1 00:00:00,080 --> 00:00:02,637 We're dealing with magic here that we I think probably all would have thought 2 00:00:02,720 --> 00:00:05,677 was impossible 5 years ago or certainly 10 years ago. This is the most amazing 3 00:00:05,760 --> 00:00:07,997 technology ever and it's moving really fast and yet we're still like really 4 00:00:08,080 --> 00:00:11,197 disappointed. Like it's not moving fast enough and like it's like maybe right on 5 00:00:11,280 --> 00:00:13,917 the verge of stalling out. We should both be like hyper excited but also on 6 00:00:14,000 --> 00:00:17,117 the verge of like slitting our wrists cuz like you know the gravy train is 7 00:00:17,200 --> 00:00:20,557 coming to an end, >> right? >> It is faster but it's not at computer 8 00:00:20,640 --> 00:00:23,597 speed, right? What we expect computer speed to be. It's sort of like watching 9 00:00:23,680 --> 00:00:24,637 a person work. 10 00:00:24,720 --> 00:00:28,477 >> It's like watching John Carmarmac 11 00:00:28,560 --> 00:00:30,637 >> the world. Okay. the world's the world's 12 00:00:30,720 --> 00:00:32,557 best programmer on a stimulus. 13 00:00:32,640 --> 00:00:36,680 >> On a stimulant. Yeah, that's right. 14 00:00:37,200 --> 00:00:41,277 >> So, let's start with um let's assume that I'm a sort of a novice programmer. 15 00:00:41,360 --> 00:00:44,957 So, maybe I'm a student um uh or maybe I'm just somebody, you know, I took a 16 00:00:45,040 --> 00:00:47,437 few coding classes and I've hacked around a little bit or like I don't 17 00:00:47,520 --> 00:00:51,277 know, I do Excel macros or something like that, but I'm like not less. I'm 18 00:00:51,360 --> 00:00:53,597 not like a master craftsman at coding. 19 00:00:53,680 --> 00:00:57,917 Um and you know people somebody tells me about replet and and specifically AI um 20 00:00:58,000 --> 00:01:01,757 AI and Replet like what's my what's my experience uh when when I launch in with 21 00:01:01,840 --> 00:01:03,677 with what replet is today with AI. 22 00:01:03,760 --> 00:01:07,597 >> Yeah I I would um I I think the experience of someone with no coding 23 00:01:07,680 --> 00:01:10,797 experience or some coding experience is largely the same when you go into 24 00:01:10,880 --> 00:01:13,997 replet. Okay. >> The first thing we try to do is get all 25 00:01:14,080 --> 00:01:17,757 the nonsense away from like setting up development environment and all of that 26 00:01:17,840 --> 00:01:21,437 stuff and just have you focus on your idea. So what do you want to build? Do 27 00:01:21,520 --> 00:01:24,637 you want to build a product? Do you want to solve a problem? Do you want to do a 28 00:01:24,720 --> 00:01:28,317 data vis visualization? So the prompt box is really open for you. You can put 29 00:01:28,400 --> 00:01:32,397 in anything there. So let's say you want to, you know, build a startup. You have 30 00:01:32,480 --> 00:01:34,557 an idea for a startup. I would I would 31 00:01:34,640 --> 00:01:37,517 start with like a paragraph long kind of 32 00:01:37,600 --> 00:01:42,477 description of what I want to build. Uh the agents will read that. It will 33 00:01:42,560 --> 00:01:45,517 >> you just type just type >> standard English. Standard English. You 34 00:01:45,600 --> 00:01:49,677 just type it in. I want to build a I want to sell I want to sell crepes. I 35 00:01:49,760 --> 00:01:53,357 want to sell crepes online. So you just like type in I want to talk. 36 00:01:53,440 --> 00:01:56,637 >> You can it literally could be that four words or five words. Okay. 37 00:01:56,720 --> 00:02:00,797 >> Or it could be if you're if you have a programming language you prefer or stack 38 00:02:00,880 --> 00:02:04,317 you prefer, you could do that. But we actually prefer not for you not to do 39 00:02:04,400 --> 00:02:07,917 that because we're going to pick the best thing for we're going to classify 40 00:02:08,000 --> 00:02:10,797 the best stack for that request. Right? 41 00:02:10,880 --> 00:02:14,397 >> It's a if it's a data app, we'll pick Python and stream whatever. If it's like 42 00:02:14,480 --> 00:02:18,237 a web app, we'll pick JavaScript and Postgress and things like that. So you 43 00:02:18,320 --> 00:02:19,997 just type that >> or you can decide you can decide you can 44 00:02:20,080 --> 00:02:23,037 say and I want to do it I know Python or I'm learning Python in school and I want 45 00:02:23,120 --> 00:02:25,437 to do it in Python. >> That's right. The cool thing about 46 00:02:25,520 --> 00:02:29,117 Replet is you know we we've been around for almost 10 years now and we built all 47 00:02:29,200 --> 00:02:32,637 this infrastructure. Replet runs any programming language. So if you're 48 00:02:32,720 --> 00:02:35,277 comfortable with Python you can go in and do that for sure. 49 00:02:35,360 --> 00:02:37,437 >> Okay. >> And then just again I know this is 50 00:02:37,520 --> 00:02:39,997 obvious people have used it but like I'm dealing in English. 51 00:02:40,080 --> 00:02:45,197 >> Yes. >> So okay go ahead. >> Yes. You're fully in English. I mean, 52 00:02:45,280 --> 00:02:46,877 you know, just a, you know, a little bit 53 00:02:46,960 --> 00:02:49,757 of of sort of background here, like when 54 00:02:49,840 --> 00:02:54,477 um when I I came here and pitched to you like 10 years ago or like whatever 7 55 00:02:54,560 --> 00:02:57,917 years ago, >> right? >> Uh what we were saying is we were 56 00:02:58,000 --> 00:03:01,757 exactly describing this future is that >> uh everyone would want to build 57 00:03:01,840 --> 00:03:04,637 software, right? >> And the thing that's kind of getting in 58 00:03:04,720 --> 00:03:09,677 in people's ways is all the uh what Fred Brooks called the accidental complexity 59 00:03:09,760 --> 00:03:12,957 of programming, right? They're like essential complexity which is like how 60 00:03:13,040 --> 00:03:16,557 do I bring my startup to market and how 61 00:03:16,640 --> 00:03:20,477 do I build a business and all of that accidental complexity is what package 62 00:03:20,560 --> 00:03:22,477 manager do I use all of that stuff we've 63 00:03:22,560 --> 00:03:24,877 been abstracting away that for so many 64 00:03:24,960 --> 00:03:30,077 years so you can just um and the last thing we had to abstract away is code 65 00:03:30,160 --> 00:03:32,637 >> right >> I had this realization last year which 66 00:03:32,720 --> 00:03:36,397 is I think we you know built an amazing platform but the business is not 67 00:03:36,480 --> 00:03:39,117 performing and the reason the business is not performing is that code is the 68 00:03:39,200 --> 00:03:41,517 bottleneck like yes all the other stuff 69 00:03:41,600 --> 00:03:46,157 is important to solve but syntax is still an issue like you know syntax is 70 00:03:46,240 --> 00:03:50,557 just an unnatural thing for people so ultimately English is the programming 71 00:03:50,640 --> 00:03:54,957 language >> right >> I I just does it work with other other 72 00:03:55,040 --> 00:03:56,717 world languages other than English at this point 73 00:03:56,800 --> 00:04:00,637 >> yes you can write in Japanese and we have a lot of users especially Japanese 74 00:04:00,720 --> 00:04:02,957 that tends to be very >> so does it support these days like for a 75 00:04:03,040 --> 00:04:06,077 does a support every language or is it still do you still have to do like 76 00:04:06,160 --> 00:04:09,677 custom work to craft a new new language >> no most most you know uh mainstream 77 00:04:09,760 --> 00:04:13,357 dream language that has like 100 million plus people that speak it. AI is pretty 78 00:04:13,440 --> 00:04:15,197 good at it. >> Okay. Yeah. >> Yeah. Wow. 79 00:04:15,280 --> 00:04:19,917 >> So, uh I I I did a bit of a bit of historical research recently for for 80 00:04:20,000 --> 00:04:22,877 some reason. I just want to just understand the moment we're in and 81 00:04:22,960 --> 00:04:24,637 because it's such a special moment. It's 82 00:04:24,720 --> 00:04:26,797 I think it's important to contextualize 83 00:04:26,880 --> 00:04:30,557 it and I I I read this quote from 84 00:04:30,640 --> 00:04:32,957 Gracehopper. So, Gracehopper invented 85 00:04:33,040 --> 00:04:36,077 the compiler as you know. uh at the time 86 00:04:36,160 --> 00:04:40,877 people were uh you know programming in machine code and that's what programmers 87 00:04:40,960 --> 00:04:45,437 do that's what the specialists do >> yes >> and she said you know specialists will 88 00:04:45,520 --> 00:04:48,557 always be a specialist they have to learn the underlying machinery of 89 00:04:48,640 --> 00:04:52,397 computers but I want to get to a world where people are programming English 90 00:04:52,480 --> 00:04:56,317 that's what she said that's before karpathy right that's like you know 75 91 00:04:56,400 --> 00:05:00,397 years ago >> uh and and and that's why I invented the 92 00:05:00,480 --> 00:05:04,477 compiler and in her mind like C programming is English >> right 93 00:05:04,560 --> 00:05:08,797 >> uh But that, you know, that really didn't uh that was just the start of it. 94 00:05:08,880 --> 00:05:13,277 You had C and then you go higher level Python and JavaScript. And I think it 95 00:05:13,360 --> 00:05:15,997 we're at a moment where it's the next step, >> right? 96 00:05:16,080 --> 00:05:19,597 >> Instead of typing syntax, you're actually typing thoughts, you know, 97 00:05:19,680 --> 00:05:21,037 which is what we ultimately want. 98 00:05:21,120 --> 00:05:22,957 >> And the machine writes the code >> and the machine writes the code, 99 00:05:23,040 --> 00:05:25,197 >> right? Right. >> Um yeah, I remember it. you're you're 100 00:05:25,280 --> 00:05:28,237 probably not old enough uh to remember but I I remember when when I was a kid 101 00:05:28,320 --> 00:05:31,517 it was um you know there there were were higher level languages you know by the 102 00:05:31,600 --> 00:05:33,757 70s like like basic and so forth and forran 103 00:05:33,840 --> 00:05:37,517 >> and C and C but um uh there were still you know you still would run into people 104 00:05:37,600 --> 00:05:40,797 who were doing assembly programming assembly language which by the way you 105 00:05:40,880 --> 00:05:43,837 still do you know like game companies or whatever still do assembly to to to get 106 00:05:43,920 --> 00:05:46,877 >> and they were hating on the kids that were doing basic. Oh well so so the 107 00:05:46,960 --> 00:05:48,957 assembly people were hating the kids doing basic but there were also older 108 00:05:49,040 --> 00:05:51,597 coders who hated on the assembly programmers for doing assembly and not 109 00:05:51,680 --> 00:05:55,357 and not and not no no no not doing direct machine code right not doing 110 00:05:55,440 --> 00:05:58,637 direct zero in one machine code because because as because assembly assembly so 111 00:05:58,720 --> 00:06:01,517 people don't know assembly language is sort of this very low-level programming 112 00:06:01,600 --> 00:06:05,357 language that sort of compiles to actual actual machine code and if and if it's 113 00:06:05,440 --> 00:06:08,237 it's it's incomprehensible gibberish to most program even most programmers 114 00:06:08,320 --> 00:06:10,477 >> you're writing an octal or something >> you're writing like very very close to 115 00:06:10,560 --> 00:06:14,317 the hardware but even still it's still a language that compiles to zeros and ones 116 00:06:14,400 --> 00:06:18,077 >> um whereas the actual real programmers actually wrote in zeros and ones. And so 117 00:06:18,160 --> 00:06:20,797 there there's always there's always this tendency, you know, for the for the for 118 00:06:20,880 --> 00:06:22,637 the pros to be, you know, look on the nose. 119 00:06:22,720 --> 00:06:24,317 >> Yeah. >> And say, you know, the new people are 120 00:06:24,400 --> 00:06:26,637 being are being, you know, basically sloppy. They don't understand what's 121 00:06:26,720 --> 00:06:28,957 happening. You know, they don't really understand the machine. And then, of 122 00:06:29,040 --> 00:06:31,597 course, you know, with with the higher level with the higher level abstractions 123 00:06:31,680 --> 00:06:35,517 do is they make they democratize. The absolute irony is I was part of the 124 00:06:35,600 --> 00:06:40,397 JavaScript revolution. I was at Facebook uh before starting repled and we built 125 00:06:40,480 --> 00:06:44,557 the modern JavaScript stack. We built ReactJS and all the tooling around it 126 00:06:44,640 --> 00:06:48,397 >> and we got a lot of hate from from the programmers that you should type you 127 00:06:48,480 --> 00:06:52,957 know vanilla JavaScript directly and >> um I was like okay whatever and then 128 00:06:53,040 --> 00:06:56,797 that you and now that's mainstream and then those guys that built their careers 129 00:06:56,880 --> 00:07:01,597 on the last wave we invented are hating on this new wave and so just you know 130 00:07:01,680 --> 00:07:04,957 people never change. Okay, got it. Okay, so you you're typing English I want to 131 00:07:05,040 --> 00:07:07,437 sell crepes online. I want to do this. I want to have a t-shirt. Whatever the 132 00:07:07,520 --> 00:07:09,037 business is. Okay. What what happens then? 133 00:07:09,120 --> 00:07:12,637 >> Yeah. and then uh uh replet agent will 134 00:07:12,720 --> 00:07:17,357 show you what it understood. So it's trying to build um a common 135 00:07:17,440 --> 00:07:20,877 understanding between you and it and I think there's a lot of things we can do 136 00:07:20,960 --> 00:07:24,557 better there in terms of UI but for now it'll show you a list of tasks. 137 00:07:24,640 --> 00:07:28,237 >> It'll tell you I'm going to go set up a database because you need to store your 138 00:07:28,320 --> 00:07:32,317 data somewhere. Uh we need to set up Shopify or Stripe because we need to 139 00:07:32,400 --> 00:07:36,637 accept payments. Uh and then it shows you this list and gives you two options 140 00:07:36,720 --> 00:07:40,077 initially. Do you want to start with a design so that we can iterate back and 141 00:07:40,160 --> 00:07:43,757 forth to get lock that design down or do you want to build a full thing? 142 00:07:43,840 --> 00:07:47,917 >> Hey, if you want to build a full thing, we'll go for 20, 30, 40 minutes. 143 00:07:48,000 --> 00:07:52,797 >> Uh, and the a and be like the agent will tell you go here, install the app. 144 00:07:52,880 --> 00:07:54,637 >> Uh, I'm going to go set up the database, 145 00:07:54,720 --> 00:07:58,637 do the migrations, write the SQL, you know, build the site. I'm going to also 146 00:07:58,720 --> 00:08:00,717 test it. So this is a recent innovation 147 00:08:00,800 --> 00:08:03,997 we did with um agent 3 is that after it 148 00:08:04,080 --> 00:08:06,397 writes the software spins up a browser 149 00:08:06,480 --> 00:08:11,117 goes around and tests in the browser and then any issue it kind of iterates kind 150 00:08:11,200 --> 00:08:14,397 of goes and fix the code. So it'll spend 20 30 minutes building that I'll send 151 00:08:14,480 --> 00:08:17,117 you a notification it'll tell you the app is ready. And so you can test it on 152 00:08:17,200 --> 00:08:21,757 your phone. You can go back to your computer. You'll see maybe you'll find a 153 00:08:21,840 --> 00:08:24,957 bug or an issue, you'll describe it to the agent, say, "Hey, it's not exactly 154 00:08:25,040 --> 00:08:29,437 doing what I expected." Uh or if it's perfect and and you're ready to go and 155 00:08:29,520 --> 00:08:32,317 that's it. You know, 20 minutes. By the way, there's a lot of examples where 156 00:08:32,400 --> 00:08:36,957 people just get their idea in 20 30 minutes, which is amazing. Um you just 157 00:08:37,040 --> 00:08:38,077 hit publish. >> Mhm. 158 00:08:38,160 --> 00:08:41,997 >> You hit you hit publish. Um 159 00:08:42,080 --> 00:08:44,237 couple clicks, you'll be up in the 160 00:08:44,320 --> 00:08:49,357 cloud. we'll set up a a virtual machine in the cloud. The database is deployed. 