --> 00:20:57.037 and it's relatively independent of the language in which we're going to say it 00:20:57.120 --> 00:21:00.557 when we when we talk about like a I don't know let's say mathematical 00:21:00.640 --> 00:21:03.437 concept or something >> the kind of thinking that we're doing 00:21:03.520 --> 00:21:06.877 and the answer that we're planning to produce 00:21:06.960 --> 00:21:08.957 is not linked to whether we're going to 00:21:09.040 --> 00:21:12.157 see it in French or Russian or English 00:21:12.240 --> 00:21:14.397 >> Chsky just rolled his But I understand. 00:21:14.480 --> 00:21:17.357 So you're saying that there's a a bigger 00:21:17.440 --> 00:21:20.237 abstraction that rep that's uh that goes 00:21:20.320 --> 00:21:23.037 before language that maps onto language. 00:21:23.120 --> 00:21:26.637 >> Right. It's certainly true for a lot of thinking that we that we do. 00:21:26.720 --> 00:21:31.517 >> Is that obvious that we don't like you're saying your thinking is same in 00:21:31.600 --> 00:21:34.717 French as it is in English? >> Yeah, pretty much. Yeah. 00:21:34.800 --> 00:21:37.597 >> Pretty much. Or is this like how how 00:21:37.680 --> 00:21:41.757 flexible are you? Like if if there's a probability distribution? 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 00:21:46.000 --> 00:21:49.597 like producing puns, I get much better in French than English about that. 00:21:49.680 --> 00:21:53.757 >> No, but is there an abstract representation of puns? Like is your 00:21:53.840 --> 00:21:57.597 humor an abstract like when you tweet and your tweets are sometimes a little 00:21:57.680 --> 00:22:01.997 bit spicy? Uh what's is there an abstract representation in your brain of 00:22:02.080 --> 00:22:04.477 a tweet before it maps onto English? 00:22:04.560 --> 00:22:07.037 There is an abstract representation of 00:22:07.120 --> 00:22:11.597 imagining the reaction of a reader to that uh text. 00:22:11.680 --> 00:22:14.797 >> You start with laughter and then figure out how to make that happen 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 00:22:19.760 --> 00:22:22.957 say it, right? So that it causes that reaction. But that's like really close 00:22:23.040 --> 00:22:27.037 to language. But think about like a mathemat mathematical concept 00:22:27.120 --> 00:22:31.277 >> or you know imagining you know something you want to build out of wood or 00:22:31.360 --> 00:22:34.477 something like this, right? the kind of thinking you're doing is absolutely 00:22:34.560 --> 00:22:37.837 nothing to do with language really. Like it's not like you have necessarily like 00:22:37.920 --> 00:22:41.677 an internal monologue in any particular language. You're you're you know 00:22:41.760 --> 00:22:44.557 imagining mental models of of the thing, 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 00:22:48.960 --> 00:22:53.517 look like if I rotate it >> 90 degrees um that has nothing to do 00:22:53.600 --> 00:22:55.277 with language. 00:22:55.360 --> 00:22:58.637 And so uh so clearly there is you know a 00:22:58.720 --> 00:23:01.517 more abstract level of representation 00:23:01.600 --> 00:23:03.837 uh in which we we do most of our 00:23:03.920 --> 00:23:06.477 thinking and we plan what we're going to 00:23:06.560 --> 00:23:10.077 say if the output is is 00:23:10.160 --> 00:23:13.757 you know uttered words as opposed to an 00:23:13.840 --> 00:23:17.357 output being uh you know muscle actions 00:23:17.440 --> 00:23:21.117 >> right um we we plan our answer before we 00:23:21.200 --> 00:23:23.437 produce it and L&Ms don't do that they 00:23:23.520 --> 00:23:26.557 just produce one word after the other 00:23:26.640 --> 00:23:28.397 instinctively if you want. It's like 00:23:28.480 --> 00:23:31.037 it's a bit like the you know 00:23:31.120 --> 00:23:35.997 subconscious uh actions where you don't like you're distracted you're doing 00:23:36.080 --> 00:23:40.157 something you're completely concentrated and someone comes to you and you know 00:23:40.240 --> 00:23:43.197 ask you a question and you kind of answer the question you don't have time 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 00:23:46.080 --> 00:23:50.157 attention you sort of respond automatically that's kind of what an NLM 00:23:50.240 --> 00:23:52.237 does right it doesn't think about its 00:23:52.320 --> 00:23:57.357 answer really uh it retrieves it because it's accumulated a lot of knowledge So 00:23:57.440 --> 00:23:59.037 it can retrieve some some things, but 00:23:59.120 --> 00:24:02.877 it's going to just spit out one token 00:24:02.960 --> 00:24:07.437 after the other without planning the answer. But 00:24:07.520 --> 00:24:10.957 you're making it sound just one token after the other, one token at a time, 00:24:11.040 --> 00:24:12.637 generation 00:24:12.720 --> 00:24:18.077 is uh bound to be simplistic. 00:24:18.160 --> 00:24:20.157 But if the world model is sufficiently 00:24:20.240 --> 00:24:24.077 sophisticated, that one token at a time, 00:24:24.160 --> 00:24:27.037 the the most likely thing it generates 00:24:27.120 --> 00:24:29.677 as a sequence of tokens is going to be a 00:24:29.760 --> 00:24:31.917 deeply profound thing. 00:24:32.000 --> 00:24:37.037 >> Okay. But then that assumes that those systems actually possess an eternal 00:24:37.120 --> 00:24:39.917 world model. >> So really it goes to the I I think the 00:24:40.000 --> 00:24:42.717 fundamental question is can you build a 00:24:42.800 --> 00:24:45.757 a really 00:24:45.840 --> 00:24:50.637 complete world model? uh complete but a uh one that has a deep understanding of 00:24:50.720 --> 00:24:56.317 the world. >> Yeah. So can you build this first of all by prediction, 00:24:56.400 --> 00:24:59.357 >> right? >> And the answer is probably yes. Can you 00:24:59.440 --> 00:25:02.477 predict can you build it by predicting 00:25:02.560 --> 00:25:05.757 words? And the answer is most probably 00:25:05.840 --> 00:25:09.837 no because language is very poor in 00:25:09.920 --> 00:25:12.957 terms or weak or low bandwidth if you want. There's just not enough 00:25:13.040 --> 00:25:15.997 information there. So building world 00:25:16.080 --> 00:25:21.197 models means observing the world and uh 00:25:21.280 --> 00:25:22.717 understanding 00:25:22.800 --> 00:25:24.797 why the world is evolving the way the 00:25:24.880 --> 00:25:30.157 way it is. And then uh the the extra 00:25:30.240 --> 00:25:34.877 component of a world model is something that can predict how the world is going 00:25:34.960 --> 00:25:39.037 to evolve as a consequence of an action you might take. Right? So world model 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 00:25:42.400 --> 00:25:44.797 action I might take. What is the 00:25:44.880 --> 00:25:47.197 predicted state of the world at time t 00:25:47.280 --> 00:25:52.077 plus one? Now that state of the world does not does does not need to represent 00:25:52.160 --> 00:25:55.757 everything about the world. It just needs to represent enough that's 00:25:55.840 --> 00:25:57.997 relevant for this planning of of the 00:25:58.080 --> 00:26:03.037 action but not necessarily all the details. Now, here is the problem. Um 00:26:03.120 --> 00:26:08.237 you're not going to be able to do this with generative models. So a generative 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 00:26:11.840 --> 00:26:15.117 a video, show a system a piece of video, 00:26:15.200 --> 00:26:19.997 and then ask it to predict the reminder of the video. Basically, predict what's 00:26:20.080 --> 00:26:22.797 going to happen >> one frame at a time. Do the same thing 00:26:22.880 --> 00:26:27.037 as sort of uh the autogressive LLMs do, but for video, 00:26:27.120 --> 00:26:30.397 >> right? Either one frame at a time or a group of friends at a time. 00:26:30.480 --> 00:26:33.517 >> Um, but yeah, uh, a large video model if 00:26:33.600 --> 00:26:38.157 you want. uh the idea of of doing this 00:26:38.240 --> 00:26:39.677 has been floating around for a long time 00:26:39.760 --> 00:26:43.837 and at at fair u some colleagues and I 00:26:43.920 --> 00:26:46.877 have been trying to do this for about 10 years 00:26:46.960 --> 00:26:49.997 um and you can't you can't really do the 00:26:50.080 --> 00:26:52.557 same trick as with LLMs because uh you 00:26:52.640 --> 00:26:55.917 know LLM as I said you can't predict 00:26:56.000 --> 00:26:58.637 exactly which word is going to follow a 00:26:58.720 --> 00:27:03.037 sequence of words but you can predict the distribution of words now if you go 00:27:03.120 --> 00:27:06.717 to video what you would have to do is predict the distribution over all 00:27:06.800 --> 00:27:09.597 possible frames in a video and we don't 00:27:09.680 --> 00:27:12.077 really know how to do that properly. 00:27:12.160 --> 00:27:16.157 Uh we we we do not know how to represent distributions over highdimensional 00:27:16.240 --> 00:27:19.597 continuous spaces in ways that are useful. 00:27:19.680 --> 00:27:22.797 >> Uh and and that's that there lies the 00:27:22.880 --> 00:27:25.757 main issue and the reason we can do this 00:27:25.840 --> 00:27:29.437 is because the world is incredibly more 00:27:29.520 --> 00:27:33.757 complicated and richer in terms of information than than text. Text is 00:27:33.840 --> 00:27:37.037 discrete. Video is high dimensional and 00:27:37.120 --> 00:27:39.997 continuous. A lot of details in this. Um 00:27:40.080 --> 00:27:43.837 so if I take a a video of this room uh 00:27:43.920 --> 00:27:46.397 and the video is you know a camera 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 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 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 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 00:28:02.480 --> 00:28:05.597 predict what the painting on the wall looks like or what the texture of the 00:28:05.680 --> 00:28:07.837 couch looks like. Certainly not the 00:28:07.920 --> 00:28:10.237 texture of the carpet. So there's no way 00:28:10.320 --> 00:28:12.797 it can predict all those details. So the 00:28:12.880 --> 00:28:16.637 the way to handle this is one way 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 00:28:20.240 --> 00:28:24.717 model that has what's called a latent variable. And the latent variable is fed 00:28:24.800 --> 00:28:28.237 to a neural net and it's supposed to represent all the information about the 00:28:28.320 --> 00:28:31.517 world that you don't perceive yet. and 00:28:31.600 --> 00:28:35.037 uh that you need to augment 00:28:35.120 --> 00:28:38.317 uh the the system for the prediction to 00:28:38.400 --> 00:28:39.997 do a good job at predicting pixels 00:28:40.080 --> 00:28:42.717 including the you know fine texture of 00:28:42.800 --> 00:28:46.157 the of the carpet and the and the couch 00:28:46.240 --> 00:28:49.677 and and the painting on the wall. um 00:28:49.760 --> 00:28:53.757 uh that has been a complete failure essentially and we've tried lots of 00:28:53.840 --> 00:28:56.317 things we tried u just straight neural 00:28:56.400 --> 00:29:00.317 nets we tried GANs we tried uh you know 00:29:00.400 --> 00:29:02.557 VAES all kinds of regularized 00:29:02.640 --> 00:29:06.877 autoenccoders we tried um many things we 00:29:06.960 --> 00:29:08.797 also tried those kind of methods to 00:29:08.880 --> 00:29:12.317 learn uh good representations of images 00:29:12.400 --> 00:29:16.317 or video um that could then be used as 00:29:16.400 --> 00:29:19.517 input to for example an image classification system. 00:29:19.600 --> 00:29:22.557 >> Mhm. >> And that also has basically failed like 00:29:22.640 --> 00:29:25.357 all the systems that attempt to predict 00:29:25.440 --> 00:29:28.797 missing parts of an image or video 00:29:28.880 --> 00:29:33.757 um you know from a corrupted version of it basically. Okay. So right take an 00:29:33.840 --> 00:29:37.517 image or a video corrupt it or transform it in some way and then try to 00:29:37.600 --> 00:29:40.237 reconstruct the complete video or image 00:29:40.320 --> 00:29:43.357 from the corrupted version 00:29:43.440 --> 00:29:47.277 and then hope that internally the system will develop a good representations of 00:29:47.360 --> 00:29:52.017 images that you can use for object recognition segmentation whatever it is 00:29:52.100 --> 00:29:55.437 [snorts] that has been essentially a complete failure 00:29:55.520 --> 00:29:59.277 and it works really well for text that's the principle that is used for LLMs 00:29:59.360 --> 00:30:02.637 right >> so where is the failure exactly is it 00:30:02.720 --> 00:30:05.517 It's very difficult to form a good 00:30:05.600 --> 00:30:10.797 representation of an image a good in like a good embedding of all all the 00:30:10.880 --> 00:30:14.237 important information in the image. Is it in terms of the consistency of image 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 00:30:19.120 --> 00:30:20.717 do a highlight reel of all the ways you 00:30:20.800 --> 00:30:23.437 failed. What's that look like? 00:30:23.520 --> 00:30:26.717 >> Okay, so the reason this doesn't work uh 00:30:26.800 --> 00:30:30.637 is first of all I have to tell you exactly what doesn't work because there 00:30:30.720 --> 00:30:32.717 is something else that does work. Mhm. 00:30:32.800 --> 00:30:35.197 >> So the thing that does not work is 00:30:35.280 --> 00:30:38.557 training a system to learn 00:30:38.640 --> 00:30:41.277 representations of images by training it 00:30:41.360 --> 00:30:43.917 to reconstruct 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. 00:30:48.400 --> 00:30:50.797 And we have a whole slew of techniques 00:30:50.880 --> 00:30:55.517 for this uh that are you know variant of den noising autoenccoders. Something 00:30:55.600 --> 00:31:00.077 called MAE developed by some of my colleagues at fair maxenccoder. So it's 00:31:00.160 --> 00:31:04.877 basically like the you know LLMs or or or things like this 00:31:04.960 --> 00:31:08.317 where you train a system by corrupting text except you corrupt images you 00:31:08.400 --> 00:31:11.927 remove patches from it and you train a gigantic neural net to reconstruct 00:31:12.010 --> 00:31:15.117 [snorts] the features you get are not good and you know they're not good 00:31:15.200 --> 00:31:17.597 because if you now train the same 00:31:17.680 --> 00:31:20.557 architecture but you train it supervised 00:31:20.640 --> 00:31:24.237 >> with uh label data with text textual 00:31:24.320 --> 00:31:27.197 descriptions of images etc. you do get 00:31:27.280 --> 00:31:29.437 good representations and the performance 00:31:29.520 --> 00:31:34.237 on recognition tasks is much better than if you do this self-supervised 00:31:34.320 --> 00:31:37.197 pre-training. >> So the architecture is good. 00:31:37.280 --> 00:31:40.397 >> The architecture is good. The architecture of the encoder is good. 00:31:40.480 --> 00:31:41.997 >> Okay. But the fact that you train the 00:31:42.080 --> 00:31:47.837 system to reconstruct images does not lead it to produce to learn good generic 00:31:47.920 --> 00:31:51.117 features of images >> when you train in a self-supervised way. 00:31:51.200 --> 00:31:53.117 >> Self-supervised by reconstruction. 00:31:53.200 --> 00:31:54.157 >> Yeah. By reconstruction. 00:31:54.240 --> 00:31:57.037 >> Okay. So what's the alternative? 00:31:57.120 --> 00:32:00.237 The alternative is joint embedding. 00:32:00.320 --> 00:32:03.437 >> What is joint embedding? What are what are these architectures that you're so 00:32:03.520 --> 00:32:05.917 excited about? >> Okay. So now instead of training a 00:32:06.000 --> 00:32:07.677 system to encode the image and then 00:32:07.760 --> 00:32:10.557 training it to reconstruct the the full 00:32:10.640 --> 00:32:13.117 image from a corrupted version, you take 00:32:13.200 --> 00:32:16.397 the full image, you take the 00:32:16.480 --> 00:32:21.517 corrupted or transformed version, you run them both through encoders, which in 00:32:21.600 --> 00:32:22.877 general are identical, but not 00:32:22.960 --> 00:32:25.517 necessarily. [snorts] And then you you 00:32:25.600 --> 00:32:29.037 train a predictor on top of those 00:32:29.120 --> 00:32:30.557 encoders 00:32:30.640 --> 00:32:33.837 um to predict the representation of the 00:32:33.920 --> 00:32:37.197 full input from the representation of 00:32:37.280 --> 00:32:42.077 the corrupted one. >> Okay. So joint embedding because you're 00:32:42.160 --> 00:32:46.877 you're taking the the full input and the corrupted version or transform version 00:32:46.960 --> 00:32:51.037 run them both through encoders. you get a joint embedding and then you and then 00:32:51.120 --> 00:32:55.197 you're you're saying can I predict the representation of the full one from the 00:32:55.280 --> 00:32:57.517 representation of the corrupted one. 00:32:57.600 --> 00:33:02.157 Okay. Um and I call this a jet. So that means joint embedding predictive 00:33:02.240 --> 00:33:04.877 architecture because this joint embedding and there is this predictor 00:33:04.960 --> 00:33:09.197 that predicts the representation of the good guy from from the bad guy. [snorts] 00:33:09.280 --> 00:33:11.517 Um and the big question is how do you 00:33:11.600 --> 00:33:14.877 train something like this? Uh and until 00:33:14.960 --> 00:33:18.957 five years ago or six years ago, we didn't have particularly good answers 00:33:19.040 --> 00:33:20.797 for how you train those things except 00:33:20.880 --> 00:33:23.677 for one um called contrastive 00:33:23.760 --> 00:33:27.837 contrastive learning where 00:33:27.920 --> 00:33:29.357 um and the idea of contractive learning 00:33:29.440 --> 00:33:31.757 is you you take a pair of images that 00:33:31.840 --> 00:33:34.877 are again an image and a corrupted 00:33:34.960 --> 00:33:37.197 version or degraded version somehow or 00:33:37.280 --> 00:33:39.837 transformed version of the original one 00:33:39.920 --> 00:33:44.397 and you train the predicted representation to the same as as as 00:33:44.480 --> 00:33:47.677 that. If you only do this, the system collapses. It basically completely 00:33:47.760 --> 00:33:49.357 ignores the input and produces 00:33:49.440 --> 00:33:52.717 representations that are constant. 00:33:52.800 --> 00:33:56.957 So the contrastive methods avoid this and and those things have been around 00:33:57.040 --> 00:33:58.797 since the early 90s. I had a paper on 00:33:58.880 --> 00:34:03.517 this in 1993. Um is you also show pairs 00:34:03.600 --> 00:34:07.037 of images that you know are different 00:34:07.120 --> 00:34:10.717 and then you push away the representations from each other. So you 00:34:10.800 --> 00:34:14.957 say not only do representations of things that we know are the same should 00:34:15.040 --> 00:34:18.157 be the same or should be similar but representation of things that we know 00:34:18.240 --> 00:34:22.157 are different should be different and that prevents the collapse but it has 00:34:22.240 --> 00:34:24.957 some limitation and there's a whole bunch of uh techniques that have 00:34:25.040 --> 00:34:28.237 appeared over the last six seven years 00:34:28.320 --> 00:34:30.317 um that can revive this this type of 00:34:30.400 --> 00:34:36.287 method um some of them from fair some of them from from Google and other places 00:34:36.370 --> 00:34:39.117 [snorts] um but there are limitations to those 00:34:39.200 --> 00:34:43.517 contrasting method What has changed in the last uh 00:34:43.600 --> 00:34:46.077 you know three four years is now now we 00:34:46.160 --> 00:34:47.757 have methods that are non-contrastive. 00:34:47.840 --> 00:34:50.637 So they don't require those negative 00:34:50.720 --> 00:34:53.037 contrastive samples of images that are 00:34:53.120 --> 00:34:58.157 that we know are different. You can only you train them only with images that are 00:34:58.240 --> 00:35:02.637 you know different versions or different views of the same thing. Uh and you rely 00:35:02.720 --> 00:35:06.317 on some other tweaks to prevent the system from collapsing. And we have half 00:35:06.400 --> 00:35:08.797 a dozen different methods for this now. 00:35:08.880 --> 00:35:10.717 So what is the fundamental difference 00:35:10.800 --> 00:35:13.597 between joint embedding architectures 00:35:13.680 --> 00:35:17.677 and LLMs? So can uh can uh jabbot take 00:35:17.760 --> 00:35:21.677 us to AGI? Whether we should say that 00:35:21.760 --> 00:35:24.957 you don't like uh the term AGI and we'll 00:35:25.040 --> 00:35:26.957 probably argue I think every single time 00:35:27.040 --> 00:35:28.717 I've talked to you with argued about the 00:35:28.800 --> 00:35:32.797 G and AGI. Yes, I get I get it. I get 00:35:32.880 --> 00:35:34.397 it. We'll we'll probably continue to 00:35:34.480 --> 00:35:36.717 argue about it. It's great. Uh you you 00:35:36.800 --> 00:35:39.917 like uh Amy this cuz cuz you like French 00:35:40.000 --> 00:35:44.557 and um Ammy is is is uh I guess friend in French. 00:35:44.640 --> 00:35:48.637 >> Yes. >> And AMI stands for advanced machine intelligence, 00:35:48.720 --> 00:35:52.557 >> right? >> Um but either way can Japa take us to 00:35:52.640 --> 00:35:55.437 that towards that advanced machine intelligence? 00:35:55.520 --> 00:35:57.677 >> Well, so it's a it's a first step. Okay. 00:35:57.760 --> 00:35:59.757 So first of all, uh what's the 00:35:59.840 --> 00:36:02.157 difference with generative architectures 00:36:02.240 --> 00:36:07.757 like LLMs? Um so LLMs um or 00:36:07.840 --> 00:36:10.477 vision systems that are trained by reconstruction 00:36:10.560 --> 00:36:13.437 generate the inputs right they generate 00:36:13.520 --> 00:36:17.997 the original input that is non-corrupted 00:36:18.080 --> 00:36:21.117 non-transformed right so you have to predict all the 00:36:21.200 --> 00:36:26.397 pixels and there is a huge amount of resources spent in the system to 00:36:26.480 --> 00:36:29.837 actually predict all those pixels all the details 00:36:29.920 --> 00:36:32.637 uh in a ja you're not trying to predict 00:36:32.720 --> 00:36:34.397 all the pixels you're only trying to 00:36:34.480 --> 00:36:37.757 predict an abstract representation of of 00:36:37.840 --> 00:36:43.197 the inputs, right? And that's much easier in many ways. So what the JPA 00:36:43.280 --> 00:36:47.037 system when it's being trained is trying to do is extract as much information as 00:36:47.120 --> 00:36:49.917 possible from the input but yet only 00:36:50.000 --> 00:36:54.477 extract information that is relatively easily predictable. 00:36:54.560 --> 00:36:56.397 >> Okay. >> So there's a lot of things in the world 00:36:56.480 --> 00:36:58.077 that we cannot predict. For example, if 00:36:58.160 --> 00:36:59.757 you have a self-driving car driving down 00:36:59.840 --> 00:37:03.517 the street or road, uh there may be uh 00:37:03.600 --> 00:37:06.637 trees around the around the road and it 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 00:37:10.720 --> 00:37:14.797 of semi chaotic random ways that you can't predict and you don't care. You 00:37:14.880 --> 00:37:18.957 don't want to predict. So what you want is your encoder to basically eliminate 00:37:19.040 --> 00:37:21.837 all those details. It will tell you there's moving leaves, but it's not 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 00:37:26.960 --> 00:37:29.997 the prediction in representation space, you're not going to have to predict 00:37:30.080 --> 00:37:32.797 every single pixel of every leaf. And 00:37:32.880 --> 00:37:35.997 that, you know, um not only is a lot 00:37:36.080 --> 00:37:38.477 simpler, but also it allows the system 00:37:38.560 --> 00:37:40.877 to essentially learn an abstract 00:37:40.960 --> 00:37:43.437 representation of of the world where you 00:37:43.520 --> 00:37:46.237 know what can be modeled and predicted 00:37:46.320 --> 00:37:51.997 is preserved and the rest is viewed as noise and eliminated by the encoder. So 00:37:52.080 --> 00:37:55.197 it kind of lifts the level of abstraction of the representation. If 00:37:55.280 --> 00:37:58.557 you think about this, this is something we do absolutely all the time. Whenever 00:37:58.640 --> 00:38:02.797 we describe a phenomenon, we describe it at a particular level of abstraction. 00:38:02.880 --> 00:38:07.277 >> And we don't always describe every natural phenomenon in terms of quantum 00:38:07.360 --> 00:38:11.357 field theory, right? That would be impossible, right? So we have multiple 00:38:11.440 --> 00:38:15.757 levels of abstraction to describe what happens in the world. you know, starting 00:38:15.840 --> 00:38:17.997 from quantum field theory to like atomic 00:38:18.080 --> 00:38:23.117 theory and molecules, you know, and chemistry, materials and, you know, all 00:38:23.200 --> 00:38:26.877 the way up to, you know, kind of concrete objects in the real world and 00:38:26.960 --> 00:38:29.117 things like that. So the we we can't 00:38:29.200 --> 00:38:32.717 just only model everything at the lowest 00:38:32.800 --> 00:38:35.197 level and that that's what the idea of 00:38:35.280 --> 00:38:38.397 JPA is really on is really about learn 00:38:38.480 --> 00:38:43.117 abstract representation in a self-supervised uh manner and you know 00:38:43.200 --> 00:38:47.277 you can do it hierarchically as well. So that I think is an essential component 00:38:47.360 --> 00:38:51.277 of an intelligent system and in language we can get away without doing this 00:38:51.360 --> 00:38:53.117 because language is already to some 00:38:53.200 --> 00:38:56.557 level abstract and already has 00:38:56.640 --> 00:38:58.557 eliminated a lot of information that is 00:38:58.640 --> 00:39:01.117 not predictable and um so we can get 00:39:01.200 --> 00:39:03.837 away without doing the joint embedding 00:39:03.920 --> 00:39:06.157 without you know lifting the abstraction 00:39:06.240 --> 00:39:09.037 level and by directly predicting words. 00:39:09.120 --> 00:39:11.517 So joint embedding, it's still 00:39:11.600 --> 00:39:16.157 generative, but it's generative in this abstract representation space. 00:39:16.240 --> 00:39:19.437 >> Yeah. >> And you're saying language, we were lazy 00:39:19.520 --> 00:39:23.437 with language cuz we already got the abstract representation for free and now 00:39:23.520 --> 00:39:27.677 we have to zoom out actually think about generally intelligent systems. We have 00:39:27.760 --> 00:39:31.597 to [sighs] deal with a full mess of physical 00:39:31.680 --> 00:39:33.917 reality of reality. And you can't you 00:39:34.000 --> 00:39:37.037 you do have to do this step of jumping 00:39:37.120 --> 00:39:39.277 from 00:39:39.360 --> 00:39:44.877 uh the full rich detailed reality to a 00:39:44.960 --> 00:39:48.717 uh abstract representation of that reality based on which you can then 00:39:48.800 --> 00:39:52.237 reason and all that kind of stuff >> right and the thing is those self-s 00:39:52.320 --> 00:39:54.317 supervised algorithm that that learn by 00:39:54.400 --> 00:39:58.157 prediction even in representation space 00:39:58.240 --> 00:40:02.957 uh they learn more concept if the input 00:40:03.040 --> 00:40:04.637 data you feed them is more redundant. 