Open AI Chief Outlines His AI Thesis in 9 Minutes (YouTube Video Transcript)

(00:00:00) Your YouTube transcript will appear here (00:00:00) So many people are trying to invest in (00:00:01) AI now. This is it's like there's (00:00:03) literally I this just feels like it's (00:00:05) like it's like a bubble time again. Like (00:00:07) there's $100 million rounds here. (00:00:08) There's $200 million rounds. There's (00:00:10) everyone we pitch who pitches us. (00:00:11) There's the AI part of the pitch. All my (00:00:13) old SAS companies, one of them was a (00:00:15) legal tech company. You probably don't (00:00:16) even know this. That's partnered with (00:00:17) OpenAI and suddenly sold for like eight (00:00:19) times the last round valuation to (00:00:21) Thompson Reuters because it was all of a (00:00:22) sudden like doing really useful things (00:00:23) for parallegals. I mean this is just (00:00:25) it's changed a lot of things. like in (00:00:26) terms of like other people doing AI like (00:00:28) what else is useful that people are (00:00:30) working on that you're excited that (00:00:31) you're seeing people do because this (00:00:32) this is outside of open AI now is there (00:00:34) infrastructure you like is there is (00:00:35) there a need for infrastructure outside (00:00:36) of open AI like like how do you see this (00:00:38) >> yeah I think I I'm always a little (00:00:39) worried about the infrastructure work (00:00:41) because I think you know you're solving (00:00:42) the problems as they exist today and (00:00:44) when GPD 4.5 and GPD 5 come out you know (00:00:47) they're going to have fundamentally (00:00:47) different use cases and fundamentally (00:00:49) different infrastructure what I like to (00:00:50) see is people who are using AI to solve (00:00:53) something that wasn't possible with AI (00:00:54) >> you like the application layers Yeah. (00:00:56) And what I what I think about is AI now (00:00:59) it's like having an infinite number of (00:01:02) interns with very short attention spans. (00:01:05) And so anything you can you could have (00:01:07) solved with interns. You know GPD3 was (00:01:09) sort of like a high school student. (00:01:11) GBD3.5 is maybe a college freshman. GBD4 (00:01:13) is maybe a college junior. GBD5 is going (00:01:17) to be something else. (00:01:18) >> Bigger intern. (00:01:18) >> And so you know if you have these (00:01:20) interns, what can you do with them? And (00:01:22) what new business does that allow you to (00:01:24) create? Like that's the thing I'm really (00:01:25) excited about. (00:01:26) >> Lots of free interns. (00:01:27) >> Yeah, (00:01:29) that's cool. Well, the app layers, you (00:01:31) know, where I've built a ton of things. (00:01:32) So, I I'm getting pitched tons of (00:01:33) infrastructure. I'm getting I tons of my (00:01:35) friends are like, you know, getting (00:01:37) billions of dollars of loans to get like (00:01:39) A100s and H100s and like doing like (00:01:41) crazy amounts of like hardware (00:01:43) infrastructure. Is that is that going to (00:01:44) be needed still? I assume that's going (00:01:46) to be needed still. Or how do you think (00:01:47) about that side? (00:01:48) >> Yeah, I think it's really hard to say. A (00:01:50) lot of people have tried building new (00:01:51) chips and you know I think what what (00:01:54) everybody has seen so far is that Nvidia (00:01:56) has just been able to continually do (00:01:58) better and better and better because (00:01:59) they have you know the big market and a (00:02:01) lot of capital to throw at it. So I (00:02:02) think I think it's pretty hard to bet (00:02:04) against Nvidia. On the other hand, a (00:02:05) small chance of a really big thing uh is (00:02:07) also something worthwhile. (00:02:08) >> Yeah. You know when I was at Stanford in (00:02:10) computer science doing graphics I was (00:02:11) obsessed with Nvidia at the time. That's (00:02:12) where I wanted to work and then PayPal (00:02:14) and Peter but that was (00:02:15) >> you probably did okay. (00:02:16) >> Probably did fine. But it's pretty funny (00:02:18) how the company I was really into ended (00:02:19) up pivoting and doing doing this. (00:02:21) >> So my PhD before I started Palunteer was (00:02:23) in AI and I left uh because I felt AI (00:02:26) wasn't happening and this was 2005 and (00:02:28) in fact AI was not happening in 2005. (00:02:30) >> Yes. So it would have been very hard to (00:02:31) work on AI 20 years ago. (00:02:32) >> The the key thing about AI was actually (00:02:34) just big data and you didn't need AI to (00:02:36) unlock it. (00:02:37) >> The lesson was that everyone who got (00:02:38) went to work on this just never it was (00:02:39) just like a bad choice forever. It was (00:02:41) exciting but then it was never the right (00:02:42) choice. (00:02:43) >> Yeah. until it was (00:02:43) >> until it was which which so how do you (00:02:45) know that it was like how do you how do (00:02:46) you have the intuition that it might be (00:02:48) >> well so what happened is uh in I think (00:02:50) 2011 uh there was a paper that was (00:02:52) published by Ilia Sutzkver and some (00:02:54) other people who IAS are chief (00:02:55) scientists at openai where they (00:02:58) basically reinvented uh neural networks (00:03:01) and they showed that if you ran them on (00:03:02) GPUs you could plug in huge amounts much (00:03:04) bigger amounts of compute and much (00:03:06) bigger amounts of data and neural (00:03:08) networks went