161 00:08:49,440 --> 00:08:51,677 Everything's done and now you have a production database. 162 00:08:51,760 --> 00:08:56,237 >> So, think about the steps needed just two or 3 years ago in order to get to 163 00:08:56,320 --> 00:08:59,277 that step. You have to set up your local development environment. You have to 164 00:08:59,360 --> 00:09:02,637 sign up for an AWS account. You have to provision the databases, the virtual 165 00:09:02,720 --> 00:09:06,957 machines, you have to create the entire pip deployment pipeline. All of that is 166 00:09:07,040 --> 00:09:09,037 done for you. And it just, you know, a 167 00:09:09,120 --> 00:09:11,277 kid can do it, a lay person can do it. 168 00:09:11,360 --> 00:09:16,237 If you're a programmer and uh you're curious about what the agent did, the 169 00:09:16,320 --> 00:09:19,517 cool thing about replet because we have this history of being an IDE, you can 170 00:09:19,600 --> 00:09:23,677 peel the layers. You can open the file tree and you could look at the files. 171 00:09:23,760 --> 00:09:26,557 You can open gits, you can push it to GitHub, you can connect it to your 172 00:09:26,640 --> 00:09:30,077 editor if you want, you can open it in Emacs. So the cool thing about Replet, 173 00:09:30,160 --> 00:09:34,157 yes, it is a vibe coding platform that abstracts away all the complexities, but 174 00:09:34,240 --> 00:09:36,077 all the layers are there for you to look at. 175 00:09:36,160 --> 00:09:40,237 >> Right. So let's go let's go back to um that was great, but let's go back to you 176 00:09:40,320 --> 00:09:43,597 said it. It it gives you that the a the agent gives you you you say I've got my 177 00:09:43,680 --> 00:09:46,317 idea. You plug it in and it says it gives you this list of things and then 178 00:09:46,400 --> 00:09:48,477 you and then when you describe it you said I'm going to do this. I'm going to 179 00:09:48,560 --> 00:09:52,477 do that. The eye there in that case was the agent as opposed to the user. Yes. 180 00:09:52,560 --> 00:09:55,517 >> And so the the agent lists the set of things that it's going to do and then 181 00:09:55,600 --> 00:09:57,197 the agent actually does those things. 182 00:09:57,280 --> 00:10:00,237 >> Agent does those things. Yeah. That that's a that's a that's a very 183 00:10:00,320 --> 00:10:03,357 important point. when we did this shift, 184 00:10:03,440 --> 00:10:07,917 we hadn't realized internally at Replet how much the actual user stopped being 185 00:10:08,000 --> 00:10:10,797 the human user and it's actually the agent programmer, >> right? 186 00:10:10,880 --> 00:10:12,957 >> So, one really uh funny thing happened 187 00:10:13,040 --> 00:10:17,997 is we had servers in Asia uh and we the reason we had servers in Asia because we 188 00:10:18,080 --> 00:10:19,917 wanted our Indian or you know Japanese 189 00:10:20,000 --> 00:10:23,357 users to be to have a you know shorter 190 00:10:23,440 --> 00:10:27,037 uh time to the servers. uh when we 191 00:10:27,120 --> 00:10:30,637 launched the agent their experience got significantly worse and we're like what 192 00:10:30,720 --> 00:10:33,757 happened like it's supposed to be faster well turns out it's worse it's because 193 00:10:33,840 --> 00:10:38,237 the AIS are sitting in uh in United States and so the the programmer is 194 00:10:38,320 --> 00:10:41,357 actually in United States it's you're sending the request to the programmer 195 00:10:41,440 --> 00:10:45,277 and the programmer is interfacing with a machine across the world and so yes 196 00:10:45,360 --> 00:10:49,997 suddenly the agent is the programmer okay so like the the ter ter you know 197 00:10:50,080 --> 00:10:53,917 new terminology agent is a software program that is basically using the rest 198 00:10:54,000 --> 00:10:57,597 of the as if it were a as if it were a human 199 00:10:57,680 --> 00:10:59,277 user, but it's not. It's a it's a bot. 200 00:10:59,360 --> 00:11:00,957 >> That's right. It has access to tools 201 00:11:01,040 --> 00:11:03,357 such as write a file, edit a file, 202 00:11:03,440 --> 00:11:06,877 delete a file, uh uh search the package 203 00:11:06,960 --> 00:11:11,677 index, install a package, uh provision a database, provision object object 204 00:11:11,760 --> 00:11:15,437 storage. It is a programmer that has the tools and interface. It has a sort of an 205 00:11:15,520 --> 00:11:18,317 interface >> that that is very similar to a human 206 00:11:18,400 --> 00:11:22,317 programmer. And then um you know we'll talk more about how this all works but a 207 00:11:22,400 --> 00:11:27,037 debate inside the AI industry um is with these was kind of this you know this 208 00:11:27,120 --> 00:11:29,917 idea now of having agents that do things on your behalf and then go out you know 209 00:11:30,000 --> 00:11:32,717 go go out and kind of accomplish missions. Um there's this you know kind 210 00:11:32,800 --> 00:11:36,237 of debate which is okay how like obviously you know it's a big deal even 211 00:11:36,320 --> 00:11:38,717 to have an AI agent that can do relatively simple things to do complex 212 00:11:38,800 --> 00:11:41,357 things of course is you know one of the great technical challenges of the last 213 00:11:41,440 --> 00:11:44,877 80 years you know to to do that and then there's sort of this question of like 214 00:11:44,960 --> 00:11:48,157 can the agent go out and run and operate on its own for 5 minutes you know for 215 00:11:48,240 --> 00:11:53,437 for 15 minutes for an hour for 8 hours and and meaning like sort of like how 216 00:11:53,520 --> 00:11:56,397 long does it maintain coherence like how long does it actually like stay in full 217 00:11:56,480 --> 00:12:00,077 control of it of its faculties and not kind of spin out because at least the 218 00:12:00,160 --> 00:12:03,197 early early agents or the the early AIS, if if you set them off to do this, they 219 00:12:03,280 --> 00:12:05,357 might be able to run for two or three minutes, then they would they would 220 00:12:05,440 --> 00:12:07,997 start to get confused and go down rabbit holes and, you know, kind of kind of 221 00:12:08,080 --> 00:12:11,997 spin out. Um, more recently, more recently, um, uh, you know, we've seen 222 00:12:12,080 --> 00:12:15,197 that that that that agents can run a lot longer and and do more complex tasks. 223 00:12:15,280 --> 00:12:19,597 Like, where are we on the curve of agents being able to run for how long 224 00:12:19,680 --> 00:12:22,477 and for what complexity tasks before before they break? 225 00:12:22,560 --> 00:12:24,557 >> That's that's absolutely the the I think 226 00:12:24,640 --> 00:12:29,197 the main metric we're looking at. even back in 2023, you know, had the idea for 227 00:12:29,280 --> 00:12:33,117 software agents, you know, four or five years ago now. The problem every time we 228 00:12:33,200 --> 00:12:37,117 attempt them, the the problem of coherence, you know, they'll they'll go 229 00:12:37,200 --> 00:12:41,357 on for a minute or two and then they'll just, you know, they compound in errors 230 00:12:41,440 --> 00:12:43,837 in a way that they just can't recover. 231 00:12:43,920 --> 00:12:45,037 >> Um, >> and you can actually see it, right? 232 00:12:45,120 --> 00:12:47,757 Because they actually they actually, if you watch watch them operate, they get 233 00:12:47,840 --> 00:12:51,037 increasingly confused and then, you know, maybe even deranged. Yeah, they 234 00:12:51,120 --> 00:12:54,237 vary the range and they they go into very weird areas and sometimes they 235 00:12:54,320 --> 00:12:56,157 start speaking Chinese and doing really 236 00:12:56,240 --> 00:12:59,437 weird things and um but I would say 237 00:12:59,520 --> 00:13:03,437 sometime around last year we maybe 238 00:13:03,520 --> 00:13:05,917 crossed the 3 four five minute mark 239 00:13:06,000 --> 00:13:09,037 >> and it felt to us that okay we're on a 240 00:13:09,120 --> 00:13:12,397 path where long re you know long horizon 241 00:13:12,480 --> 00:13:17,197 reasoning is getting solved >> uh and so we made we made a bet and I I 242 00:13:17,280 --> 00:13:20,157 tell my team >> so long horizon reasoning meaning 243 00:13:20,240 --> 00:13:23,837 reasoning meaning like dealing in like facts and logic 244 00:13:23,920 --> 00:13:28,157 >> um in a in a sort of complex way and then long horizon being over a long 245 00:13:28,240 --> 00:13:31,037 period of time. Yes. >> With many many steps to a reasoning process. 246 00:13:31,120 --> 00:13:34,157 >> Yeah, that's right. So if you think about the way large language models work 247 00:13:34,240 --> 00:13:37,757 is that they have a context. This context is basically the memory all the 248 00:13:37,840 --> 00:13:42,797 text all your prompt and also all the internal talk that the AI is doing as 249 00:13:42,880 --> 00:13:45,437 it's reasoning. So when the AI is reasoning it's actually talking to 250 00:13:45,520 --> 00:13:49,357 itself. It's like oh now I need to go set up a database. Well, what what kind 251 00:13:49,440 --> 00:13:53,277 of tool do I have? Oh, there's a tool here that says Postgress. Okay, let me 252 00:13:53,360 --> 00:13:57,677 try using that. Okay, I use that. I got feedback. Let me look at the feedback 253 00:13:57,760 --> 00:13:58,877 and read it. And it'll read the 254 00:13:58,960 --> 00:14:01,997 feedback. And so the that that prompt 255 00:14:02,080 --> 00:14:06,877 box or context is where both the user 256 00:14:06,960 --> 00:14:11,757 input, the environment input, and the internal thoughts of the machine are all 257 00:14:11,840 --> 00:14:15,837 within. It's sort of like a program memory in in memory space. And so 258 00:14:15,920 --> 00:14:17,917 reasoning over that was the challenge 259 00:14:18,000 --> 00:14:20,797 for a long time. That's when AIs just 260 00:14:20,880 --> 00:14:23,677 like went off track and now they're able 261 00:14:23,760 --> 00:14:25,197 to kind of think through this entire 262 00:14:25,280 --> 00:14:27,917 thing and and maintain coherence. And 263 00:14:28,000 --> 00:14:30,717 there's there's now techniques around uh 264 00:14:30,800 --> 00:14:33,437 compression of contacts. So there still 265 00:14:33,520 --> 00:14:37,917 context length is still a problem, right? So I would say LM today, you 266 00:14:38,000 --> 00:14:40,077 know, they're marketed as a million uh 267 00:14:40,160 --> 00:14:42,477 token uh length, which is like a million 268 00:14:42,560 --> 00:14:44,957 words almost. uh in reality it's about 269 00:14:45,040 --> 00:14:47,437 200,000 and then they start to struggle. 270 00:14:47,520 --> 00:14:49,997 So we do a lot of uh you know we stop we 271 00:14:50,080 --> 00:14:54,957 compress the memory. So if a memory if if a portion of the memory is saying 272 00:14:55,040 --> 00:14:59,597 that I'm getting all the logs from the database you can summarize you know 273 00:14:59,680 --> 00:15:03,997 paragraphs of logs with one statement or the database setup that's it right and 274 00:15:04,080 --> 00:15:08,317 so every once in a while we'll compress the context so that we make sure we 275 00:15:08,400 --> 00:15:11,997 maintain coherence so that there's a lot of innovation happened outside of the 276 00:15:12,080 --> 00:15:16,397 foundation models as well in order to to enable that long context coherence. So 277 00:15:16,480 --> 00:15:19,117 what was the what was the key technical breakthrough at the in the foundation 278 00:15:19,200 --> 00:15:20,957 models that made this possible do you think? 279 00:15:21,040 --> 00:15:22,877 >> I think it's RL I think it's uh 280 00:15:22,960 --> 00:15:24,957 reinforcement learning. So the way 281 00:15:25,040 --> 00:15:28,477 pre-training works is you know uh they 282 00:15:28,560 --> 00:15:31,997 uh pre-training is a uh the first step 283 00:15:32,080 --> 00:15:35,917 of training a large language model. It reads a piece of text. It covers the 284 00:15:36,000 --> 00:15:37,997 last words and tries to guess it. That's 285 00:15:38,080 --> 00:15:41,037 how it's trained. That doesn't really 286 00:15:41,120 --> 00:15:44,317 imply long context reasoning. it it you 287 00:15:44,400 --> 00:15:47,037 know it it turns out to be very very 288 00:15:47,120 --> 00:15:48,557 effective. It can learn language that 289 00:15:48,640 --> 00:15:51,597 way. But the reason we weren't able to 290 00:15:51,680 --> 00:15:54,957 move past that limitation is that that 291 00:15:55,040 --> 00:15:57,357 modality of training just wasn't good 292 00:15:57,440 --> 00:15:59,997 enough. And what you want is you want a 293 00:16:00,080 --> 00:16:03,597 type of problem solving over a uh over 294 00:16:03,680 --> 00:16:05,837 long context. So what reinforcement 295 00:16:05,920 --> 00:16:08,717 learning uh uh especially from code 296 00:16:08,800 --> 00:16:12,557 execution gave us is the ability to for 297 00:16:12,640 --> 00:16:16,317 the machine to for the LLM to roll out 298 00:16:16,400 --> 00:16:18,397 what we call trajectories in AI. So 299 00:16:18,480 --> 00:16:22,717 trajectory is a uh stepbystep reasoning 300 00:16:22,800 --> 00:16:27,677 chain in order to reach a solution. So uh the way uh as I understand 301 00:16:27,760 --> 00:16:29,517 reinforcement learning works is they put 302 00:16:29,600 --> 00:16:31,917 the LM in a programming environment like 303 00:16:32,000 --> 00:16:35,757 replet and say hey here's a pro here's a 304 00:16:35,840 --> 00:16:38,077 codebase here's a bug in the codebase 305 00:16:38,160 --> 00:16:43,197 and we want you to solve it. Um now the human trainer already knows what the 306 00:16:43,280 --> 00:16:46,237 solution would look like. So we have a pull request that we have on GitHub so 307 00:16:46,320 --> 00:16:50,157 we know exactly or we have a unit test that we can run and verify the solution. 308 00:16:50,240 --> 00:16:53,517 So what it does is it rolls out a lot of different trajectories. Those they 309 00:16:53,600 --> 00:16:55,917 sample the model and maybe one of those 310 00:16:56,000 --> 00:16:57,837 trajectories will reach and a lot of 311 00:16:57,920 --> 00:17:00,877 them will just go go off off track but 312 00:17:00,960 --> 00:17:05,517 one of them will reach the solution by solving the bug and it reinforces on 313 00:17:05,600 --> 00:17:09,357 that. So that that gets a reward and the model gets trained that okay you know 314 00:17:09,440 --> 00:17:12,557 this is how you solve these type of problems. So that's how we're able to 315 00:17:12,640 --> 00:17:13,917 extend these reasoning chains. 316 00:17:14,000 --> 00:17:15,917 >> Got it. and and how it's a two-part 317 00:17:16,000 --> 00:17:18,557 question is how how how good how good 318 00:17:18,640 --> 00:17:23,757 are the models now at long long long reasoning and and I would say and how do 319 00:17:23,840 --> 00:17:25,197 we know like how how is that 320 00:17:25,280 --> 00:17:27,757 established? Um 321 00:17:27,840 --> 00:17:30,877 there is a nonprofit called meter um 322 00:17:30,960 --> 00:17:34,077 that is uh measuring 323 00:17:34,160 --> 00:17:37,837 uh useful has a benchmark to measure uh 324 00:17:37,920 --> 00:17:43,037 how long a model runs while maintaining coherence and doing useful useful things 325 00:17:43,120 --> 00:17:46,797 whether it's programming or other benchmark tasks that they've done. uh 326 00:17:46,880 --> 00:17:49,357 and they put up a paper I think uh late 327 00:17:49,440 --> 00:17:52,877 last year that said every seven months 328 00:17:52,960 --> 00:17:57,277 >> uh the the minutes that a model can run 329 00:17:57,360 --> 00:18:01,517 is doubling. >> So you go from 2 minutes to you know 4 330 00:18:01,600 --> 00:18:05,197 minutes in 7 months I think they vastly underestimated that. 331 00:18:05,280 --> 00:18:06,717 >> Is that right? Vastly it's doubling. 332 00:18:06,800 --> 00:18:08,157 It's doubling more often than 7 months. 