00:40:04.720 --> 00:40:08.237 The the more redundancy there is in the data, the more they're able to capture 00:40:08.320 --> 00:40:10.797 some internal structure of it. And so 00:40:10.880 --> 00:40:13.357 there there is way more redundancy and 00:40:13.440 --> 00:40:16.797 structure in perceptual uh inputs, 00:40:16.880 --> 00:40:22.157 sensory input like like like vision than there is in uh text which is not nearly 00:40:22.240 --> 00:40:25.597 as redundant. This is back to the question you were asking a few minutes 00:40:25.680 --> 00:40:28.877 ago. Language might represent more information really because it's already 00:40:28.960 --> 00:40:31.597 compressed. You're you're right about that. But that means it's also less 00:40:31.680 --> 00:40:34.717 redundant and so self-supervised 00:40:34.800 --> 00:40:36.317 learning will not work as well. 00:40:36.400 --> 00:40:40.717 >> Is it possible to join the 00:40:40.800 --> 00:40:43.037 self-supervised training on visual data 00:40:43.120 --> 00:40:45.757 and self-supervised training on language 00:40:45.840 --> 00:40:48.157 data? There is a huge amount of 00:40:48.240 --> 00:40:50.397 knowledge. Even though you talk down 00:40:50.480 --> 00:40:55.437 about those 10 to the 13 tokens, those 10 to the 13 tokens represent the 00:40:55.520 --> 00:40:57.197 entirety, 00:40:57.280 --> 00:41:00.317 a large fraction of what us humans have 00:41:00.400 --> 00:41:03.037 figured out. Both the talk on 00:41:03.120 --> 00:41:05.757 Reddit and the contents of all the books 00:41:05.840 --> 00:41:08.317 and the articles and the full spectrum 00:41:08.400 --> 00:41:12.157 of human uh intellectual creation. So is 00:41:12.240 --> 00:41:14.957 it possible to join those two together? 00:41:15.040 --> 00:41:18.557 >> Well, eventually yes. But I think uh if 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, 00:41:23.440 --> 00:41:26.477 that's what people are doing at the moment with vision language model. We're 00:41:26.560 --> 00:41:29.837 basically cheating. We're using uh 00:41:29.920 --> 00:41:31.997 language as a crutch to help the 00:41:32.080 --> 00:41:35.837 deficiencies of our uh vision systems to 00:41:35.920 --> 00:41:37.837 kind of learn good representations from 00:41:37.920 --> 00:41:40.237 uh images and video. And uh the problem 00:41:40.320 --> 00:41:43.277 with this is that we might you know 00:41:43.360 --> 00:41:45.997 improve our uh vision language system a 00:41:46.080 --> 00:41:48.397 bit I mean our language models by you 00:41:48.480 --> 00:41:51.197 know feeding them images but we're not 00:41:51.280 --> 00:41:53.757 going to get to the level of even the 00:41:53.840 --> 00:41:58.397 intelligence or level of understanding of the world of a cat or dog which 00:41:58.480 --> 00:42:02.317 doesn't have language you know they don't have language and they understand 00:42:02.400 --> 00:42:04.957 the world much better than any LLM they 00:42:05.040 --> 00:42:08.077 can plan really complex actions and sort 00:42:08.160 --> 00:42:09.597 of imagine the result of a bunch of 00:42:09.680 --> 00:42:12.317 actions. How do we get machines to learn 00:42:12.400 --> 00:42:16.797 that before we combine that with language? Obviously, if we combine this 00:42:16.880 --> 00:42:19.597 with language, this is going to be a winner. 00:42:19.680 --> 00:42:22.717 >> Um, but but before that, we have to 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 00:42:26.720 --> 00:42:31.997 of joint embedding predictive architecture for you that's going to be 00:42:32.080 --> 00:42:33.597 able to learn something like common 00:42:33.680 --> 00:42:37.037 sense, something like what a cat uses to 00:42:37.120 --> 00:42:40.397 predict how to mess with its owner most 00:42:40.480 --> 00:42:42.717 optimally by knocking over a thing. 00:42:42.800 --> 00:42:45.917 >> That's that's the hope. In fact, the techniques we're using are 00:42:46.000 --> 00:42:49.277 non-contrastive. Uh so not only is the architecture 00:42:49.360 --> 00:42:51.917 non-generative, the learning procedures 00:42:52.000 --> 00:42:57.037 we're using are non-contrastive. So we have two two sets of techniques. One set 00:42:57.120 --> 00:42:58.877 is based on distillation and there's a 00:42:58.960 --> 00:43:01.757 number of uh methods that use this 00:43:01.840 --> 00:43:05.437 principle. Uh one by deep mind called by 00:43:05.520 --> 00:43:09.037 uh a couple by by fair one one called uh 00:43:09.120 --> 00:43:13.917 vikra and another one called Jeppa and Vikra I 00:43:14.000 --> 00:43:19.037 should say is not a distillation method actually but IA and B certainly are. And 00:43:19.120 --> 00:43:24.557 there's another one also called Dino or Dino also produced from uh at fair. And 00:43:24.640 --> 00:43:25.757 the idea of those things is that you 00:43:25.840 --> 00:43:28.157 take the full input let's say an image. 00:43:28.240 --> 00:43:31.757 Uh you run it through an encoder 00:43:31.840 --> 00:43:36.237 uh produces a representation and then you corrupt that input or transform it 00:43:36.320 --> 00:43:39.517 run it through the essentially what amounts to the same encoder with some 00:43:39.600 --> 00:43:44.157 minor differences and then train a predictor. Sometimes the predictor is 00:43:44.240 --> 00:43:47.197 very simple, sometime doesn't exist. But train a predictor to predict a 00:43:47.280 --> 00:43:50.317 representation of the first uh 00:43:50.400 --> 00:43:52.477 uncorrupted input from the corrupted 00:43:52.560 --> 00:43:56.557 input. Um but you only train the the 00:43:56.640 --> 00:43:59.357 second branch. Um you only train the 00:43:59.440 --> 00:44:04.557 part of the network that is fed with the corrupted input. The other network you 00:44:04.640 --> 00:44:07.757 don't you don't train. But since they share the same weight, when you modify 00:44:07.840 --> 00:44:12.397 the first one, it also modifies the second one. uh and with various tricks 00:44:12.480 --> 00:44:16.237 you can prevent the system from collapsing uh with the collapse of the 00:44:16.320 --> 00:44:17.757 type I was explaining before where the 00:44:17.840 --> 00:44:20.317 system basically ignores the input 00:44:20.400 --> 00:44:24.317 um so that works very well the the 00:44:24.400 --> 00:44:26.077 technique we've the two techniques we've 00:44:26.160 --> 00:44:30.237 developed at fair dino and uh and Jeppa 00:44:30.320 --> 00:44:34.637 work really well for that >> so what kind of data are we talking about here 00:44:34.720 --> 00:44:36.957 >> so this there's several scenario one uh 00:44:37.040 --> 00:44:40.717 one scenario is you take an image You 00:44:40.800 --> 00:44:45.357 corrupt it by um changing the cropping, 00:44:45.440 --> 00:44:48.477 for example, changing the size a little bit, maybe changing the orientation, 00:44:48.560 --> 00:44:52.717 blurring it, changing the colors, doing all kinds of horrible things to it, 00:44:52.800 --> 00:44:55.757 >> but basic horrible things. >> Basic horrible things that sort of 00:44:55.840 --> 00:44:57.197 degrade the quality a little bit and 00:44:57.280 --> 00:44:59.517 change the framing, uh you know, crop 00:44:59.600 --> 00:45:04.157 the image. Um or and in some cases, in 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 00:45:07.280 --> 00:45:09.517 some parts of it, right? You just 00:45:09.600 --> 00:45:12.717 basically remove some regions like a big 00:45:12.800 --> 00:45:15.597 block essentially [snorts] and and then 00:45:15.680 --> 00:45:19.837 you know run through the encoders um and train the entire system encoder and 00:45:19.920 --> 00:45:21.597 predictor to predict the representation 00:45:21.680 --> 00:45:23.117 of the good one from the representation 00:45:23.200 --> 00:45:27.597 of the corrupted one. Um so that's the 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 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 00:45:36.480 --> 00:45:39.437 know it's an image because you need to do things like you know geometry 00:45:39.520 --> 00:45:44.317 transformation and blurring and things like that that are really image specific. 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 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 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 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 00:46:01.200 --> 00:46:05.197 video over the entire video. >> And that tube is like statically 00:46:05.280 --> 00:46:08.717 positioned throughout the frames. Just literally straight tube 00:46:08.800 --> 00:46:12.717 >> the the tube. Yeah. Typically is 16 frames or something. And we mask the 00:46:12.800 --> 00:46:15.037 same region over the entire 16 frames. 00:46:15.120 --> 00:46:16.957 It's a different one for every video 00:46:17.040 --> 00:46:20.237 obviously. And um and then again train 00:46:20.320 --> 00:46:22.717 that system. so as to predict the 00:46:22.800 --> 00:46:27.277 representation of the full video from the partially masked video. Uh and that 00:46:27.360 --> 00:46:31.357 works really well. It's the first system that we have that learns good 00:46:31.440 --> 00:46:33.197 representations of video so that when 00:46:33.280 --> 00:46:38.397 you feed those representations to a supervised uh classifier head, it can it 00:46:38.480 --> 00:46:41.917 can tell you what action is taking place in a video with you know pretty good 00:46:42.000 --> 00:46:45.997 accuracy. So that that's it's the first time we 00:46:46.080 --> 00:46:50.317 get something of that uh of that quality. So that's a good test that a 00:46:50.400 --> 00:46:53.117 good representation is formed. That means there's something to this. 00:46:53.200 --> 00:46:57.037 >> Yeah. Um we also preliminary result that 00:46:57.120 --> 00:46:59.117 seem to indicate that the representation 00:46:59.200 --> 00:47:02.157 allows us allow our system to tell 00:47:02.240 --> 00:47:05.357 whether the video is physically possible 00:47:05.440 --> 00:47:09.517 or completely impossible because some object disappeared or an object, you 00:47:09.600 --> 00:47:11.117 know, suddenly jumped from one location 00:47:11.200 --> 00:47:13.917 to another or or change shape or 00:47:14.000 --> 00:47:19.677 something. So it's able to capture some physical con some physics based 00:47:19.760 --> 00:47:22.077 constraints about the reality represented in the video. 00:47:22.160 --> 00:47:25.917 >> Yeah. >> About the appearance and the disappearance of objects. 00:47:26.000 --> 00:47:28.557 >> Yeah. That's really new. 00:47:28.640 --> 00:47:32.957 >> Okay. But c can this actually 00:47:33.040 --> 00:47:37.437 get us to this kind of uh world model that 00:47:37.520 --> 00:47:41.357 understands enough about the world to be able to drive a car? 00:47:41.440 --> 00:47:43.597 >> Uh possibly. I mean this is going to 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 00:47:48.720 --> 00:47:50.077 you know robotic systems that are based 00:47:50.160 --> 00:47:54.637 on this uh idea. Uh and uh what you need 00:47:54.720 --> 00:47:57.117 for this is a slightly modified version 00:47:57.200 --> 00:48:01.517 of this where um imagine that you have 00:48:01.600 --> 00:48:05.437 uh a video and a complete video. And 00:48:05.520 --> 00:48:10.237 what you're doing to this video is that you're either translating it in time 00:48:10.320 --> 00:48:13.197 towards the future. So you only see the beginning of the video, but you don't 00:48:13.280 --> 00:48:15.437 see the latter part of it that is in the 00:48:15.520 --> 00:48:18.077 original one. Or you just mask the 00:48:18.160 --> 00:48:20.477 second half of the video, for example. 00:48:20.560 --> 00:48:25.117 Um, and then you you train a a JPA system of the type I described to 00:48:25.200 --> 00:48:26.477 predict the representation of the full 00:48:26.560 --> 00:48:29.677 video from the the shifted one, but you 00:48:29.760 --> 00:48:32.477 also feed the predictor with an action. 00:48:32.560 --> 00:48:34.237 For example, you know, the wheel is 00:48:34.320 --> 00:48:36.957 turned 10° to the to the right or 00:48:37.040 --> 00:48:39.677 something, right? So, if [snorts] it's 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 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 00:48:47.680 --> 00:48:51.597 happen to what you see. uh you're not going to be able to 00:48:51.680 --> 00:48:55.037 predict all the details of you know objects that appear in the view 00:48:55.120 --> 00:48:59.517 obviously but at a abstract representation level you can you can 00:48:59.600 --> 00:49:04.237 probably predict what's going to happen >> so now what you have is 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 00:49:08.080 --> 00:49:11.037 here is an action I'm taking here's a 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 00:49:16.000 --> 00:49:17.917 whatever it is if you have a model of 00:49:18.000 --> 00:49:19.757 this type you can use is it for 00:49:19.840 --> 00:49:23.197 planning. So now you can do what LLMs 00:49:23.280 --> 00:49:25.837 cannot do which is planning what you're 00:49:25.920 --> 00:49:28.077 going to do so as to arrive at a 00:49:28.160 --> 00:49:31.597 particular uh outcome or satisfy a 00:49:31.680 --> 00:49:34.077 particular objective right so you can 00:49:34.160 --> 00:49:37.037 have a number of objectives 00:49:37.120 --> 00:49:39.677 um right if you know I can I can predict 00:49:39.760 --> 00:49:42.317 that uh if I have 00:49:42.400 --> 00:49:44.637 an object like this right and I open my 00:49:44.720 --> 00:49:49.917 hand it's going to fall right >> and uh and if I push it with a 00:49:50.000 --> 00:49:52.477 particular force on the table, it's going to move. If I push the table 00:49:52.560 --> 00:49:54.397 itself, it's probably not going to move 00:49:54.480 --> 00:49:57.677 uh with the same force. Um so we have we 00:49:57.760 --> 00:49:59.677 have this internal model of the world in 00:49:59.760 --> 00:50:03.197 our in our mind uh which allows us to 00:50:03.280 --> 00:50:05.437 plan sequences of actions to arrive at a 00:50:05.520 --> 00:50:10.077 particular goal. Um and so um so now if 00:50:10.160 --> 00:50:15.037 you have this world model we can imagine a sequence of actions predict what the 00:50:15.120 --> 00:50:16.637 outcome of the sequence of action is 00:50:16.720 --> 00:50:20.317 going to be measure to what extent the 00:50:20.400 --> 00:50:22.637 final state satisfies a particular 00:50:22.720 --> 00:50:28.677 objective like you know moving the bottle to the left of the table 00:50:28.760 --> 00:50:31.117 >> [snorts] >> um and then plan a sequence of actions 00:50:31.200 --> 00:50:34.797 that will minimize this objective at runtime. We're not talking about 00:50:34.880 --> 00:50:38.957 learning. talking about inference time, right? So, this is planning really. And 00:50:39.040 --> 00:50:42.397 in optimal control, this is a very classical thing. It's called uh model 00:50:42.480 --> 00:50:46.797 predictive control. You have a model of the system you want to control that you 00:50:46.880 --> 00:50:51.757 know can predict the sequence of states corresponding to a sequence of commands 00:50:51.840 --> 00:50:53.757 and you're planning a sequence of 00:50:53.840 --> 00:50:56.637 commands so that according to your world 00:50:56.720 --> 00:50:58.797 model the the the end state of the 00:50:58.880 --> 00:51:02.797 system will uh satisfy an objective that 00:51:02.880 --> 00:51:05.997 you fix. This is the way 00:51:06.080 --> 00:51:10.237 uh you know rocket trajectories have been planned since computers have been 00:51:10.320 --> 00:51:12.797 around so since the early 60s essentially. 00:51:12.880 --> 00:51:18.157 >> So yes for model predictive control but you also often talk about hierarchical 00:51:18.240 --> 00:51:21.677 planning. >> Yeah. >> Can hierarchical planning emerge from this somehow? 00:51:21.760 --> 00:51:26.477 >> Well so no you you will have to build specific architecture to allow for 00:51:26.560 --> 00:51:29.917 hierarchical planning. So hierarchical planning is absolutely necessary if you 00:51:30.000 --> 00:51:35.117 want to plan complex actions. uh if I want to go from let's say from 00:51:35.200 --> 00:51:36.797 New York to Paris, it's the example I 00:51:36.880 --> 00:51:39.597 use all the time and I'm sitting uh in 00:51:39.680 --> 00:51:42.557 my office at NYU. My objective that I 00:51:42.640 --> 00:51:43.997 need to minimize is my distance to 00:51:44.080 --> 00:51:47.117 Paris. At a high level, a very abstract 00:51:47.200 --> 00:51:50.157 representation of my uh my location, I 00:51:50.240 --> 00:51:54.637 would have to decompose this into two sub goals. First one is um go to the 00:51:54.720 --> 00:51:57.357 airport. Second one is catch your plane to Paris. 00:51:57.440 --> 00:52:00.157 >> Okay. So my sub goal is now 00:52:00.240 --> 00:52:02.557 uh going to the airport. My objective 00:52:02.640 --> 00:52:05.357 function is my distance to the airport. 00:52:05.440 --> 00:52:07.277 How do I go to the airport where I have 00:52:07.360 --> 00:52:10.957 to go in the street and hail the taxi 00:52:11.040 --> 00:52:14.157 >> which you can do in New York? Um, okay. 00:52:14.240 --> 00:52:18.957 Now, I have another sub goal. Go down on the street. Uh, well, that means uh 00:52:19.040 --> 00:52:21.117 going to the elevator. Going down the 00:52:21.200 --> 00:52:23.997 elevator, walk out the street. How do I 00:52:24.080 --> 00:52:27.837 go to the elevator? I have to uh 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 00:52:32.480 --> 00:52:34.317 the button. How do I get up from my 00:52:34.400 --> 00:52:39.677 chair? Like, you know, you can imagine going down all the way down to basically 00:52:39.760 --> 00:52:43.197 what amounts to millisecond by millisecond muscle control. 00:52:43.280 --> 00:52:46.157 >> Okay? And obviously, you're not going to 00:52:46.240 --> 00:52:48.397 plan your entire trip from New York to 00:52:48.480 --> 00:52:53.357 Paris in terms of millisecond by millisecond muscle control. First, that 00:52:53.440 --> 00:52:56.557 would be incredibly expensive. But it will also be completely impossible 00:52:56.640 --> 00:53:00.797 because you don't know all the conditions of what's going to happen. Uh 00:53:00.880 --> 00:53:02.317 you know how long it's going to take to 00:53:02.400 --> 00:53:05.757 catch a taxi um or to go to the airport 00:53:05.840 --> 00:53:10.157 with traffic you know uh I mean you you would have to know exactly the condition 00:53:10.240 --> 00:53:13.597 of everything to be able to do this planning and you don't have the 00:53:13.680 --> 00:53:17.357 information. So you you have to do this hierarchical planning so that you can 00:53:17.440 --> 00:53:19.437 start acting and then sort of replanning 00:53:19.520 --> 00:53:23.277 as you go. And nobody really knows how 00:53:23.360 --> 00:53:27.277 to do this in AI. Um nobody knows how to 00:53:27.360 --> 00:53:29.757 train a system to learn the appropriate 00:53:29.840 --> 00:53:34.317 multiple levels of representation so that hierarchical planning works. Does 00:53:34.400 --> 00:53:35.997 something like that already emerge? So, 00:53:36.080 --> 00:53:39.517 like, can you use an LLM, 00:53:39.600 --> 00:53:42.317 state-of-the-art LLM to get you from New 00:53:42.400 --> 00:53:46.157 York to Paris by doing exactly the kind 00:53:46.240 --> 00:53:48.237 of detailed set of questions that you 00:53:48.320 --> 00:53:50.477 just did, which is, can you give me a 00:53:50.560 --> 00:53:54.157 high a list of 10 steps I need to do to 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 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 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 00:54:06.720 --> 00:54:09.597 your mus individual muscles? Uh maybe 00:54:09.680 --> 00:54:14.237 not whatever you can actually act upon using your own mind. 00:54:14.320 --> 00:54:17.357 >> Right? So there's a lot of questions that are sort of implied by this. Right? 00:54:17.440 --> 00:54:19.837 So the first thing is uh LLM will be 00:54:19.920 --> 00:54:25.117 able to answer some of those questions down to some level of abstraction 00:54:25.200 --> 00:54:29.037 >> under the condition that they've been trained with similar scenarios in their 00:54:29.120 --> 00:54:30.877 training set. They would be able to 00:54:30.960 --> 00:54:36.317 answer all of those questions, but some of them may be hallucinated, meaning 00:54:36.400 --> 00:54:38.237 non-factual. >> Yeah, true. I mean, they will probably 00:54:38.320 --> 00:54:40.877 produce some answer, except they're not going to be able to really kind of 00:54:40.960 --> 00:54:44.397 produce millisecond by millisecond muscle control of how you how you stand 00:54:44.480 --> 00:54:47.757 up from your chair, >> right? So, but down to some level of 00:54:47.840 --> 00:54:51.357 abstraction where you can describe things by words, they might be able to 00:54:51.440 --> 00:54:54.317 give you a plan, but only under the condition that they've been trained to 00:54:54.400 --> 00:54:56.477 produce those kind of plans. >> Mhm. 00:54:56.560 --> 00:55:00.797 >> Right. They're not going to be able to plan for situations where that that they 00:55:00.880 --> 00:55:04.237 never encountered before. They basically are going to have to regurgitate a 00:55:04.320 --> 00:55:05.757 template that they've been trained on. 00:55:05.840 --> 00:55:07.837 >> But where like just for the example of 00:55:07.920 --> 00:55:09.837 New York to Paris is is going to start 00:55:09.920 --> 00:55:14.877 getting into trouble like at which layer layer of abstraction do you think you'll 00:55:14.960 --> 00:55:18.797 start cuz like I can imagine almost every single part of that an LLM would 00:55:18.880 --> 00:55:21.837 be able to answer somewhat accurately especially when you're talking about New 00:55:21.920 --> 00:55:24.397 York and Paris major cities. So I mean 00:55:24.480 --> 00:55:27.037 certainly uh LM would be able to solve 00:55:27.120 --> 00:55:29.357 that problem if you fine tune it for it 00:55:29.440 --> 00:55:33.677 you know just uh and and so uh I can't 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 00:55:37.280 --> 00:55:40.637 question uh down to a certain level 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 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 00:55:48.720 --> 00:55:51.117 chair in terms of uh words like you you 00:55:51.200 --> 00:55:54.157 can't you can't do it Um you you you 00:55:54.240 --> 00:55:59.197 need that's one of the reasons you need experience of the physical world which 00:55:59.280 --> 00:56:03.757 is much higher bandwidth than what you can express in words in human language. 00:56:03.840 --> 00:56:08.157 So everything we've been talking about on the joint embedding space, is it 00:56:08.240 --> 00:56:11.677 possible that that's what we need for like interaction with physical reality 00:56:11.760 --> 00:56:14.077 for on the robotics front and then just 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 00:56:20.480 --> 00:56:22.957 >> Yeah. >> the fact that I need to book a plane 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. 00:56:26.560 --> 00:56:28.877 >> Sure. And you know a lot of plans that 00:56:28.960 --> 00:56:32.077 people know about um that are relatively 00:56:32.160 --> 00:56:33.757 high level are actually learned. They're 00:56:33.840 --> 00:56:36.237 not people mo most people don't invent 00:56:36.320 --> 00:56:42.157 the you know plans um uh they they 00:56:42.240 --> 00:56:44.397 by themselves. they uh you know we have 00:56:44.480 --> 00:56:46.477 some ability to do this of course uh 00:56:46.560 --> 00:56:51.757 obviously but um but most plans that people use are plants that they've been 00:56:51.840 --> 00:56:54.477 trained on like they've seen other people use those plans or they've been 00:56:54.560 --> 00:56:57.197 told how to do things right um like you 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 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 00:57:06.640 --> 00:57:09.837 be able to kind of you know deconstruct the whole plan unless they've seen 00:57:09.920 --> 00:57:12.077 examples of that before um so certainly 00:57:12.160 --> 00:57:14.157 LM going to be able to do this. But but 00:57:14.240 --> 00:57:17.677 then um how you link this from the the 00:57:17.760 --> 00:57:21.917 low level of of of actions uh that needs 00:57:22.000 --> 00:57:25.997 to be done with things like like JPA that basically lift the abstraction 00:57:26.080 --> 00:57:29.357 level of the representation without attempting to reconstruct every detail 00:57:29.440 --> 00:57:33.437 of the situation. That's why we need JPA for 00:57:33.520 --> 00:57:36.317 >> I would love to sort of linger on your 00:57:36.400 --> 00:57:38.717 skepticism around 00:57:38.800 --> 00:57:41.277 uh autogressive LLMs. 00:57:41.360 --> 00:57:44.077 So one way I I would like to test that 00:57:44.160 --> 00:57:50.280 skepticism is everything you say makes a lot of sense. 00:57:50.480 --> 00:57:53.277 But if I apply everything you said today 00:57:53.360 --> 00:57:56.397 and in general to like I don't know 10 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 00:58:01.440 --> 00:58:03.197 to predict 00:58:03.280 --> 00:58:07.437 the uh success of LLMs. So, does it make 00:58:07.520 --> 00:58:10.077 sense to you that autogressive LLMs are 00:58:10.160 --> 00:58:13.437 able to be so damn good? 00:58:13.520 --> 00:58:17.437 >> Yes. >> Can you explain your intuition? Because 00:58:17.520 --> 00:58:21.837 if I were to take your wisdom and intuition 00:58:21.920 --> 00:58:26.557 at face value, I would say there's no way autogressive LLMs, one token at a 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 00:58:29.760 --> 00:58:34.557 thing that autogressive LLM uh or that LLM in general, not just the auto 00:58:34.640 --> 00:58:36.637 reggressive one, but including the bird 00:58:36.720 --> 00:58:40.557 style birectional ones u are exploiting 00:58:40.640 --> 00:58:43.917 and it's self-supervised learning and I've been a very very strong advocate of 00:58:44.000 --> 00:58:46.157 self-supervised learning for many years. 