333 00:18:08,240 --> 00:18:10,957 >> We so Asian 3 we measure that you know 334 00:18:11,040 --> 00:18:16,157 very closely uh and we measure that in real tasks from real users. So we're not 335 00:18:16,240 --> 00:18:19,357 doing benchmarking. We're actually doing AB tests and we're looking at the data 336 00:18:19,440 --> 00:18:21,677 that how users are successful or not. 337 00:18:21,760 --> 00:18:24,477 >> For us, the the absolute sign of success 338 00:18:24,560 --> 00:18:26,157 is you made an app and you published it. 339 00:18:26,240 --> 00:18:28,717 Because when you publish it, you're paying extra money. You're saying this 340 00:18:28,800 --> 00:18:32,637 app is economically useful. I'm going to publish it. So that's as clear-cut as 341 00:18:32,720 --> 00:18:37,357 possible. And so what we're seeing is in agent one, the agent could run for 2 342 00:18:37,440 --> 00:18:40,317 minutes >> and then and then perhaps struggle. 343 00:18:40,400 --> 00:18:45,597 Agent two came out in February, it ran for 20 minutes. Agent 3 200 minutes. 344 00:18:45,680 --> 00:18:48,797 >> Okay, >> 200. Some users are pushing it to like 345 00:18:48,880 --> 00:18:50,797 12 hours and things like that. I'm less 346 00:18:50,880 --> 00:18:53,437 confident that it is as good and when it 347 00:18:53,520 --> 00:18:56,317 goes to these stratospheres, but at like 348 00:18:56,400 --> 00:18:59,117 2 3 hours timeline, it is really it's 349 00:18:59,200 --> 00:19:01,997 it's it's it's insanely good. And and 350 00:19:02,080 --> 00:19:03,677 the main innovation outside of the 351 00:19:03,760 --> 00:19:07,117 models is a verification loop. Actually, 352 00:19:07,200 --> 00:19:09,997 uh I remember reading um a research 353 00:19:10,080 --> 00:19:12,797 paper from Nvidia. So what Nvidia did is 354 00:19:12,880 --> 00:19:15,597 they're trying to uh write um GPU 355 00:19:15,680 --> 00:19:18,877 kernels uh using deepseek and that was 356 00:19:18,960 --> 00:19:23,597 like perhaps 7 months ago when deepseek came out and what they found is that if 357 00:19:23,680 --> 00:19:27,837 we add a verifier in the loop if we can run the kernel and verify it's working 358 00:19:27,920 --> 00:19:31,677 we're able to run deepseeek for like 20 minutes and it it was generating 359 00:19:31,760 --> 00:19:34,077 actually optimized kernels 360 00:19:34,160 --> 00:19:36,397 >> and so I was like okay the next thing 361 00:19:36,480 --> 00:19:39,677 for us obviously as an as a sort of a 362 00:19:39,760 --> 00:19:43,037 agent lab or like applay our company. 363 00:19:43,120 --> 00:19:46,157 We're not doing the foundation model stuff, but we're doing a lot of research 364 00:19:46,240 --> 00:19:50,397 on top of that. And so, okay, we know that agents can run for 10 20 minutes 365 00:19:50,480 --> 00:19:52,637 now or LLMs can stay coherent for 366 00:19:52,720 --> 00:19:55,997 longer, but for you to push them to 200, 367 00:19:56,080 --> 00:19:58,157 300 minutes, you need a verifier in the 368 00:19:58,240 --> 00:19:59,917 loop. So, that's why we spend all our 369 00:20:00,000 --> 00:20:03,757 time uh creating scaffolds to make it so 370 00:20:03,840 --> 00:20:08,797 that the agent can spin up a browser and do computer use style testing. So once 371 00:20:08,880 --> 00:20:10,397 you put that in the middle, what's 372 00:20:10,480 --> 00:20:12,877 happening is it works for 20 minutes uh 373 00:20:12,960 --> 00:20:16,157 spin up another agent uh spins up a a 374 00:20:16,240 --> 00:20:18,077 browser tests the work of the previous 375 00:20:18,160 --> 00:20:20,157 agent. So it's a multi- aent system 376 00:20:20,240 --> 00:20:24,557 >> and if it is uh if it founds a bug it 377 00:20:24,640 --> 00:20:29,037 starts a new trajectory and says okay good work let's summarize what you did 378 00:20:29,120 --> 00:20:32,717 the last 20 minutes >> now that be that plus what the bug that 379 00:20:32,800 --> 00:20:34,957 we found that's a prompt for a new trajectory right 380 00:20:35,040 --> 00:20:37,837 >> so you stack those on each other and you can go endlessly but 381 00:20:37,920 --> 00:20:41,037 >> so it's like a mar like setting up a marathon or like a relay race 382 00:20:41,120 --> 00:20:43,837 >> as long as as long as each step is done properly you could do in sort of an 383 00:20:43,920 --> 00:20:46,077 infinite number of steps >> that's right that's right you can always 384 00:20:46,160 --> 00:20:49,197 compress the previous step into a paragraph And that becomes a prompt. So 385 00:20:49,280 --> 00:20:51,277 it's it's an agent prompting the next agent. 386 00:20:51,360 --> 00:20:54,237 >> Right. Right. Right. That's amazing. So um and then when when an agent like when 387 00:20:54,320 --> 00:20:57,677 a modern agent like running on modern modern LM that are trained this way when 388 00:20:57,760 --> 00:21:01,117 it let's say it runs for 200 minutes like when you watch the agent run is it 389 00:21:01,200 --> 00:21:05,037 like running is it like processing through like logic and tasks at the same 390 00:21:05,120 --> 00:21:07,197 pace that like a human being is or 391 00:21:07,280 --> 00:21:08,797 slower or faster? I 392 00:21:08,880 --> 00:21:12,637 >> it's actually I would say it is faster 393 00:21:12,720 --> 00:21:14,957 but not that much significantly faster. 394 00:21:15,040 --> 00:21:18,397 It's not at computer speed, right? What we expect computer speed to be. 395 00:21:18,480 --> 00:21:20,877 >> It's like watching a per like if you watch if you if it's describing what 396 00:21:20,960 --> 00:21:23,037 it's doing, it's sort of like watching a person work. 397 00:21:23,120 --> 00:21:26,877 >> It's like watching John Carmarmacine work. 398 00:21:26,960 --> 00:21:31,277 >> The world Okay. The world's the world's best programmer. 399 00:21:31,360 --> 00:21:33,597 >> Yeah. >> The world's best programmer on a stim on a stimulant. 400 00:21:33,680 --> 00:21:35,597 >> On a stimulant. Yeah, that's right. 401 00:21:35,680 --> 00:21:37,037 >> Working for you. Working for you. 402 00:21:37,120 --> 00:21:41,277 >> Yeah. There. So, it's very fast and you can see the uh file diffs running 403 00:21:41,360 --> 00:21:44,557 through, but every once in a while it'll stop and it'll start thinking. I'll show 404 00:21:44,640 --> 00:21:47,757 you the reasoning. Yeah. It's like, I did this and I did this. Am I on the 405 00:21:47,840 --> 00:21:50,317 right track? It kind of really tries to reflect, right? 406 00:21:50,400 --> 00:21:54,557 >> Uh and then it might review its work and decide the next step or it might kick 407 00:21:54,640 --> 00:21:59,677 into the testing agent or you know, so so you're seeing it do all of that and 408 00:21:59,760 --> 00:22:03,117 every once in a while it calls the tool for example, it stops and says, well, we 409 00:22:03,200 --> 00:22:05,597 ran into an issue. you know, Postgress 410 00:22:05,680 --> 00:22:11,437 um 15 is not um compatible with this, 411 00:22:11,520 --> 00:22:14,717 you know, database ORM package that that I have. 412 00:22:14,800 --> 00:22:17,837 >> Um okay, this is a problem I haven't seen before. I'm going to go search the 413 00:22:17,920 --> 00:22:21,997 web. So, it has a web search tool. Go do that. And so, it looks like a human 414 00:22:22,080 --> 00:22:24,477 programmer, right? >> And it's really fascinating to watch. 415 00:22:24,560 --> 00:22:28,237 It's one of my favorite things to do is just to watch the tool chain and 416 00:22:28,320 --> 00:22:30,477 reasoning chain and the testing chain. 417 00:22:30,560 --> 00:22:34,477 And it's yeah it is like watching a hyperproductive programmer 418 00:22:34,560 --> 00:22:37,437 >> right so you know we're kind of getting into here kind of the holy grail of AI 419 00:22:37,520 --> 00:22:41,757 which is sort of you know generalized reasoning um you know by the machine um 420 00:22:41,840 --> 00:22:46,477 so uh you mentioned this a couple times but this idea of a of a verification so 421 00:22:46,560 --> 00:22:49,997 so just for folks on the listening to podcast who maybe aren't in the details 422 00:22:50,080 --> 00:22:53,357 let me try to describe this and see see if I have it right so like a just a just 423 00:22:53,440 --> 00:22:55,997 a large language model the way you would exper you would have experienced with 424 00:22:56,080 --> 00:22:59,357 like Chad GPT out of the gate two years ago or whatever would have been it's 425 00:22:59,440 --> 00:23:03,757 like And it's incredible how fluid uh it is at language. Um it's incredible how 426 00:23:03,840 --> 00:23:06,877 good it is at like writing Shakespearean sonnetss or rap lyrics. It's it's 427 00:23:06,960 --> 00:23:09,917 amazing how good it is at human conversation. But if you start to ask it 428 00:23:10,000 --> 00:23:14,157 like problems that involve like rational thinking uh or problem solving all of a 429 00:23:14,240 --> 00:23:17,677 sudden like you math or the math the whole show and and in the very beginning 430 00:23:17,760 --> 00:23:20,557 it was you could ask if you ask it very basic math problems, you know, it would 431 00:23:20,640 --> 00:23:21,757 it would not be able to do them. 432 00:23:21,840 --> 00:23:24,397 >> That's right. Uh but then even when it got better at those, if you started to 433 00:23:24,480 --> 00:23:27,117 ask it to like, you know, it it could maybe add two small numbers together, 434 00:23:27,200 --> 00:23:29,837 but it couldn't add two large numbers together. Or if it could add two large 435 00:23:29,920 --> 00:23:31,197 numbers, it couldn't multiply them. 436 00:23:31,280 --> 00:23:32,877 Yeah. >> And it's just like, all right, this is 437 00:23:32,960 --> 00:23:35,517 And then it had this there was this famous the famous was the straw the 438 00:23:35,600 --> 00:23:37,997 strawberry test, the famous strawberry test, which is how many Rs are in the 439 00:23:38,080 --> 00:23:39,037 word strawberry. >> That's right. 440 00:23:39,120 --> 00:23:41,837 >> And there was this long period where it it kept it would it would just guess 441 00:23:41,920 --> 00:23:44,797 wrong. It would say there were only two Rs in the word strawberry. And it turns 442 00:23:44,880 --> 00:23:49,757 out there are three. Um, so, um, so it it was this thing and so people were and 443 00:23:49,840 --> 00:23:52,637 there was even this term that was being used kind of the the slur that was being 444 00:23:52,720 --> 00:23:54,237 used at the time was stoastic parrot. 445 00:23:54,320 --> 00:23:55,837 >> Yeah, >> I was thinking clanker. 446 00:23:55,920 --> 00:23:59,997 >> Well, well, clanker is the is the new slur. Clank clanker. Clanker is just the 447 00:24:00,080 --> 00:24:02,477 full-on racial slur against AI as a 448 00:24:02,560 --> 00:24:07,437 species. Um, but the technical critique was so-called stoastic par stoastic 449 00:24:07,520 --> 00:24:10,637 means random. Uh so sort of random parrot me meaning basically that this 450 00:24:10,720 --> 00:24:13,357 thing was sort of the large language models were like a they were like a 451 00:24:13,440 --> 00:24:15,997 mirage where they were like repeating back to you things that they thought 452 00:24:16,080 --> 00:24:19,597 that you wanted to hear but they didn't >> in a way it's true in the in the pure 453 00:24:19,680 --> 00:24:22,157 pre-training LLM world >> right for the for the very basic layer 454 00:24:22,240 --> 00:24:24,877 but then what happened is as you said over the last year or something there 455 00:24:24,960 --> 00:24:28,237 there was this layering in of of reinforcement learning and then but the 456 00:24:28,320 --> 00:24:32,157 key to >> it's not new crucially it's like it's alpha go right so 457 00:24:32,240 --> 00:24:33,597 >> describe so describe that for a second. 458 00:24:33,680 --> 00:24:35,197 Yeah. So we we had this breakthrough 459 00:24:35,280 --> 00:24:39,597 before in uh 2015 was the Alph Go 460 00:24:39,680 --> 00:24:42,477 breakthrough. I think 2015 2016 where it 461 00:24:42,560 --> 00:24:47,357 is emerging of sort of uh you know the the the you would know a lot better than 462 00:24:47,440 --> 00:24:51,997 me the old AI debate between the connectionists uh the the people who who 463 00:24:52,080 --> 00:24:53,757 thinks neuronet networks are the true 464 00:24:53,840 --> 00:24:56,957 sort of way of doing AI and the symbolic 465 00:24:57,040 --> 00:25:01,837 systems I think or like the people that think that you know discrete reasonings 466 00:25:01,920 --> 00:25:05,517 fates and knowledge bases whatever this is the way to go and so there was there 467 00:25:05,600 --> 00:25:09,677 was a merging of these two worlds where the way AlphaG go worked is it had a 468 00:25:09,760 --> 00:25:13,997 neural network but it had a Monte Carlo research algorithm on top of that. So 469 00:25:14,080 --> 00:25:18,317 the neural network would generate uh would would like uh generate a list of 470 00:25:18,400 --> 00:25:20,957 potential moves uh and then you had a 471 00:25:21,040 --> 00:25:23,597 more discrete algorithm sort those moves 472 00:25:23,680 --> 00:25:26,797 and find the best based on just uh tree 473 00:25:26,880 --> 00:25:29,357 search based on just trying to verify 474 00:25:29,440 --> 00:25:30,877 again this sort of a verifier in the 475 00:25:30,960 --> 00:25:34,557 loop trying to verify which move might 476 00:25:34,640 --> 00:25:38,077 yield the best based on more classical 477 00:25:38,160 --> 00:25:40,877 way of doing algorithms. Um, and so that 478 00:25:40,960 --> 00:25:43,437 that's a resurgence of of that movement 479 00:25:43,520 --> 00:25:46,637 where we have this amazing generative uh 480 00:25:46,720 --> 00:25:51,997 neural network that is the the LLM and now let's layer on more discrete ways of 481 00:25:52,080 --> 00:25:55,197 trying to verify whether it's doing the right thing or not and let's put that in 482 00:25:55,280 --> 00:25:57,117 a training loop and once you do that the 483 00:25:57,200 --> 00:26:02,237 LLM will start gaining new capabilities such as uh reasoning over math and code 484 00:26:02,320 --> 00:26:04,317 and things like that. >> Exactly. Right. Okay. And then that's 485 00:26:04,400 --> 00:26:07,677 great. And then and then the the key thing there though for for RL to work 486 00:26:07,760 --> 00:26:12,717 for LMS to reason the key is that it be a a problem statement that there is a 487 00:26:12,800 --> 00:26:16,477 defined and verifiable answer. That's right. Is that right? And so and and and 488 00:26:16,560 --> 00:26:19,117 you might think about this as like let's give a bunch of examples like in 489 00:26:19,200 --> 00:26:22,637 medicine this might be like um you know a diagnosis that like a panel of human 490 00:26:22,720 --> 00:26:26,077 doctors agrees with um or or or by the way or a diagnosis that actually you 491 00:26:26,160 --> 00:26:31,037 know solves the condition. Um in law this would be a um you know a a argument 492 00:26:31,120 --> 00:26:34,957 that in front of a jury actually results in an acquitt or or something like that. 493 00:26:35,040 --> 00:26:39,037 Um in u math it's an equation that actually solves properly. Uh in physics 494 00:26:39,120 --> 00:26:41,277 it's a result that actually works in the real world. 495 00:26:41,360 --> 00:26:44,077 >> I don't know in civil engineering it's a bridge that doesn't collapse. Right. So 496 00:26:44,160 --> 00:26:45,837 so so there there there's always some 497 00:26:45,920 --> 00:26:49,117 some test is that the first two do not 498 00:26:49,200 --> 00:26:51,597 work very well just yet. like the the 499 00:26:51,680 --> 00:26:55,997 like I would say uh law and healthcare 500 00:26:56,080 --> 00:27:01,277 they're still a little too squishy a little too soft it's unlike math or code 501 00:27:01,360 --> 00:27:04,397 like the way that they're training on math they're using this uh sort of like 502 00:27:04,480 --> 00:27:08,637 a program language uh provable language called lean for proofs right so you can 503 00:27:08,720 --> 00:27:13,357 run a lean statement you can run a computer code uh perhaps you can run a 504 00:27:13,440 --> 00:27:17,677 physics simulation or civil engineering uh sort of physics simulation but you 505 00:27:17,760 --> 00:27:21,437 can't run a diagnosis >> okay So I would say the 506 00:27:21,520 --> 00:27:23,757 >> but you could verify it with human answers or or not. 507 00:27:23,840 --> 00:27:26,557 >> Yeah. So that that's a more RL HF in a 508 00:27:26,640 --> 00:27:29,277 way. So it is not the like sort of 509 00:27:29,360 --> 00:27:32,157 autonomous RL train like fully scalable 510 00:27:32,240 --> 00:27:35,357 autonomous which is why coding is moving 511 00:27:35,440 --> 00:27:37,357 faster than any other domain is because 512 00:27:37,440 --> 00:27:40,877 we can we we can generate these problems 513 00:27:40,960 --> 00:27:44,397 and verify them on the fly. But there's two but with coding as anybody who's 514 00:27:44,480 --> 00:27:46,717 coded knows there's coding there's two tests which is one is does the code 515 00:27:46,800 --> 00:27:48,717 compile >> right >> and then the other is does it produce 516 00:27:48,800 --> 00:27:51,437 the right output and just because it compiles doesn't mean it produces the 517 00:27:51,520 --> 00:27:54,237 right output and I you tell me but verifying that it's the correct output 518 00:27:54,320 --> 00:27:55,197 is harder 519 00:27:55,280 --> 00:27:56,957 >> yeah sobbench 520 00:27:57,040 --> 00:28:02,877 is a collection of uh verified pull 521 00:28:02,960 --> 00:28:06,157 request end state uh so so it is it is 522 00:28:06,240 --> 00:28:09,197 not just about compiling we so they they 523 00:28:09,280 --> 00:28:11,437 group of scientists sobbench is the main 524 00:28:11,520 --> 00:28:14,877 benchmark used to test whether AI is 525 00:28:14,960 --> 00:28:16,877 good at software engineering tasks and 526 00:28:16,960 --> 00:28:19,197 we're almost saturating that. So last 527 00:28:19,280 --> 00:28:23,277 year we're at like maybe 5% early 24 or 528 00:28:23,360 --> 00:28:25,997 less and now we're like 82% or something 529 00:28:26,080 --> 00:28:30,477 like that with cloth on at 4.5 that's state-of-the-art and that's like a 530 00:28:30,560 --> 00:28:34,557 really nice health climb that's happening right now. uh and basically 531 00:28:34,640 --> 00:28:38,557 they went and looked on GitHub. They found the the you know most complex 532 00:28:38,640 --> 00:28:41,197 repositories. They found bug statements 533 00:28:41,280 --> 00:28:45,917 that are very clear uh and they found ProQuest that actually solve those bug 534 00:28:46,000 --> 00:28:49,757 statements with unit tests and everything. So there is an existing 535 00:28:49,840 --> 00:28:54,557 corpus on GitHub of tasks that that the AIS can can solve and you can also 536 00:28:54,640 --> 00:28:56,557 generate them. Those are not too hard to 537 00:28:56,640 --> 00:29:00,317 to generate uh you know what's called 538 00:29:00,400 --> 00:29:04,157 synthetic uh data. Uh uh but but you're 539 00:29:04,240 --> 00:29:09,437 right it's not infinitely scalable um because you you some human verifiers 540 00:29:09,520 --> 00:29:12,557 still need to kind of look at the at the task but maybe the foundation models 541 00:29:12,640 --> 00:29:15,677 have found a way to have the synthetic training go all the way 542 00:29:15,760 --> 00:29:17,997 >> right and then what's happening I think I think because what's happening is the 543 00:29:18,080 --> 00:29:20,877 foundation model companies are in some cases they are hire they're actually 544 00:29:20,960 --> 00:29:23,517 hiring human experts to generate new training data. Yes. 545 00:29:23,600 --> 00:29:25,997 >> So they're actually hiring mathematicians and physicists and coders 546 00:29:26,080 --> 00:29:29,917 to basically sit and you know they're they're hiring they're they're hiring 547 00:29:30,000 --> 00:29:31,917 human programmers putting them on the cocaine. Yes. 548 00:29:32,000 --> 00:29:37,437 >> Um and having them probably coffee um uh and having them actually write code and 549 00:29:37,520 --> 00:29:39,997 then and then write code in a way where there's a known result of the code 550 00:29:40,080 --> 00:29:43,277 running such that the this RO loop can be trained properly. That's right. And 551 00:29:43,360 --> 00:29:46,077 then the other the other and then the other thing these companies are doing is 552 00:29:46,160 --> 00:29:49,677 as you said they're building systems where the software itself generates the 553 00:29:49,760 --> 00:29:53,277 training data, generates the tests, generates the valid the validated 554 00:29:53,360 --> 00:29:55,677 results and and that's soal synthetic training data. 555 00:29:55,760 --> 00:30:00,317 >> That's right. And but yeah, but but again those work in the very hard 556 00:30:00,400 --> 00:30:03,677 domains. It works to some extent in the software domains 557 00:30:03,760 --> 00:30:06,397 >> and I think there's some transfer learning we can you can see the 558 00:30:06,480 --> 00:30:10,157 reasoning work when it comes to you know tools like deep research and things like 559 00:30:10,240 --> 00:30:12,557 that but we're not making as fast as 560 00:30:12,640 --> 00:30:15,117 progress in the in the more soft domain. 561 00:30:15,200 --> 00:30:17,437 >> So so say softer domains meaning like 562 00:30:17,520 --> 00:30:22,157 domains in which it's harder harder or even impossible to actually verify 563 00:30:22,240 --> 00:30:26,317 correctness of of result in a sort of a deterministic factual grounded 564 00:30:26,400 --> 00:30:28,077 >> non-controversial way. Like if you have 565 00:30:28,160 --> 00:30:29,917 a a chronic disease, you could you could 566 00:30:30,000 --> 00:30:35,197 have you know you have a POTS or uh you 567 00:30:35,280 --> 00:30:38,157 know whatever EDS syndrome or and 568 00:30:38,240 --> 00:30:42,397 they're all they're all clusters and it's because it it is the domain of 569 00:30:42,480 --> 00:30:46,797 abstraction. It is not as concrete as code and math and things like that. So I 570 00:30:46,880 --> 00:30:48,397 think there's still a long ways to go there. 571 00:30:48,480 --> 00:30:51,677 >> Right. So sort of the more concrete the problem like it's the concretness of the 572 00:30:51,760 --> 00:30:54,477 problem that is the key variable not the difficulty of the problem. Would that be 573 00:30:54,560 --> 00:30:57,837 a way to think about it? >> Yeah. Yeah. I think the the uh 574 00:30:57,920 --> 00:31:02,237 concreteness in a sense of can you get a true or false ver verifiable 575 00:31:02,320 --> 00:31:05,277 >> right but like in any domain in any domain of human effort in which there's 576 00:31:05,360 --> 00:31:07,677 a verifiable answer we should expect extremely rapid progress. 577 00:31:07,760 --> 00:31:09,997 >> Yes. >> Right. >> Yes. Absolutely. And I I think that's what we're saying. 578 00:31:10,080 --> 00:31:12,477 >> Right. And that and that for sure includes math. That for sure includes 579 00:31:12,560 --> 00:31:15,837 physics for sure includes chemistry. For sure includes >> large areas of code. 580 00:31:15,920 --> 00:31:18,077 >> That's right. >> Right. What what else does that include do you think? 581 00:31:18,160 --> 00:31:22,237 >> Bio like we're seeing with a protein >> genomic. Yeah. Okay. Right. 582 00:31:22,320 --> 00:31:23,917 >> Yeah. Yeah. Things like that. I think 583 00:31:24,000 --> 00:31:28,237 some some areas of robotics, right? Um 584 00:31:28,320 --> 00:31:30,637 there's a clear outcome, right? 585 00:31:30,720 --> 00:31:34,717 >> Uh but but it's not that many. I mean, surprisingly, >> well, it depends. 586 00:31:34,800 --> 00:31:36,157 >> Yeah, depends on your point of view. 587 00:31:36,240 --> 00:31:39,917 That's some people might say that's a lot. Um so, uh and then um you you 588 00:31:40,000 --> 00:31:42,557 mentioned that we you mentioned the pace of improvement. So, what would you 589 00:31:42,640 --> 00:31:45,277 expect from the pace of improvement going forward for this? 590 00:31:45,360 --> 00:31:47,437 >> I I think we're we're ripping on coding. 591 00:31:47,520 --> 00:31:50,077 Like I think I think it's just it's 592 00:31:50,160 --> 00:31:52,637 going like I think it's going to be like 593 00:31:52,720 --> 00:31:55,277 what we're working on with with agent 594 00:31:55,360 --> 00:31:59,837 floor right now um is by by next year we think you're going to be sitting instead 595 00:31:59,920 --> 00:32:02,237 of rep in front of replet and you're 596 00:32:02,320 --> 00:32:04,477 shooting off multiple agents at a time. 597 00:32:04,560 --> 00:32:07,437 You're like planning a new feature. Um 598 00:32:07,520 --> 00:32:10,157 so I I want you know social network on 599 00:32:10,240 --> 00:32:12,637 top of my storefront and another one is 600 00:32:12,720 --> 00:32:16,157 like hey um refactor the database. Hey, 601 00:32:16,240 --> 00:32:17,997 in and you're running parallel agents. 602 00:32:18,080 --> 00:32:20,797 So, you have five 10 agents kind of working in the background and they're 603 00:32:20,880 --> 00:32:25,037 merging the code and taking care of all of that, but you also have a really nice 604 00:32:25,120 --> 00:32:28,317 interface on top of that that you're doing design and you're interacting with 605 00:32:28,400 --> 00:32:32,637 AI in a more creative way. Uh maybe using visuals and charts and things like 606 00:32:32,720 --> 00:32:34,957 that. So, there's a multimodal angle of 607 00:32:35,040 --> 00:32:37,597 that of that interaction. So I think you 608 00:32:37,680 --> 00:32:40,237 know creating software is going to be 609 00:32:40,320 --> 00:32:42,557 such an exciting 610 00:32:42,640 --> 00:32:45,597 uh area and and and I think that the lay 611 00:32:45,680 --> 00:32:48,957 person will be as good as a what a 612 00:32:49,040 --> 00:32:50,477 senior software engineer that works at 613 00:32:50,560 --> 00:32:52,637 Google uh is today. So I think I think 614 00:32:52,720 --> 00:32:56,477 that's happening very soon. Um but but 615 00:32:56,560 --> 00:33:00,877 you know I don't see them and be curious about your point of view but like my 616 00:33:00,960 --> 00:33:03,277 experience between as as a sort of a you 617 00:33:03,360 --> 00:33:06,077 know on the let's say healthcare side or 618 00:33:06,160 --> 00:33:10,797 more you know write me an essay side or more creative side haven't seen as much 619 00:33:10,880 --> 00:33:14,077 of a rapid improvement as what we're seeing in code. So so I think I think 620 00:33:14,160 --> 00:33:15,677 code is going to go to the moon. Math is 621 00:33:15,760 --> 00:33:18,557 probably as well some some you know 622 00:33:18,640 --> 00:33:22,877 scientific domains bio things like that those are are going to move really fast. 623 00:33:22,960 --> 00:33:26,077 >> Yeah. So there's this there's this there's this weird dynamic see if you 624 00:33:26,160 --> 00:33:28,957 agree with this and Eric also curious your point of view on this like there's 625 00:33:29,040 --> 00:33:31,357 this weird dynamic that we have and we have this in the office here a lot and I 626 00:33:31,440 --> 00:33:33,837 also have this with like the leading of entrepreneurs a lot which is this thing 627 00:33:33,920 --> 00:33:36,317 of like >> like wow this is the most amazing 628 00:33:36,400 --> 00:33:38,637 technology ever and it's moving really fast and yet we're still like really 629 00:33:38,720 --> 00:33:43,437 disappointed um and like it's not moving fast enough and like it's like maybe 630 00:33:43,520 --> 00:33:46,637 right on the verge of stalling out >> and like you know we should both be like 631 00:33:46,720 --> 00:33:49,517 hyper excited but also on the verge of like slitting our wrists because like 632 00:33:49,600 --> 00:33:51,517 you know the gravy train is coming to an end, 633 00:33:51,600 --> 00:33:53,757 >> right? And and I always wonder it's like 634 00:33:53,840 --> 00:33:56,557 you know on the one hand it's like okay like you know not all I don't know 635 00:33:56,640 --> 00:33:59,277 ladders go to the moon like just because something you know looks like it works 636 00:33:59,360 --> 00:34:01,517 or you know doesn't mean it's going to you know be able to you're going to be 637 00:34:01,600 --> 00:34:05,357 able to scale it up and have it work you know to the fullest extent. Um uh you 638 00:34:05,440 --> 00:34:08,397 know so like it's important to like recognize practical limits and not just 639 00:34:08,480 --> 00:34:11,677 extrapolate everything to infinity. Um on the other hand like you know we're 640 00:34:11,760 --> 00:34:14,237 dealing with magic here that we I think probably all would have thought was 641 00:34:14,320 --> 00:34:16,717 impossible 5 years ago or certainly 10 years ago. 642 00:34:16,800 --> 00:34:19,597 >> Like I I didn't you know look I I you know I got my CS degree in the late ' 643 00:34:19,680 --> 00:34:22,637 80s early 90s. I I never I didn't think I would live to see any of this, right? 644 00:34:22,720 --> 00:34:26,077 Like this is just amazing that this is actually happening in in in my lifetime. 645 00:34:26,160 --> 00:34:28,797 >> Um >> but but there's a huge bet on AGI, 646 00:34:28,880 --> 00:34:32,317 right? like whether it's the foundation models uh I think you know now the 647 00:34:32,400 --> 00:34:34,877 entire US economy is sort of a bet on 648 00:34:34,960 --> 00:34:37,357 AGI and and there are crucial questions 649 00:34:37,440 --> 00:34:42,317 to ask whether are we on track to AGI or not because there are some ways that I 650 00:34:42,400 --> 00:34:46,957 can tell you it doesn't seem like we're on track to AGI because we uh because 651 00:34:47,040 --> 00:34:50,557 there doesn't seem to be transfer learning across these domains that are 652 00:34:50,640 --> 00:34:55,677 that are you know significance right so if we get a lot better at code 653 00:34:55,760 --> 00:34:59,757 we're not immediately getting better at like generalized reasoning we need to go 654 00:34:59,840 --> 00:35:01,757 also you know get training data and 655 00:35:01,840 --> 00:35:03,757 create RL environment for bio or 656 00:35:03,840 --> 00:35:06,957 chemistry or physics or math or law or 657 00:35:07,040 --> 00:35:09,677 so so and this this has been the sort of 658 00:35:09,760 --> 00:35:11,277 point of discussion now in the AI 659 00:35:11,360 --> 00:35:14,077 community after the Darkish and Richard 660 00:35:14,160 --> 00:35:17,197 Sutton uh interview where uh you know 661 00:35:17,280 --> 00:35:19,197 Richard Sutton kind of poured this cold 662 00:35:19,280 --> 00:35:22,157 water on the on the bitter lesson. So 663 00:35:22,240 --> 00:35:24,397 everyone was using this uh essay that he 664 00:35:24,480 --> 00:35:26,237 wrote called the bitter lesson. The idea 665 00:35:26,320 --> 00:35:29,597 is that there are um infinitely scalable 666 00:35:29,680 --> 00:35:34,397 ways of uh doing uh uh AI research and 667 00:35:34,480 --> 00:35:37,437 and and and anytime you can pour more 668 00:35:37,520 --> 00:35:41,597 compute and more data and go more performance out you're just you know 669 00:35:41,680 --> 00:35:43,277 that's the ultimate way of getting to 670 00:35:43,360 --> 00:35:47,437 AGI and some people you know interpreted 671 00:35:47,520 --> 00:35:50,157 that interview that perhaps he's 672 00:35:50,240 --> 00:35:52,717 doubtful that even we're even on a on a 673 00:35:52,800 --> 00:35:56,317 bitter uh lessened path here and perhaps 674 00:35:56,400 --> 00:35:58,797 the current training regime is actually 675 00:35:58,880 --> 00:36:01,277 very much the opposite in which we we 676 00:36:01,360 --> 00:36:03,837 are so dependent on human data and human 677 00:36:03,920 --> 00:36:05,917 annotation and and all of that stuff. So 678 00:36:06,000 --> 00:36:09,037 I think the I I agree with you. I mean 679 00:36:09,120 --> 00:36:13,197 as a company we're we're excited about where things are headed but but there's 680 00:36:13,280 --> 00:36:15,277 there's a question of like are we on 681 00:36:15,360 --> 00:36:20,397 track to AGI or not and be curious what you think. So, so and you know Ilia I 682 00:36:20,480 --> 00:36:23,037 think you know Ilioskcover makes a makes a specific form of this argument which 683 00:36:23,120 --> 00:36:25,357 is basically like we're just literally running out of training data. It's a 684 00:36:25,440 --> 00:36:28,557 fossil fuel argument right like if we slurped all the training fundamentally 685 00:36:28,640 --> 00:36:30,797 we've slurped all the data off the internet that is where almost all the 686 00:36:30,880 --> 00:36:33,597 data is at this point. There's a little bit more data that's in like you know 687 00:36:33,680 --> 00:36:36,157 private dark pool somewhere that we're going to go get but like 688 00:36:36,240 --> 00:36:38,477 >> we have it all and then right we're we're in this business now trying to 689 00:36:38,560 --> 00:36:41,037 generate new data but generating new data is hard and expensive you know 690 00:36:41,120 --> 00:36:43,517 compared to just like slurping things off the internet. So 691 00:36:43,600 --> 00:36:46,557 >> there are these arguments. Um you know having said that you know you get into 692 00:36:46,640 --> 00:36:48,797 definitional questions here really quick which are kind of a rabbit hole but 693 00:36:48,880 --> 00:36:51,117 having said that like you mentioned transfer learning. So transfer learning 694 00:36:51,200 --> 00:36:54,237 is the ability of the machine to right to be an expert in one domain and then 695 00:36:54,320 --> 00:36:56,397 and then generalize that into another domain. 