00:58:46.240 --> 00:58:50.557 So those things are a incredibly 00:58:50.640 --> 00:58:52.237 impressive demonstration that self-s 00:58:52.320 --> 00:58:54.877 supervised learning actually works. uh 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 00:59:00.720 --> 00:59:05.037 it was really kind of a good demonstration with this. So the the the 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 00:59:08.000 --> 00:59:09.517 some gigantic neural net to reconstruct 00:59:09.600 --> 00:59:12.317 the parts that are missing. Um that has 00:59:12.400 --> 00:59:16.237 been an enormous uh uh 00:59:16.320 --> 00:59:18.077 produced an enormous amount of benefits. 00:59:18.160 --> 00:59:20.317 Uh it allowed allowed us to create 00:59:20.400 --> 00:59:22.797 systems that understand understand 00:59:22.880 --> 00:59:26.157 language. Uh systems that can translate 00:59:26.240 --> 00:59:29.837 um hundreds of languages in any direction. Systems that are 00:59:29.920 --> 00:59:33.757 multilingual. So they're not it's a single system that can be trained to 00:59:33.840 --> 00:59:36.077 understand hundreds of languages and 00:59:36.160 --> 00:59:39.357 translate in any direction. Um and 00:59:39.440 --> 00:59:42.397 produce summaries um and then answer 00:59:42.480 --> 00:59:46.717 questions and produce text. And then there's a special case of it where you 00:59:46.800 --> 00:59:48.797 know you which is the auto reggressive 00:59:48.880 --> 00:59:51.277 uh trick where you constrain the system 00:59:51.360 --> 00:59:54.477 to not elaborate a representation of the 00:59:54.560 --> 00:59:58.477 text from looking at the entire text but only 00:59:58.560 --> 01:00:02.157 predicting a word from the words that are come before right and you do this by 01:00:02.240 --> 01:00:05.917 the constraining the architecture of the network and that's what you can build an 01:00:06.000 --> 01:00:08.477 auto reggressive LLM from. So there was 01:00:08.560 --> 01:00:11.837 a surprise many years ago with what's 01:00:11.920 --> 01:00:14.797 called decoder only LLM. So since you 01:00:14.880 --> 01:00:16.397 know systems of this type that are just 01:00:16.480 --> 01:00:19.757 trying to produce uh words from the from 01:00:19.840 --> 01:00:22.077 the previous one and and the fact that 01:00:22.160 --> 01:00:26.397 when you scale them up they they tend to 01:00:26.480 --> 01:00:30.317 really kind of understand more about the about language when you train them on 01:00:30.400 --> 01:00:33.517 lots of data and you make them really big. That was kind of a surprise and 01:00:33.600 --> 01:00:35.357 that surprise occurred quite a while 01:00:35.440 --> 01:00:40.317 back like you know uh with uh work from 01:00:40.400 --> 01:00:43.117 uh you know Google Meta, OpenAI etc. you 01:00:43.200 --> 01:00:48.797 know going back to you know the GPT kind of uh work general pre-trained 01:00:48.880 --> 01:00:51.837 transformers. >> You mean like GPT2 like there's a 01:00:51.920 --> 01:00:54.237 certain place where you start to realize 01:00:54.320 --> 01:00:58.157 scaling might actually keep giving us an 01:00:58.240 --> 01:01:01.597 emergent benefit. Yeah, I mean there were there were work from from various 01:01:01.680 --> 01:01:05.117 places but uh uh if if you want to kind 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 01:01:10.480 --> 01:01:14.397 you know. >> Well I just cuz you said it you you're 01:01:14.480 --> 01:01:18.637 so charismatic and you said so many words but self-supervised learning. Yes. 01:01:18.720 --> 01:01:20.957 But again the same intuition you're 01:01:21.040 --> 01:01:23.757 applying to saying that autogressive 01:01:23.840 --> 01:01:26.637 LLMs cannot have a deep understanding of 01:01:26.720 --> 01:01:29.837 the world. If we just apply that same 01:01:29.920 --> 01:01:33.117 intuition, does it make sense to you 01:01:33.200 --> 01:01:37.517 that they're able to form enough of a representation of the world to be damn 01:01:37.600 --> 01:01:42.637 convincing? Essentially passing the original touring test with 01:01:42.720 --> 01:01:45.597 flying colors. >> Well, we're fooled by their fluency, 01:01:45.680 --> 01:01:47.757 right? We just assume that if a system 01:01:47.840 --> 01:01:50.637 is fluent in manipulating language, then 01:01:50.720 --> 01:01:52.877 it has all the characteristics of human 01:01:52.960 --> 01:01:58.237 intelligence. But that impression is false. We we're we're really fooled by 01:01:58.320 --> 01:02:01.357 it. Um >> what do you think Alan Turing would say 01:02:01.440 --> 01:02:04.077 it without understanding anything just hanging out with it? 01:02:04.160 --> 01:02:07.837 >> Alan Turing would decide that a touring test is a really bad test. 01:02:07.920 --> 01:02:12.157 >> Okay, this is what the AI community has decided many years ago that the touring 01:02:12.240 --> 01:02:14.637 test was a really bad test of intelligence. 01:02:14.720 --> 01:02:18.717 >> What would Hans Marvc say about the about the large language models? 01:02:18.800 --> 01:02:22.477 >> Well, Hence Marv would say the Marvc paradox still applies. 01:02:22.560 --> 01:02:26.317 >> Okay. >> Okay. [snorts] Okay. We can pass. I don't think he would be really 01:02:26.400 --> 01:02:28.077 impressed. >> No, of course everybody would be 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 01:02:33.040 --> 01:02:37.197 a question of knowing what the limit of those systems can do. Like there again 01:02:37.280 --> 01:02:40.637 they are impressive. They can do a lot of useful things. There's a whole 01:02:40.720 --> 01:02:44.637 industry that is being built around them. They're going to make progress. Uh 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 01:02:47.760 --> 01:02:51.597 cannot do and uh and then figure out you 01:02:51.680 --> 01:02:53.357 know how we get there. And you know and 01:02:53.440 --> 01:02:55.917 and I'm not seeing this I'm seeing this 01:02:56.000 --> 01:02:58.797 from basically you know 10 years of of 01:02:58.880 --> 01:03:03.277 research uh on on 01:03:03.360 --> 01:03:04.957 the idea of self-s supervised learning. 01:03:05.040 --> 01:03:07.677 Actually that's going back more than 10 years but the idea of self-s supervised 01:03:07.760 --> 01:03:09.917 learning. So basically capturing the 01:03:10.000 --> 01:03:14.877 internal structure of a piece of uh of a set of inputs without training the 01:03:14.960 --> 01:03:18.307 system for any particular task right learning representations. 01:03:18.390 --> 01:03:20.397 >> [snorts] >> um you know the the conference I 01:03:20.480 --> 01:03:23.917 co-ounded 14 years ago is called inter international conference on learning 01:03:24.000 --> 01:03:27.917 representations that's the entire issue that deep learning is is dealing with 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 01:03:32.800 --> 01:03:36.557 learning representation is really the thing uh for the longest time we could 01:03:36.640 --> 01:03:40.717 only do this with supervised learning and then we started working on uh you 01:03:40.800 --> 01:03:44.957 know what we used to call unsupervised learning [snorts] uh 01:03:45.040 --> 01:03:48.237 and sort of revive the idea of unsupervised learning 01:03:48.320 --> 01:03:53.197 in the early 2000s with Yosha Benju and Jeffon then discovered that supervised 01:03:53.280 --> 01:03:57.837 learning actually works pretty well if you can collect enough data and so the 01:03:57.920 --> 01:03:59.597 whole idea of you know unsupervised 01:03:59.680 --> 01:04:02.157 self-supervised learning kind of took a 01:04:02.240 --> 01:04:04.957 a backseat for for a bit and then I kind 01:04:05.040 --> 01:04:09.357 of tried to revive it um uh in a big way 01:04:09.440 --> 01:04:11.837 you know starting in 2014 basically when 01:04:11.920 --> 01:04:15.277 we started fair and uh and really 01:04:15.360 --> 01:04:19.597 pushing for like finding new new methods to do self-supervised learning both for 01:04:19.680 --> 01:04:23.757 text and for images and for video and audio and some of that work has been 01:04:23.840 --> 01:04:26.397 incredibly successful I mean the reason 01:04:26.480 --> 01:04:28.877 why we have multilingual translation 01:04:28.960 --> 01:04:33.517 system you know things to do content moderation on on meta for example on 01:04:33.600 --> 01:04:37.597 Facebook that are multilingual that understand whether piece of text is HP 01:04:37.680 --> 01:04:41.277 speech or not or something is due to that progress using self-supervised 01:04:41.360 --> 01:04:45.917 learning for NLP combining this with you know transformer architectures and and 01:04:46.000 --> 01:04:48.957 blah blah blah So that's the big success of self-supervised learning. We had 01:04:49.040 --> 01:04:53.357 similar success in speech recognition, a system called wave tovec, which is also 01:04:53.440 --> 01:04:56.557 a joint embedding architecture by the way, trained with contrastive learning 01:04:56.640 --> 01:04:58.957 and and that that system also can 01:04:59.040 --> 01:05:01.917 produce um speech recognition systems 01:05:02.000 --> 01:05:06.717 that are multilingual with mostly unlabelled data and only need a few 01:05:06.800 --> 01:05:10.317 minutes of label data to actually do speech recognition. That's that's 01:05:10.400 --> 01:05:13.597 amazing. Um we have systems now based on 01:05:13.680 --> 01:05:17.917 those combination of ideas that can do realtime translation of hundreds of 01:05:18.000 --> 01:05:20.877 languages into each other. Uh speech to speech 01:05:20.960 --> 01:05:23.357 >> speech to speech even including which is 01:05:23.440 --> 01:05:27.677 fascinating languages that uh don't have written forms. That's right. They're 01:05:27.760 --> 01:05:29.517 spoken only. >> That's right. We don't go through text. 01:05:29.600 --> 01:05:32.957 It goes directly from from speech to speech using an internal representation 01:05:33.040 --> 01:05:36.397 of kind of speech units that are discrete. But it's um it's called 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 01:05:40.800 --> 01:05:44.877 incredible success there and then you know for 10 years we tried to apply this 01:05:44.960 --> 01:05:47.677 idea to learning representations of 01:05:47.760 --> 01:05:51.917 images by training a system to predict videos learning intuitive physics by 01:05:52.000 --> 01:05:56.397 training a system to predict what's going to happen in the video and tried 01:05:56.480 --> 01:06:00.157 and tried and failed and failed with generative models with models that 01:06:00.240 --> 01:06:04.317 predict pixels. Uh we could not get them to learn good 01:06:04.400 --> 01:06:07.997 representations of images. We could not get them to learn good representations 01:06:08.080 --> 01:06:11.997 of videos. And we tried many times. We published lots of papers on it. You 01:06:12.080 --> 01:06:13.757 know, they kind of sort of worked but 01:06:13.840 --> 01:06:16.717 not really great. They started working. 01:06:16.800 --> 01:06:20.237 We we abandoned this idea of predicting 01:06:20.320 --> 01:06:23.917 every pixel and basically just doing the joint embedding and predicting in 01:06:24.000 --> 01:06:26.317 representation space that works. 01:06:26.400 --> 01:06:29.917 >> Mhm. >> So there's ample evidence that we're not 01:06:30.000 --> 01:06:31.837 going to be able to learn good 01:06:31.920 --> 01:06:37.277 representations of the real world using generative model. So I'm telling people 01:06:37.360 --> 01:06:40.957 everybody is talking about generative AI. If you're really interested in human 01:06:41.040 --> 01:06:44.237 level AI, abandon the idea of generative AI. 01:06:44.320 --> 01:06:46.557 >> Okay. But you you you really think it's 01:06:46.640 --> 01:06:49.117 possible to get far with the joint 01:06:49.200 --> 01:06:51.357 embedding representation. So like 01:06:51.440 --> 01:06:54.397 there's common sense reasoning and then 01:06:54.480 --> 01:06:59.677 there's high level reasoning. Like I I feel like those are 01:06:59.760 --> 01:07:03.997 two the kind of reasoning that LLMs are able 01:07:04.080 --> 01:07:07.357 to do. Okay. Let me not use the word reasoning, but the kind of stuff that 01:07:07.440 --> 01:07:11.437 LMs are able to do seems fundamentally different than the common sense 01:07:11.520 --> 01:07:13.597 reasoning we use to navigate the world. 01:07:13.680 --> 01:07:15.277 >> It seems like we're going to need both. 01:07:15.360 --> 01:07:17.757 You're not sure. Would you be able to get 01:07:17.840 --> 01:07:19.517 >> with the joint embedding? Would you jump 01:07:19.600 --> 01:07:24.587 type of approach looking at video? Would you be able to learn [sighs] 01:07:24.670 --> 01:07:27.117 [gasps] let's see well how to get from New York 01:07:27.200 --> 01:07:30.957 to Paris or um 01:07:31.040 --> 01:07:34.397 how to uh state understand the state of 01:07:34.480 --> 01:07:38.157 politics in the world today [laughter] right these these are things where 01:07:38.240 --> 01:07:40.397 various humans generate a lot of 01:07:40.480 --> 01:07:42.477 language and opinions on in the space of 01:07:42.560 --> 01:07:45.357 language but don't visually represent 01:07:45.440 --> 01:07:50.717 that in any clearly uh compressible way >> right well there's a lot of situations 01:07:50.800 --> 01:07:52.317 that you know might 01:07:52.400 --> 01:07:55.197 difficult to for a purely language based 01:07:55.280 --> 01:07:59.277 system to um to know like okay you can 01:07:59.360 --> 01:08:03.197 probably learn from reading text the entirety of the public publicly 01:08:03.280 --> 01:08:06.637 available text in the world that I cannot get from New York to Paris by 01:08:06.720 --> 01:08:10.077 snapping my fingers >> that's not going to work right >> yes 01:08:10.160 --> 01:08:12.397 >> uh but there's you know probably sort of 01:08:12.480 --> 01:08:15.117 more complex scenarios of this type 01:08:15.200 --> 01:08:18.397 which an NLM may never have encountered 01:08:18.480 --> 01:08:20.637 and may not be able to determine whether 01:08:20.720 --> 01:08:23.917 it's possible or not. Um so 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 01:08:28.480 --> 01:08:29.677 thing is that the high level that 01:08:29.760 --> 01:08:33.517 language expresses is based on a common 01:08:33.600 --> 01:08:36.797 experience of the low level which LLMs 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 01:08:41.760 --> 01:08:43.357 common experience of the of the world 01:08:43.440 --> 01:08:45.917 like you know a lot of it is 01:08:46.000 --> 01:08:50.957 >> is similar u and 01:08:51.040 --> 01:08:54.797 don't have that but see there it's present you and I have a common 01:08:54.880 --> 01:08:56.157 experience of the world in terms of the 01:08:56.240 --> 01:08:58.957 physics of how gravity works and stuff 01:08:59.040 --> 01:09:01.917 like this and that 01:09:02.000 --> 01:09:04.877 common knowledge of the world I feel 01:09:04.960 --> 01:09:08.877 like is there in the language we don't 01:09:08.960 --> 01:09:11.277 explicitly express it. But if you have a 01:09:11.360 --> 01:09:15.037 huge amount of text, you're going to get 01:09:15.120 --> 01:09:16.877 this stuff that's between the lines. 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 01:09:21.920 --> 01:09:23.837 going to have to understand how gravity 01:09:23.920 --> 01:09:29.037 works even if you don't have an explicit explanation of gravity. So even though 01:09:29.120 --> 01:09:30.477 there in the case of gravity, there is 01:09:30.560 --> 01:09:31.997 explicit explanations of gravity on 01:09:32.080 --> 01:09:36.317 Wikipedia. But uh you're like the stuff 01:09:36.400 --> 01:09:40.557 that we think of as common sense reasoning. I feel like to generate 01:09:40.640 --> 01:09:45.437 language correctly, you're going to have to figure that out. Now, you could say 01:09:45.520 --> 01:09:47.437 as you there's not enough text. Sorry. 01:09:47.520 --> 01:09:50.477 Okay. So, what [laughter] you don't think so? 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 01:09:53.680 --> 01:09:58.717 level um common sense, to have high level common sense, you need to have the 01:09:58.800 --> 01:10:00.797 low-level common sense to build on top of. 01:10:00.880 --> 01:10:04.797 >> Yeah. Um >> but that's not there >> and that's not there in LLMs. LLMs are 01:10:04.880 --> 01:10:08.797 purely trained from text. So so then the other statement you made um I would not 01:10:08.880 --> 01:10:11.117 I would not agree with the fact that 01:10:11.200 --> 01:10:16.077 implicit in all languages in the world is the underlying reality. There's a lot 01:10:16.160 --> 01:10:19.597 about underlying reality which is not expressed in language. 01:10:19.680 --> 01:10:20.797 >> Is that obvious to you? 01:10:20.880 --> 01:10:23.037 >> Yeah, totally. 01:10:23.120 --> 01:10:26.637 So like all all the conversations we 01:10:26.720 --> 01:10:31.197 have what okay there's the dark web meaning uh whatever the private 01:10:31.280 --> 01:10:32.877 conversations like DMs and stuff like 01:10:32.960 --> 01:10:38.797 this which is much much larger probably than what's available what what LLMs are 01:10:38.880 --> 01:10:42.797 trained on >> you don't need to communicate the stuff that is common 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 01:10:47.120 --> 01:10:49.757 through like you like if I accidentally 01:10:49.840 --> 01:10:51.757 uh knock this over you'll probably make 01:10:51.840 --> 01:10:54.077 fun of me and in the content of the you 01:10:54.160 --> 01:10:57.437 making fun of me will be uh explanation 01:10:57.520 --> 01:11:00.877 of the fact that cups fall and then you 01:11:00.960 --> 01:11:04.957 know gravity works in this way and then you you'll have some very vague 01:11:05.040 --> 01:11:09.677 information about what kind of things explode when they hit the ground and 01:11:09.760 --> 01:11:12.957 then maybe you'll make a joke about entropy or something like this and we'll 01:11:13.040 --> 01:11:17.357 never be able to reconstruct this again like okay you'll make a a little joke 01:11:17.440 --> 01:11:21.197 like this and there'll be trillion of other jokes and from the jokes you can 01:11:21.280 --> 01:11:25.117 piece together. The fact that gravity works and mugs can break and all this 01:11:25.200 --> 01:11:28.077 kind of stuff. You don't need to see uh 01:11:28.160 --> 01:11:32.317 it'll be very inefficient. It's easier for like 01:11:32.400 --> 01:11:34.877 not knock the thing over. Yeah. 01:11:34.960 --> 01:11:37.277 >> But uh I feel like it would be there if 01:11:37.360 --> 01:11:39.357 you have enough of that data. 01:11:39.440 --> 01:11:42.477 I just think that most of the 01:11:42.560 --> 01:11:45.037 information of this type that we have 01:11:45.120 --> 01:11:47.357 accumulated when when we were babies is 01:11:47.440 --> 01:11:51.037 just not present in uh in in text any in 01:11:51.120 --> 01:11:55.037 any description essentially >> and the sensory data is much is a much 01:11:55.120 --> 01:11:58.477 richer source for getting that kind of understanding. I mean that's the 16,000 01:11:58.560 --> 01:12:01.757 hours of of wake time of a 4-year-old 01:12:01.840 --> 01:12:06.397 and 10 to the 15 bytes you know going through vision just vision right there 01:12:06.480 --> 01:12:08.477 is a similar 01:12:08.560 --> 01:12:11.517 uh bandwidth you know of touch and uh a 01:12:11.600 --> 01:12:13.837 little less through audio and then text 01:12:13.920 --> 01:12:15.597 doesn't language doesn't come in until 01:12:15.680 --> 01:12:19.677 like you know a year uh in in life and 01:12:19.760 --> 01:12:24.157 by the time you are 9 years old you've learned about gravity you know about 01:12:24.240 --> 01:12:27.757 inertia you know about gravity you know the stability you 01:12:27.840 --> 01:12:29.277 you know about the distinction between 01:12:29.360 --> 01:12:31.597 animate and inanimate objects. you know 01:12:31.680 --> 01:12:34.957 by 18 months you know about like uh why 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 01:12:39.040 --> 01:12:42.477 there's a lot of things that you learn mostly by observation 01:12:42.560 --> 01:12:46.637 >> really uh not even through interaction in the first few months of life babies 01:12:46.720 --> 01:12:50.717 don't don't really have any influence on the world they can only observe right 01:12:50.800 --> 01:12:55.037 and [snorts] you accumulate like a gigantic amount of uh of knowledge just 01:12:55.120 --> 01:13:00.157 just from that so that that's what we're missing from uh current AI systems 01:13:00.240 --> 01:13:01.757 >> I I think in one of your slides you have 01:13:01.840 --> 01:13:04.557 this nice plot that is one of the ways 01:13:04.640 --> 01:13:08.877 you show that LLMs are limited. I wonder if you could talk about hallucinations 01:13:08.960 --> 01:13:11.037 from your perspectives. [snorts] 01:13:11.120 --> 01:13:15.437 The why hallucinations happen from large 01:13:15.520 --> 01:13:20.477 language models and why and to what degree is that a fundamental flaw of 01:13:20.560 --> 01:13:22.237 large language models. 01:13:22.320 --> 01:13:24.957 >> Right? So because of the auto 01:13:25.040 --> 01:13:27.597 reggressive prediction, every time an 01:13:27.680 --> 01:13:31.277 L&M produces a token or a word, uh there 01:13:31.360 --> 01:13:35.677 is some level of probability for that word to take you out of the set of 01:13:35.760 --> 01:13:39.997 reasonable answers. Uh and if you assume which is a very 01:13:40.080 --> 01:13:42.317 strong assumption that the probability 01:13:42.400 --> 01:13:44.717 of such error 01:13:44.800 --> 01:13:47.997 um is that those errors are independent 01:13:48.080 --> 01:13:52.237 across a a sequence of tokens being produced. Mhm. 01:13:52.320 --> 01:13:55.357 >> What [clears throat] that means is that every time you produce a token, the 01:13:55.440 --> 01:13:59.117 probability that you rest you you stay within the the set of correct answer 01:13:59.200 --> 01:14:01.357 decreases and it decreases exponentially. 01:14:01.440 --> 01:14:05.837 >> So there's a strong like you said assumption there that if uh there's a 01:14:05.920 --> 01:14:09.757 non-zero probability of making a mistake which there appears to be then there's 01:14:09.840 --> 01:14:11.517 going to be a kind of drift. 01:14:11.600 --> 01:14:14.157 >> Yeah. And that drift is exponential. 01:14:14.240 --> 01:14:16.957 It's like errors accumulate. Right. So, 01:14:17.040 --> 01:14:19.757 so the probability that an answer would 01:14:19.840 --> 01:14:24.157 be nonsensical increases exponentially with the number of tokens. 01:14:24.240 --> 01:14:26.557 >> Is that obvious to you by the way? Like 01:14:26.640 --> 01:14:29.677 well so mathematically speaking maybe. 01:14:29.760 --> 01:14:31.677 But like isn't there a kind of 01:14:31.760 --> 01:14:33.917 gravitational pull towards the truth 01:14:34.000 --> 01:14:37.437 because on on average hopefully the 01:14:37.520 --> 01:14:40.637 truth is well represented in the uh 01:14:40.720 --> 01:14:43.277 training set. No, it's basically a 01:14:43.360 --> 01:14:47.437 struggle against uh the curse of 01:14:47.520 --> 01:14:50.797 dimensionality. So the way you can correct for this is that you fine-tune 01:14:50.880 --> 01:14:55.837 the system by having it produce answers for all kinds of questions that people 01:14:55.920 --> 01:15:00.397 might come up with. >> Mhm. >> And people are people. So they a lot of 01:15:00.480 --> 01:15:03.517 the questions that they have are very similar to each other. So you can 01:15:03.600 --> 01:15:06.397 probably cover you know 80% or whatever 01:15:06.480 --> 01:15:09.917 of questions that people will will ask 01:15:10.000 --> 01:15:13.997 um by you know collecting data and then 01:15:14.080 --> 01:15:18.077 um and then you fine-tune the system to produce good answers for all of those 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 01:15:21.920 --> 01:15:24.957 lot of capacity to to learn. Um but then 01:15:25.040 --> 01:15:29.597 there is you know the enormous set of 01:15:29.680 --> 01:15:33.997 prompts that you have not covered during training and that set is enormous like 01:15:34.080 --> 01:15:35.917 within the set of all possible prompts 01:15:36.000 --> 01:15:39.037 the proportion of prompts that have been 01:15:39.120 --> 01:15:42.397 uh used for training is absolutely tiny. 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 01:15:47.200 --> 01:15:50.717 system will behave properly on the prompts that has been either trained 01:15:50.800 --> 01:15:52.877 pre-trained or fine-tuned. 01:15:52.960 --> 01:15:55.917 [snorts] Um but then there is an entire 01:15:56.000 --> 01:15:58.877 space of things that it cannot possibly 01:15:58.960 --> 01:16:00.477 have been trained on because it's just 01:16:00.560 --> 01:16:03.997 the the number is gigantic. So um so 01:16:04.080 --> 01:16:07.037 whatever training the system 01:16:07.120 --> 01:16:09.757 uh has been subject to to produce 01:16:09.840 --> 01:16:12.317 appropriate tensors, you can break it by 01:16:12.400 --> 01:16:17.357 finding out a prompt that will be outside of the the the set of prompts 01:16:17.440 --> 01:16:21.757 has been trained on or things that are similar and then it will just spew 01:16:21.840 --> 01:16:23.597 complete nonsense. Do when you say 01:16:23.680 --> 01:16:26.397 prompt, do you mean that exact prompt or 01:16:26.480 --> 01:16:29.357 do you mean a prompt that's like in many 01:16:29.440 --> 01:16:32.157 parts very different than like is it 01:16:32.240 --> 01:16:36.077 that easy to ask a question or to say a 01:16:36.160 --> 01:16:38.477 thing that hasn't been said before on 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 01:16:43.680 --> 01:16:47.117 essentially a random sequence of characters in a prompt and that's enough 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 01:16:52.000 --> 01:16:56.077 answer something completely different than it would have answered without this. Mhm. 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 01:16:59.680 --> 01:17:01.997 its uh of its conditioning right 01:17:02.080 --> 01:17:03.277 >> so that that's a very clear 01:17:03.360 --> 01:17:06.717 demonstration of it but of course 01:17:06.800 --> 01:17:09.517 uh you know that's uh that goes outside 01:17:09.600 --> 01:17:12.397 of what is designed to do right if you 01:17:12.480 --> 01:17:17.597 actually stitch together reasonably grammatical sentences is that is it that 01:17:17.680 --> 01:17:21.757 easy to break it >> yeah some people have done things like 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 01:17:26.080 --> 01:17:30.077 in English and it it produces a perfectly fine answer and then you just 01:17:30.160 --> 01:17:35.277 substitute a few words >> by the same word in another language and 01:17:35.360 --> 01:17:37.117 all of a sudden the answer is complete nonsense. 