696 00:36:56,480 --> 00:36:58,957 >> My answer to that is like have you met people? 697 00:36:59,040 --> 00:37:03,037 >> Um and how many people do you know are able to do transfer learning? 698 00:37:03,120 --> 00:37:05,437 >> Not many. Right. Well because there's 699 00:37:05,520 --> 00:37:07,037 >> quite the opposite actually. The nerdier 700 00:37:07,120 --> 00:37:11,517 they are in a certain domain the kind of you know often they have blind spots. We 701 00:37:11,600 --> 00:37:14,797 joke about how everyone's just in one area or they make some like 702 00:37:14,880 --> 00:37:18,157 massive mistake and and like don't trust them on this but on this other topic you 703 00:37:18,240 --> 00:37:20,717 know >> right? Yeah. Well and this is a well-known thing among like for example 704 00:37:20,800 --> 00:37:22,717 public intellect. So this happens there's actually been whole books 705 00:37:22,800 --> 00:37:25,197 written about this on so-called public intellectuals. So you get these people 706 00:37:25,280 --> 00:37:28,077 who show up on TV and they're experts and what happens is they're like an 707 00:37:28,160 --> 00:37:30,717 expert in economics right and then they show up on TV and they talk about 708 00:37:30,800 --> 00:37:33,757 politics and they don't know anything about politics right or they don't know 709 00:37:33,840 --> 00:37:36,397 anything about like medicine or they don't know anything about the law or 710 00:37:36,480 --> 00:37:38,797 they don't know anything about computers. You know, this is the Paul 711 00:37:38,880 --> 00:37:41,197 Gregman talking about how the internet's going to be no more significant than the 712 00:37:41,280 --> 00:37:43,197 fax machine. >> Facts. >> He's a brilliant economist. He has no 713 00:37:43,280 --> 00:37:44,237 idea what a how a computer works. 714 00:37:44,320 --> 00:37:45,517 >> Is he a brilliant economist? 715 00:37:45,600 --> 00:37:49,437 >> Well, at one at one point at one point 716 00:37:49,520 --> 00:37:54,397 at one point, let's get even if even if he's a brilliant Well, this is the thing 717 00:37:54,480 --> 00:37:56,877 like what does that mean? Like should a brilliant economist be able to 718 00:37:56,960 --> 00:38:00,717 extrapolate, you know, the internet is is a good question. But um but the point 719 00:38:00,800 --> 00:38:03,837 being like even if he is a you know, take any take anybody Oh, by the way, or 720 00:38:03,920 --> 00:38:07,757 like Ein like Einstein's like actually my favorite example. I think you'd agree 721 00:38:07,840 --> 00:38:09,357 Einstein was a brilliant physicist. 722 00:38:09,440 --> 00:38:11,437 >> He was like a he was he was a Stalinist. 723 00:38:11,520 --> 00:38:14,237 Like he was just he was Yeah. He was a socialist and he was a Stalinist and he 724 00:38:14,320 --> 00:38:16,317 was like he thought like Stalin was fantastic. 725 00:38:16,400 --> 00:38:18,797 >> Out still. >> Yeah. Okay. All right. 726 00:38:18,880 --> 00:38:24,077 >> True socialism. >> All right. All right. Einstein, you know, I'll I'll 727 00:38:24,160 --> 00:38:26,957 I'll take your word for it. But like once he got into politics, he was just 728 00:38:27,040 --> 00:38:29,837 like totally loopy or or you know, even right or wrong. It's just he just 729 00:38:29,920 --> 00:38:32,557 sounded like all of a sudden like an undergraduate lunatic, like somebody in 730 00:38:32,640 --> 00:38:35,357 a dorm room. Like he there was no transfer learning from physics into 731 00:38:35,440 --> 00:38:39,197 politics. like he he didn't listen right or wrong he didn't there was no there 732 00:38:39,280 --> 00:38:42,477 was clearly there was nothing new in his political analysis it was the same wrote 733 00:38:42,560 --> 00:38:45,757 routine you get out of >> you know yeah so so in a way the 734 00:38:45,840 --> 00:38:50,077 argument you're making is like we maybe already a human level AI I mean perhaps 735 00:38:50,160 --> 00:38:54,157 the definition of AGI is is is something totally different is like above human 736 00:38:54,240 --> 00:38:57,677 level that something that totally generalizes across domains it's it's not 737 00:38:57,760 --> 00:39:00,317 something that we've seen >> yeah like we've ideal yeah I was saying 738 00:39:00,400 --> 00:39:02,877 we we've we've and you know look we should we should shoot big but we we've 739 00:39:02,960 --> 00:39:07,437 idealized a a we've idealized a goal um that may be idealized in a way that 740 00:39:07,520 --> 00:39:10,637 like number one it's just it it's it's like so far beyond what people can do 741 00:39:10,720 --> 00:39:13,277 that it's you're no longer it's no longer relevant comparison to people and 742 00:39:13,360 --> 00:39:16,397 and usually AGI is defined as you know able to do everything better than a 743 00:39:16,480 --> 00:39:18,877 person can >> and it's like well okay so if doing 744 00:39:18,960 --> 00:39:21,357 everything better than a person can it's like if a person can't do any transfer 745 00:39:21,440 --> 00:39:24,717 learning at all >> right doing even a little little bit a 746 00:39:24,800 --> 00:39:27,357 marginal bit might might actually be better or it might not matter just 747 00:39:27,440 --> 00:39:30,077 because no no human can do it and so therefore you just you just stack up the 748 00:39:30,160 --> 00:39:33,357 domains there's also this well-known phenomenon in AI with you know t 749 00:39:33,440 --> 00:39:36,317 typically this works the other way which there's a phenomenon AI AI engineers 750 00:39:36,400 --> 00:39:38,957 always complain about and scientists always complain about which is the 751 00:39:39,040 --> 00:39:42,477 definition of AI is always the next thing that that the machine can't do and 752 00:39:42,560 --> 00:39:45,677 so like the definition for of AI for a long time was like can it beat humans at 753 00:39:45,760 --> 00:39:47,517 chess >> and then the minute it could beat humans 754 00:39:47,600 --> 00:39:50,717 at chess that was no longer AI that was just like oh that's just like boring 755 00:39:50,800 --> 00:39:53,437 >> that's computer chess it became >> computer chess it's just like boring and 756 00:39:53,520 --> 00:39:56,477 now it's an app on your iPhone and nobody nobody and nobody cares right and 757 00:39:56,560 --> 00:40:00,317 it's immediately then >> cheering test was the test and then we passed it and nobody 758 00:40:00,400 --> 00:40:02,957 >> we blew this is a really big deal >> there was no celebration 759 00:40:03,040 --> 00:40:06,317 >> there was no parties That's exactly right. There was no for 80 years the 760 00:40:06,400 --> 00:40:09,117 Turing test I mean they made a movie about it like the whole thing that was 761 00:40:09,200 --> 00:40:11,917 the thing and like we blew right through it and nobody even registered it. Nobody 762 00:40:12,000 --> 00:40:14,317 cares. It gets no credit for it. We're just like ah it's still you know 763 00:40:14,400 --> 00:40:17,997 complete p piece of like >> right and so there's this thing where so 764 00:40:18,080 --> 00:40:21,757 the AI scientists are are are used to complaining basically that they're that 765 00:40:21,840 --> 00:40:24,157 they're they're being they're always being judged against the next thing as 766 00:40:24,240 --> 00:40:26,317 opposed to all the things they've already they've already solved. 767 00:40:26,400 --> 00:40:29,757 >> Um uh but but that's maybe the other side of it which is they're also putting 768 00:40:29,840 --> 00:40:33,757 out for themselves um an unreasonable goal. an an unreasonable goal and then 769 00:40:33,840 --> 00:40:36,957 doing this sort of self flagagillation kind of along the way and and and I I 770 00:40:37,040 --> 00:40:39,277 kind of wonder yeah I I wonder kind of which way that cuts. 771 00:40:39,360 --> 00:40:41,037 >> Yeah. Yeah. It's an interesting question 772 00:40:41,120 --> 00:40:42,557 like I started thinking about this idea 773 00:40:42,640 --> 00:40:45,117 of like it doesn't matter whether it's 774 00:40:45,200 --> 00:40:48,237 truly AGI and the way I define AGI is 775 00:40:48,320 --> 00:40:51,197 that you put in a AI system in any 776 00:40:51,280 --> 00:40:53,677 environment and efficiently learns right 777 00:40:53,760 --> 00:40:55,437 >> um you know it doesn't have to have that 778 00:40:55,520 --> 00:40:58,557 much prior knowledge in order to kind of learn something but also can transfer 779 00:40:58,640 --> 00:41:00,877 that knowledge across different domains. 780 00:41:00,960 --> 00:41:03,677 Um but you know we can get to like 781 00:41:03,760 --> 00:41:06,157 functional AGI and what functional AGI 782 00:41:06,240 --> 00:41:09,357 is is just yeah collect data on every uh 783 00:41:09,440 --> 00:41:12,637 useful uh economic activity in uh in the 784 00:41:12,720 --> 00:41:15,517 world today and train an LLM on top of 785 00:41:15,600 --> 00:41:19,677 that or train the same foundation model on top of that and and we we'll go we'll 786 00:41:19,760 --> 00:41:23,757 target every sector economy and and you can automate a big part of labor that 787 00:41:23,840 --> 00:41:27,757 way. So I think I think yeah I think we're on that track for sure. 788 00:41:27,840 --> 00:41:31,197 >> Right. um you tweeted after GBG5 came out that you were feeling the 789 00:41:31,280 --> 00:41:35,437 diminishing returns. Yeah. What were you expecting and but and and what needs to 790 00:41:35,520 --> 00:41:38,077 be done? Do we need another breakthrough to get back to the pace of growth or 791 00:41:38,160 --> 00:41:40,637 what are your thoughts there? >> I mean this this whole discussion is is 792 00:41:40,720 --> 00:41:42,557 sort of about that and and my feeling is 793 00:41:42,640 --> 00:41:46,877 that uh you know GPT5 uh got good at 794 00:41:46,960 --> 00:41:51,837 verifiable domains. It didn't feel that much better at anything else. the more 795 00:41:51,920 --> 00:41:53,437 human angle of it. It felt like it 796 00:41:53,520 --> 00:41:55,677 regressed and like you had this uh sort 797 00:41:55,760 --> 00:41:59,837 of uh Reddit pitchfork uh sort of uh 798 00:41:59,920 --> 00:42:02,317 movement against against Sam and Open AI 799 00:42:02,400 --> 00:42:04,157 because they felt like they lost a 800 00:42:04,240 --> 00:42:06,877 friend. Gupta felt a lot more human and 801 00:42:06,960 --> 00:42:10,957 closer uh whereas GT5 felt a lot more 802 00:42:11,040 --> 00:42:14,557 robotic, you know, very in its head kind of trying to think through through 803 00:42:14,640 --> 00:42:17,437 everything. And um and so I I I would 804 00:42:17,520 --> 00:42:19,117 have just expected like when we went 805 00:42:19,200 --> 00:42:23,037 from GP2 2 to 3, it was clear it was 806 00:42:23,120 --> 00:42:25,757 getting a lot more human. It was uh a 807 00:42:25,840 --> 00:42:30,557 lot closer to our experience. It can you can feel like it's actually all it gets 808 00:42:30,640 --> 00:42:32,397 me like there's something about it that 809 00:42:32,480 --> 00:42:34,557 understands the world better. Similarly 810 00:42:34,640 --> 00:42:39,277 3 to four to five didn't feel like it 811 00:42:39,360 --> 00:42:42,717 was a better overall 812 00:42:42,800 --> 00:42:45,597 being as it were. But is that is that is 813 00:42:45,680 --> 00:42:47,917 that is that a is the question there 814 00:42:48,000 --> 00:42:50,717 like is it emotionality? Is it partly 815 00:42:50,800 --> 00:42:53,437 emotionality but but again partly like I 816 00:42:53,520 --> 00:42:55,437 like to ask models like very 817 00:42:55,520 --> 00:42:58,557 controversial uh things. Um can it 818 00:42:58,640 --> 00:43:01,117 reason through 819 00:43:01,200 --> 00:43:06,717 uh I don't know how deep we want to go here but like um what happened with 820 00:43:06,800 --> 00:43:09,837 World Trade 7, >> right? >> Sure. 821 00:43:09,920 --> 00:43:11,197 >> It's an interesting question, right? 822 00:43:11,280 --> 00:43:15,437 Like I'm not I'm not putting out a theory, but like it's interesting like 823 00:43:15,520 --> 00:43:17,517 how did it you know and and can it can 824 00:43:17,600 --> 00:43:19,917 it think through controversial questions 825 00:43:20,000 --> 00:43:22,557 >> in the same way that it can go think 826 00:43:22,640 --> 00:43:27,597 through a coding problem and there hasn't been any movement there like the 827 00:43:27,680 --> 00:43:30,797 all the reasoning and all of that stuff I haven't se and not just that you know 828 00:43:30,880 --> 00:43:36,237 that's a cute example but like um co right like you know the origins of co 829 00:43:36,320 --> 00:43:41,197 right >> you know go you know dig up GPT4 or other models 830 00:43:41,280 --> 00:43:44,957 and go to GPT5, you're not going to find that much difference of okay, let's 831 00:43:45,040 --> 00:43:48,157 reason together. Let's try to figure out what was the origins of CO because it's 832 00:43:48,240 --> 00:43:52,237 still an unanswered question, you know, and I don't see them making progress in 833 00:43:52,320 --> 00:43:53,517 that. I mean, you play a lot with them. 834 00:43:53,600 --> 00:43:55,277 Do you feel like >> I use it differently? I don't know, 835 00:43:55,360 --> 00:43:59,597 maybe I have different expectations. Um, I I'm I the way I my main use case 836 00:43:59,680 --> 00:44:03,437 actually is sort of sort of PhD and everything at my beck and call. Um, and 837 00:44:03,520 --> 00:44:05,997 so I'm I'm trying to get it to explain things to me more than I'm trying to 838 00:44:06,080 --> 00:44:09,357 like, you know, have conversations with it. Maybe maybe I'm just unusual with 839 00:44:09,440 --> 00:44:12,317 that. But >> and that that that gets back >> well. So what I what I what I found 840 00:44:12,400 --> 00:44:16,397 specifically is uh a combination of like GPT5 Pro plus deep reasoning or like 841 00:44:16,480 --> 00:44:21,277 Rock 4 heavy like the you know the the highest end models um u like that um you 842 00:44:21,360 --> 00:44:24,797 know they now basically generate 30 to 40 page you know essentially books on 843 00:44:24,880 --> 00:44:29,677 demand on any topic. Um and so anytime I get curious about something you just 844 00:44:29,760 --> 00:44:32,397 take it maybe it's my version of it but it's something like I don't like a good 845 00:44:32,480 --> 00:44:36,157 here's a good example. um when when when a when an advanced economy puts a tariff 846 00:44:36,240 --> 00:44:38,957 on on on a raw m you know on a raw material or on a finished good like who 847 00:44:39,040 --> 00:44:42,477 pays >> you know is it is it the consumer is it the is it the importer is it the 848 00:44:42,560 --> 00:44:45,677 exporter or is it the producer and and this actually a very complicated it 849 00:44:45,760 --> 00:44:48,557 turns out very complicated question it's a big big big thing that economists 850 00:44:48,640 --> 00:44:51,757 study a lot and it's just like okay who you know who pays and what I found like 851 00:44:51,840 --> 00:44:53,677 for that kind of thing is it's outstanding 852 00:44:53,760 --> 00:44:57,437 >> well well but but it's outstanding at um sort of going out of the web getting 853 00:44:57,520 --> 00:45:00,637 information synthesizing it >> correct it it gives me it gives me a 854 00:45:00,720 --> 00:45:02,957 synthesized 20 30 40 p basically tops 855 00:45:03,040 --> 00:45:05,437 out 40 p 40 40 pages of PDF. Yeah. 856 00:45:05,520 --> 00:45:09,037 >> Um uh but I can get I can get up to 40 pages of PDF but it's a completely 857 00:45:09,120 --> 00:45:12,317 coherent and as far as I can tell for everything I've cross-cheed a completely 858 00:45:12,400 --> 00:45:15,997 like it like world class like if I hired you know for a question like that if I 859 00:45:16,080 --> 00:45:20,317 hired like a great you know econ posttock at Stanford who just like went 860 00:45:20,400 --> 00:45:22,877 out and did that work like it would maybe be that good. 861 00:45:22,960 --> 00:45:24,477 >> Yeah. Um but then but then of course the 862 00:45:24,560 --> 00:45:28,077 significance is it's like it's like you know at least for this is true for many 863 00:45:28,160 --> 00:45:29,917 domains you know kind of PhD and everything and so 864 00:45:30,000 --> 00:45:33,277 >> but but this is synthesizing knowledge not trying to create new knowledge. 