01:17:37.200 --> 01:17:39.277 >> Yeah. So so I guess what I'm saying is 01:17:39.360 --> 01:17:42.237 like which fraction of prompts that 01:17:42.320 --> 01:17:45.677 humans are likely to generate are going 01:17:45.760 --> 01:17:49.277 to break the system. >> So the the problem is that there is a 01:17:49.360 --> 01:17:53.597 long tail. >> Yes. Uh this is a an issue that a lot of 01:17:53.680 --> 01:17:57.677 people have realized you know in social networks and stuff like that which is uh 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 01:18:02.240 --> 01:18:04.637 fine-tune the system for the 80% or 01:18:04.720 --> 01:18:07.357 whatever of uh of the things that most 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 01:18:12.080 --> 01:18:15.437 going to be able to fetune the system for all the conditions and in the end 01:18:15.520 --> 01:18:19.757 the system ends up being kind of a giant lookup table right essentially which is 01:18:19.840 --> 01:18:22.877 not really what you want you want systems that can reason certainly that 01:18:22.960 --> 01:18:27.197 can plan. So the type of reasoning that takes place in LLM is very very 01:18:27.280 --> 01:18:29.357 primitive and the reason you can tell is 01:18:29.440 --> 01:18:31.597 primitive is because the amount of 01:18:31.680 --> 01:18:34.397 computation that is spent per token 01:18:34.480 --> 01:18:36.717 produced is constant. 01:18:36.800 --> 01:18:38.957 So if you ask a question and that 01:18:39.040 --> 01:18:43.917 question has an answer in a given number of token the amount of computation 01:18:44.000 --> 01:18:45.837 devoted to computing that answer can be 01:18:45.920 --> 01:18:48.957 exactly estimated. It's like you know 01:18:49.040 --> 01:18:51.677 it's h it's the the size of the 01:18:51.760 --> 01:18:53.917 prediction network you know with it 36 01:18:54.000 --> 01:18:56.637 layers or 92 layers or whatever it is uh 01:18:56.720 --> 01:18:58.877 multiply by number of tokens that's it 01:18:58.960 --> 01:19:01.997 and so essentially it doesn't matter if 01:19:02.080 --> 01:19:06.317 the question being asked is is 01:19:06.400 --> 01:19:09.437 simple to answer complicated to answer 01:19:09.520 --> 01:19:14.397 impossible to answer because it's a decidable or something um the amount of 01:19:14.480 --> 01:19:18.077 computation the system will be able to devote to that to the answer is constant 01:19:18.160 --> 01:19:20.157 or is proportional to number of token 01:19:20.240 --> 01:19:22.477 produced in the answer. Right? This is 01:19:22.560 --> 01:19:25.517 not the way we work. The way we reason 01:19:25.600 --> 01:19:29.117 is that when we're faced with a complex 01:19:29.200 --> 01:19:33.997 problem or complex question, we spend more time trying to solve it and answer 01:19:34.080 --> 01:19:36.477 it, right? >> Because it's more difficult. 01:19:36.560 --> 01:19:38.957 >> There's a prediction element. There's a 01:19:39.040 --> 01:19:43.357 iterative element where you're like uh 01:19:43.440 --> 01:19:47.597 adjusting your understanding of a thing by going o over and over and over. 01:19:47.680 --> 01:19:51.357 There's a hierarchical element. So on does this mean it's a fundamental flaw 01:19:51.440 --> 01:19:54.477 of LLMs or does it mean that [laughter] 01:19:54.560 --> 01:19:57.517 there's more part to that question? 01:19:57.600 --> 01:19:59.817 Now you're just behaving like an LLM 01:19:59.900 --> 01:20:03.117 [laughter] immediately answering no that 01:20:03.200 --> 01:20:06.877 that it's just the low-level world model 01:20:06.960 --> 01:20:10.157 on top of which we can then build some 01:20:10.240 --> 01:20:11.837 of these kinds of mechanisms like you 01:20:11.920 --> 01:20:15.837 said persistent long-term memory or 01:20:15.920 --> 01:20:18.957 uh reasoning so on but we need that 01:20:19.040 --> 01:20:21.437 world model that comes from language is 01:20:21.520 --> 01:20:24.157 it maybe it is not so difficult to build 01:20:24.240 --> 01:20:27.197 this kind of uh reasoning system on top 01:20:27.280 --> 01:20:29.597 of a well constructed world model. 01:20:29.680 --> 01:20:32.077 >> Okay. Whether it's difficult or not, the 01:20:32.160 --> 01:20:36.637 near future will will say because a lot of people are working on reasoning and 01:20:36.720 --> 01:20:39.837 planning abilities for for dialogue systems. 01:20:39.920 --> 01:20:42.237 >> Um I mean if we're even if we restrict 01:20:42.320 --> 01:20:44.477 ourselves to language 01:20:44.560 --> 01:20:47.117 uh just having the ability to plan your 01:20:47.200 --> 01:20:51.837 answer before you answer >> uh in terms that are not necessarily 01:20:51.920 --> 01:20:55.357 linked with the language you're going to use to produce the answer, right? So the 01:20:55.440 --> 01:20:58.477 this idea of this mental model that allows you to plan what you're going to 01:20:58.560 --> 01:21:00.477 say before you say it. >> Mhm. 01:21:00.560 --> 01:21:05.117 >> Um that is very important. I think there's 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 01:21:08.640 --> 01:21:11.757 this capability but the blueprint of 01:21:11.840 --> 01:21:13.197 those systems will be extremely 01:21:13.280 --> 01:21:19.277 different from autogressive LLMs. So um 01:21:19.360 --> 01:21:23.277 it's the same difference as the difference between what psychologist 01:21:23.360 --> 01:21:27.117 call system one and system two in humans, right? So system one is the type 01:21:27.200 --> 01:21:29.037 of task that you can accomplish without 01:21:29.120 --> 01:21:30.957 like deliberately consciously think 01:21:31.040 --> 01:21:36.077 about how you do them. You just do them [snorts] you've done them enough that 01:21:36.160 --> 01:21:39.437 you can just do it subconsciously, right? Without thinking about them. If 01:21:39.520 --> 01:21:43.677 you're an experienced driver, you can drive without really thinking about it 01:21:43.760 --> 01:21:45.117 and you can talk to someone at the same 01:21:45.200 --> 01:21:47.357 time or listen to the radio, right? 01:21:47.440 --> 01:21:52.557 >> Um, if you are a very experienced chess player, you can play against a 01:21:52.640 --> 01:21:55.597 non-experienced chess player without really thinking either. You just 01:21:55.680 --> 01:21:57.197 recognize the pattern and you play. 01:21:57.280 --> 01:22:01.277 >> Mhm. >> Right. That's system one. Um, so all the 01:22:01.360 --> 01:22:05.517 things that you do instinctively without really having to deliberately plan and 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 01:22:08.400 --> 01:22:13.677 you are a not so experienced uh chess player or you are experienced but you 01:22:13.760 --> 01:22:16.797 play against another experienced chess player, you think about all kinds of 01:22:16.880 --> 01:22:21.357 options, right? You you think about it for a while, right? And you you you're 01:22:21.440 --> 01:22:22.957 much better if you have time to think 01:22:23.040 --> 01:22:24.557 about it than you are if you are if you 01:22:24.640 --> 01:22:28.877 play blitz uh with limited time. So and 01:22:28.960 --> 01:22:32.077 um so this type of deliberate uh 01:22:32.160 --> 01:22:34.717 planning which uses your internal world 01:22:34.800 --> 01:22:39.357 model um that's system two. This is what LM currently cannot do. 01:22:39.440 --> 01:22:40.957 >> So how how do we get them to do this 01:22:41.040 --> 01:22:43.357 right? H how do we build a system that 01:22:43.440 --> 01:22:47.357 can do this kind of planning that or 01:22:47.440 --> 01:22:50.317 reasoning that devotes more resources to 01:22:50.400 --> 01:22:54.717 complex problems than to simple problems and it's not going to be auto 01:22:54.800 --> 01:22:56.797 reggressive prediction of tokens. is 01:22:56.880 --> 01:23:03.217 going to be more something akin to inference of latent variables in 01:23:03.300 --> 01:23:06.637 [snorts] um you know what used to be called 01:23:06.720 --> 01:23:08.957 probabistic models or graphical models 01:23:09.040 --> 01:23:10.877 and things of that type. So basically 01:23:10.960 --> 01:23:16.637 the principle is like this you you know the prompt is like observed uh 01:23:16.720 --> 01:23:17.517 variables. 01:23:17.600 --> 01:23:22.317 >> Mhm. And what your what the model does 01:23:22.400 --> 01:23:26.237 is that it's basically a measure of it 01:23:26.320 --> 01:23:30.717 can measure to what extent an answer is a good answer for a prompt. 01:23:30.800 --> 01:23:33.437 >> Mhm. >> Okay. So think of it as some gigantic 01:23:33.520 --> 01:23:35.277 neural net but it's got only one output 01:23:35.360 --> 01:23:38.157 and that output is a scalar number which 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 01:23:43.440 --> 01:23:46.957 number if the answer is not a good answer for the question. Imagine you had 01:23:47.040 --> 01:23:49.517 this model. If you had such a model, you 01:23:49.600 --> 01:23:51.597 could use it to produce good answers. 01:23:51.680 --> 01:23:55.917 The way you would do is, you know, produce the prompt and then 01:23:56.000 --> 01:23:59.837 search through the space of possible answers for one that minimizes that 01:23:59.920 --> 01:24:01.357 number. >> Mhm. [clears throat] 01:24:01.440 --> 01:24:04.397 >> Um that's called an energy based model. 01:24:04.480 --> 01:24:07.437 >> But that energy based model would need 01:24:07.520 --> 01:24:11.277 the the model constructed by the LLM. 01:24:11.360 --> 01:24:13.677 Well, so uh really what you need to do 01:24:13.760 --> 01:24:16.797 would be to not uh search over possible 01:24:16.880 --> 01:24:19.357 strings of text that minimize that uh 01:24:19.440 --> 01:24:24.237 energy. But what you would do is do this in abstract representation space. So in 01:24:24.320 --> 01:24:26.237 in sort of the space of abstract 01:24:26.320 --> 01:24:29.197 thoughts, you would elaborate a thought, 01:24:29.280 --> 01:24:32.877 right? Using this process of minimizing 01:24:32.960 --> 01:24:35.357 the output of your your model, okay, 01:24:35.440 --> 01:24:40.157 which is just a scalar. Um it's an optimization process, right? So now the 01:24:40.240 --> 01:24:44.797 the way the system produces its answer is through optimization 01:24:44.880 --> 01:24:49.517 um by you know minimizing an objective function basically right uh and this is 01:24:49.600 --> 01:24:52.477 we're talking about inference we're not talking about training right the system 01:24:52.560 --> 01:24:56.157 has been trained already so now we have an abstract representation of the 01:24:56.240 --> 01:24:58.397 thought of the answer representation of 01:24:58.480 --> 01:25:01.117 the answer we feed that to basically an 01:25:01.200 --> 01:25:03.837 autographive decoder uh which can be 01:25:03.920 --> 01:25:06.237 very simple that turns this into a text 01:25:06.320 --> 01:25:08.877 that expresses this thought okay so that 01:25:08.960 --> 01:25:11.037 That in my opinion is the blueprint of 01:25:11.120 --> 01:25:14.797 future dialogue systems. Um they will 01:25:14.880 --> 01:25:16.637 think about their answer, plan their 01:25:16.720 --> 01:25:19.357 answer by optimization before turning it 01:25:19.440 --> 01:25:23.997 into text. Uh and that is turning complete. 01:25:24.080 --> 01:25:28.157 >> Can you explain exactly what the optimization problem there is? Like 01:25:28.240 --> 01:25:30.637 what's the objective function? Just 01:25:30.720 --> 01:25:35.277 linger on it. You you kind of briefly described it, but over what space are 01:25:35.360 --> 01:25:38.717 you optimizing? the space of representations, 01:25:38.800 --> 01:25:39.917 >> those abstract representations. 01:25:40.000 --> 01:25:43.997 >> Abstract represent. So you have an abstract representation inside the 01:25:44.080 --> 01:25:47.197 system. You have a prompt. The prompt goes to an encoder, produces a 01:25:47.280 --> 01:25:50.317 representation, perhaps goes to a predictor that predicts a representation 01:25:50.400 --> 01:25:52.637 of the answer of the proper answer, but 01:25:52.720 --> 01:25:57.997 that representation may not be a good answer because there might there might 01:25:58.080 --> 01:25:59.597 be some complicated reasoning you need 01:25:59.680 --> 01:26:02.957 to do, right? So um so then you have 01:26:03.040 --> 01:26:05.517 another process that takes the 01:26:05.600 --> 01:26:07.357 representation of the answers and 01:26:07.440 --> 01:26:11.357 modifies it so as to minimize 01:26:11.440 --> 01:26:13.597 uh a cost function that measures to what 01:26:13.680 --> 01:26:15.037 extent the answer is a good answer for 01:26:15.120 --> 01:26:18.637 the question. Now we we sort of [snorts] 01:26:18.720 --> 01:26:21.437 ignore the the fact for I mean the issue 01:26:21.520 --> 01:26:23.437 for a moment of how you train that 01:26:23.520 --> 01:26:26.957 system to measure whether an answer is a 01:26:27.040 --> 01:26:31.597 good answer for for a question. But suppose such a system could be created. 01:26:31.680 --> 01:26:35.117 But what's the process? This kind of search like process 01:26:35.200 --> 01:26:37.117 >> it's a optimization process. You can do 01:26:37.200 --> 01:26:38.957 this if if the entire system is 01:26:39.040 --> 01:26:44.477 differentiable that scalar output is the result of you know running through some 01:26:44.560 --> 01:26:47.757 neural net. >> Mhm. >> Uh running the answer the representation 01:26:47.840 --> 01:26:52.237 of the answer to some neural net then by gradient descent by back proping back 01:26:52.320 --> 01:26:55.997 propagating gradients you can figure out like how to modify the representation of 01:26:56.080 --> 01:26:59.357 the answer so as to minimize that. So that's still a gradient based 01:26:59.440 --> 01:27:03.197 >> it's gradient based inference. So now you have a representation of the answer 01:27:03.280 --> 01:27:06.957 in abstract space. Now you can turn it into text. 01:27:07.040 --> 01:27:10.557 >> Mhm. >> Right. And the cool thing about this is 01:27:10.640 --> 01:27:15.597 that the representation now can be optimized through gradient descent but 01:27:15.680 --> 01:27:20.317 also is independent of the language in which you're going to express the answer. 01:27:20.400 --> 01:27:23.277 >> Right? So you're operating in the abstract representation. I mean this 01:27:23.360 --> 01:27:25.357 goes backs to the joint embedding 01:27:25.440 --> 01:27:27.997 >> that it's better to work in the uh in 01:27:28.080 --> 01:27:31.357 the space of I don't know to romanticize 01:27:31.440 --> 01:27:34.317 the notion like space of concepts versus 01:27:34.400 --> 01:27:38.477 >> the space of concrete >> sensory information 01:27:38.560 --> 01:27:42.637 >> right >> okay but is can can this do something 01:27:42.720 --> 01:27:44.637 like reasoning which is what we're talking about 01:27:44.720 --> 01:27:48.397 >> well not really in a only in a very simple way I mean basically you can 01:27:48.480 --> 01:27:52.717 think of those things as doing the kind of optimization I was I was talking 01:27:52.800 --> 01:27:55.997 about except they optimize in the discrete space which is the space of 01:27:56.080 --> 01:28:00.877 possible sequences of of tokens and they do it they do this optimization in a 01:28:00.960 --> 01:28:04.317 horribly inefficient way which is generate a lot of hypothesis and then 01:28:04.400 --> 01:28:06.077 select the best ones 01:28:06.160 --> 01:28:11.437 >> and that's incredibly wasteful in terms of uh computation 01:28:11.520 --> 01:28:12.717 because you have you run you basically 01:28:12.800 --> 01:28:15.277 have to run your LLM for like every 01:28:15.360 --> 01:28:18.477 possible you know generated sequence um 01:28:18.560 --> 01:28:21.917 and it's incredibly wasteful um so So 01:28:22.000 --> 01:28:26.877 it's much better to do an optimization in continuous space where you can do 01:28:26.960 --> 01:28:29.917 gradient and descent as opposed to like generate tons of things and then select 01:28:30.000 --> 01:28:33.117 the best. You just iteratively refine 01:28:33.200 --> 01:28:36.957 your answer to to go towards the best, right? That's much more efficient. But 01:28:37.040 --> 01:28:40.877 you can only do this in continuous spaces with differentiable functions. 01:28:40.960 --> 01:28:43.277 You're talking about the reasoning like 01:28:43.360 --> 01:28:46.237 ability to think deeply or to reason 01:28:46.320 --> 01:28:47.917 deeply. 01:28:48.000 --> 01:28:50.797 How do you know what 01:28:50.880 --> 01:28:56.717 is an answer uh that's better or worse based on deep 01:28:56.800 --> 01:28:59.277 reasoning? >> Right? So then we're asking the question 01:28:59.360 --> 01:29:03.117 of conceptually how do you train an energy based model? Right? So energy 01:29:03.200 --> 01:29:05.277 based model is a function with a scalar 01:29:05.360 --> 01:29:07.597 output just a number. 01:29:07.680 --> 01:29:10.797 >> You give it two inputs X and Y 01:29:10.880 --> 01:29:12.317 >> and it tells you whether Y is compatible 01:29:12.400 --> 01:29:14.717 with X or not. X you observe. Let's say 01:29:14.800 --> 01:29:16.317 it's a prompt, an image, a video, 01:29:16.400 --> 01:29:20.077 whatever. and y is a proposal for an 01:29:20.160 --> 01:29:23.597 answer, a continuation of video, um you 01:29:23.680 --> 01:29:28.557 know, whatever. And it tells you whether Y is compatible with X. And the way it 01:29:28.640 --> 01:29:31.917 tells you that Y is compatible with X is that the output of that function would 01:29:32.000 --> 01:29:33.997 be zero if Y is compatible with X and 01:29:34.080 --> 01:29:37.677 would be a a positive number non zero if 01:29:37.760 --> 01:29:40.557 Y is not compatible with X. Okay. How do 01:29:40.640 --> 01:29:42.797 you train a system like this at a 01:29:42.880 --> 01:29:46.717 completely general level is you show it 01:29:46.800 --> 01:29:51.357 pairs of X and Y that are compatible a question and the corresponding answer 01:29:51.440 --> 01:29:53.357 and you train the parameters of the big 01:29:53.440 --> 01:29:57.037 neural net inside um to produce zero. 01:29:57.120 --> 01:30:01.597 Mhm. Okay. Now that doesn't completely work because the system might decide 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 01:30:05.440 --> 01:30:08.237 process to make sure that for a a wrong 01:30:08.320 --> 01:30:13.277 Y the energy would be larger than zero and [snorts] there you have two options. 01:30:13.360 --> 01:30:16.877 One is contrastive method. So contrastive method is you show an X and 01:30:16.960 --> 01:30:20.957 a bad Y and you tell the system well that's you 01:30:21.040 --> 01:30:23.997 know give a high energy to this like push up the energy right change the 01:30:24.080 --> 01:30:28.877 weights in the neural net that computes the energy so that it goes up 01:30:28.960 --> 01:30:30.557 u so that's contrasting methods. The 01:30:30.640 --> 01:30:33.197 problem with this is if the space of Y 01:30:33.280 --> 01:30:37.037 is large, the number of such contrastive 01:30:37.120 --> 01:30:41.277 samples you're going to have to show is gigantic. 01:30:41.360 --> 01:30:43.917 But people do this. They they do this 01:30:44.000 --> 01:30:46.397 when you train a system with RHF. 01:30:46.480 --> 01:30:50.717 Basically, what you're training is what's called a reward model, which is 01:30:50.800 --> 01:30:53.197 basically an objective function that 01:30:53.280 --> 01:30:58.237 tells you whether an answer is good or bad. And that's basically exactly what 01:30:58.320 --> 01:31:01.277 what this is. Mhm. >> So we already do this to some extent. 01:31:01.360 --> 01:31:02.717 We're just not using it for inference. 01:31:02.800 --> 01:31:04.637 We're just using it for training. 01:31:04.720 --> 01:31:06.237 >> Um 01:31:06.320 --> 01:31:08.797 uh there is another set of methods which 01:31:08.880 --> 01:31:11.757 are non-contrastive and I prefer those. 01:31:11.840 --> 01:31:13.677 Uh and those non-contrastive method 01:31:13.760 --> 01:31:15.757 basically say 01:31:15.840 --> 01:31:19.277 uh okay the energy function needs to 01:31:19.360 --> 01:31:23.837 have low energy on pairs of XY that are compatible that come from your training 01:31:23.920 --> 01:31:27.117 set. How do you make sure that the energy is 01:31:27.200 --> 01:31:29.917 going to be higher everywhere else? 01:31:30.000 --> 01:31:34.717 And the way you do this is by um having a 01:31:34.800 --> 01:31:36.877 regularizer, a criterion, a term in your 01:31:36.960 --> 01:31:39.357 cost function that basically minimizes 01:31:39.440 --> 01:31:44.477 the volume of space that can take low energy. 01:31:44.560 --> 01:31:47.837 And the precise way to do this is all kinds of different specific ways to do 01:31:47.920 --> 01:31:51.997 this depending on the architecture. But that's the basic principle. So that if 01:31:52.080 --> 01:31:56.957 you push down the energy function for particular regions in the XY space, it 01:31:57.040 --> 01:31:58.957 will automatically go up in other places 01:31:59.040 --> 01:32:01.757 because there's only a limited volume of 01:32:01.840 --> 01:32:04.877 space that can take low energy. Okay, by 01:32:04.960 --> 01:32:06.717 the construction of the system or by the 01:32:06.800 --> 01:32:09.677 regularizer regularizing function. We've 01:32:09.760 --> 01:32:12.237 been talking very generally, but what is 01:32:12.320 --> 01:32:14.717 a good X and a good Y? What is a good 01:32:14.800 --> 01:32:18.877 representation of X and Y? because we've 01:32:18.960 --> 01:32:23.437 been talking about language and if you just take language directly that 01:32:23.520 --> 01:32:27.117 presumably is not good. So there has to be some kind of abstract representation 01:32:27.200 --> 01:32:28.877 of ideas. 01:32:28.960 --> 01:32:30.637 >> Yeah. So you I mean you can do this with 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 01:32:35.920 --> 01:32:40.877 that text. >> Yes. >> Um or X is a question Y is the answer. 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 01:32:43.200 --> 01:32:46.157 do what LLMs are doing. Well, no. It 01:32:46.240 --> 01:32:50.397 depends on how you how the internal structure of the system is built. If the 01:32:50.480 --> 01:32:53.757 if the internal structure of the system is built in such a way that inside of 01:32:53.840 --> 01:32:57.597 the system there is a latent variable let's call it Z 01:32:57.680 --> 01:32:59.917 >> that uh 01:33:00.000 --> 01:33:03.277 you can manipulate so as to minimize the 01:33:03.360 --> 01:33:07.837 output energy then that Z can be viewed as a 01:33:07.920 --> 01:33:11.677 representation of a good answer that you can translate into a Y that is a good 01:33:11.760 --> 01:33:15.357 answer. >> [snorts] >> So this kind of system could be trained 01:33:15.440 --> 01:33:18.717 in a very similar way, >> very similar way, but you have to have 01:33:18.800 --> 01:33:21.357 this way of preventing collapse of of 01:33:21.440 --> 01:33:24.077 ensuring that you know there is high 01:33:24.160 --> 01:33:26.477 energy for things you don't train it on. 01:33:26.560 --> 01:33:30.477 Um and and currently it's it's very 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 01:33:33.680 --> 01:33:36.317 is it is being done is is due to the 01:33:36.400 --> 01:33:38.077 fact that when you give a high 01:33:38.160 --> 01:33:41.997 probability to a word 01:33:42.080 --> 01:33:46.477 automatically you give low probability to other words because you only have a 01:33:46.560 --> 01:33:50.397 finite amount of probability to go around right there to sum to one 01:33:50.480 --> 01:33:52.157 >> so when you minimize the cross entropy 01:33:52.240 --> 01:33:54.877 or whatever when you train the your LLM 01:33:54.960 --> 01:34:00.317 to produce the to predict the next word uh you're increasing the probability 01:34:00.400 --> 01:34:03.037 your system will give to the correct word, but you're also decreasing the 01:34:03.120 --> 01:34:03.917 probability it will give to the 01:34:04.000 --> 01:34:07.597 incorrect words. Now, indirectly that 01:34:07.680 --> 01:34:11.757 gives a low probability to a high probability to sequences of words that 01:34:11.840 --> 01:34:14.557 are good and low probability to sequences of words that are bad. But 01:34:14.640 --> 01:34:18.477 it's very indirect >> and it's not it's not obvious why this 01:34:18.560 --> 01:34:23.117 actually works at all. that uh because you're not doing it on a joint 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 01:34:27.200 --> 01:34:31.757 uh you sort of factoriize that probability in terms of conditional 01:34:31.840 --> 01:34:34.077 probabilities over successive tokens. 01:34:34.160 --> 01:34:36.637 >> So how do you do this for visual data? 01:34:36.720 --> 01:34:38.317 So we've been doing this with OJA 01:34:38.400 --> 01:34:40.877 architectures basically joint embedding. 01:34:40.960 --> 01:34:44.557 So uh there the compatibility between 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 01:34:49.840 --> 01:34:51.357 shifted or transformed version of that 01:34:51.440 --> 01:34:55.117 image or video or masked. Okay. And then 01:34:55.200 --> 01:34:57.997 uh the energy of the system is the 01:34:58.080 --> 01:35:02.157 prediction error of the 01:35:02.240 --> 01:35:06.877 representation uh the the predicted representation of 01:35:06.960 --> 01:35:10.397 the good thing versus the actual representation of the good thing. Right? 01:35:10.480 --> 01:35:13.437 So so you run the corrupted image to the 01:35:13.520 --> 01:35:17.837 system, predict the representation of the the good input uncorrupted and then 01:35:17.920 --> 01:35:19.597 compute the prediction error. That's the 01:35:19.680 --> 01:35:21.757 energy of the system. So the system will 01:35:21.840 --> 01:35:26.477 tell you this is a good 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 01:35:30.080 --> 01:35:34.957 you zero energy if those two things are effectively one of them is a corrupted 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 01:35:38.400 --> 01:35:41.517 completely different >> and hopefully that whole process gives 01:35:41.600 --> 01:35:46.877 you a really nice compressed representation of of reality of visual 01:35:46.960 --> 01:35:49.197 reality >> and we know it does because then we use 01:35:49.280 --> 01:35:52.397 those representations as input to a classification system. 01:35:52.480 --> 01:35:53.997 >> That classification system works really 01:35:54.080 --> 01:35:57.597 nicely. Okay. Well, so to summarize, you 01:35:57.680 --> 01:36:01.677 recommend in a in a in a spicy way that 01:36:01.760 --> 01:36:05.997 only Yun can. You recommend that we abandon generative models in favor of 01:36:06.080 --> 01:36:07.677 joint embedding architectures. 