865 00:45:33,360 --> 00:45:35,037 >> Well but this this gets to the this sort 866 00:45:35,120 --> 00:45:37,837 of you know of course the you get into the angels dancing on the head of a pin 867 00:45:37,920 --> 00:45:40,797 thing which is like what what you know what's the difference how many how much 868 00:45:40,880 --> 00:45:44,157 new knowledge ever actually is there anyway? What do you actually expect from 869 00:45:44,240 --> 00:45:47,437 people when you ask them questions? Um, and so what what I'm looking for is 870 00:45:47,520 --> 00:45:50,157 like, yes, explain this to me in like the the the clearest, most 871 00:45:50,240 --> 00:45:53,677 sophisticated, most complex, most like complete way that it's possible for 872 00:45:53,760 --> 00:45:57,757 somebody to, you know, for a real expert to be able to to to explain things to me. 873 00:45:57,840 --> 00:46:00,077 >> Um, and that's what I use it for. And again, as far as I can tell from the 874 00:46:00,160 --> 00:46:02,877 crossing, like I'm getting, you know, like almost like basically 100 out of 875 00:46:02,960 --> 00:46:07,517 100, like I don't even think I've had an issue in months where it's like had had 876 00:46:07,600 --> 00:46:09,917 a problem in it. >> And it's like, yeah, you can say, yeah, 877 00:46:10,000 --> 00:46:12,237 synthesizing is supposed to create new information, but like it's it's 878 00:46:12,320 --> 00:46:15,037 generating a 40 p. He's basically generating a 40-page book. 879 00:46:15,120 --> 00:46:18,157 >> That's amazing. >> That's like incredibly like fluid. It's, 880 00:46:18,240 --> 00:46:22,077 you know, it's it's it's it's you know, the the logical coherence of the entire 881 00:46:22,160 --> 00:46:26,477 like it's it's a great writing. Like if if you if you evaluated an a a human 882 00:46:26,560 --> 00:46:29,677 author on it, you would say, "Wow, that's a great author." You know, do are 883 00:46:29,760 --> 00:46:32,877 people who write books, you know, creating new knowledge? Well, yeah. 884 00:46:32,960 --> 00:46:35,997 Well, sort of not because a lot of what they're doing is building on everything 885 00:46:36,080 --> 00:46:39,677 that came before them is synthesizing a mind, but also like a book is a creative 886 00:46:39,760 --> 00:46:43,517 accomplishment, right? And so, >> yeah, one of the thing I'm I'm I'm I'm 887 00:46:43,600 --> 00:46:45,917 interested in I'm hoping AI could help 888 00:46:46,000 --> 00:46:49,357 us solve is just like how confusing the 889 00:46:49,440 --> 00:46:53,037 information ecosystem right now. You know, everything feels like propaganda. 890 00:46:53,120 --> 00:46:55,997 Like it doesn't feel like you're getting real information from anywhere. So, I I 891 00:46:56,080 --> 00:46:59,837 really want an AI that could help me reason from first principles about 892 00:46:59,920 --> 00:47:04,237 what's happening in the world for me to actually get real information. and and 893 00:47:04,320 --> 00:47:08,877 maybe that's an unreasonable sort of ask of of the AI researchers, but but I 894 00:47:08,960 --> 00:47:12,077 don't think we're we have made any progress there. So maybe I'm over focus. 895 00:47:12,160 --> 00:47:15,277 Yeah, maybe I'm over in being in my my line or maybe I'm over focused on ar 896 00:47:15,360 --> 00:47:18,877 arguing at people as opposed to um trying to get as trying to get underline 897 00:47:18,960 --> 00:47:21,917 truth. But well here here's the thing I I do a lot with this is I just say like 898 00:47:22,000 --> 00:47:25,837 take take a provocative point of view um and then steel man the position take 899 00:47:25,920 --> 00:47:29,277 your co thing steel so I often I often pair these steel man the position that 900 00:47:29,360 --> 00:47:32,877 it was a lab leak um and steel man the position that it was natural origins 901 00:47:32,960 --> 00:47:34,317 >> um and and again like is this creativity 902 00:47:34,400 --> 00:47:37,357 or not? I don't know. But like what comes back is like 30 pages each of like 903 00:47:37,440 --> 00:47:40,237 wow like that is like the most compelling case in the world I can 904 00:47:40,320 --> 00:47:42,877 imagine with like every you know everything marshaled against it like the 905 00:47:42,960 --> 00:47:44,557 argument structured in like the most possible 906 00:47:44,640 --> 00:47:47,037 >> part of the reason that started happening is because it stopped being 907 00:47:47,120 --> 00:47:51,277 taboo to talk about a human origin when it was taboo 908 00:47:51,360 --> 00:47:53,917 >> the the AIS would like talk will you 909 00:47:54,000 --> 00:47:55,757 know talk down to like oh you're a 910 00:47:55,840 --> 00:47:58,957 conspiracy theorist and so yes uh so 911 00:47:59,040 --> 00:48:02,877 there's a there's a you know period of time and so it takes something truly 912 00:48:02,960 --> 00:48:05,757 controversial and they actually they they can't reason about it because of 913 00:48:05,840 --> 00:48:09,677 all RLHF and answers all the limitations and as as you know I won't pick no 914 00:48:09,760 --> 00:48:12,237 specific ones here but like there there are certain certain big models that will 915 00:48:12,320 --> 00:48:14,877 still lecture you >> that you're a bad person for asking that 916 00:48:14,960 --> 00:48:17,917 question but but you know like I just there some of them are just like really 917 00:48:18,000 --> 00:48:19,437 really open now to you know being able 918 00:48:19,520 --> 00:48:20,797 to do these things 919 00:48:20,880 --> 00:48:25,517 >> um and then um uh yeah so um okay uh 920 00:48:25,600 --> 00:48:29,517 yeah so okay so yeah so there's this yeah so so basically like ultimately 921 00:48:29,600 --> 00:48:32,717 what you're looking for like the ultimate thing would be if there's 922 00:48:32,800 --> 00:48:34,957 something that's like I don't I think anybody's really defined this really 923 00:48:35,040 --> 00:48:37,597 well because it's not because again it's like the conventional all the 924 00:48:37,680 --> 00:48:40,077 conventional definitions of AGI are like basically comparing to people. 925 00:48:40,160 --> 00:48:41,757 >> Yeah. >> And there there and there it's always 926 00:48:41,840 --> 00:48:44,317 like you know it's the conventional 927 00:48:44,400 --> 00:48:49,277 explanations of of of um of uh of AGI always for me struck me a lot like the 928 00:48:49,360 --> 00:48:51,677 debate around like whether a self-driving car works or not which is 929 00:48:51,760 --> 00:48:55,277 is does a self-driving car work because it's a perfect driver uh or does it work 930 00:48:55,360 --> 00:48:58,877 because it is a is better than the human driver and better than the human driver 931 00:48:58,960 --> 00:49:01,757 I think is actually quite you know just like with the the chess thing and the go 932 00:49:01,840 --> 00:49:04,477 thing. I actually think like that that that's like a real thing. And then and 933 00:49:04,560 --> 00:49:07,277 then and then and then there's the like is it a perfect driver which is you know 934 00:49:07,360 --> 00:49:09,597 what they're obviously the the self-driving car companies are working 935 00:49:09,680 --> 00:49:12,557 for but then I think you're looking for something beyond the perfect driver. 936 00:49:12,640 --> 00:49:16,077 You're looking for the car who like knows where to go. So I I I I'm of two 937 00:49:16,160 --> 00:49:20,157 minds, right? So one mind is the sort of practical entrepreneur, right? 938 00:49:20,240 --> 00:49:22,797 >> Uh and I just I have so many toys to 939 00:49:22,880 --> 00:49:27,277 play with to build like stop AI progress today and replet will continue to get 940 00:49:27,360 --> 00:49:30,477 better for the next 5 years. like there's so much we can do just on the 941 00:49:30,560 --> 00:49:33,197 app uh app layer and the infrastructure layer. 942 00:49:33,280 --> 00:49:37,757 >> So you know I but but I think that will you know the the foundation models will 943 00:49:37,840 --> 00:49:40,877 continue to get better as well and so it's it's a very exciting time in our 944 00:49:40,960 --> 00:49:45,277 industry. Um the other mind is more academic because as a kid I've always 945 00:49:45,360 --> 00:49:49,197 been interested in the nature of consciousness, nature of intelligence. I 946 00:49:49,280 --> 00:49:51,277 was always interested in AI and reading 947 00:49:51,360 --> 00:49:56,397 the literature there and I would point to the RL uh literature. So Richard 948 00:49:56,480 --> 00:50:00,397 Soden, there's another guy I think co-founder of deep mind Shane Lag wrote 949 00:50:00,480 --> 00:50:02,797 wrote a paper trying to define what AGI 950 00:50:02,880 --> 00:50:06,237 is. Um and in there I think that the 951 00:50:06,320 --> 00:50:08,717 definition of AI I think is the is the 952 00:50:08,800 --> 00:50:12,317 original perhaps correct one which is uh 953 00:50:12,400 --> 00:50:16,237 efficient continual learning. >> Okay. Like if you if you truly want to 954 00:50:16,320 --> 00:50:20,557 build an artificial general intelligence that you can drop in any domain, you can 955 00:50:20,640 --> 00:50:23,117 drop in a car without that much prior 956 00:50:23,200 --> 00:50:27,437 knowledge about cars and within um you 957 00:50:27,520 --> 00:50:31,117 know how long does it take a human to to learn how to drive you within months be 958 00:50:31,200 --> 00:50:35,437 able to drive a car really well, you know, generalized skill sort of 959 00:50:35,520 --> 00:50:39,437 generalized skill acquisition, generalized understanding acquisition, 960 00:50:39,520 --> 00:50:43,277 generalized reasoning acquisition. And I think that's the thing that will like 961 00:50:43,360 --> 00:50:47,437 truly change the world. That's the thing that would give us a better 962 00:50:47,520 --> 00:50:49,597 understanding of of the human mind of 963 00:50:49,680 --> 00:50:51,917 human consciousness and that's the thing 964 00:50:52,000 --> 00:50:54,637 that will like propel us to the next uh 965 00:50:54,720 --> 00:50:57,437 level of human civilization. 966 00:50:57,520 --> 00:50:59,757 on a civilization level that's a really 967 00:50:59,840 --> 00:51:04,760 deep question but separ 968 00:51:05,600 --> 00:51:07,517 but but there's an academic aspect of it that I'm really 969 00:51:07,600 --> 00:51:10,477 >> so what and what odds what if we're on if we're on Kelsey today what what odds 970 00:51:10,560 --> 00:51:11,677 do we place on that 971 00:51:11,760 --> 00:51:15,197 >> I I I'm kind of bearish on on on true 972 00:51:15,280 --> 00:51:17,757 AGI breakthrough because 973 00:51:17,840 --> 00:51:20,797 >> what we built is so useful and 974 00:51:20,880 --> 00:51:23,997 economically valuable uh so in a way 975 00:51:24,080 --> 00:51:28,157 >> good enough good enough is the enemy Yeah. Yeah. Do do you remember that 976 00:51:28,240 --> 00:51:29,997 essay? Um, >> worse is better. 977 00:51:30,080 --> 00:51:32,397 >> Worse is better. >> Worse is better. Worse is better. And and 978 00:51:32,480 --> 00:51:35,117 >> so there's like a local there's like a trap. There's like a local local maximum 979 00:51:35,200 --> 00:51:38,557 trap. We're in a local maximum >> local maximum trap where it's because 980 00:51:38,640 --> 00:51:41,757 it's because it's good enough for so much economically productive work. 981 00:51:41,840 --> 00:51:44,477 >> Yes. >> It relieves the pressure um in the 982 00:51:44,560 --> 00:51:46,077 system to create the generalized answer. 983 00:51:46,160 --> 00:51:49,197 >> Yes. And and then you have the weirdos like Rich Sutton and others that are 984 00:51:49,280 --> 00:51:51,917 still trying to go that down that path and maybe they'll succeed, 985 00:51:52,000 --> 00:51:55,997 >> right? Uh but there's enormous optimization energy behind the current 986 00:51:56,080 --> 00:51:58,637 thing that we're hell climbing on this like local maximum. 987 00:51:58,720 --> 00:52:01,037 >> Right. Right. Right. And and the irony of it is everybody's worried about like 988 00:52:01,120 --> 00:52:03,517 the you know the gazillions of dollars going into building out all this stuff 989 00:52:03,600 --> 00:52:06,797 and and so the the the most ironic thing in the world would be if the gazillions 990 00:52:06,880 --> 00:52:08,477 of dollars are going into the local maximum. 991 00:52:08,560 --> 00:52:10,557 >> That's right. >> As as opposed to a counterfactual world 992 00:52:10,640 --> 00:52:12,397 in which they're going into solving the general problem. 993 00:52:12,480 --> 00:52:15,197 >> But but it's also potentially irrational. Like maybe the general 994 00:52:15,280 --> 00:52:20,077 problem is actually you know not within our lifetimes. Who knows? Right. Um, h 995 00:52:20,160 --> 00:52:22,877 >> how much further do you think like do you think we squeeze most of the juice 996 00:52:22,960 --> 00:52:26,077 out of out of LLMs in general then? Or are there any other research directions 997 00:52:26,160 --> 00:52:29,597 that you're particularly um excited about? 