01:36:07.760 --> 01:36:10.237 >> Yes. >> Abandon auto reggressive generation. 01:36:10.320 --> 01:36:14.467 >> Yes. >> Abandon pro, [laughter] this feels like court testimony. 01:36:14.550 --> 01:36:18.317 [snorts] Uh abandon probabilistic models in favor of energy based models as we 01:36:18.400 --> 01:36:20.077 talked about. abandon contrastive 01:36:20.160 --> 01:36:22.957 methods in favor of regularized methods. 01:36:23.040 --> 01:36:25.917 And uh let me ask you about this. 01:36:26.000 --> 01:36:29.197 You've been for a while a critic of reinforcement learning. 01:36:29.280 --> 01:36:32.237 >> Yes. >> So what uh the last recommendation is 01:36:32.320 --> 01:36:35.517 that we abandon RL in favor of model 01:36:35.600 --> 01:36:37.117 predictive control as you were talking 01:36:37.200 --> 01:36:40.077 about and only use RL when planning 01:36:40.160 --> 01:36:44.077 doesn't yield the predicted outcome. And 01:36:44.160 --> 01:36:46.317 uh we use RL in that case to adjust the 01:36:46.400 --> 01:36:48.877 world model or the critic. Yes. 01:36:48.960 --> 01:36:53.197 >> So, uh you mentioned uh RLHF, 01:36:53.280 --> 01:36:57.677 reinforcement learning with human feedback. Uh why do you still hate uh 01:36:57.760 --> 01:36:59.917 reinforcement learning? >> I don't hate reinforcement learning and 01:37:00.000 --> 01:37:03.357 I think I think it should not be uh 01:37:03.440 --> 01:37:07.997 abandoned completely, but I think it its use should be minimized because it's 01:37:08.080 --> 01:37:09.837 incredibly inefficient in terms of 01:37:09.920 --> 01:37:12.717 samples. And so the the proper way to 01:37:12.800 --> 01:37:17.117 train a system is to first have it learn 01:37:17.200 --> 01:37:19.677 uh good representations of the world and 01:37:19.760 --> 01:37:24.237 world models from mostly observation maybe a little bit of interactions 01:37:24.320 --> 01:37:27.357 >> and then steered based on that. If the representation is good then the 01:37:27.440 --> 01:37:29.677 adjustments should be minimal. 01:37:29.760 --> 01:37:31.677 >> Yeah. Now there's two things you can use 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 01:37:35.040 --> 01:37:36.477 of actions to arrive at a particular 01:37:36.560 --> 01:37:41.997 objective. Mhm. >> You don't need R out unless the way you 01:37:42.080 --> 01:37:45.037 measure whether you succeed might be inexact. 01:37:45.120 --> 01:37:47.837 Your idea of, you know, whether you were 01:37:47.920 --> 01:37:51.117 going to fall from your bike [gasps] 01:37:51.200 --> 01:37:55.437 might be wrong or whether the person you're fighting with MMA was going to do 01:37:55.520 --> 01:37:57.357 something and then do something else. 01:37:57.440 --> 01:38:00.957 Um, so there uh 01:38:01.040 --> 01:38:02.237 so there's two ways you can be wrong. 01:38:02.320 --> 01:38:05.917 either your your objective function does 01:38:06.000 --> 01:38:09.757 not reflect the actual objective function you want to optimize or your 01:38:09.840 --> 01:38:14.237 world model is inaccurate right so you didn't you the prediction 01:38:14.320 --> 01:38:18.717 you were making about what was going to happen in the world is inaccurate so if 01:38:18.800 --> 01:38:20.237 you want to adjust your world model 01:38:20.320 --> 01:38:23.837 while you are operating in the world or 01:38:23.920 --> 01:38:26.957 your objective function that is 01:38:27.040 --> 01:38:31.837 basically in the realm of RL this is what RL deal deals with uh to some 01:38:31.920 --> 01:38:33.677 extent Right? So adjust your world model 01:38:33.760 --> 01:38:35.757 and the way to adjust your world model 01:38:35.840 --> 01:38:40.957 even in advance uh is to explore parts of the space where your world model 01:38:41.040 --> 01:38:42.317 where you know that your world model is 01:38:42.400 --> 01:38:44.877 inaccurate. That's called curiosity 01:38:44.960 --> 01:38:48.557 basically or play right when you play 01:38:48.640 --> 01:38:51.197 you kind of explore part of the state 01:38:51.280 --> 01:38:54.157 space that um [snorts] 01:38:54.240 --> 01:38:57.037 you know you don't want to do in for 01:38:57.120 --> 01:38:59.037 real because it might be dangerous but 01:38:59.120 --> 01:39:01.517 uh but you can adjust your world model 01:39:01.600 --> 01:39:04.877 uh without killing yourself basically. 01:39:04.960 --> 01:39:07.677 Um, so that's what you want to use RL 01:39:07.760 --> 01:39:11.997 for when when when it comes time to learning a particular task. You already 01:39:12.080 --> 01:39:14.957 have all the good representations. You already have your world model, but you 01:39:15.040 --> 01:39:18.637 want you need to adjust it for the situation at hand. That's when you use 01:39:18.720 --> 01:39:23.277 RL. >> Why do you think RHF works so well? This 01:39:23.360 --> 01:39:26.637 reinforcement learning with human feedback, why did it have such a 01:39:26.720 --> 01:39:28.957 transformational effect on large 01:39:29.040 --> 01:39:33.837 language models that before? What's had the transformational effect is human 01:39:33.920 --> 01:39:36.797 feedback. >> There is many ways to use it and some of 01:39:36.880 --> 01:39:38.557 it is just purely supervised actually. 01:39:38.640 --> 01:39:40.157 It's not really reinforcement learning. 01:39:40.240 --> 01:39:43.277 >> So it's the a the HF. >> It's the HF. 01:39:43.360 --> 01:39:45.197 >> Uh and then there is various ways to use 01:39:45.280 --> 01:39:49.997 human feedback, right? So you can uh you can ask humans to rate answers uh 01:39:50.080 --> 01:39:52.077 multiple answers that are produced by a 01:39:52.160 --> 01:39:56.557 world model and uh and and then what you 01:39:56.640 --> 01:39:58.557 do is you train an objective function to 01:39:58.640 --> 01:40:02.237 predict that rating and then you can use 01:40:02.320 --> 01:40:06.637 that objective function to predict you know whether an answer is good and you 01:40:06.720 --> 01:40:09.757 can back propagate gradient through this to fine-tune your system so that it only 01:40:09.840 --> 01:40:13.517 produces high highly rated answers. 01:40:13.600 --> 01:40:17.437 Okay, so that's one way. So that's like 01:40:17.520 --> 01:40:22.637 in RL that means uh training what's called a reward model right so something 01:40:22.720 --> 01:40:25.997 that you know basically a small neural net that estimates to what extent an 01:40:26.080 --> 01:40:30.397 answer is good right it's very similar to the objective I was I was talking 01:40:30.480 --> 01:40:33.837 about talking about earlier for planning except now it's not used for planning 01:40:33.920 --> 01:40:37.037 it's it's used for fine-tuning your system 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 01:40:42.160 --> 01:40:45.917 currently it's used to fine-tune the parameters of the them. Now there's 01:40:46.000 --> 01:40:50.477 several ways to do this. Um you know some of some of them are supervised. You 01:40:50.560 --> 01:40:52.717 just you know ask a human person like 01:40:52.800 --> 01:40:55.037 what is a good answer for this right? 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 01:40:59.760 --> 01:41:03.437 ways that those systems are are being adjusted. 01:41:03.520 --> 01:41:05.837 Now, a lot of people have been very 01:41:05.920 --> 01:41:08.397 critical of the recently released 01:41:08.480 --> 01:41:11.757 Google's Gemini 1.5 01:41:11.840 --> 01:41:14.237 for essentially, in my words, I could 01:41:14.320 --> 01:41:17.357 say super woke. Woke in the uh negative 01:41:17.440 --> 01:41:19.757 connotation of that word. Uh there is 01:41:19.840 --> 01:41:22.397 some almost hilariously absurd things 01:41:22.480 --> 01:41:25.517 that it does like it modifies history uh 01:41:25.600 --> 01:41:29.117 like generating images of a um black 01:41:29.200 --> 01:41:33.117 George Washington or um perhaps more 01:41:33.200 --> 01:41:34.797 seriously something that you commented 01:41:34.880 --> 01:41:39.277 on Twitter which is refusing to comment 01:41:39.360 --> 01:41:42.877 on or generate images of u or even 01:41:42.960 --> 01:41:46.797 descriptions of Tianaan Square or the 01:41:46.880 --> 01:41:49.197 the tank man. one of the most sort of 01:41:49.280 --> 01:41:54.237 legendary protest images in history. Of 01:41:54.320 --> 01:41:58.077 course, these images are highly censored 01:41:58.160 --> 01:42:01.037 by the Chinese government and therefore 01:42:01.120 --> 01:42:02.717 everybody started asking questions of 01:42:02.800 --> 01:42:06.157 what is the process of uh 01:42:06.240 --> 01:42:10.317 designing these LLMs, what is what is what is the role of censorship in these 01:42:10.400 --> 01:42:11.757 all that kind of stuff. 01:42:11.840 --> 01:42:15.357 >> So you uh commented on Twitter saying 01:42:15.440 --> 01:42:18.397 that open source is the answer. Yeah, 01:42:18.480 --> 01:42:22.317 >> essentially. So, um, can you explain? 01:42:22.400 --> 01:42:25.997 >> I I actually made that comment on just about every social network I can and 01:42:26.080 --> 01:42:29.037 I've [laughter] I I've uh I've made that 01:42:29.120 --> 01:42:31.117 point multiple times in in various 01:42:31.200 --> 01:42:34.477 forums. Um, uh, here's my my point of 01:42:34.560 --> 01:42:37.757 view on this. Uh people can complain 01:42:37.840 --> 01:42:40.797 that AI systems are biased and they 01:42:40.880 --> 01:42:43.917 generally are biased by the distribution 01:42:44.000 --> 01:42:45.917 of the training data that they've been 01:42:46.000 --> 01:42:49.677 trained on. Um that reflects biases in 01:42:49.760 --> 01:42:52.637 society. Um 01:42:52.720 --> 01:42:55.997 and that is potentially offensive to 01:42:56.080 --> 01:43:00.637 some people or potentially not. and and 01:43:00.720 --> 01:43:03.917 some techniques to debias then become 01:43:04.000 --> 01:43:07.037 offensive to some people 01:43:07.120 --> 01:43:11.997 um because of you know historical uh 01:43:12.080 --> 01:43:14.637 incorrectness and things like that 01:43:14.720 --> 01:43:18.317 >> um and so you can ask the question you 01:43:18.400 --> 01:43:22.397 can ask two questions the first question is is it possible to produce an AI 01:43:22.480 --> 01:43:24.157 system that is not biased and the answer 01:43:24.240 --> 01:43:26.637 is absolutely not and it's not because 01:43:26.720 --> 01:43:29.837 of technological 01:43:29.920 --> 01:43:34.397 challenges although there are technological challenges to that it's 01:43:34.480 --> 01:43:39.357 because bias is in the eye of the beholder 01:43:39.440 --> 01:43:41.037 um different people may have different 01:43:41.120 --> 01:43:44.397 ideas about what constitutes bias um you 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] 01:43:49.040 --> 01:43:51.277 indisputable but there are a lot of 01:43:51.360 --> 01:43:56.237 opinions or or things that can be expressed in different ways and so you 01:43:56.320 --> 01:43:57.917 cannot have an unbiased system that's 01:43:58.000 --> 01:44:00.317 just an impossib ility. 01:44:00.400 --> 01:44:03.997 Um, and so what's the 01:44:04.080 --> 01:44:05.917 what's the answer to this? And the the 01:44:06.000 --> 01:44:09.197 answer is the same answer that we found 01:44:09.280 --> 01:44:13.517 in liberal democracy about the press. 01:44:13.600 --> 01:44:16.317 The press needs to be free and uh 01:44:16.400 --> 01:44:18.077 diverse. 01:44:18.160 --> 01:44:21.117 We have free speech for a good reason is 01:44:21.200 --> 01:44:24.077 because uh we don't want all of our 01:44:24.160 --> 01:44:27.757 information to be uh to come from a 01:44:27.840 --> 01:44:30.317 unique source. Um because that's 01:44:30.400 --> 01:44:32.717 opposite to the whole idea of democracy 01:44:32.800 --> 01:44:36.397 and uh you know progress of ideas and 01:44:36.480 --> 01:44:38.797 even science right in in science people 01:44:38.880 --> 01:44:43.677 have to argue for different opinions and and science makes progress when people 01:44:43.760 --> 01:44:47.437 disagree and they come up with an answer and you know a consensus forms right and 01:44:47.520 --> 01:44:49.917 it's true in all democracies around the 01:44:50.000 --> 01:44:53.597 world. So there is a a future 01:44:53.680 --> 01:44:56.797 which is already happening where every 01:44:56.880 --> 01:44:59.277 single one of our interaction with the 01:44:59.360 --> 01:45:05.017 digital world will be mediated by AI AI systems AI assistants right 01:45:05.100 --> 01:45:07.677 [snorts and gasps] we're going to have smart glasses you 01:45:07.760 --> 01:45:12.957 can already buy them from meta the Rayban Meta where um you know you can 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 01:45:17.200 --> 01:45:19.437 question you have or you can be looking 01:45:19.520 --> 01:45:22.477 at a monument and there is a camera in 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 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 01:45:31.200 --> 01:45:34.477 a foreign language and I think we'll translate it for you or we can do 01:45:34.560 --> 01:45:39.117 realtime translation if you speak different languages so a lot of our 01:45:39.200 --> 01:45:42.797 interactions with the digital world are going to be mediated by those systems in 01:45:42.880 --> 01:45:48.157 the near future um you know increasingly the 01:45:48.240 --> 01:45:51.037 search engines that we're going to are not going to be search engines. They're 01:45:51.120 --> 01:45:54.157 going to be uh dialog systems 01:45:54.240 --> 01:45:57.357 >> that will just ask a question and it 01:45:57.440 --> 01:46:02.317 will answer and then point you to perhaps appropriate reference for it. 01:46:02.400 --> 01:46:04.077 But here's the thing. We cannot afford 01:46:04.160 --> 01:46:06.317 those systems to come from a handful of 01:46:06.400 --> 01:46:09.997 companies on the west coast of the US 01:46:10.080 --> 01:46:14.797 because those systems will constitute the repository of all human knowledge 01:46:14.880 --> 01:46:18.797 and we cannot have that be controlled by a 01:46:18.880 --> 01:46:22.877 small number of people right it has to be diverse for the same reason the press 01:46:22.960 --> 01:46:24.957 has to be diverse. 01:46:25.040 --> 01:46:27.597 So how do we get a diverse set of AI 01:46:27.680 --> 01:46:30.477 assistants? [snorts] um it's very 01:46:30.560 --> 01:46:34.957 expensive and difficult to train a base model right a based LLM at the moment 01:46:35.040 --> 01:46:37.837 you know in the future it might be something different but at the moment 01:46:37.920 --> 01:46:41.597 that's an LLM uh so only a few companies can do this 01:46:41.680 --> 01:46:45.037 properly and 01:46:45.120 --> 01:46:47.757 if some of those subsystems are open 01:46:47.840 --> 01:46:50.477 source anybody can use them anybody can 01:46:50.560 --> 01:46:53.677 fine-tune them um if we put in place 01:46:53.760 --> 01:46:57.597 some systems that allows any group of 01:46:57.680 --> 01:46:59.037 people 01:46:59.120 --> 01:47:02.877 whether they are um individual citizens, 01:47:02.960 --> 01:47:05.117 groups of citizens, 01:47:05.200 --> 01:47:08.237 government organizations, NOS's, 01:47:08.320 --> 01:47:12.237 um companies, whatever to take those 01:47:12.320 --> 01:47:17.277 open source u systems, AI systems, and fine-tune 01:47:17.360 --> 01:47:18.877 them for their own purpose on their own 01:47:18.960 --> 01:47:21.437 data. Then we're going to have a very 01:47:21.520 --> 01:47:24.477 large diversity of uh different AI 01:47:24.560 --> 01:47:29.357 systems that are specialized for all of those things, right? So I tell you I 01:47:29.440 --> 01:47:32.877 talked to the French government quite a bit and the French government will not 01:47:32.960 --> 01:47:36.237 accept that the digital diet of all 01:47:36.320 --> 01:47:40.717 their citizen be controlled by three companies on the west coast of the US. 01:47:40.800 --> 01:47:44.877 That's just not acceptable. It's a danger to democracy regardless of how 01:47:44.960 --> 01:47:46.797 well-intentioned those companies are. 01:47:46.880 --> 01:47:51.117 Right. Um and so u and it's also a 01:47:51.200 --> 01:47:56.077 danger to local culture to values to 01:47:56.160 --> 01:48:01.277 language right I was talking with um uh 01:48:01.360 --> 01:48:05.837 the uh founder of Infosys in India um 01:48:05.920 --> 01:48:09.357 he's funding a project to fine-tune Latu 01:48:09.440 --> 01:48:11.437 the open source model produced by by 01:48:11.520 --> 01:48:14.877 Meta so that Latu speaks all 22 official 01:48:14.960 --> 01:48:19.917 languages in India it's very important for people in India. I was talking to a 01:48:20.000 --> 01:48:24.077 former colleague of mine Mustafa Asis who used to be a scientist at fair and 01:48:24.160 --> 01:48:28.077 then moved back to Africa created a research lab for Google in Africa and 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 01:48:32.640 --> 01:48:36.797 have LLM that speak the local languages in Sineagal so that people can have 01:48:36.880 --> 01:48:40.397 access to medical information because they don't have access to doctors. It's 01:48:40.480 --> 01:48:42.397 a very small number of doctors per per 01:48:42.480 --> 01:48:47.037 capita in the in Sineagal. Um I mean you 01:48:47.120 --> 01:48:51.277 can't have any of this unless you have open source platforms. So with open 01:48:51.360 --> 01:48:55.197 source platforms you can have AI systems that are not only diverse in terms of 01:48:55.280 --> 01:48:56.957 political opinions or things of that 01:48:57.040 --> 01:49:01.757 type but in terms of uh uh language, 01:49:01.840 --> 01:49:05.117 culture, value systems, political 01:49:05.200 --> 01:49:08.237 opinions um 01:49:08.320 --> 01:49:11.597 technical abilities in various domains 01:49:11.680 --> 01:49:15.757 and you can have an industry an ecosystem of companies that fine-tune 01:49:15.840 --> 01:49:20.237 those open source systems for vertical applications in industry. Right? you you 01:49:20.320 --> 01:49:23.997 have I don't know a publisher has thousands of books and they want to 01:49:24.080 --> 01:49:28.477 build a system that allows a customer to just ask a question about any but the 01:49:28.560 --> 01:49:31.437 content of any of their books you [snorts] need to train on their 01:49:31.520 --> 01:49:34.077 proprietary data right um you have a 01:49:34.160 --> 01:49:38.957 company we have one within meta it's called metamate and it's basically an 01:49:39.040 --> 01:49:40.557 LLM that can answer any question about 01:49:40.640 --> 01:49:43.757 internal uh stuff about about the 01:49:43.840 --> 01:49:46.797 company u very useful a lot of companies 01:49:46.880 --> 01:49:51.037 want this right a lot of companies want is not just for their employees but also 01:49:51.120 --> 01:49:55.117 for their customers to take care of their customers. So the only way you're 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 01:49:59.280 --> 01:50:02.077 are not uniquely biased is if you have 01:50:02.160 --> 01:50:05.037 open source platforms on top of which uh 01:50:05.120 --> 01:50:08.957 any group can u build specialized 01:50:09.040 --> 01:50:13.837 systems. So the the direction of of 01:50:13.920 --> 01:50:16.637 inevitable direction of history is that 01:50:16.720 --> 01:50:19.117 the vast majority of AI systems will be 01:50:19.200 --> 01:50:21.197 built on top of open source platforms. 01:50:21.280 --> 01:50:23.837 >> So that's a beautiful vision. So meaning 01:50:23.920 --> 01:50:25.917 like 01:50:26.000 --> 01:50:30.637 a company like Meta or Google or so on 01:50:30.720 --> 01:50:32.637 should take only minimal fine-tuning 01:50:32.720 --> 01:50:35.757 steps after the building the foundation 01:50:35.840 --> 01:50:41.697 pre-trained model as few steps as possible >> basically [gasps] 01:50:41.780 --> 01:50:44.557 [sighs] >> can Meta afford to do that no 01:50:44.640 --> 01:50:46.637 >> so I don't know if you you know this but 01:50:46.720 --> 01:50:48.077 companies are supposed to make money 01:50:48.160 --> 01:50:52.157 somehow and uh open source is is is like 01:50:52.240 --> 01:50:54.637 giving away. I don't know, Mark made a 01:50:54.720 --> 01:50:57.517 video, Mark Zuckerberg, 01:50:57.600 --> 01:51:00.557 uh, very sexy video talking about 01:51:00.640 --> 01:51:04.957 350,000 Nvidia H100s. 01:51:05.040 --> 01:51:08.157 >> Yeah, >> the the math of that is just for the 01:51:08.240 --> 01:51:11.437 GPUs, that's a 100 billion 01:51:11.520 --> 01:51:14.237 um, plus the infrastructure for training 01:51:14.320 --> 01:51:19.037 everything. So, I'm no business guy, but 01:51:19.120 --> 01:51:22.317 how do you make money on that? So the vision you paint is a really powerful 01:51:22.400 --> 01:51:25.437 one. But how is it possible to make money? 01:51:25.520 --> 01:51:29.437 >> Okay, so you have several business models, right? 01:51:29.520 --> 01:51:31.917 >> The business model that uh Meta is built 01:51:32.000 --> 01:51:35.037 around is 01:51:35.120 --> 01:51:39.677 um you offer a service and the the 01:51:39.760 --> 01:51:44.797 financing of that service is uh either through ads or through business 01:51:44.880 --> 01:51:46.477 customers. So for example, if you have 01:51:46.560 --> 01:51:49.917 an LLM that uh you know can help a mom 01:51:50.000 --> 01:51:53.997 and pop pizza place um by you know 01:51:54.080 --> 01:51:58.477 talking to the customers through WhatsApp and so the customers can just 01:51:58.560 --> 01:52:01.997 order a pizza and the system will just you know ask them like what topping do 01:52:02.080 --> 01:52:04.557 you want or what size blah blah blah [gasps] 01:52:04.640 --> 01:52:07.197 um the business will pay for that. Okay, 01:52:07.280 --> 01:52:11.960 that's a model. Um, 01:52:12.160 --> 01:52:16.157 and otherwise, you know, if it's a system that is on the more kind of 01:52:16.240 --> 01:52:20.477 classical services, it can be uh ad supported or, you know, there's several 01:52:20.560 --> 01:52:23.037 models. But the point is 01:52:23.120 --> 01:52:26.477 uh if you have a big enough um potential 01:52:26.560 --> 01:52:30.157 customer base and you need to build that system 01:52:30.240 --> 01:52:34.317 anyway for them, it doesn't hurt you to actually 01:52:34.400 --> 01:52:36.477 distribute it in open source. Again, I'm 01:52:36.560 --> 01:52:39.517 no business guy, but if you release the 01:52:39.600 --> 01:52:42.077 open source model, then other people can 01:52:42.160 --> 01:52:45.277 do the same kind of task and compete on 01:52:45.360 --> 01:52:48.637 it. Basically, provide fine-tuned models 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 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 01:52:57.440 --> 01:52:58.397 job of it. 01:52:58.480 --> 01:53:01.917 >> Well, no. the the bet is is more we have 01:53:02.000 --> 01:53:05.837 we already have a huge user base and customer base. 01:53:05.920 --> 01:53:07.677 >> Ah right. >> Right. So it's going to be useful to 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 01:53:12.560 --> 01:53:14.077 revenue from this. 01:53:14.160 --> 01:53:18.237 >> Uh and it doesn't hurt that you know we 01:53:18.320 --> 01:53:20.557 provide that system 01:53:20.640 --> 01:53:22.317 or the base the base model right the 01:53:22.400 --> 01:53:25.677 foundation model uh in open source for 01:53:25.760 --> 01:53:29.677 others to build applications on top of it too. If those applications turn out 01:53:29.760 --> 01:53:33.117 to be useful for our customers, we can just buy it from them. 01:53:33.200 --> 01:53:37.677 >> Um, uh, it could be that they will improve the platform. In fact, we see 01:53:37.760 --> 01:53:39.997 this already. Um, I mean there is, you 01:53:40.080 --> 01:53:45.197 know, literally millions of downloads of Latu and thousands of people who have, 01:53:45.280 --> 01:53:46.957 you know, provided ideas about how to 01:53:47.040 --> 01:53:49.677 make it better. Um, so, you know, this 01:53:49.760 --> 01:53:52.397 this clearly accelerates progress to 01:53:52.480 --> 01:53:55.757 make the system available to a sort of a 01:53:55.840 --> 01:54:00.077 a wide community of people. and and there's literally thousands of 01:54:00.160 --> 01:54:01.597 businesses who are building applications 01:54:01.680 --> 01:54:06.397 with it. So um so 01:54:06.480 --> 01:54:10.397 our ability to meta's ability to derive 01:54:10.480 --> 01:54:12.637 revenue from this technology is not 01:54:12.720 --> 01:54:14.317 impaired 01:54:14.400 --> 01:54:17.757 uh by the distribution of it of base 01:54:17.840 --> 01:54:22.077 models in open source. The fundamental criticism that Gemini is getting is that 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 01:54:26.720 --> 01:54:28.957 the east coast where I would suppose 01:54:29.040 --> 01:54:32.237 Meta AI headquarters would be. 01:54:32.320 --> 01:54:34.957 So there uh strong words about the West 01:54:35.040 --> 01:54:37.837 Coast, but uh I guess the issue that 01:54:37.920 --> 01:54:40.877 happens is I think it's fair to say that 01:54:40.960 --> 01:54:44.717 most tech people have a political 01:54:44.800 --> 01:54:49.597 affiliation with the left wing. They're they lean left. And so the problem that 01:54:49.680 --> 01:54:53.837 people are criticizing Gemini with is that there's in that debiasing process 01:54:53.920 --> 01:54:58.157 that you mentioned that their ideological 01:54:58.240 --> 01:55:00.317 lean 01:55:00.400 --> 01:55:02.397 becomes obvious. 01:55:02.480 --> 01:55:07.677 Uh is this something that could be escaped? 01:55:07.760 --> 01:55:11.197 You're saying open source is the only way? Have have you witnessed this kind 01:55:11.280 --> 01:55:16.157 of ideological lean that makes engineering difficult? No, I don't think 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 01:55:20.000 --> 01:55:22.477 people designing those systems. It has 01:55:22.560 --> 01:55:26.637 to do with the uh acceptability or 01:55:26.720 --> 01:55:29.677 political leanings of the their customer 01:55:29.760 --> 01:55:32.397 base or audience, right? So, a big 01:55:32.480 --> 01:55:35.357 company cannot afford to offend too many 01:55:35.440 --> 01:55:39.117 people. So, they're going to make sure 01:55:39.200 --> 01:55:41.597 that whatever product they put out is 01:55:41.680 --> 01:55:45.677 safe, whatever that means. and 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 01:55:51.120 --> 01:55:54.237 impossible to do it properly for everyone. You're not going to satisfy 01:55:54.320 --> 01:55:56.557 everyone. So that's what I said before. 01:55:56.640 --> 01:55:59.997 You cannot have a system that is unbiased is perceived as unbiased by 01:56:00.080 --> 01:56:03.037 everyone. It's going to be, you know, 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 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 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 01:56:15.120 --> 01:56:18.397 system perhaps a little too far in one direction, it's going to be non-factual, 01:56:18.480 --> 01:56:21.117 right? You're going to have, you know, 01:56:21.200 --> 01:56:24.237 you know, black Nazi soldiers in 01:56:24.320 --> 01:56:25.837 >> Yeah. We should we should mention image 01:56:25.920 --> 01:56:29.597 generation of of uh black Nazi soldiers, 01:56:29.680 --> 01:56:31.677 which is not factually accurate, 01:56:31.760 --> 01:56:34.157 >> right? And can be offensive for for some 01:56:34.240 --> 01:56:39.437 people as well, right? So uh so you know it's going to be impossible 01:56:39.520 --> 01:56:43.037 to kind of produce systems that are unbiased for everyone. So the only 01:56:43.120 --> 01:56:45.997 solution that I see is diversity and 01:56:46.080 --> 01:56:50.717 diversity the full meaning of that word diversity of in every possible way. 01:56:50.800 --> 01:56:51.197 >> Yeah. 01:56:51.280 --> 01:56:56.877 >> Um Mark Andre just tweeted today. 01:56:56.960 --> 01:56:59.837 Let me do a TLDDR. The conclusion is 01:56:59.920 --> 01:57:02.557 only startups and open source can avoid 01:57:02.640 --> 01:57:04.317 the issue that he's highlighting with 01:57:04.400 --> 01:57:07.357 big tech. He's asking can big tech 01:57:07.440 --> 01:57:10.317 actually field generative AI products. 01:57:10.400 --> 01:57:14.717 One ever escalating demands from internal activists, employee mobs, 01:57:14.800 --> 01:57:17.197 crazed executives, broken boards, 01:57:17.280 --> 01:57:19.117 pressure groups, extremist regulators, 01:57:19.200 --> 01:57:21.757 government agencies, the press in quotes 01:57:21.840 --> 01:57:26.157 experts and everything uh corrupting the 01:57:26.240 --> 01:57:29.037 output. Two, constant risk of generating 01:57:29.120 --> 01:57:31.757 a bad answer or drawing a bad picture or 01:57:31.840 --> 01:57:35.197 rendering a bad video. Who knows what is 01:57:35.280 --> 01:57:40.477 going to say or do at any moment? Three, legal exposure, product liability, 01:57:40.560 --> 01:57:42.157 slander, election law, many other 01:57:42.240 --> 01:57:45.677 things, and so on. Anything that makes 01:57:45.760 --> 01:57:50.077 Congress mad. for continuous attempts to tighten grip 01:57:50.160 --> 01:57:52.477 on acceptable output degrade the model 01:57:52.560 --> 01:57:55.437 like how good it actually is. Uh in 01:57:55.520 --> 01:57:59.917 terms of usable and pleasant to use and effective and all that kind of stuff and 01:58:00.000 --> 01:58:02.397 five publicity of bad text, images, 01:58:02.480 --> 01:58:05.117 video actual puts those examples into 01:58:05.200 --> 01:58:07.757 the training data for the next version 01:58:07.840 --> 01:58:09.917 so on. So he just highlights how 01:58:10.000 --> 01:58:15.117 difficult this is from all kinds of people being unhappy. You said you can't 01:58:15.200 --> 01:58:17.517 create a system that makes everybody happy. 01:58:17.600 --> 01:58:20.237 >> Yes. >> Uh so if you're going to do the 01:58:20.320 --> 01:58:24.797 finetuning yourself and keep a close source 01:58:24.880 --> 01:58:27.837 essentially the problem there is then trying to minimize the number of people 01:58:27.920 --> 01:58:30.317 who are going to be unhappy. >> Y 01:58:30.400 --> 01:58:35.357 >> um and you're saying like the only that that almost impossible to do right and 01:58:35.440 --> 01:58:38.397 that's the better way is to do open source 01:58:38.480 --> 01:58:41.437 >> basically. Yeah. I mean he's Mark is 01:58:41.520 --> 01:58:43.837 right about uh a number of things that 01:58:43.920 --> 01:58:47.517 he lists that indeed scare large 01:58:47.600 --> 01:58:51.277 companies. Uh you know certainly congressional 01:58:51.360 --> 01:58:53.517 investigations is one of them. legal 01:58:53.600 --> 01:58:57.837 liability. Uh you know uh making things 01:58:57.920 --> 01:59:00.237 that get people to you know hurt 01:59:00.320 --> 01:59:03.437 themselves or hurt others like you know 01:59:03.520 --> 01:59:05.357 um big companies are really careful 01:59:05.440 --> 01:59:07.597 about not um producing things of this 01:59:07.680 --> 01:59:12.797 type and um [snorts] because they have you know they don't 01:59:12.880 --> 01:59:15.677 want to hurt anyone first of all and then second they want to preserve their 01:59:15.760 --> 01:59:18.077 business. So um it's essentially 01:59:18.160 --> 01:59:23.117 impossible for systems like this that can inevitably formulate political 01:59:23.200 --> 01:59:26.637 opinions and you know opinions about various things that may be political or 01:59:26.720 --> 01:59:28.877 not but that people may disagree about 01:59:28.960 --> 01:59:33.037 about you know moral issues and you know 01:59:33.120 --> 01:59:35.757 um things about like questions about 01:59:35.840 --> 01:59:40.077 religion and things like that, right? Or or cultural issues that people from 01:59:40.160 --> 01:59:43.277 different communities would disagree with in the first place. Um so there's 01:59:43.360 --> 01:59:44.957 only kind of a relatively small number 01:59:45.040 --> 01:59:48.797 of things that people will uh sort of 01:59:48.880 --> 01:59:52.637 agree on you know basic principles but beyond that if you if you want those 01:59:52.720 --> 01:59:54.957 systems to be useful they will 01:59:55.040 --> 01:59:58.877 necessarily have to uh offend 01:59:58.960 --> 02:00:03.837 a number of people inevitably >> and so open source is just better and 02:00:03.920 --> 02:00:08.317 then >> diversity is better right >> and open source enables diversity 02:00:08.400 --> 02:00:10.477 >> that's right open source enables diversity 02:00:10.560 --> 02:00:14.877 >> that that's going to be fascinating world where if it's true that the open 02:00:14.960 --> 02:00:18.637 source world if metal leads the way and creates this kind of open- source 02:00:18.720 --> 02:00:20.957 foundation model world there's going to 02:00:21.040 --> 02:00:23.357 be like governments will have a fine new 02:00:23.440 --> 02:00:25.997 model and 02:00:26.080 --> 02:00:28.557 >> and and then potentially 02:00:28.640 --> 02:00:33.037 uh you know people that vote left and right will have their own model and 02:00:33.120 --> 02:00:35.117 preference and be able to choose and it 02:00:35.200 --> 02:00:37.997 will potentially divide us even more but 02:00:38.080 --> 02:00:41.037 that's on us humans we get to figure out 02:00:41.120 --> 02:00:43.037 basically the techn technology enables 02:00:43.120 --> 02:00:47.277 humans to human more effectively and all 02:00:47.360 --> 02:00:49.277 the difficult ethical questions that 02:00:49.360 --> 02:00:51.997 humans raise will just 02:00:52.080 --> 02:00:55.357 um leave it up to us to figure that out. 02:00:55.440 --> 02:00:58.237 Yeah. I mean there are some limits to what you know the same way there are 02:00:58.320 --> 02:00:59.917 limits to free speech there has to be 02:01:00.000 --> 02:01:01.837 some limit to the kind of stuff that 02:01:01.920 --> 02:01:05.517 those systems might uh be authorized to 02:01:05.600 --> 02:01:08.717 um to produce um you know some guard 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 02:01:12.960 --> 02:01:14.237 type of architecture that we were 02:01:14.320 --> 02:01:17.917 discussing before where the output of a 02:01:18.000 --> 02:01:20.557 system is the result of an inference to 02:01:20.640 --> 02:01:23.517 satisfy an objective that objective can 02:01:23.600 --> 02:01:25.677 include guardrails 02:01:25.760 --> 02:01:29.197 >> and uh we can put guardrails in open 02:01:29.280 --> 02:01:31.277 source systems. I mean if we eventually 02:01:31.360 --> 02:01:36.237 have systems that are built with this blueprint uh we can put guard rails uh 02:01:36.320 --> 02:01:40.157 in those systems that guarantee that there is sort of a a minimum [snorts] 02:01:40.240 --> 02:01:44.157 set of guardrails that make the system non-dangerous and non-toxic etc. You 02:01:44.240 --> 02:01:49.197 know basic things that everybody would agree on. Um and and then you know the 02:01:49.280 --> 02:01:52.717 the finetuning that people will add or the additional guardrails that people 02:01:52.800 --> 02:01:55.917 will add will kind of cater to their um 02:01:56.000 --> 02:01:59.837 community whatever it is >> and yeah the fine tuning will be more 02:01:59.920 --> 02:02:03.597 about the gray areas of what is hate speech what is dangerous and all that 02:02:03.680 --> 02:02:06.237 kind of stuff I mean you've >> different value systems 02:02:06.320 --> 02:02:09.757 >> different value systems I mean like uh but still even with the objectives of 02:02:09.840 --> 02:02:13.677 how to build a boweapon for example I think something you've commented on or 02:02:13.760 --> 02:02:17.677 at least there's a paper where collection of researchers is 02:02:17.760 --> 02:02:20.077 trying to understand the social impacts 02:02:20.160 --> 02:02:22.237 of these LLMs 02:02:22.320 --> 02:02:24.877 and I guess one threshold is nice is 02:02:24.960 --> 02:02:27.997 like does the LLM make it any easier 02:02:28.080 --> 02:02:31.357 than a than a search would like a Google 02:02:31.440 --> 02:02:32.317 search would 02:02:32.400 --> 02:02:35.917 >> right so the increasing uh number of 02:02:36.000 --> 02:02:40.317 studies on this seems to point to 02:02:40.400 --> 02:02:43.117 the fact that it doesn't help so having 02:02:43.200 --> 02:02:45.357 an LLM doesn't help you 02:02:45.440 --> 02:02:48.797 >> uh design or built a boweapon or a 02:02:48.880 --> 02:02:50.717 chemical weapon if you already have 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 02:02:56.160 --> 02:02:59.117 sort of increased information you get or the ease with which you get it doesn't 02:02:59.200 --> 02:03:03.437 really help you. Um that's the first thing. The second thing is it's one 02:03:03.520 --> 02:03:08.317 thing to have a list of instructions of how to make a [snorts] chemical weapon 02:03:08.400 --> 02:03:10.317 for example or boweapon. It's another 02:03:10.400 --> 02:03:13.517 thing to actually build it and it's much 02:03:13.600 --> 02:03:18.637 harder than you might think. and an NLM will not help you with that. Um, in 02:03:18.720 --> 02:03:21.437 fact, you know, nobody in the world, not even like, you know, countries use 02:03:21.520 --> 02:03:24.397 bioweapons because most of the time they 02:03:24.480 --> 02:03:28.797 have no idea how to protect their own populations against it. So, um, so it's 02:03:28.880 --> 02:03:30.797 too dangerous actually to kind of ever 02:03:30.880 --> 02:03:34.877 use um, and it's in fact banned by uh, 02:03:34.960 --> 02:03:37.357 uh, international treaties. Um, chemical 02:03:37.440 --> 02:03:42.557 weapons is different. It's also banned by treaties but um, [snorts] uh, but 02:03:42.640 --> 02:03:44.317 it's the same problem. It's difficult to 02:03:44.400 --> 02:03:47.357 use in situations that doesn't turn 02:03:47.440 --> 02:03:49.917 against the perpetrators. But we could 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 02:03:54.720 --> 02:03:56.317 how you build a rocket engine. 02:03:56.400 --> 02:04:00.397 >> Mhm. >> And even if you have a team of 50 engineers that are really experienced 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 02:04:03.360 --> 02:04:05.117 get one that works. 02:04:05.200 --> 02:04:08.657 >> Um, and you know, it's the same with u, 02:04:08.740 --> 02:04:12.717 [snorts] you know, the chemical weapons or bio weapons or things like this. It 02:04:12.800 --> 02:04:17.117 requires expertise, you know, in the in the real world that another line is not 02:04:17.200 --> 02:04:20.317 going to help you with. >> And it requires even the common sense 02:04:20.400 --> 02:04:21.837 expertise that we've been talking about, 02:04:21.920 --> 02:04:26.077 which is how to take uh language based 02:04:26.160 --> 02:04:28.477 instructions and materialize them in the 02:04:28.560 --> 02:04:33.117 physical world requires a lot of knowledge that's not in the 02:04:33.200 --> 02:04:36.157 instructions. >> Yeah, exactly. A lot of biologists have 02:04:36.240 --> 02:04:39.677 posted on this actually in response to those things saying like do you realize 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 02:04:43.120 --> 02:04:44.797 trivial. 02:04:44.880 --> 02:04:47.997 >> Yeah. And that's Hans Marvik comes comes 02:04:48.080 --> 02:04:52.877 to light once again. Uh just to linger on llama you know Mark announced that 02:04:52.960 --> 02:04:56.957 Llama 3 is coming out eventually. I don't think there's a release date but 02:04:57.040 --> 02:04:59.677 what what are you most excited about? 02:04:59.760 --> 02:05:03.517 First of all, Llama 2 that's already out there and maybe the future Alama 3, 02:05:03.600 --> 02:05:06.637 four, five, six, 10, just the the future 02:05:06.720 --> 02:05:09.917 of the open source under meta. 02:05:10.000 --> 02:05:11.997 Well, a number of things. So, uh there's 02:05:12.080 --> 02:05:14.157 going to be like various versions of of 02:05:14.240 --> 02:05:17.837 llama that are uh you know improvements 02:05:17.920 --> 02:05:21.197 of previous llamas, bigger, better, 02:05:21.280 --> 02:05:25.757 multimodel, things like that. And then in future generations, systems that are 02:05:25.840 --> 02:05:29.757 capable of planning, that really understand how the world works. um maybe 02:05:29.840 --> 02:05:33.997 are trained from video so they have some world model maybe you know capable of 02:05:34.080 --> 02:05:37.517 the type of reasoning and planning I was talking about earlier like how long is 02:05:37.600 --> 02:05:40.877 that going to take like when is the research that is doing going in that 02:05:40.960 --> 02:05:43.677 direction going to sort of feed into the 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 02:05:48.240 --> 02:05:49.437 you know a few breakthroughs that we 02:05:49.520 --> 02:05:53.357 have to basically uh go through before 02:05:53.440 --> 02:05:56.077 we can get there but you'll be able to 02:05:56.160 --> 02:05:58.957 monitor our progress because we publish 02:05:59.040 --> 02:06:01.917 our research, right? So, you know, if 02:06:02.000 --> 02:06:04.797 last week we published the Vijpa work, 02:06:04.880 --> 02:06:09.197 which is sort of a first step towards training systems from video um and then 02:06:09.280 --> 02:06:11.357 the next step is going to be world 02:06:11.440 --> 02:06:14.077 models based on kind of this type of 02:06:14.160 --> 02:06:16.477 idea training training from video. 02:06:16.560 --> 02:06:18.477 There's similar work at uh at Deepind 02:06:18.560 --> 02:06:21.917 also and um taking place people and also 02:06:22.000 --> 02:06:26.237 at UC Berkeley on uh world models from 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 02:06:29.200 --> 02:06:32.637 coming are appearing. My bet is that 02:06:32.720 --> 02:06:34.557 those systems are going to be JEAL like 02:06:34.640 --> 02:06:36.237 they're not going to be gener generative 02:06:36.320 --> 02:06:39.997 models. Um [snorts] and uh we'll see 02:06:40.080 --> 02:06:43.597 what the future will tell. Um there's 02:06:43.680 --> 02:06:47.437 really good work at uh um a gentleman 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 02:06:51.280 --> 02:06:52.877 this type that learn representations and 02:06:52.960 --> 02:06:55.437 then use them for planning or learning 02:06:55.520 --> 02:06:59.037 tasks by reinforcement learning. Um and 02:06:59.120 --> 02:07:02.397 a lot of work at Berkeley by Peter 02:07:02.480 --> 02:07:04.877 Savine bunch of other people of that 02:07:04.960 --> 02:07:07.677 type I'm collaborating with actually in 02:07:07.760 --> 02:07:10.637 the context of some grants with my NYU 02:07:10.720 --> 02:07:14.077 hat. Um and then collaborations also 02:07:14.160 --> 02:07:17.037 through meta because the the lab at 02:07:17.120 --> 02:07:19.517 Berkeley is associated with meta in some 02:07:19.600 --> 02:07:22.557 way. So with fair so I I think uh it's 02:07:22.640 --> 02:07:25.277 very exciting you know I I think I'm 02:07:25.360 --> 02:07:27.677 super excited about I I haven't been 02:07:27.760 --> 02:07:32.557 that excited about like the direction of machine learning and AI you know since 02:07:32.640 --> 02:07:34.637 uh you know 10 years ago when fair was 02:07:34.720 --> 02:07:37.597 started and before that um 30 years ago 02:07:37.680 --> 02:07:40.797 when we were working on 35 on on 02:07:40.880 --> 02:07:43.597 commercial nets and and and the early 02:07:43.680 --> 02:07:46.877 days of neural nets so um I'm super 02:07:46.960 --> 02:07:50.237 excited because I see a path towards 02:07:50.320 --> 02:07:52.877 potentially human level intelligence 02:07:52.960 --> 02:07:55.197 uh with you know systems that can 02:07:55.280 --> 02:07:57.997 understand the world remember plan 02:07:58.080 --> 02:08:01.517 reason um there there is some some set 02:08:01.600 --> 02:08:05.357 of ideas to make progress there that might have a chance of working and I'm 02:08:05.440 --> 02:08:08.077 really excited about this what I like is 02:08:08.160 --> 02:08:11.517 that you know it uh 02:08:11.600 --> 02:08:15.837 somewhat we we get onto like a good direction and perhaps succeed before my 02:08:15.920 --> 02:08:18.317 brain turns to a white sauce or or 02:08:18.400 --> 02:08:21.277 before I need to retire. [laughter] 02:08:21.360 --> 02:08:24.797 >> Yeah. Yeah. You're also excited by are 02:08:24.880 --> 02:08:27.037 you 02:08:27.120 --> 02:08:29.117 is it beautiful to you just the amount 02:08:29.200 --> 02:08:32.557 of GPUs involved sort of the the the 02:08:32.640 --> 02:08:34.397 whole training process on this much 02:08:34.480 --> 02:08:37.277 compute. It's just zooming out just 02:08:37.360 --> 02:08:39.837 looking at earth and humans together 02:08:39.920 --> 02:08:42.797 have built these computing devices and 02:08:42.880 --> 02:08:46.317 are able to train this one brain. 02:08:46.400 --> 02:08:50.317 Then then we then open source [laughter] 02:08:50.400 --> 02:08:53.037 like giving birth to this open-source 02:08:53.120 --> 02:08:56.157 brain trained on this gigantic compute 02:08:56.240 --> 02:09:01.197 system. There's just the details of how to train on that, how to build the 02:09:01.280 --> 02:09:05.277 infrastructure and the the hardware, the cooling, all of this kind of stuff. 02:09:05.360 --> 02:09:06.717 >> Are you just still most of your 02:09:06.800 --> 02:09:08.717 excitement is in the the theory aspect 02:09:08.800 --> 02:09:12.397 of it, the meaning like the software? 02:09:12.480 --> 02:09:14.397 >> Well, I used to be a hardware guy many years ago. 02:09:14.480 --> 02:09:15.117 >> Yes. [laughter] Yes. That's right. 02:09:15.200 --> 02:09:19.837 decades ago >> hardware has improved a little bit, changed a little bit. 02:09:19.920 --> 02:09:25.197 >> Yeah. >> I mean certainly scale is necessary but not sufficient. 02:09:25.280 --> 02:09:27.677 >> Absolutely. >> So we certainly need computation. I mean 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 02:09:32.640 --> 02:09:36.637 match the compute power of the human brain. Um you know this may occur in the 02:09:36.720 --> 02:09:41.117 next couple decades but um but we're still some ways away and certainly in 02:09:41.200 --> 02:09:43.117 terms of power efficiency we're really 02:09:43.200 --> 02:09:45.917 far. Um so there's a lot of progress to 02:09:46.000 --> 02:09:50.077 make in uh in in in hardware and you 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 02:09:54.000 --> 02:09:56.957 from silicon technology but a lot of it 02:09:57.040 --> 02:09:59.277 coming from architectural innovation and 02:09:59.360 --> 02:10:04.157 quite a bit coming from uh like more efficient ways of you know implementing 02:10:04.240 --> 02:10:05.837 the architectures that have become 02:10:05.920 --> 02:10:08.717 popular basically combination of 02:10:08.800 --> 02:10:12.477 transformers and comnets right and uh so 02:10:12.560 --> 02:10:15.037 you know there's still some ways to go 02:10:15.120 --> 02:10:17.197 until 02:10:17.280 --> 02:10:20.557 uh we're going to saturate. We're going 02:10:20.640 --> 02:10:22.397 to have to come up with like new new 02:10:22.480 --> 02:10:24.717 principles, new fabrication technology, 02:10:24.800 --> 02:10:27.997 new uh basic components. 02:10:28.080 --> 02:10:32.157 Um perhaps you know based on sort of different principles than those 02:10:32.240 --> 02:10:34.717 classical digital semos. 02:10:34.800 --> 02:10:37.677 >> Interesting. So you think in order to 02:10:37.760 --> 02:10:42.557 build AMI and me we need we potentially 02:10:42.640 --> 02:10:45.517 might need some hardware innovation too. 02:10:45.600 --> 02:10:49.677 >> Well if you want to make it um ubiquitous yeah certainly because we're 02:10:49.760 --> 02:10:51.917 going to have to reduce the 02:10:52.000 --> 02:10:55.037 you know comput the power consumption. A 02:10:55.120 --> 02:10:57.837 GPU today right is half a kilowatt to a 02:10:57.920 --> 02:10:59.597 kilowatt. 02:10:59.680 --> 02:11:02.717 >> Human brain is about 25 watts. 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 02:11:06.800 --> 02:11:08.877 like 100,000 or a million to match it. 02:11:08.960 --> 02:11:11.997 So uh so you know we are off by a huge 02:11:12.080 --> 02:11:14.237 factor here. 02:11:14.320 --> 02:11:17.117 You often say that AGI 02:11:17.200 --> 02:11:20.157 is not coming soon meaning like not this 02:11:20.240 --> 02:11:24.637 year not the next few years potentially 02:11:24.720 --> 02:11:26.797 farther away. What's your basic 02:11:26.880 --> 02:11:29.357 intuition behind that? So first of all 02:11:29.440 --> 02:11:33.837 it's not going to be an event right the idea somehow which you know is 02:11:33.920 --> 02:11:37.997 popularized by science fiction and Hollywood that you know somehow somebody 02:11:38.080 --> 02:11:40.077 is going to discover the secret the 02:11:40.160 --> 02:11:43.677 secret to AGI or human level AI or AMI 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 02:11:47.200 --> 02:11:49.437 have a GI that's just not going to 02:11:49.520 --> 02:11:52.477 happen it's not going to be an event 02:11:52.560 --> 02:11:56.237 it's going to be gradual progress 02:11:56.320 --> 02:11:58.797 are we going to have systems that and 02:11:58.880 --> 02:12:03.357 learn from video how the world works and learn good representations. Yeah. Before 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 02:12:07.200 --> 02:12:08.797 take quite a while. It's not going to 02:12:08.880 --> 02:12:11.917 happen in one day. Um 02:12:12.000 --> 02:12:16.957 uh are we going to get systems that can uh have large amount of associative 02:12:17.040 --> 02:12:20.797 memory so they can they can remember stuff? Yeah, but same it's not going to 02:12:20.880 --> 02:12:23.757 happen tomorrow. I mean there is some basic techniques that need to be 02:12:23.840 --> 02:12:27.197 developed. We have a lot of them, but like you know to get this to work 02:12:27.280 --> 02:12:30.717 together with full system is another story. Are we going to have systems that 02:12:30.800 --> 02:12:32.797 can reason and plan perhaps along the 02:12:32.880 --> 02:12:37.757 lines of objective driven AI architectures that I I described before. 02:12:37.840 --> 02:12:40.637 Yeah. But like before we get this to work, you know, properly, it's going to 02:12:40.720 --> 02:12:44.717 take a while. So, and before we get all those things to work together and then 02:12:44.800 --> 02:12:47.917 on top of this have systems that can learn like hierarchical planning, 02:12:48.000 --> 02:12:52.477 hierarchical representations, systems that can be configured for a lot of 02:12:52.560 --> 02:12:54.077 different situation at hands. the way 02:12:54.160 --> 02:12:57.997 the human brain can. Um, you know, all 02:12:58.080 --> 02:13:01.757 of this is going to take, you know, at least a decade and probably much more 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 02:13:06.560 --> 02:13:08.717 not encountered and so we don't know if 02:13:08.800 --> 02:13:10.637 there is a easy solution within this 02:13:10.720 --> 02:13:15.117 framework. Um, so you know it's it's not 02:13:15.200 --> 02:13:18.797 just around the corner. I mean, I've I've been hearing people for the last 02:13:18.880 --> 02:13:21.357 12, 15 years claiming that, you know, 02:13:21.440 --> 02:13:25.837 AGI is just around the corner and being systematically wrong, and I knew they 02:13:25.920 --> 02:13:29.277 were wrong when they were saying it. I call their 02:13:29.360 --> 02:13:31.997 >> Why do you think people have been calling? First of all, I mean, from the 02:13:32.080 --> 02:13:36.237 beginning of from the birth of the term artificial intelligence, there has been 02:13:36.320 --> 02:13:38.957 a eternal optimism 02:13:39.040 --> 02:13:40.397 that's perhaps unlike other 02:13:40.480 --> 02:13:45.437 technologies. Is it Maravic paradox? is 02:13:45.520 --> 02:13:49.677 the explanation for why people are so optimistic about AGI. 02:13:49.760 --> 02:13:51.437 >> I don't think it's just morxic paradox. 02:13:51.520 --> 02:13:53.117 Morex paradox is a consequence of 02:13:53.200 --> 02:13:55.997 realizing that the world is not as easy 02:13:56.080 --> 02:13:59.277 as we think. So first of all um 02:13:59.360 --> 02:14:01.437 intelligence is not a linear thing that 02:14:01.520 --> 02:14:03.597 you can measure with a scalar with a 02:14:03.680 --> 02:14:06.877 single number. Um you know can you say 02:14:06.960 --> 02:14:09.677 that humans are smarter than 02:14:09.760 --> 02:14:13.197 orangutong? In some ways, yes. But in 02:14:13.280 --> 02:14:17.037 some ways, orongs are smarter than humans in a lot of domains that allows 02:14:17.120 --> 02:14:21.117 them to survive in the forest, for example. So, IQ is a very limited 02:14:21.200 --> 02:14:25.757 measure of intelligence. Human intelligence is bigger than what IQ, for 02:14:25.840 --> 02:14:29.357 example, measures. >> Well, IQ can measure, you know, 02:14:29.440 --> 02:14:31.677 approximately something for humans. 02:14:31.760 --> 02:14:35.517 >> Mhm. But um because humans kind of, you 02:14:35.600 --> 02:14:40.557 know, come in relatively kind of uniform 02:14:40.640 --> 02:14:42.317 form, right? 