998 00:52:29,680 --> 00:52:31,837 >> Well, that's the thing. I think the 999 00:52:31,920 --> 00:52:33,997 problem is there aren't that that many. 1000 00:52:34,080 --> 00:52:38,717 I I think the the the breakthroughs in RL are incredibly exciting, but we also 1001 00:52:38,800 --> 00:52:42,637 knew about them now for like over 10 years where you marry generative uh 1002 00:52:42,720 --> 00:52:47,357 systems with uh with tree search and things like that. Um but but there's a 1003 00:52:47,440 --> 00:52:51,997 lot more to go there and I think again the the the the original minds behind 1004 00:52:52,080 --> 00:52:55,517 reinforcement learning are trying to go down that path and try to kind of 1005 00:52:55,600 --> 00:53:00,557 bootstrap intelligence from scratch. Uh Carmarmac is is going down that path as 1006 00:53:00,640 --> 00:53:04,237 as far as I understand Carmarmac you guys may be invested but the the you 1007 00:53:04,320 --> 00:53:07,277 know they're they're not trying to go down the LLM path. So there are people 1008 00:53:07,360 --> 00:53:11,997 that are trying to do that but I'm not seeing a lot of progress or outcome 1009 00:53:12,080 --> 00:53:13,837 there but I watch it kind of from far. 1010 00:53:13,920 --> 00:53:16,237 Although, you know, for all we know, it's already there's already a bot on X 1011 00:53:16,320 --> 00:53:19,837 somewhere. >> You know, you know, you never know. It 1012 00:53:19,920 --> 00:53:21,917 might not be a big announcement. It might just be a you know, one day 1013 00:53:22,000 --> 00:53:24,477 there's just like a bot on X that starts winning all the arguments. 1014 00:53:24,560 --> 00:53:28,957 >> Yeah, it could be >> or a code a user read and all of a 1015 00:53:29,040 --> 00:53:31,677 sudden it's >> generating incredible software. Um, 1016 00:53:31,760 --> 00:53:34,477 okay. Let's uh let's spend our remaining minutes. Let's let's let's talk about 1017 00:53:34,560 --> 00:53:38,717 you. So, uh so uh so how so yeah, take us start from the beginning with your uh 1018 00:53:38,800 --> 00:53:41,197 with your life and how how did you get how did you get from being born to being 1019 00:53:41,280 --> 00:53:45,597 in Silicon Valley? >> Okay. um >> in two minutes. Yeah, 1020 00:53:45,680 --> 00:53:47,197 >> I'm just I'm joking. But 1021 00:53:47,280 --> 00:53:50,557 >> yeah, I I got introduced to computers uh 1022 00:53:50,640 --> 00:53:53,437 very very early on. And so for whatever 1023 00:53:53,520 --> 00:53:58,557 reason, so I was born in Aman, Jordan >> and for whatever reason, my my dad who 1024 00:53:58,640 --> 00:54:00,397 was just a government engineer at the 1025 00:54:00,480 --> 00:54:03,597 time uh decided that computers were 1026 00:54:03,680 --> 00:54:07,757 important and he didn't have a lot of money took out of that, bought a 1027 00:54:07,840 --> 00:54:10,477 computer. It was the first computer in their in our neighborhood. first 1028 00:54:10,560 --> 00:54:14,317 computer of anyone I know. And I just 1029 00:54:14,400 --> 00:54:18,077 one of my earliest memories I was 6 years old just watching my dad unpack 1030 00:54:18,160 --> 00:54:20,877 this machine and sort of open up this 1031 00:54:20,960 --> 00:54:24,317 huge manual and kind of finger type CD 1032 00:54:24,400 --> 00:54:29,197 LS MKDIR and like I would, you know, be behind his shoulder and just like 1033 00:54:29,280 --> 00:54:33,197 watching him, you know, type these commands and seeing the sort of machine 1034 00:54:33,280 --> 00:54:36,157 kind of respond and do exactly what he's 1035 00:54:36,240 --> 00:54:41,680 asked it to do. Um, >> pop in Tylenol as your 1036 00:54:41,840 --> 00:54:47,757 Exactly. Autism activated. >> Of course, you have to. 1037 00:54:47,840 --> 00:54:51,517 >> You have to. >> Exactly. What kind of um what kind of 1038 00:54:51,600 --> 00:54:55,917 computer was it? >> Uh it was uh an IBM as far as I 1039 00:54:56,000 --> 00:55:00,557 remember. It was IBM PC. >> What year was this? Uh 1993. 1040 00:55:00,640 --> 00:55:03,917 >> 1993. Okay. So, it's DOS. So, did it have Windows at that point or 1041 00:55:04,000 --> 00:55:05,197 >> No, it didn't have Windows. 1042 00:55:05,280 --> 00:55:07,917 >> Right before Windows. right before Windows, but I think Windows had been 1043 00:55:08,000 --> 00:55:11,277 out, but you would add >> you it was an add-on. You wouldn't boot 1044 00:55:11,360 --> 00:55:13,597 it up. So, we I think we bought the disc 1045 00:55:13,680 --> 00:55:16,397 for uh for uh for Windows and you had to 1046 00:55:16,480 --> 00:55:20,797 kind of uh bootloaded, you know, from the disk and and it will open Windows 1047 00:55:20,880 --> 00:55:24,077 and you can click around. It wasn't that interesting cuz there wasn't a lot on 1048 00:55:24,160 --> 00:55:25,917 it. So, a lot of time I just spend in 1049 00:55:26,000 --> 00:55:28,477 DOSs and writing batch files and opening 1050 00:55:28,560 --> 00:55:31,597 games and messing around with that. Um 1051 00:55:31,680 --> 00:55:36,637 but it wasn't until Visual Basic that I started. So like after Windows 95 that I 1052 00:55:36,720 --> 00:55:38,317 started making real software, right? 1053 00:55:38,400 --> 00:55:41,997 >> Uh and the first idea I had was um I I 1054 00:55:42,080 --> 00:55:44,477 used to be a huge gamer. So I used I 1055 00:55:44,560 --> 00:55:49,437 used to go to these uh Lang gaming cafes and play Counter Strike and I would go 1056 00:55:49,520 --> 00:55:52,797 there and you know the whole thing is full of computers but they don't use any 1057 00:55:52,880 --> 00:55:56,477 software to run their business. M it was just like people running around just 1058 00:55:56,560 --> 00:56:00,317 like writing down your machine number, how much time you spend on it and how 1059 00:56:00,400 --> 00:56:03,277 much did you pay and kind of tapping your shoulders like hey you need to pay 1060 00:56:03,360 --> 00:56:06,797 a little more for that. And I asked him like why don't you like just build a 1061 00:56:06,880 --> 00:56:09,997 piece of software that allows me to log in and have a time or whatever. And I 1062 00:56:10,080 --> 00:56:13,117 was like yeah we don't know how to do that. And I was like okay I think I know 1063 00:56:13,200 --> 00:56:17,357 how to do that. So I spent I was like 12 or something like that. I spent like 2 1064 00:56:17,440 --> 00:56:21,517 years building that uh and then went out and tried to sell it and was able to 1065 00:56:21,600 --> 00:56:23,597 sell it uh and was making so much money. 1066 00:56:23,680 --> 00:56:26,637 I remember McDonald's opened uh in 1067 00:56:26,720 --> 00:56:30,717 Jordan uh around the time when I was 13 14. I took my entire class to 1068 00:56:30,800 --> 00:56:32,237 McDonald's. It was very expensive, but I 1069 00:56:32,320 --> 00:56:35,677 was balling it all this money and I was 1070 00:56:35,760 --> 00:56:39,677 showing off um and uh and so that was 1071 00:56:39,760 --> 00:56:43,517 the first uh business that I uh created. 1072 00:56:43,600 --> 00:56:46,397 And then when it came to and at the time 1073 00:56:46,480 --> 00:56:48,877 I started kind of learning about AI, you 1074 00:56:48,960 --> 00:56:53,757 know, reading sci-fi and all of that stuff. And when it came time to go to 1075 00:56:53,840 --> 00:56:57,917 college, uh I didn't want to go to computer science because I felt like 1076 00:56:58,000 --> 00:57:02,877 coding is on its way to get automated. I remember using these um wizards. Do you 1077 00:57:02,960 --> 00:57:05,997 remember wizards? >> Yes. >> Wizards basically. It's like extremely 1078 00:57:06,080 --> 00:57:09,197 crude early bots or that generate code. 1079 00:57:09,280 --> 00:57:10,797 Yeah. >> Yeah. And I remember you could like, you 1080 00:57:10,880 --> 00:57:13,837 know, type in a few things like here's my project, here's what it does, 1081 00:57:13,920 --> 00:57:17,197 whatever, and then click click click and just like scaffold a lot of code. I was 1082 00:57:17,280 --> 00:57:18,637 like, oh, I think that's the future. 1083 00:57:18,720 --> 00:57:22,637 Like coding is such a >> it's almost >> yeah it's solved you know why why should 1084 00:57:22,720 --> 00:57:27,197 I go into coding I was okay if AI can do the code what should I do well someone 1085 00:57:27,280 --> 00:57:30,317 needs to build and maintain the computers and so I went to the computer 1086 00:57:30,400 --> 00:57:32,477 engineering and and and did that for a 1087 00:57:32,560 --> 00:57:35,757 while uh but then rediscovered my love 1088 00:57:35,840 --> 00:57:38,397 for for programming uh reading program 1089 00:57:38,480 --> 00:57:41,197 essays on lisp and things like that and 1090 00:57:41,280 --> 00:57:45,597 uh started messing around with scheme and programming languages like that um 1091 00:57:45,680 --> 00:57:49,277 but then I found it incredibly difficult to just like learn different programming 1092 00:57:49,360 --> 00:57:53,277 languages. I didn't have a laptop at the time. And so every time I go to like 1093 00:57:53,360 --> 00:57:56,077 wanting to learn Python or Java, I would 1094 00:57:56,160 --> 00:57:58,717 go to the computer lab, download 1095 00:57:58,800 --> 00:58:01,517 gigabytes of software, try to set it up, 1096 00:58:01,600 --> 00:58:03,677 type a little bit of code, try to run 1097 00:58:03,760 --> 00:58:06,797 it, you know, run into missing DL issue 1098 00:58:06,880 --> 00:58:10,077 or and I was like, man, this is so 1099 00:58:10,160 --> 00:58:13,037 primitive. Like at the time it was 2008 1100 00:58:13,120 --> 00:58:16,157 something like that you know we had uh 1101 00:58:16,240 --> 00:58:18,317 Google Docs, we had Gmail, you could 1102 00:58:18,400 --> 00:58:21,357 like open the browser uh and partly 1103 00:58:21,440 --> 00:58:23,917 thanks to you and be able to kind of uh 1104 00:58:24,000 --> 00:58:28,397 use software on the internet and I thought the web is the ultimate software 1105 00:58:28,480 --> 00:58:31,917 platform like everything should go on the web. Okay, who's building an online 1106 00:58:32,000 --> 00:58:35,837 development environment, right? And and no one, right? And it felt like I w I 1107 00:58:35,920 --> 00:58:40,637 found like $100 bill on the, you know, on the floor of Grand Association. Like 1108 00:58:40,720 --> 00:58:44,237 surely someone should be building this, but no, no one was building this. And so 1109 00:58:44,320 --> 00:58:46,557 I was like, okay, I'll I'll try to build it. 1110 00:58:46,640 --> 00:58:51,437 >> And I got something done in like a couple hours. Uh, which was a text box. 1111 00:58:51,520 --> 00:58:54,797 You type in some JavaScript. We And there's a there's a button that says 1112 00:58:54,880 --> 00:58:57,597 eval. You click eval and evaluates. It 1113 00:58:57,680 --> 00:58:59,757 shows you in a in an alert box, right? 1114 00:58:59,840 --> 00:59:03,917 Right. So oneplus 1 2 I was like oh I have a programming environment. I showed 1115 00:59:04,000 --> 00:59:07,197 it to my friends people started using it. I added a few additional things like 1116 00:59:07,280 --> 00:59:11,037 saving the program. I was like okay all right this is there's there's a real 1117 00:59:11,120 --> 00:59:15,277 idea here. People love it. And then again it took me two two or three years 1118 00:59:15,360 --> 00:59:18,877 to actually be able to build anything because you know the browser can only 1119 00:59:18,960 --> 00:59:21,117 run JavaScript. And it took a 1120 00:59:21,200 --> 00:59:24,077 breakthrough at the time. Uh, Moisella 1121 00:59:24,160 --> 00:59:26,397 had a research project called mcripton 1122 00:59:26,480 --> 00:59:30,077 that allowed you to uh compile different 1123 00:59:30,160 --> 00:59:32,237 uh programming languages like C, C++ 1124 00:59:32,320 --> 00:59:35,597 into JavaScript. And for the browser to 1125 00:59:35,680 --> 00:59:39,437 be able to run something like Python, I needed to compile C, Python to 1126 00:59:39,520 --> 00:59:44,237 JavaScript. So I was the first to do it in the world. uh so built uh contributed 1127 00:59:44,320 --> 00:59:45,677 to that project and built a lot of the 1128 00:59:45,760 --> 00:59:49,357 scaffolding around it and we uh my 1129 00:59:49,440 --> 00:59:53,917 friends and I compiled Python into JavaScript and I was like okay we did it 1130 00:59:54,000 --> 00:59:58,077 for Python let's do it for Ruby let's do it for Lo and that's how the emergence 1131 00:59:58,160 --> 01:00:01,837 of the idea for replet came is that when you need a ripple you should get it you 1132 01:00:01,920 --> 01:00:05,437 should replet it and so ripple is is the most primitive programming environment 1133 01:00:05,520 --> 01:00:08,877 possible so I added all these programming languages and again all this 1134 01:00:08,960 --> 01:00:11,597 time my friends were using it and excited about 1135 01:00:11,680 --> 01:00:15,997 And I was on GitHub at the time and just my standard thing is like when I make a 1136 01:00:16,080 --> 01:00:19,677 piece of software is open source it. And so I was open sourcing all the things I 1137 01:00:19,760 --> 01:00:23,517 was you know years building just like this underlying infrastructure to be 1138 01:00:23,600 --> 01:00:25,757 able to just run code in the browser 1139 01:00:25,840 --> 01:00:28,157 >> and then it went viral uh went viral in 1140 01:00:28,240 --> 01:00:31,357 hacker news and it coincided with the 1141 01:00:31,440 --> 01:00:34,397 MOO era. So massively online courses 1142 01:00:34,480 --> 01:00:38,797 Udacity was coming online Corsera and and most famously Code Academy. Right. 1143 01:00:38,880 --> 01:00:42,717 So, Code Academy uh was the first kind of website that allowed you to code in 1144 01:00:42,800 --> 01:00:46,477 the browser interactively and learn how to code. And they built a lot of it on 1145 01:00:46,560 --> 01:00:49,597 my software that I was open sourcing all the way from Jordan. And so, I remember 1146 01:00:49,680 --> 01:00:52,237 seeing them on Hacker News and they were going super viral. I was like, "Hey, 1147 01:00:52,320 --> 01:00:54,317 that's, you know, I I recognize this. 1148 01:00:54,400 --> 01:00:56,797 What are you using?" And so, I left the Hacker News comments. I was like, "Oh, 1149 01:00:56,880 --> 01:00:59,117 you're using my my open source package." 1150 01:00:59,200 --> 01:01:02,077 And so, they reached out to me. They uh they're like, "Hey, would like to hire 1151 01:01:02,160 --> 01:01:04,877 you." I was like, "I'm not interested. I I want to start a startup. I want to 1152 01:01:04,960 --> 01:01:09,197 start this thing called Replet." and and they're like, "Well, no, you know, you 1153 01:01:09,280 --> 01:01:12,717 should come work with us. We can we can do the same stuff." And I kept saying, 1154 01:01:12,800 --> 01:01:16,157 "No." I was like, "Okay, I'll contract with you." They were paying me $12 an 1155 01:01:16,240 --> 01:01:17,517 hour. I was really excited about it. 1156 01:01:17,600 --> 01:01:21,357 Back from Oman. >> Um, but they came out to their to their 1157 01:01:21,440 --> 01:01:24,397 credit. They came out of Jordan to recruit me and spend a few a few days 1158 01:01:24,480 --> 01:01:28,157 there. And then I, you know, I kept saying no. And in the end, they gave me 1159 01:01:28,240 --> 01:01:33,037 an offer I can't refuse. Um, and they got me an O1 visa. Came to the United States. 1160 01:01:33,120 --> 01:01:37,037 >> That's when you moved. So when when was the first cuz you were born what year to 1161 01:01:37,120 --> 01:01:39,357 >> 1987 >> 87. What was the first year that you 1162 01:01:39,440 --> 01:01:41,917 could remember where you had the idea that you might not live your life in 1163 01:01:42,000 --> 01:01:44,317 Jordan and that you might you might actually move to the US? 1164 01:01:44,400 --> 01:01:46,557 >> Uh when I watched Pirates of Silicon Valley. 1165 01:01:46,640 --> 01:01:47,837 >> Is that right? Okay. Got it. All right. 1166 01:01:47,920 --> 01:01:52,157 >> Uh maybe 98 or 99. I don't know when it came out. 1167 01:01:52,240 --> 01:01:53,917 >> Okay. That might be a good place to >> Yeah. 1168 01:01:54,000 --> 01:01:56,397 >> Is it worth telling the hacker story because there's a version of the world 1169 01:01:56,480 --> 01:01:59,677 where you didn't actually like if that changed maybe you wouldn't have gone to 1170 01:01:59,760 --> 01:02:03,197 America. >> Right. Right. Yeah. So uh in in school I 1171 01:02:03,280 --> 01:02:06,717 was programming the whole time you so I I just want to start businesses. I just 1172 01:02:06,800 --> 01:02:09,997 like I'm exploding with ideas all the time. And like the reason Replet exists 1173 01:02:10,080 --> 01:02:13,357 is because I have ideas all the time. I just want to go type on the computer and 1174 01:02:13,440 --> 01:02:17,117 like build them. Um so I wasn't going to school. It was like incredibly boring 1175 01:02:17,200 --> 01:02:21,117 for me. Uh and part of the reason why Replet has a mobile app today is because 1176 01:02:21,200 --> 01:02:25,037 I always wanted to program under the desk just to do things. 1177 01:02:25,120 --> 01:02:27,917 >> Um and so the at school they kept 1178 01:02:28,000 --> 01:02:30,557 failing me uh for attendance. you know, 1179 01:02:30,640 --> 01:02:35,037 so I would get A's, but I just didn't show up and so they they would fail me. 1180 01:02:35,120 --> 01:02:37,517 And so I felt it was incredibly unfair. 1181 01:02:37,600 --> 01:02:39,517 And all my friends were graduating now. 1182 01:02:39,600 --> 01:02:43,997 This year was like 2011. I've been like for 6 years in college. It should be 1183 01:02:44,080 --> 01:02:45,917 like a three or four year. And I was 1184 01:02:46,000 --> 01:02:47,517 like incredibly depressed. I really 1185 01:02:47,600 --> 01:02:51,037 wanted to be in Silicon Valley. And so I 1186 01:02:51,120 --> 01:02:54,877 was like, "Oh, what if I changed my grades?" 1187 01:02:54,960 --> 01:02:55,917 >> There we go. 1188 01:02:56,000 --> 01:02:59,517 >> The university database. And um and and 1189 01:02:59,600 --> 01:03:05,277 so I went into my parents uh uh basement 1190 01:03:05,360 --> 01:03:10,797 uh and uh implemented uh the polyphasic sleep. Are you familiar with that? 1191 01:03:10,880 --> 01:03:14,477 >> I I I I am >> uh Leonardo da Vinci's uh polyphysic 1192 01:03:14,560 --> 01:03:18,077 sleep. I didn't hear from Ronaldo da Vinci. I heard it from Seinfeld cuz uh 1193 01:03:18,160 --> 01:03:20,957 there's an episode where John Kamemer goes on on 1194 01:03:21,040 --> 01:03:25,917 >> poly sleep what 20 minutes every four hours. 20 minutes every 24 hours. And 1195 01:03:26,000 --> 01:03:27,997 yes, and this this somehow is going to work well. And it 1196 01:03:28,080 --> 01:03:30,477 >> Yeah. And and hacking, if you've ever done anything, 1197 01:03:30,560 --> 01:03:33,277 >> as the meme goes, this this has never worked for anybody else, but it might 1198 01:03:33,360 --> 01:03:36,397 but it might work for me. >> Yes. 1199 01:03:36,480 --> 01:03:40,557 >> And a lot of what hacking is is that you're you're coming up with ideas for 1200 01:03:40,640 --> 01:03:43,597 like finding certain security holes and like writing a script and then running 1201 01:03:43,680 --> 01:03:46,877 that script and that script will take like a 20 30 minutes to run and so 1202 01:03:46,960 --> 01:03:50,317 you'll take that, you know, 20 30 minutes to sleep and go on. So I spent 1203 01:03:50,400 --> 01:03:52,077 two weeks just going mad like trying to 1204 01:03:52,160 --> 01:03:54,877 hack into the university database and uh 1205 01:03:54,960 --> 01:03:58,477 finally I found um a way I found a SQL 1206 01:03:58,560 --> 01:04:02,557 injection somewhere on the site uh and I found a way to like be able to edit the 1207 01:04:02,640 --> 01:04:06,797 the records but I didn't want to risk it. So I went to my neighbor who's going 1208 01:04:06,880 --> 01:04:10,637 to the same school. Uh I think till this day no one caught him. But I went to him 1209 01:04:10,720 --> 01:04:15,357 and I said um hey uh I have this way to change grades like would you want to be 1210 01:04:15,440 --> 01:04:18,397 my guinea pig? And I was honest about it. I was like I'm not going to do it. 1211 01:04:18,480 --> 01:04:22,077 are you open to doing? He's like, "Yeah, yeah, yeah." They call 1212 01:04:22,160 --> 01:04:24,637 his human trials. 1213 01:04:24,720 --> 01:04:27,677 This is how medicine works. 1214 01:04:27,760 --> 01:04:32,397 So, so we we went and and and uh we went and changed his grades and he he went 1215 01:04:32,480 --> 01:04:36,957 and pulled his transcript and the you know, the update wasn't wasn't there and 1216 01:04:37,040 --> 01:04:41,517 went back to the basement. Well, turned out that I had access to the uh slave 1217 01:04:41,600 --> 01:04:43,997 database. I didn't have access to ambassador database. 1218 01:04:44,080 --> 01:04:48,397 >> So, find a way through the network privilege escalation. It was an Oracle 1219 01:04:48,480 --> 01:04:52,317 database that had a vulnerability and then found the real database and then I 1220 01:04:52,400 --> 01:04:54,877 just, you know, did it for myself. Uh, 1221 01:04:54,960 --> 01:04:58,557 changed the grades and went and pulled 1222 01:04:58,640 --> 01:05:00,797 my scrcripts and sure enough it actually 1223 01:05:00,880 --> 01:05:04,077 changed. Went and bought the the the 1224 01:05:04,160 --> 01:05:07,197 gown, went to the graduation parties, 1225 01:05:07,280 --> 01:05:10,637 uh, did all that. We're graduating. Um, 1226 01:05:10,720 --> 01:05:13,357 and then one day I'm at home. It's like 1227 01:05:13,440 --> 01:05:15,757 maybe 6:00 or 7:00 p.m. I get a you know 1228 01:05:15,840 --> 01:05:19,677 the the telephone at home rings 1229 01:05:19,760 --> 01:05:23,357 ominous ominous ring >> Santa 1230 01:05:23,440 --> 01:05:27,357 um hello and he's like hey this is the university registration system and I 1231 01:05:27,440 --> 01:05:31,277 knew the guy that run it. Uh he's like look you know we we we're having this 1232 01:05:31,360 --> 01:05:36,157 problem. The system's been down all day and it keeps coming back to your record. 1233 01:05:36,240 --> 01:05:38,157 there's an anomaly in your record where 1234 01:05:38,240 --> 01:05:40,237 you're both pass you have a passing 1235 01:05:40,320 --> 01:05:43,277 grade but you're also banned from that 1236 01:05:43,360 --> 01:05:47,837 uh final exam of subject I was like oh well turns out the database is not 1237 01:05:47,920 --> 01:05:50,157 normalized so typically that when they 1238 01:05:50,240 --> 01:05:54,877 ban you from an exam the the grades resets to 35 out of 100 but apparently 1239 01:05:54,960 --> 01:05:57,997 there's a boolean flag and by the way all the column names in the database are 1240 01:05:58,080 --> 01:06:02,797 single single letters that was the hardest thing is security by obscurity >> right 1241 01:06:02,880 --> 01:06:04,317 >> and turns out there's a flag that I 1242 01:06:04,400 --> 01:06:06,237 didn't track so when when when when you 1243 01:06:06,320 --> 01:06:09,917 go over attendance um uh when you don't 1244 01:06:10,000 --> 01:06:13,677 attend and they they they want to fail you, they they ban you from the final 1245 01:06:13,760 --> 01:06:18,557 exam. So, I changed the grades and that that that created uh an issue and 1246 01:06:18,640 --> 01:06:22,877 brought down the system. So, they were calling me and I thought at the time I 1247 01:06:22,960 --> 01:06:25,037 was like, you know, I could I could 1248 01:06:25,120 --> 01:06:27,197 potentially lie and I'll it'll be a huge 1249 01:06:27,280 --> 01:06:30,717 issue or I just like I'll just I'll just 1250 01:06:30,800 --> 01:06:34,637 fess up. Yeah. So, I said, hey, listen, look, um yeah, I might know something 1251 01:06:34,720 --> 01:06:38,957 about it. Hey, let me let me come uh tomorrow and kind of talk to you about 1252 01:06:39,040 --> 01:06:43,037 what happened. So, I go in and I open the door and it's the deans of all the 1253 01:06:43,120 --> 01:06:46,317 all the schools. It's a computer science computer. They were all working on it 1254 01:06:46,400 --> 01:06:50,717 for like days because it's like it's like it's a very computerheavy, you 1255 01:06:50,800 --> 01:06:53,197 know, university and it was like a problem 1256 01:06:53,280 --> 01:06:55,517 >> and they're all kind of really intrigued 1257 01:06:55,600 --> 01:06:59,597 about what happened. And so I pull up a whiteboard and started explaining what I 1258 01:06:59,680 --> 01:07:03,517 did and and everyone was engaged. I gave them a lecture basically. your oral exam 1259 01:07:03,600 --> 01:07:07,197 for your PhD. This is great. >> They were they were they were really 1260 01:07:07,280 --> 01:07:11,277 excited and uh and I think I it was endearing to them. I was like, "Oh, wow. 1261 01:07:11,360 --> 01:07:13,597 This is a this is a very interesting problem." 1262 01:07:13,680 --> 01:07:15,997 >> Um and then I was like, "Okay, great. 1263 01:07:16,080 --> 01:07:19,117 Thank you." And I was like, "Hey, wait, 1264 01:07:19,200 --> 01:07:20,957 >> wait. We don't know what to do with you. 1265 01:07:21,040 --> 01:07:24,397 Do we send you to jail? Do we 1266 01:07:24,480 --> 01:07:28,317 >> And uh I was like, "Hey, we have to escalate to the university um uh 1267 01:07:28,400 --> 01:07:30,717 president." and and he he was a great 1268 01:07:30,800 --> 01:07:35,757 man and I think uh he gave me a second chance in life and I went to him and I 1269 01:07:35,840 --> 01:07:39,117 uh you know I I explained the situation I said like I'm really frustrated. I 1270 01:07:39,200 --> 01:07:42,797 need to graduate. I need to get on with my life. I've been here for six years 1271 01:07:42,880 --> 01:07:45,837 and I just can't sit in in in school the 1272 01:07:45,920 --> 01:07:50,157 stuff I already know. I'm a really good programmer. Uh and and he gave me a 1273 01:07:50,240 --> 01:07:52,957 Spider-Man line at the time. was like with great power comes great 1274 01:07:53,040 --> 01:07:55,357 responsibility and you have a great power and you know and it really 1275 01:07:55,440 --> 01:08:00,637 affected me and I think he he was right at the moment and and so he said well 1276 01:08:00,720 --> 01:08:03,597 we're going to let you go but you're going to have to help the system 1277 01:08:03,680 --> 01:08:07,997 administrators secure the system >> uh for the summer I was like happy to do 1278 01:08:08,080 --> 01:08:12,477 it and I show up and all the programmers there hate me hate my gut% 1279 01:08:12,560 --> 01:08:15,917 >> and uh they they would lock me out like I would see them they would be outside I 1280 01:08:16,000 --> 01:08:18,397 would knock on the door and nobody would listen it's like they don't want to let 1281 01:08:18,480 --> 01:08:20,477 me in I try to help them a little bit 1282 01:08:20,560 --> 01:08:25,677 they theyen't and collaborative and so I was like all right whatever. Uh and so 1283 01:08:25,760 --> 01:08:27,117 it came time for me to actually 1284 01:08:27,200 --> 01:08:30,397 graduate. It was the final project and 1285 01:08:30,480 --> 01:08:34,237 one of the computer science dean came to me and he said look I I need to call a 1286 01:08:34,320 --> 01:08:38,237 favor. I was a big part of the reason we kind of let you go and we didn't kind of 1287 01:08:38,320 --> 01:08:43,357 prosecute you. Uh so I want you to work with me on the um on the final project 1288 01:08:43,440 --> 01:08:47,037 and it's going to be around security and hacking. I was like no I'm I'm done with 1289 01:08:47,120 --> 01:08:50,317 that like I just want to I just want to build programming environments 1290 01:08:50,400 --> 01:08:54,317 and things like that. Uh and he's like no you have to do it. I was like okay. 1291 01:08:54,400 --> 01:08:57,117 So I I thought I' do something more productive. So I wrote a security 1292 01:08:57,200 --> 01:09:00,957 scanner uh that was very proud of that that kind of crawls the different side 1293 01:09:01,040 --> 01:09:05,277 that tries to do SQL injection and all sorts of things. Um and actually my 1294 01:09:05,360 --> 01:09:07,677 security scanner found another vulnerability in the system. 1295 01:09:07,760 --> 01:09:10,157 >> Amazing. >> And so I went to the defense and he's 1296 01:09:10,240 --> 01:09:14,157 like you need to run this security scanner live and show that there's a 1297 01:09:14,240 --> 01:09:17,037 vulnerability. And I didn't understand what was going on at the time, but I 1298 01:09:17,120 --> 01:09:21,757 just okay. So I gave the presentation about how the system works and I was 1299 01:09:21,840 --> 01:09:24,557 like, oh, let's run it. And it showed that there's security vulnerability. 1300 01:09:24,640 --> 01:09:28,077 Okay, let's get let's try to get a shell. So the system automatically runs 1301 01:09:28,160 --> 01:09:29,997 all the security stuff and it gets you 1302 01:09:30,080 --> 01:09:33,197 gets you a shell. And then the other 1303 01:09:33,280 --> 01:09:38,157 dean that turned out he was giving the mandate to secure the system. And now I 1304 01:09:38,240 --> 01:09:40,157 started to realize I'm a pawn in some 1305 01:09:40,240 --> 01:09:44,557 kind of rivalry here. and and his his 1306 01:09:44,640 --> 01:09:48,237 face turned red and he's like, "No, it's impossible. You know, we secured the 1307 01:09:48,320 --> 01:09:53,277 system. You're lying." I was like, "You know, you're accusing me of lying." All 1308 01:09:53,360 --> 01:09:57,997 right. What should we know? Should we know your um uh your salary or your 1309 01:09:58,080 --> 01:10:01,037 password? What do you want me to look up? And I was like, "Yeah, I look up my 1310 01:10:01,120 --> 01:10:04,957 password." So, I I look up his his password. Uh and it was like gibberish. 1311 01:10:05,040 --> 01:10:07,357 It was encrypted. And I was like, "Oh, that's not my password. See, you're 1312 01:10:07,440 --> 01:10:10,557 lying." I was like, "Well, there's a decrypt function that the programmers 1313 01:10:10,640 --> 01:10:15,117 put in there." So I I do decrypt and it shows his password. It was something 1314 01:10:15,200 --> 01:10:19,997 embarrassing. I forgot I forgot what it was. And so he gets up really angry, 1315 01:10:20,080 --> 01:10:24,477 shakes my hand and leaves to change his password. Uh so that I I was able to 1316 01:10:24,560 --> 01:10:26,557 hack into the university another time. 1317 01:10:26,640 --> 01:10:29,997 Luckily I I was able to graduate, gave them the software, they secured the 1318 01:10:30,080 --> 01:10:33,357 system. But um but yeah, later on I 1319 01:10:33,440 --> 01:10:37,037 would realize that yeah, he wanted to embarrass the other guy, which was why I 1320 01:10:37,120 --> 01:10:39,357 was in the middle. >> Politics. Well, I think the the moral 1321 01:10:39,440 --> 01:10:42,157 the moral of the story is if if you can successfully hack into your school 1322 01:10:42,240 --> 01:10:45,277 system and change your grade, you deserve the grade and you deserve to graduate. 1323 01:10:45,360 --> 01:10:47,677 >> I I think so. >> And and and just for any for any parents 1324 01:10:47,760 --> 01:10:50,477 out there just children out there, you can just you can see you can site you 1325 01:10:50,560 --> 01:10:53,917 can sight site me as the moral you can site you can set out to me as the moral 1326 01:10:54,000 --> 01:10:55,597 moral authority moral authority on this. 1327 01:10:55,680 --> 01:10:58,477 One maybe lesson I think that is very 1328 01:10:58,560 --> 01:11:02,717 relevant for the AI age. Uh I think that 1329 01:11:02,800 --> 01:11:04,957 the traditional sort of more conformous 1330 01:11:05,040 --> 01:11:08,477 path is paying less and less dividends 1331 01:11:08,560 --> 01:11:10,717 and I think uh you know kids coming up 1332 01:11:10,800 --> 01:11:13,117 today should use all the tools available 1333 01:11:13,200 --> 01:11:15,837 to be able to discover and chart their 1334 01:11:15,920 --> 01:11:18,557 own paths cuz I feel like just you know 1335 01:11:18,640 --> 01:11:21,357 listening to the traditional advice and 1336 01:11:21,440 --> 01:11:23,517 doing the same things that people have 1337 01:11:23,600 --> 01:11:29,117 always done is just not as it's not working out as much as as we'd like. 1338 01:11:29,200 --> 01:11:33,800 >> Thanks for coming on the podcast. Thank you, man. Fantastic. 1339 01:11:41,280 --> 01:11:44,280 Wow. 1340 01:11:45,040 --> 01:11:48,600 Wow. Wow.