02:14:42.400 --> 02:14:45.917 >> Uh but it only measures one type of uh 02:14:46.000 --> 02:14:47.837 ability that you know may be relevant 02:14:47.920 --> 02:14:51.837 for some tasks but not others. And 02:14:51.920 --> 02:14:53.437 uh but then if you are talking about 02:14:53.520 --> 02:14:57.917 other intelligent entities for which the 02:14:58.000 --> 02:14:59.517 you know the basic things that are easy 02:14:59.600 --> 02:15:02.557 to them is very different then it 02:15:02.640 --> 02:15:05.597 doesn't mean anything. So intelligence 02:15:05.680 --> 02:15:10.797 is a collection of skills 02:15:10.880 --> 02:15:14.807 and an ability to acquire new skills efficiently. 02:15:14.890 --> 02:15:19.277 [clears throat] >> Right. And the collection of skills that an 02:15:19.360 --> 02:15:23.837 intelligent particular intelligent entity possess or is capable of learning 02:15:23.920 --> 02:15:28.397 quickly is different from the collection of skills of another one. And because 02:15:28.480 --> 02:15:32.397 it's a multi-dimensional thing, the set of skills is high dimensional space. You 02:15:32.480 --> 02:15:36.237 can't measure, you can compare, you cannot compare two things as to whether 02:15:36.320 --> 02:15:38.397 one is more intelligent than the other. 02:15:38.480 --> 02:15:41.437 It's multi-dimensional. 02:15:41.520 --> 02:15:44.717 So you push back against 02:15:44.800 --> 02:15:48.957 what are called AI doomers a lot. Uh can 02:15:49.040 --> 02:15:52.477 you explain their perspective and why you think they're wrong? 02:15:52.560 --> 02:15:55.037 >> Okay. So AI doomers imagine all kinds of 02:15:55.120 --> 02:15:58.397 catastrophe scenarios of how AI could 02:15:58.480 --> 02:16:01.997 escape our control and basically kill us 02:16:02.080 --> 02:16:03.517 all. [laughter] 02:16:03.600 --> 02:16:06.557 Uh and that relies on a whole bunch of 02:16:06.640 --> 02:16:09.357 assumptions that are mostly false. 02:16:09.440 --> 02:16:13.037 So the first assumption is that the emergence of super intelligence is going 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 02:16:17.360 --> 02:16:21.997 the secret and we'll turn on a machine that is super intelligent and because 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. 02:16:25.760 --> 02:16:27.437 That is false. It's not going to be an 02:16:27.520 --> 02:16:29.597 event. We're going to have systems that 02:16:29.680 --> 02:16:33.037 are like as smart as a cat. has all the 02:16:33.120 --> 02:16:37.757 have all the characteristics of, you know, human level intelligence, but 02:16:37.840 --> 02:16:39.117 their level of intelligence would be 02:16:39.200 --> 02:16:42.157 like a cat or a parrot maybe or 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 02:16:47.360 --> 02:16:49.997 things more intelligent. And as we make them more intelligent, we're also going 02:16:50.080 --> 02:16:53.277 to put some guardrails in them and learn how to kind of put some guardrails so 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 02:16:56.880 --> 02:16:59.517 going to be one effort, but it's going to be lots of different people doing 02:16:59.600 --> 02:17:03.437 this. And some of them are going to succeed at making intelligent systems 02:17:03.520 --> 02:17:07.757 that are uh controllable and safe and have the right guardrails. And if some 02:17:07.840 --> 02:17:10.317 other goes rogue, then we can use the 02:17:10.400 --> 02:17:12.077 the good ones to go against the rogue 02:17:12.160 --> 02:17:14.797 ones. So it's going to be my, you know, 02:17:14.880 --> 02:17:17.597 smart AI police against your rogue AI. 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 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. 02:17:24.640 --> 02:17:28.317 Now there is another fallacy which is the fact that because a system is 02:17:28.400 --> 02:17:31.117 intelligent it necessarily wants to take over. 02:17:31.200 --> 02:17:36.077 >> Mhm. >> Um and there is several arguments 02:17:36.160 --> 02:17:38.077 that make people scare of this which I 02:17:38.160 --> 02:17:41.277 think are completely false uh as well. 02:17:41.360 --> 02:17:45.037 So uh one of them is um you know in 02:17:45.120 --> 02:17:47.357 nature it seems to be that the more 02:17:47.440 --> 02:17:49.037 intelligent species are the one that end 02:17:49.120 --> 02:17:54.157 up dominating the other and uh and even 02:17:54.240 --> 02:17:58.797 you know extinguishing the others sometimes by design sometimes just by 02:17:58.880 --> 02:18:02.237 mistake and 02:18:02.320 --> 02:18:05.437 and so you know there is sort of a 02:18:05.520 --> 02:18:08.317 thinking by which you say well if AI 02:18:08.400 --> 02:18:12.717 systems are more intelligent than us surely they're going uh eliminate us if 02:18:12.800 --> 02:18:17.517 not by design simply because they don't care about us. And that's just 02:18:17.600 --> 02:18:21.677 preposteruous for for a number of reasons. Um first reason is they're not 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. 02:18:25.920 --> 02:18:29.437 They're not going to have the desire to dominate because the desire to dominate 02:18:29.520 --> 02:18:31.357 is something that has to be hardwired 02:18:31.440 --> 02:18:35.197 into an intelligent system. Uh it is 02:18:35.280 --> 02:18:38.317 hardwired in humans. It is hardwired in 02:18:38.400 --> 02:18:41.517 baboons, in chimpanzees, in wolves, 02:18:41.600 --> 02:18:43.997 not in or wrongs. 02:18:44.080 --> 02:18:47.357 The species in which this desire to 02:18:47.440 --> 02:18:50.797 dominate or submit or or attain status 02:18:50.880 --> 02:18:53.037 in other ways 02:18:53.120 --> 02:18:55.997 is is specific to social species. 02:18:56.080 --> 02:18:59.117 Non-social species like don't have it, 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 02:19:04.080 --> 02:19:05.597 significant incentive for humans to 02:19:05.680 --> 02:19:08.557 encode that into the AI systems. And to 02:19:08.640 --> 02:19:11.117 the degree they do, there will be other 02:19:11.200 --> 02:19:14.797 AIs that um sort of punish them for it. 02:19:14.880 --> 02:19:17.197 I'll compete them over. >> Well, there's all kinds of incentive to 02:19:17.280 --> 02:19:19.917 make AI systems submissive to humans, right? 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, 02:19:23.680 --> 02:19:26.797 so then people say, "Oh, but look at LLMs. LM are not controllable." And 02:19:26.880 --> 02:19:28.477 they're right. OLMs are not 02:19:28.560 --> 02:19:30.717 controllable. But objectived driven AI, 02:19:30.800 --> 02:19:34.557 so systems that derive their answers by 02:19:34.640 --> 02:19:38.797 optimization of an objective means they have to optimize its objective. And that 02:19:38.880 --> 02:19:41.197 objective can include guardrails. One 02:19:41.280 --> 02:19:45.757 guardrail is uh obey humans. Another 02:19:45.840 --> 02:19:50.477 guardrail is don't obey humans if it's hurting other humans within. 02:19:50.560 --> 02:19:52.397 >> I've heard that before somewhere. I don't remember. 02:19:52.480 --> 02:19:54.797 >> Yes. [laughter] >> Maybe in a book. 02:19:54.880 --> 02:19:58.157 >> Yeah. Uh but speaking of that book, what 02:19:58.240 --> 02:20:01.917 is could there be unintended consequences also from all of this? 02:20:02.000 --> 02:20:06.477 >> No, of course. Uh so this is not a simple problem, right? I mean designing 02:20:06.560 --> 02:20:10.077 those guardrails so that the system behaves properly is not going to be a 02:20:10.160 --> 02:20:14.557 simple u issue that for which there is a silver 02:20:14.640 --> 02:20:18.637 bullet for which you have a mathematical proof that the system can be safe. It's 02:20:18.720 --> 02:20:23.197 going to be very progressive iterative design system where we put those 02:20:23.280 --> 02:20:26.477 guardrails in such a way that the system behave properly and sometimes they're 02:20:26.560 --> 02:20:30.877 going to do something that you know was unexpected because the guardrail wasn't 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 02:20:34.400 --> 02:20:36.477 somehow that we can't get it slightly 02:20:36.560 --> 02:20:41.517 wrong because if we get it slightly wrong we all die is is ridiculous. 02:20:41.600 --> 02:20:45.197 um we we're just going to go progressively and it's it's just going 02:20:45.280 --> 02:20:47.597 to be the the analogy I've used many 02:20:47.680 --> 02:20:53.197 times is um is uh turbo jet design. Um 02:20:53.280 --> 02:20:55.357 how how did we figure out how to make 02:20:55.440 --> 02:20:58.797 turbo jets so unbelievably reliable, 02:20:58.880 --> 02:21:03.437 right? Uh I mean those are like, you know, incredibly complex pieces of 02:21:03.520 --> 02:21:08.157 hardware that run at really high temperatures for, you know, 20 hours at 02:21:08.240 --> 02:21:10.397 a time sometimes. 02:21:10.480 --> 02:21:13.037 And we can, you know, fly halfway around 02:21:13.120 --> 02:21:17.837 the world with a on a two engine uh uh 02:21:17.920 --> 02:21:19.917 jetliner at near the speed of sound. 02:21:20.000 --> 02:21:21.997 Like how incredible is this? Is just 02:21:22.080 --> 02:21:25.517 unbelievable, right? And 02:21:25.600 --> 02:21:29.437 did we do this because we invented like a general principle of how to make turbo 02:21:29.520 --> 02:21:31.997 jet safe? No. We it took decades to kind 02:21:32.080 --> 02:21:33.677 of fine-tune the design of those systems 02:21:33.760 --> 02:21:37.997 so that they they were safe. Is there a 02:21:38.080 --> 02:21:42.877 separate uh group within General Electric or Sneakma or whatever 02:21:42.960 --> 02:21:44.397 that 02:21:44.480 --> 02:21:47.757 is specialized in turbo jet safety? No, 02:21:47.840 --> 02:21:51.597 it's the design is all about safety 02:21:51.680 --> 02:21:56.077 because a better turbo jet is also a safer turbo jet. So, um a more reliable 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 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 02:22:04.240 --> 02:22:06.317 because they are designed to be more 02:22:06.400 --> 02:22:09.197 useful uh and more controllable. 02:22:09.280 --> 02:22:11.757 >> So, let's imagine a system AI system 02:22:11.840 --> 02:22:15.997 that's able to be incredibly convincing 02:22:16.080 --> 02:22:19.117 and can convince you of anything. I I 02:22:19.200 --> 02:22:21.997 can at least imagine such a system. 02:22:22.080 --> 02:22:25.437 And I can see such a system be 02:22:25.520 --> 02:22:29.757 weapon-like because it can control people's minds. We're pretty gullible. 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 02:22:33.760 --> 02:22:37.597 and you could see governments using that as a weapon. 02:22:37.680 --> 02:22:41.437 So do you think if you imagine such a system 02:22:41.520 --> 02:22:46.077 there's any parallel to something like nuclear weapons? 02:22:46.160 --> 02:22:51.037 >> No. >> So is why why why is that technology 02:22:51.120 --> 02:22:54.237 different? So you're saying there's going to be gradual development. 02:22:54.320 --> 02:22:56.717 >> Yeah. is going to be I mean it might be 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 02:23:01.520 --> 02:23:01.917 so on. 02:23:02.000 --> 02:23:04.237 >> So that AI system designed by Vladimir 02:23:04.320 --> 02:23:09.517 Putin or whatever or his uh minions uh 02:23:09.600 --> 02:23:12.397 you know is going to be uh like talking 02:23:12.480 --> 02:23:17.837 to trying to talk to every American to uh convince them to vote for you know 02:23:17.920 --> 02:23:21.757 >> whoever >> whoever pleases Putin. 02:23:21.840 --> 02:23:23.037 >> Sure. 02:23:23.120 --> 02:23:26.717 uh or whatever or or you know or rile 02:23:26.800 --> 02:23:29.757 people up against each other um as 02:23:29.840 --> 02:23:33.757 they've been trying to do. They're not going to be talking to you. 02:23:33.840 --> 02:23:36.317 They're going to be talking to your AI assistant. 02:23:36.400 --> 02:23:39.837 >> Mhm. Which is going to be as smart as theirs. 02:23:39.920 --> 02:23:43.117 >> Mhm. >> Right. That AI because as I said in the 02:23:43.200 --> 02:23:46.317 future every single one of your interaction with the digital world will 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 02:23:50.480 --> 02:23:52.717 this a scam? Yeah, >> like is this thing like telling me the 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 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 02:24:01.520 --> 02:24:03.357 going to be like a spam filter, right? 02:24:03.440 --> 02:24:07.597 You're not even seeing the email, the spam email, right? It's automatically 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 02:24:12.160 --> 02:24:15.357 system that tries to convince you of something is going to be talking to AI 02:24:15.440 --> 02:24:19.197 assistant, which is going to be at least as smart as it. 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 02:24:24.240 --> 02:24:27.117 your attention. >> So to you it's very difficult for any 02:24:27.200 --> 02:24:29.277 one AI system to take such a big leap 02:24:29.360 --> 02:24:31.597 ahead to where it can convince even the 02:24:31.680 --> 02:24:34.797 the other AI systems. So like it there's 02:24:34.880 --> 02:24:36.877 always going to be this kind of race 02:24:36.960 --> 02:24:42.157 where nobody's way ahead. That's the history of the world. History of the 02:24:42.240 --> 02:24:45.437 world is you know whenever there is a progress some some place there is a 02:24:45.520 --> 02:24:48.957 countermeasure and and you know it's a 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 02:24:53.200 --> 02:24:55.357 weapons are so interesting because that 02:24:55.440 --> 02:24:58.477 was such a powerful weapon that it 02:24:58.560 --> 02:25:01.197 mattered who got it first 02:25:01.280 --> 02:25:07.960 that you know you could imagine Hitler, Stalin, 02:25:08.240 --> 02:25:10.717 Mao getting the weapon first and that 02:25:10.800 --> 02:25:12.717 that having a different kind of impact 02:25:12.800 --> 02:25:15.437 on the world than than the United States 02:25:15.520 --> 02:25:18.317 getting the weapon first. to you nuclear 02:25:18.400 --> 02:25:22.717 weapons is like you you don't imagine a 02:25:22.800 --> 02:25:25.357 uh breakthrough discovery and then 02:25:25.440 --> 02:25:28.637 Manhattan project like effort for AI. 02:25:28.720 --> 02:25:31.357 >> No, as I said it's not going to be an 02:25:31.440 --> 02:25:33.357 event. It's going to be you know 02:25:33.440 --> 02:25:37.597 continuous progress and and whenever you 02:25:37.680 --> 02:25:40.717 know one breakthrough occurs it's going to be widely disseminated really 02:25:40.800 --> 02:25:44.077 quickly. Yeah, >> probably first within industry. I mean, 02:25:44.160 --> 02:25:48.557 this is not a domain where, you know, government or military organizations are 02:25:48.640 --> 02:25:51.037 particularly innovative and they're in 02:25:51.120 --> 02:25:56.077 fact way behind. Um, and so this is going to come from industry and and this 02:25:56.160 --> 02:25:59.517 kind of information disseminates extremely quickly. We've seen this over 02:25:59.600 --> 02:26:04.157 the last few years, right? Where you have a new like, you know, even take 02:26:04.240 --> 02:26:07.597 AlphaGo. This was reproduced within three months 02:26:07.680 --> 02:26:10.717 >> even [clears throat] without like particularly detailed information, right? 02:26:10.800 --> 02:26:13.997 >> Yeah. This is an industry that's not good at secrecy. [laughter] 02:26:14.080 --> 02:26:17.437 >> No, but even even if there is, just the fact that you know that something is 02:26:17.520 --> 02:26:22.077 possible >> Yeah. >> Uh makes you like realize that it's 02:26:22.160 --> 02:26:25.437 worth investing the time to actually do it. You you may be the second person to 02:26:25.520 --> 02:26:27.677 do it, but you know, you'll you'll do 02:26:27.760 --> 02:26:31.917 it. uh and you know same for you know 02:26:32.000 --> 02:26:34.957 all the innovations uh of you know 02:26:35.040 --> 02:26:39.277 self-s supervised learning transformers dec decoder only architectures LM I mean 02:26:39.360 --> 02:26:42.877 those things you don't need to know exactly the details of how they work to 02:26:42.960 --> 02:26:46.717 know that you know it's possible u because it's deployed and then it's 02:26:46.800 --> 02:26:49.837 getting reproduced and then you know 02:26:49.920 --> 02:26:53.037 people who work for those companies move 02:26:53.120 --> 02:26:55.837 they go from one company to another and 02:26:55.920 --> 02:26:57.837 you know the information disseminates 02:26:57.920 --> 02:27:01.757 What makes the success of the the US 02:27:01.840 --> 02:27:04.957 tech industry and Silicon Valley in particular is exactly that is because 02:27:05.040 --> 02:27:09.197 information circulates really really quickly and this you know disseminates 02:27:09.280 --> 02:27:11.757 very quickly and so you know the the 02:27:11.840 --> 02:27:14.957 whole region sort of is ahead because of 02:27:15.040 --> 02:27:17.757 that circulation of information. So 02:27:17.840 --> 02:27:19.917 maybe I just to linger on the psychology 02:27:20.000 --> 02:27:23.117 of AI doomers. You give uh in the 02:27:23.200 --> 02:27:25.277 classic Yan Lagoon way, a pretty good 02:27:25.360 --> 02:27:28.237 example of just when a new technology 02:27:28.320 --> 02:27:32.237 comes to be. You say uh engineer says, 02:27:32.320 --> 02:27:35.437 "I invented this new thing. I call it a 02:27:35.520 --> 02:27:40.797 ball pen." And then the Twitter sphere responds, "OMG, people could write 02:27:40.880 --> 02:27:43.837 horrible things with it like misinformation, propaganda, hate speech, 02:27:43.920 --> 02:27:47.037 ban it now." Then writing doomers come 02:27:47.120 --> 02:27:48.557 in 02:27:48.640 --> 02:27:51.117 akin to the AI doomers. Imagine if 02:27:51.200 --> 02:27:55.757 everyone can get a ball pen. This could destroy society. There should be a law 02:27:55.840 --> 02:27:57.517 against using ball pen to write hate 02:27:57.600 --> 02:28:00.317 speech. Regulate ball pens now. And then 02:28:00.400 --> 02:28:03.517 the pencil industry mogul says, "Yeah, 02:28:03.600 --> 02:28:06.317 ball pens are very dangerous. Unlike 02:28:06.400 --> 02:28:08.797 pencil writing, which is erasable, ball 02:28:08.880 --> 02:28:13.597 pen writing stays forever. Government should require a license for a pen 02:28:13.680 --> 02:28:15.277 manufacturer. 02:28:15.360 --> 02:28:19.197 I mean this does seem to be part of um 02:28:19.280 --> 02:28:22.957 human psychology when when it comes up 02:28:23.040 --> 02:28:25.117 against new technology. 02:28:25.200 --> 02:28:28.797 So what what deep insights can you speak 02:28:28.880 --> 02:28:31.437 to about this? Well, there is a a 02:28:31.520 --> 02:28:36.317 natural fear of uh new technology and 02:28:36.400 --> 02:28:38.317 the impact it can have on society and 02:28:38.400 --> 02:28:41.597 people have kind of instinctive reaction 02:28:41.680 --> 02:28:45.917 to um you know the world they know being 02:28:46.000 --> 02:28:48.957 threatened by major transformations um 02:28:49.040 --> 02:28:51.277 that are either cultural phenomena or 02:28:51.360 --> 02:28:54.797 technological um revolutions. 02:28:54.880 --> 02:28:57.517 and they fear for their culture, they 02:28:57.600 --> 02:28:59.517 feel for their job, they feel for they 02:28:59.600 --> 02:29:01.917 fear for their you know their future of 02:29:02.000 --> 02:29:05.597 their children um and uh their way of 02:29:05.680 --> 02:29:09.277 life, right? So, so any change um is 02:29:09.360 --> 02:29:12.317 feared and and you see this you know 02:29:12.400 --> 02:29:14.797 along history like any technological 02:29:14.880 --> 02:29:17.357 revolution or cultural phenomenon was 02:29:17.440 --> 02:29:21.197 always accompanied by uh you know groups 02:29:21.280 --> 02:29:25.437 or reaction in the media uh that that 02:29:25.520 --> 02:29:29.517 basically attributed the all the problems the current 02:29:29.600 --> 02:29:31.997 problems of society to that particular 02:29:32.080 --> 02:29:34.717 change right electricity was going to 02:29:34.800 --> 02:29:40.077 kill everyone And at some point, you know, you the train was going to be a 02:29:40.160 --> 02:29:44.957 horrible thing because, you know, you can't breathe past 50 km an hour. Um, 02:29:45.040 --> 02:29:49.117 and so there's a wonderful website called the pessimist archive, 02:29:49.200 --> 02:29:52.397 >> right? >> Which has all those newspaper clips of 02:29:52.480 --> 02:29:54.637 all the horrible things people imagine 02:29:54.720 --> 02:29:58.077 would would arrive because of uh either 02:29:58.160 --> 02:30:01.597 a technological uh innovation or uh a 02:30:01.680 --> 02:30:05.677 cultural phenomenon. Um, you know, 02:30:05.760 --> 02:30:09.837 the this is wonderful examples of uh, 02:30:09.920 --> 02:30:12.557 you know, jazz or comic books being 02:30:12.640 --> 02:30:17.197 blamed for, uh, unemployment or or, you know, young 02:30:17.280 --> 02:30:19.277 people not wanting to work anymore and 02:30:19.360 --> 02:30:21.117 things like that, right? And and that 02:30:21.200 --> 02:30:24.637 has existed for for centuries. 02:30:24.720 --> 02:30:28.637 Um, and it's, you know, knee-jerk 02:30:28.720 --> 02:30:32.637 reactions. Um the question is you know do we 02:30:32.720 --> 02:30:34.797 embrace change 02:30:34.880 --> 02:30:38.797 uh or do we resist it and 02:30:38.880 --> 02:30:40.477 what are the real dangers as opposed to 02:30:40.560 --> 02:30:44.557 the imagined uh imagined ones. 02:30:44.640 --> 02:30:48.957 So people worry about I think one thing they worry about with big tech something 02:30:49.040 --> 02:30:51.277 we've been talking about over and over 02:30:51.360 --> 02:30:53.837 but I think worth [snorts] 02:30:53.920 --> 02:30:56.317 mentioning again they worry about how 02:30:56.400 --> 02:30:59.837 powerful AI will be and they worry about 02:30:59.920 --> 02:31:02.477 it being in the hands of one centralized 02:31:02.560 --> 02:31:05.277 power of just a handful of central 02:31:05.360 --> 02:31:09.597 control. And so that's the skepticism with big tech. You can make these 02:31:09.680 --> 02:31:11.037 companies can make a huge amount of 02:31:11.120 --> 02:31:15.677 money and control this technology and by 02:31:15.760 --> 02:31:19.437 so doing you know take advantage uh 02:31:19.520 --> 02:31:21.757 abuse the little guy in society. 02:31:21.840 --> 02:31:24.637 >> Well that's exactly why we need open source platforms. 02:31:24.720 --> 02:31:29.357 >> Yeah. I [snorts] just wanted to [laughter] nail the point home more and 02:31:29.440 --> 02:31:29.837 more. 02:31:29.920 --> 02:31:33.517 >> Yes. Um, so let me ask you on your, like 02:31:33.600 --> 02:31:36.957 I said, you do get a little bit uh um, 02:31:37.040 --> 02:31:39.517 yeah, flavorful on the internet. Uh, 02:31:39.600 --> 02:31:42.157 Yosha Bach tweeted something that you 02:31:42.240 --> 02:31:46.157 lolled at uh, in reference to HAL 9000. 02:31:46.240 --> 02:31:50.717 Quote, I appreciate your argument and I fully understand your frustration, but 02:31:50.800 --> 02:31:52.877 whether the pod bay doors should be 02:31:52.960 --> 02:31:54.877 opened or closed is a complex and 02:31:54.960 --> 02:31:58.557 nuanced issue. So, you're at the head of 02:31:58.640 --> 02:32:02.637 Meta AI. Um, you know, this is something 02:32:02.720 --> 02:32:06.957 that really worries me that AI, our AI 02:32:07.040 --> 02:32:09.677 overlords will speak down to us with 02:32:09.760 --> 02:32:13.677 corporate speak um of this nature and 02:32:13.760 --> 02:32:15.677 you sort of resist that with your way of 02:32:15.760 --> 02:32:18.797 being. Um is this something you can just 02:32:18.880 --> 02:32:21.037 comment on sort of working at a big 02:32:21.120 --> 02:32:27.357 company how you can avoid the 02:32:27.440 --> 02:32:32.317 overfearing I suppose the through 02:32:32.400 --> 02:32:34.157 caution create harm. 02:32:34.240 --> 02:32:36.717 >> Yeah. Again, I think the answer to this 02:32:36.800 --> 02:32:39.517 is open source platforms and then 02:32:39.600 --> 02:32:42.477 enabling a widely diverse set of people 02:32:42.560 --> 02:32:44.397 to 02:32:44.480 --> 02:32:47.837 build AI assistants that represent the 02:32:47.920 --> 02:32:51.997 diversity of uh cultures, opinions, languages, and value systems across the 02:32:52.080 --> 02:32:54.157 world. Um, so that you're not bound to 02:32:54.240 --> 02:32:57.887 just u you know be brainwashed by 02:32:57.970 --> 02:32:59.837 [laughter] a particular way of thinking 02:32:59.920 --> 02:33:04.157 because of single AI entity. Um, so I 02:33:04.240 --> 02:33:05.837 mean I I think it's really really 02:33:05.920 --> 02:33:08.237 important question for society and the 02:33:08.320 --> 02:33:13.037 problem I'm seeing is um is that um 02:33:13.120 --> 02:33:14.797 which is why I've been so vocal and 02:33:14.880 --> 02:33:17.917 sometimes a little sardonic about it. 02:33:18.000 --> 02:33:21.277 >> Never stop. Never stop. Yan, [laughter] 02:33:21.360 --> 02:33:24.157 >> we love it. is because I see the danger 02:33:24.240 --> 02:33:26.237 of this concentration of power through 02:33:26.320 --> 02:33:29.997 to through proprietary AI systems as a 02:33:30.080 --> 02:33:32.637 much bigger danger than everything else 02:33:32.720 --> 02:33:36.477 that if we really want you know uh 02:33:36.560 --> 02:33:40.237 diversity of opinion uh AI systems that 02:33:40.320 --> 02:33:42.317 you know in in this future that where 02:33:42.400 --> 02:33:44.477 we'll all be interacting through AI 02:33:44.560 --> 02:33:47.277 systems we need those to be diverse for 02:33:47.360 --> 02:33:49.997 the preservation of uh uh [snorts] 02:33:50.080 --> 02:33:53.757 diversity of ideas and you know creeds 02:33:53.840 --> 02:33:57.757 and political opinions and and and whatever 02:33:57.840 --> 02:34:00.637 uh and the preservation of democracy. 02:34:00.720 --> 02:34:02.477 And 02:34:02.560 --> 02:34:05.677 what works against this is people who 02:34:05.760 --> 02:34:08.877 think that for reasons of security we 02:34:08.960 --> 02:34:11.277 should keep AI systems under lock and 02:34:11.360 --> 02:34:16.717 key because it's too dangerous to put it in the hands of of everybody 02:34:16.800 --> 02:34:17.917 um because it could be used by 02:34:18.000 --> 02:34:21.597 terrorists or something. um 02:34:21.680 --> 02:34:24.157 that would lead to 02:34:24.240 --> 02:34:27.517 uh you know potentially a a 02:34:27.600 --> 02:34:30.557 very bad future in which all of our 02:34:30.640 --> 02:34:34.717 information diet is controlled by a small number of uh companies through 02:34:34.800 --> 02:34:37.037 proprietary systems. 02:34:37.120 --> 02:34:40.397 >> Do you trust humans with this technology 02:34:40.480 --> 02:34:44.797 to uh to build systems that are on the 02:34:44.880 --> 02:34:49.197 whole good for humanity? Isn't that what democracy and free speech is all about? 02:34:49.280 --> 02:34:53.037 >> I think so. >> Do you trust institutions to do the right thing? 02:34:53.120 --> 02:34:57.197 >> Do you trust people to do the right thing? And and yeah, there's bad people 02:34:57.280 --> 02:35:00.077 who are going to do bad things, but they're not going to have superior 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 02:35:04.400 --> 02:35:06.877 bad AI, right? I mean, it's the examples 02:35:06.960 --> 02:35:09.437 that we were just talking about of, you 02:35:09.520 --> 02:35:12.477 know, maybe some rogue country will 02:35:12.560 --> 02:35:15.117 build, you know, some AI system that's 02:35:15.200 --> 02:35:18.397 going to try to convince everybody to 02:35:18.480 --> 02:35:20.317 go into a civil war or something or or 02:35:20.400 --> 02:35:26.237 or elect favorable ruler and um but then 02:35:26.320 --> 02:35:30.637 they will have to go past our AI systems, >> right? an AI system with a strong 02:35:30.720 --> 02:35:32.957 Russian accent will be trying to convince our [laughter] 02:35:33.040 --> 02:35:35.357 >> and doesn't put any uh articles in their 02:35:35.440 --> 02:35:37.517 sentences. 02:35:37.600 --> 02:35:40.717 >> Um well, it'll be at the very least 02:35:40.800 --> 02:35:45.437 absurdly comedic. Okay. Uh so I uh since 02:35:45.520 --> 02:35:47.357 we talked about sort of the uh physical 02:35:47.440 --> 02:35:50.237 reality, I'd love to ask your vision of 02:35:50.320 --> 02:35:55.277 the future with with robots in in this physical reality. So many of the kinds 02:35:55.360 --> 02:35:56.957 of intelligence you've been speaking 02:35:57.040 --> 02:36:00.077 about would empower robots to be more 02:36:00.160 --> 02:36:03.277 effective collaborators with us humans. 02:36:03.360 --> 02:36:07.117 So um since uh Tesla's Optimus uh team 02:36:07.200 --> 02:36:11.197 has been showing off some progress on humanoid robots, I think it really 02:36:11.280 --> 02:36:15.357 reinvigorated the whole industry now that I think Boston Dynamics has been 02:36:15.440 --> 02:36:18.397 leading for a very very long time. So now there's all kinds of companies. 02:36:18.480 --> 02:36:21.757 Figure AI obviously Boston Dynamics. Uh 02:36:21.840 --> 02:36:25.997 >> Union Tree >> Union Tree. Uh but there's like a lot of 02:36:26.080 --> 02:36:28.637 them. It's great. It's great. I mean I 02:36:28.720 --> 02:36:33.837 love it. >> Uh so do you think there'll be uh 02:36:33.920 --> 02:36:36.797 millions of humanoid robots walking around soon? 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 02:36:40.160 --> 02:36:41.437 think, is going to be really interesting 02:36:41.520 --> 02:36:45.437 in robots. like the the emergence of the 02:36:45.520 --> 02:36:48.157 robotics industry has been in the 02:36:48.240 --> 02:36:52.957 waiting for you know 10 20 years without really emerging other than for like you 02:36:53.040 --> 02:36:54.637 know kind of pre-programmed behavior and 02:36:54.720 --> 02:36:58.717 stuff like that. Um and uh and the main 02:36:58.800 --> 02:37:02.877 issue is again the Maric paradox like you know how do we get those systems to 02:37:02.960 --> 02:37:06.557 understand how the world works and and kind of you know plan actions and so we 02:37:06.640 --> 02:37:09.277 can do it for really specialized tasks. 02:37:09.360 --> 02:37:13.357 Um and uh the way Boston dynamics goes 02:37:13.440 --> 02:37:15.197 about it is you know basically with a 02:37:15.280 --> 02:37:18.477 lot of um handcrafted dynamical models 02:37:18.560 --> 02:37:22.077 and careful planning uh in advance which 02:37:22.160 --> 02:37:26.397 is very classical robotics with a lot of innovation a little bit of perception. 02:37:26.480 --> 02:37:29.037 Um but it's still not like they can't 02:37:29.120 --> 02:37:34.077 build a domestic robot right. Um, and 02:37:34.160 --> 02:37:38.397 you know, we're still some distance away from completely autonomous level five 02:37:38.480 --> 02:37:42.477 driving. [clears throat] >> And we're certainly very far away from 02:37:42.560 --> 02:37:45.117 having uh, you know, level five 02:37:45.200 --> 02:37:47.197 autonomous driving by a system that can 02:37:47.280 --> 02:37:51.037 train itself by driving 20 hours like 02:37:51.120 --> 02:37:53.357 any 17-year-old. [gasps] 02:37:53.440 --> 02:37:57.597 Uh so until we have 02:37:57.680 --> 02:38:01.437 uh again world models, systems that can 02:38:01.520 --> 02:38:05.277 train themselves to understand how the world works, 02:38:05.360 --> 02:38:07.037 uh we're not going to we're not going to 02:38:07.120 --> 02:38:09.757 have significant progress in robotics. 02:38:09.840 --> 02:38:12.317 So a lot of the people working on 02:38:12.400 --> 02:38:14.957 robotic hardware at the moment are are 02:38:15.040 --> 02:38:19.757 betting or banking on the fact that >> AI is going to make sufficient progress 02:38:19.840 --> 02:38:22.877 towards that >> and they're hoping to discover a product 02:38:22.960 --> 02:38:25.277 in it too is uh 02:38:25.360 --> 02:38:27.037 >> before you have a really strong world 02:38:27.120 --> 02:38:29.917 model there'll be an almost strong world 02:38:30.000 --> 02:38:32.797 model and um 02:38:32.880 --> 02:38:37.117 >> people are trying to find a product in a clumsy robot I suppose like not a 02:38:37.200 --> 02:38:41.197 perfectly efficient robot so there's the factory setting where uh humanoid robots 02:38:41.280 --> 02:38:43.997 can help automate some aspects of the factory. 02:38:44.080 --> 02:38:47.917 >> I think that's a crazy difficult task cuz of all the safety required and all 02:38:48.000 --> 02:38:51.677 this kind of stuff. I think in the home is more interesting, but then you start 02:38:51.760 --> 02:38:55.997 to think I think you mentioned loading the dishwasher, right? 02:38:56.080 --> 02:38:59.117 >> Yeah. >> Like I suppose that's one of the main 02:38:59.200 --> 02:39:03.837 problems you're working on. >> I mean there's, you know, uh cleaning up 02:39:03.920 --> 02:39:08.157 >> Yeah. Yeah, >> cleaning the house, uh, clearing up the 02:39:08.240 --> 02:39:12.717 table after a meal, washing the dishes, you know, all those 02:39:12.800 --> 02:39:16.077 tasks, you know, cooking. I mean, all the tasks that, you know, in principle 02:39:16.160 --> 02:39:19.677 could be automated, but are actually incredibly sophisticated, really 02:39:19.760 --> 02:39:23.037 complicated, >> but even just basic navigation around an 02:39:23.120 --> 02:39:24.877 un space full of uncertainty 02:39:24.960 --> 02:39:26.877 >> that sort of works. Like you can sort of 02:39:26.960 --> 02:39:30.397 do this now. Navigation is fine. Well, 02:39:30.480 --> 02:39:35.677 navigation in a way that's compelling to us humans is is a different thing. 02:39:35.760 --> 02:39:38.957 >> Yeah. It's not going to be, you know, necessarily. I mean, we have demos 02:39:39.040 --> 02:39:41.357 actually because, you know, there is a 02:39:41.440 --> 02:39:44.877 so-called embodied AI group at at fair 02:39:44.960 --> 02:39:48.717 and uh, you know, they've been not building their own robots, but using 02:39:48.800 --> 02:39:53.757 commercial robots. Um, and you can you can tell the robot dog like, you know, 02:39:53.840 --> 02:39:56.237 go to the fridge >> and they can actually open the fridge 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 02:40:00.800 --> 02:40:05.517 bring it to you like, you know, so it can navigate, it can grab objects as 02:40:05.600 --> 02:40:08.717 long as it's been trained to recognize them, which, you know, vision systems 02:40:08.800 --> 02:40:11.997 work pretty well nowadays. Um, but but 02:40:12.080 --> 02:40:15.837 it's not like a completely, you know, general 02:40:15.920 --> 02:40:17.437 robot that would be, you know, 02:40:17.520 --> 02:40:20.477 sophisticated enough to do things like 02:40:20.560 --> 02:40:23.597 clearing up the dinner table. 02:40:23.680 --> 02:40:27.757 Yeah. Yeah, to me that's an exciting future of getting humanoid robots, 02:40:27.840 --> 02:40:31.517 robots in general in the home more and more because that gets uh humans to 02:40:31.600 --> 02:40:36.637 really directly interact with AI systems in the physical space and in so doing it 02:40:36.720 --> 02:40:39.597 allows us to philosophically psychologically explore our 02:40:39.680 --> 02:40:41.997 relationships with robots. It's going to be really 02:40:42.080 --> 02:40:44.237 >> really really interesting. So I hope you 02:40:44.320 --> 02:40:47.117 make progress on the whole uh Japa thing soon. 02:40:47.200 --> 02:40:49.117 >> Well I I mean I hope I hope things kind 02:40:49.200 --> 02:40:53.117 of you know work as uh as planned. Um 02:40:53.200 --> 02:40:56.637 I mean again we've been like kind of working on this idea of self-supervised 02:40:56.720 --> 02:40:59.437 learning of uh from video for for 10 02:40:59.520 --> 02:41:04.157 years and and you know only made significant progress in the last two or 02:41:04.240 --> 02:41:06.237 three >> and actually you've you've mentioned 02:41:06.320 --> 02:41:08.957 that there's a lot of interesting breakthroughs that can happen without 02:41:09.040 --> 02:41:11.117 having access to a lot of compute. Yeah. 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 02:41:15.360 --> 02:41:18.877 possibilities still >> to do innovative work. So like what 02:41:18.960 --> 02:41:22.957 advice would you give to a undergrad that's looking to uh go to grad school 02:41:23.040 --> 02:41:25.197 and do a PhD? 02:41:25.280 --> 02:41:27.597 >> So basically I've listed them already. 02:41:27.680 --> 02:41:31.357 Uh this idea of how do you train a world model by observation. 02:41:31.440 --> 02:41:33.517 >> Mhm. And you don't have to train 02:41:33.600 --> 02:41:38.237 necessarily on gigantic data sets or 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 02:41:41.760 --> 02:41:45.197 have emergent properties like like we have with LLMs. But I think there's a 02:41:45.280 --> 02:41:49.677 lot of good ideas that can be done without necessarily scaling up. Then 02:41:49.760 --> 02:41:54.397 there is how do you do planning with a learn world model. If the world the 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 02:41:59.000 --> 02:42:00.877 [snorts] the internet or you know some 02:42:00.960 --> 02:42:04.237 sort of uh world of where an action 02:42:04.320 --> 02:42:06.317 consists in doing a search in a search 02:42:06.400 --> 02:42:08.797 engine or interrogating a data database 02:42:08.880 --> 02:42:13.517 or running a simulation or calling a calculator or solving a differential 02:42:13.600 --> 02:42:17.117 equation. How do you get a system to actually plan 02:42:17.200 --> 02:42:19.677 a sequence of actions to you know give 02:42:19.760 --> 02:42:24.237 the solution to a problem? Um and so the 02:42:24.320 --> 02:42:28.477 question of planning is not just a question of planning physical actions 02:42:28.560 --> 02:42:31.037 could be you know planning actions to 02:42:31.120 --> 02:42:37.277 use tools for a dialog system or for any kind of intelligence system and um 02:42:37.360 --> 02:42:40.077 there's some [clears throat] work on this but not like not a huge amount some 02:42:40.160 --> 02:42:43.597 work at fair um one called tool former 02:42:43.680 --> 02:42:46.477 which was couple years ago and some more 02:42:46.560 --> 02:42:49.837 recent work on planning uh but um but I 02:42:49.920 --> 02:42:53.757 don't think we have like a good solution for any of that. Then 02:42:53.840 --> 02:42:56.077 there is the question of hierarchical 02:42:56.160 --> 02:42:58.877 planning. So as the example I I 02:42:58.960 --> 02:43:03.277 mentioned of you know planning a trip from New York to Paris 02:43:03.360 --> 02:43:07.197 >> that's hierarchical but almost every action that we take in involves 02:43:07.280 --> 02:43:09.197 hierarchical planning in some in some 02:43:09.280 --> 02:43:12.317 sense and we really have absolutely no 02:43:12.400 --> 02:43:14.557 idea how to do this like there's zero 02:43:14.640 --> 02:43:17.997 demonstration of hierarchical planning 02:43:18.080 --> 02:43:20.877 uh in AI 02:43:20.960 --> 02:43:25.517 where the various levels of 02:43:25.600 --> 02:43:27.677 representations that are necessary have 02:43:27.760 --> 02:43:32.557 been learned. We can do like two-le hierarchy hierarchical planning. When we 02:43:32.640 --> 02:43:34.157 design the two the two levels. So for 02:43:34.240 --> 02:43:37.197 example, you have like a dog like robot, 02:43:37.280 --> 02:43:40.077 right? You wanted to go from the living 02:43:40.160 --> 02:43:41.917 room to the kitchen. You can plan a path 02:43:42.000 --> 02:43:45.997 that avoids the obstacle and then um you 02:43:46.080 --> 02:43:49.437 can send this to a lower lower level planner that figures out how to move the 02:43:49.520 --> 02:43:52.317 legs to kind of follow that trajectories. Right? 02:43:52.400 --> 02:43:54.077 >> So that works. But that two-level 02:43:54.160 --> 02:43:58.237 planning is designed by hand right um we 02:43:58.320 --> 02:44:01.837 specify what the proper levels of 02:44:01.920 --> 02:44:06.077 abstraction the representation at each level of abstraction has have to be how 02:44:06.160 --> 02:44:09.837 do you learn this how do you learn that hierarchical representation of action 02:44:09.920 --> 02:44:14.797 plans right we you know with coets and deep 02:44:14.880 --> 02:44:18.397 learning we we can train the system to learn hierarchical representations of 02:44:18.480 --> 02:44:21.837 percepts >> what is the equivalent when what you're 02:44:21.920 --> 02:44:23.517 trying to represent for action plans. 02:44:23.600 --> 02:44:25.597 >> for action plans. Yeah. So, you want you 02:44:25.680 --> 02:44:27.837 want basically a robot dog or humanoid 02:44:27.920 --> 02:44:33.757 robot that turns on and travels from New York to Paris all by itself, [snorts] 02:44:33.840 --> 02:44:37.117 >> for example. >> All right. They might have some uh 02:44:37.200 --> 02:44:40.077 trouble at the at the TSA. But >> yeah. 02:44:40.160 --> 02:44:43.197 >> No, but even doing something fairly simple like a household task. 02:44:43.280 --> 02:44:46.557 >> Sure. >> Like, you know, uh cooking or something. 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 02:44:50.560 --> 02:44:53.437 once again we take it for granted. 02:44:53.520 --> 02:44:58.797 >> What hope do you have for um the future of humanity? 02:44:58.880 --> 02:45:01.277 We're talking about so many exciting technologies, so many exciting 02:45:01.360 --> 02:45:03.357 possibilities. What gives you hope when 02:45:03.440 --> 02:45:06.717 you look out over the next 10, 20, 50, 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 02:45:10.640 --> 02:45:13.677 going on. There's division. Uh there's 02:45:13.760 --> 02:45:15.837 hatred. All this kind of stuff that's 02:45:15.920 --> 02:45:18.397 also part of humanity. But amidst all 02:45:18.480 --> 02:45:21.757 that, what gives you hope? 02:45:21.840 --> 02:45:26.440 >> I love that question. Uh, 02:45:27.280 --> 02:45:32.077 we can make humanity smarter with AI. 02:45:32.160 --> 02:45:35.917 Okay? I mean, AI basically will amplify 02:45:36.000 --> 02:45:37.997 human intelligence. 02:45:38.080 --> 02:45:41.917 It's as if every one of us will have a 02:45:42.000 --> 02:45:45.277 staff of smart AI assistants. They might 02:45:45.360 --> 02:45:51.000 be smarter than us. They'll do our bidding, 02:45:51.680 --> 02:45:55.117 perhaps execute uh tasks in ways that 02:45:55.200 --> 02:45:59.197 are much better than we could do ourselves because they' be smarter than 02:45:59.280 --> 02:46:02.077 us. And so it's like everyone would be 02:46:02.160 --> 02:46:06.477 the the boss of a staff of super smart 02:46:06.560 --> 02:46:10.557 virtual people. So we shouldn't feel threatened by by 02:46:10.640 --> 02:46:15.437 this any more than we should feel threatened by being the manager of a 02:46:15.520 --> 02:46:19.917 group of people some of whom are more intelligent than us. 02:46:20.000 --> 02:46:21.437 I certainly have a lot of experience 02:46:21.520 --> 02:46:25.597 with this of uh you know having people 02:46:25.680 --> 02:46:27.597 working with me who are smarter than me. 02:46:27.680 --> 02:46:30.397 Um that's actually a wonderful thing. So 02:46:30.480 --> 02:46:32.797 uh having machines that are smarter than 02:46:32.880 --> 02:46:37.437 us that assist us in our all of our tasks our daily lives whether it's 02:46:37.520 --> 02:46:41.277 professional or personal I think would be a absolutely wonderful thing because 02:46:41.360 --> 02:46:44.157 intelligence is the most 02:46:44.240 --> 02:46:48.557 um is the commodity that is most in demand that that's really what I mean 02:46:48.640 --> 02:46:51.917 all the mistakes that humanity makes is because of lack of intelligence really 02:46:52.000 --> 02:46:53.997 or lack of knowledge which is you know 02:46:54.080 --> 02:46:55.837 related. 02:46:55.920 --> 02:46:58.957 So um making people smarter we just can 02:46:59.040 --> 02:47:02.877 only be better. I mean for the same reason that you know public education is 02:47:02.960 --> 02:47:08.237 a good thing and books are a good thing and the 02:47:08.320 --> 02:47:11.517 internet is also a good thing intrinsically and even social networks 02:47:11.600 --> 02:47:13.357 are a good thing if you run them 02:47:13.440 --> 02:47:15.677 properly. It's difficult but you know 02:47:15.760 --> 02:47:18.557 you can um 02:47:18.640 --> 02:47:22.877 uh because you know it it 02:47:22.960 --> 02:47:26.637 helps the communication of information and knowledge and the transmission of 02:47:26.720 --> 02:47:30.637 knowledge. So AI is going to make humanity smarter. 02:47:30.720 --> 02:47:33.437 And the analogy I've been using 02:47:33.520 --> 02:47:37.517 is the fact that perhaps an equivalent 02:47:37.600 --> 02:47:41.677 event in the history of humanity to what 02:47:41.760 --> 02:47:45.037 might be provided by journalization of 02:47:45.120 --> 02:47:47.037 AI assistant is the invention of the 02:47:47.120 --> 02:47:48.717 printing the printing press. It made 02:47:48.800 --> 02:47:52.077 everybody smarter. The fact that 02:47:52.160 --> 02:47:55.437 people could uh have access to um to 02:47:55.520 --> 02:47:58.077 books books were a lot cheaper than they 02:47:58.160 --> 02:48:01.597 were before. And so a lot more people 02:48:01.680 --> 02:48:03.757 had an incentive to learn to read, which 02:48:03.840 --> 02:48:09.517 wasn't the case before. Um, and 02:48:09.600 --> 02:48:13.357 people became smarter. It it enabled the enlightenment, 02:48:13.440 --> 02:48:16.317 right? There [snorts] wouldn't be an enlightenment without the printing 02:48:16.400 --> 02:48:19.117 press. It enabled 02:48:19.200 --> 02:48:20.877 uh 02:48:20.960 --> 02:48:23.677 philosophy, rationalism, 02:48:23.760 --> 02:48:27.837 u escape from religious doctrine, 02:48:27.920 --> 02:48:30.557 um democracy, 02:48:30.640 --> 02:48:32.957 science, 02:48:33.040 --> 02:48:38.077 uh and certainly without this it wouldn't be there wouldn't have been the 02:48:38.160 --> 02:48:41.997 American Revolution or the French Revolution and so would still be under 02:48:42.080 --> 02:48:43.757 feudal 02:48:43.840 --> 02:48:47.597 regimes perhaps. Um and so it completely 02:48:47.680 --> 02:48:51.437 transformed the the world because people 02:48:51.520 --> 02:48:53.517 became smarter and kind of learned 02:48:53.600 --> 02:48:56.637 learned about things. Now it also 02:48:56.720 --> 02:49:01.597 created 200 years of essentially religious conflicts in Europe, right? 02:49:01.680 --> 02:49:03.517 Because the first thing that people read 02:49:03.600 --> 02:49:08.797 was the Bible and uh realized that perhaps there was a different 02:49:08.880 --> 02:49:12.877 interpretation of the Bible than what the priests were telling them. And so 02:49:12.960 --> 02:49:15.357 that created the Protestant movement. 02:49:15.440 --> 02:49:19.277 and created a rift. And in fact, the Catholic school, the Catholic Church 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 02:49:23.440 --> 02:49:26.557 had some bad effects and some some good effects. I don't think anyone today 02:49:26.640 --> 02:49:30.237 would say that the invention of the printing press had a overall negative 02:49:30.320 --> 02:49:33.197 effect despite the fact that it created 02:49:33.280 --> 02:49:36.077 200 years of religious conflicts 02:49:36.160 --> 02:49:40.077 in Europe. Now compare this and I I 02:49:40.160 --> 02:49:44.877 thought uh I was very proud of myself to come up with this analogy but realized 02:49:44.960 --> 02:49:49.197 someone else u came with the same idea before me. Um compare this with what 02:49:49.280 --> 02:49:51.757 happened in the Ottoman Empire. The 02:49:51.840 --> 02:49:55.357 Ottoman Empire banned the printing press 02:49:55.440 --> 02:49:57.997 for 200 years. 02:49:58.080 --> 02:50:02.477 Uh and it didn't ban it uh for all 02:50:02.560 --> 02:50:04.957 languages only for Arabic. You could 02:50:05.040 --> 02:50:08.557 actually print books in Latin or Hebrew 02:50:08.640 --> 02:50:10.557 or whatever in the Ottoman [snorts] 02:50:10.640 --> 02:50:13.437 Empire, just not in Arabic. 02:50:13.520 --> 02:50:16.477 And uh 02:50:16.560 --> 02:50:19.437 I thought it was because the rulers just 02:50:19.520 --> 02:50:21.517 wanted to preserve the control over the 02:50:21.600 --> 02:50:23.917 population and the dogma, religious 02:50:24.000 --> 02:50:26.637 dogma and everything. But after talking 02:50:26.720 --> 02:50:31.997 with the uh UAE Minister of AI, uh Omar 02:50:32.080 --> 02:50:34.477 Alma 02:50:34.560 --> 02:50:36.477 Um he told me no there was another 02:50:36.560 --> 02:50:40.877 reason uh and the other reason was that 02:50:40.960 --> 02:50:44.237 uh it was to preserve the corporation of 02:50:44.320 --> 02:50:46.237 calligraphers 02:50:46.320 --> 02:50:51.437 right there's like an art form which is you know writing those beautiful 02:50:51.520 --> 02:50:55.597 >> yes >> uh you know Arabic uh poems or whatever 02:50:55.680 --> 02:50:57.997 religious text in in this thing and it 02:50:58.080 --> 02:51:00.237 was very powerful corporation of scribes 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 02:51:05.760 --> 02:51:07.197 couldn't put them out of business. So 02:51:07.280 --> 02:51:13.800 they you know banned the printing press in part to protect that business. 02:51:14.160 --> 02:51:15.917 >> Now what's the analogy for AI today? 02:51:16.000 --> 02:51:19.197 Like who are we protecting by banning AI? Like who are the people who are 02:51:19.280 --> 02:51:22.157 asking that AI be regulated to protect 02:51:22.240 --> 02:51:25.037 their their jobs? And of course, you 02:51:25.120 --> 02:51:26.797 know, there's it's a it's a it's a real 02:51:26.880 --> 02:51:29.837 question of what is going to be the 02:51:29.920 --> 02:51:33.037 effect of uh you know, technological 02:51:33.120 --> 02:51:35.757 transformation like AI on the on the job 02:51:35.840 --> 02:51:38.957 market and the labor market. And there 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 02:51:42.880 --> 02:51:47.117 them, they they tell us, you know, we're not going to run out of 02:51:47.200 --> 02:51:51.197 job. This this not this is not going to cause mass unemployment. this this is 02:51:51.280 --> 02:51:55.837 just going to be gradual uh shift of different professions. The professions 02:51:55.920 --> 02:51:58.077 that are going to be hot 10 or 15 years 02:51:58.160 --> 02:52:00.877 from now, we have no idea today what 02:52:00.960 --> 02:52:03.037 they're going to be. The same way if we 02:52:03.120 --> 02:52:05.357 go back 20 years in the past, like who 02:52:05.440 --> 02:52:07.997 could have thought 20 years ago that 02:52:08.080 --> 02:52:11.357 like the hottest job even like 5 10 02:52:11.440 --> 02:52:16.157 years ago was mobile app developer. Like smartphones weren't invented. 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. 02:52:20.880 --> 02:52:23.677 Yeah, >> but the point is you can't possibly 02:52:23.760 --> 02:52:28.957 predict. But you're right. I mean, you made a lot of strong points and 02:52:29.040 --> 02:52:30.237 I I believe that people are 02:52:30.320 --> 02:52:34.157 fundamentally good. And so if AI, 02:52:34.240 --> 02:52:38.797 especially open source AI, can um [clears throat] make them smarter, it 02:52:38.880 --> 02:52:41.117 just empowers the goodness in humans. 02:52:41.200 --> 02:52:43.837 >> So I I share that feeling. Okay. I think 02:52:43.920 --> 02:52:47.517 people are probably good. Uh and in fact 02:52:47.600 --> 02:52:49.757 a lot of doomers are doomers because 02:52:49.840 --> 02:52:51.357 they don't think that people are 02:52:51.440 --> 02:52:55.277 fundamentally good. Uh and they either 02:52:55.360 --> 02:52:57.837 don't trust people or they don't trust 02:52:57.920 --> 02:53:00.717 the institution to do the right thing so 02:53:00.800 --> 02:53:03.357 that people behave properly. 02:53:03.440 --> 02:53:05.437 Well, I think both you and I believe in 02:53:05.520 --> 02:53:07.997 humanity and I think I speak for a lot 02:53:08.080 --> 02:53:10.557 of people in saying thank you for 02:53:10.640 --> 02:53:13.117 pushing the open source movement pushing 02:53:13.200 --> 02:53:16.237 to making both researching AI open 02:53:16.320 --> 02:53:20.717 source, making it available to people and also the models themselves making 02:53:20.800 --> 02:53:24.957 that open source. So, thank you for that and uh thank you for speaking your mind 02:53:25.040 --> 02:53:28.957 in such colorful, beautiful ways on the internet. I hope you never stop. You're 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 02:53:33.600 --> 02:53:36.717 for speaking to me once again and thank you for being you. 02:53:36.800 --> 02:53:39.437 >> Thank you. Next, >> thanks for listening to this 02:53:39.520 --> 02:53:43.117 conversation with Yan Lun. To support this podcast, please check out our 02:53:43.200 --> 02:53:47.117 sponsors in the description. And now, let me leave you with some words from 02:53:47.200 --> 02:53:49.757 Arthur C. Clark. 02:53:49.840 --> 02:53:51.437 The only way to discover the limits of 02:53:51.520 --> 02:53:55.117 the possible is to go beyond them into 02:53:55.200 --> 02:53:57.357 the impossible. 02:53:57.440 --> 02:54:03.080 Thank you for listening and hope to see you next time. 02:54:09.330 --> 02:54:11.350 >> [music]