from being oh that one (00:03:10) thing that you can kind of use for (00:03:11) handwriting to something that was (00:03:13) actually the best way to identify (00:03:14) images. (00:03:14) >> And for some of our listeners who aren't (00:03:15) as technical, we're talking about a (00:03:17) neural network. That's like that's (00:03:18) something that gives that gives (00:03:19) iterative feedback on things. Explain (00:03:21) explain how would you explain neural (00:03:22) network? (00:03:22) >> So a neural network um the the analogy (00:03:25) that most people use is that it's like (00:03:26) the brain and it is but at a very very (00:03:28) high level. So a neuron is basically um (00:03:32) you can think of it as as having (00:03:34) connections to other neurons and the (00:03:36) strength of those connections determines (00:03:38) when you know the first neuron the lower (00:03:40) level neuron fires then that makes the (00:03:42) the higher level neuron fire. And if you (00:03:45) train these um you do it by showing them (00:03:48) the right answer and then basically (00:03:50) doing what's called back propagating the (00:03:51) error, you know, from sort of the top of (00:03:54) the network where the answer is all the (00:03:56) way back down um through the whole (00:03:58) network. (00:03:58) >> So these things are learning patterns (00:04:00) and and is it a pattern detection type (00:04:02) of situation then or what is it with the (00:04:03) with the flashing what they're (00:04:04) detecting? (00:04:04) >> Yeah, you're building hierarchical (00:04:06) representation. So at the very bottom, (00:04:07) think about think about a vision network (00:04:08) because it's easy to visualize. So at (00:04:09) the very bottom you're the neurons are (00:04:11) detecting edges and then you they go up (00:04:13) a little bit higher and they're (00:04:14) detecting corners and if you go high (00:04:16) enough they're detecting wheels and then (00:04:18) above that they're detecting cars (00:04:19) >> quarters in this way then it's just then (00:04:20) it's this a square if this and this this (00:04:22) could be a wheel and then it could be a (00:04:23) car. Got it. (00:04:23) >> Yeah. And and at some point there's a (00:04:24) Joe Londale neuron (00:04:26) >> um that you know is out there you know (00:04:28) and and individual people you can (00:04:30) actually find these neurons in the (00:04:31) networks. (00:04:31) >> I remember reading on intelligence by (00:04:33) Hawkins who didn't actually end up (00:04:34) solving all these problems but he (00:04:35) believe they said there's like six (00:04:36) layers of the neoortex that is part of (00:04:38) the vision system. Is that is that right (00:04:39) or is there actually a lot more than (00:04:40) that? (00:04:41) >> You know, I'm not I'm not a cognitive (00:04:42) scientist. So I the other thing is I (00:04:44) think these the analogy between the (00:04:46) brain and the neural networks it it's (00:04:47) it's very inspirational but you can take (00:04:49) it too far. (00:04:49) >> It's very imprecise. What OpenAI did is (00:04:51) say but you basically didn't base it on (00:04:52) the brain. You based it on based it on (00:04:54) like building it up from scratch from (00:04:55) from just first principles. (00:04:56) >> So I think that the thesis behind open (00:04:58) AI when I joined in 2017 was that neural (00:05:01) networks were the final architecture (00:05:03) that could take AI all the way to human (00:05:05) level intelligence or AGI. But wasn't (00:05:08) there like a trans wasn't there a (00:05:09) transformer breakthrough that was really (00:05:10) important though right around 2017? (00:05:12) >> Yeah, that's right. So um you know the (00:05:14) first couple years of of open AI we were (00:05:17) using sort of standard neural networks (00:05:19) and then there was a a paper at Google (00:05:21) some people came up with an idea called (00:05:22) transformers that allow you to um take (00:05:26) better understanding of the context. Uh (00:05:29) so for example you could look at (00:05:31) documents and now you could really (00:05:34) understand the preceding three or four (00:05:36) pages when the prior architectures had (00:05:38) struggled just to understand like the (00:05:40) the previous 10 words. (00:05:41) >> What's the intuition for why (00:05:42) transformers worked better? Like what's (00:05:44) going on there? (00:05:45) >> It's basically because that they have a (00:05:46) notion of attention. So you can pay (00:05:48) attention to particular words in a (00:05:51) document versus uh the previous (00:05:53) architecture had a notion of memory. So (00:05:56) you could sort of me remember the words (00:05:58) that you had seen, but it's much easier (00:06:00) to be able to look at a sheet of paper (00:06:02) and sort of see the different words and (00:06:04) think about, you know, reading those (00:06:05) words. (00:06:05) >> Pay attention to what matters. (00:06:06) >> Pay attention to what matters, (00:06:07) >> which is kind of just how the brain (00:06:09) works. Things jump out at us obviously (00:06:10) when we look at things. (00:06:11) >> Yeah, it's very intuitive. The the (00:06:13) intuition, I know this is not based, (00:06:15) you're not a cognitive scientist, but I (00:06:16) want to try to build intuition here for (00:06:18) people. My understanding, the big (00:06:20) breakthrough uh we had for how the brain (00:06:21) works is you're constantly predicting (00:06:22) what you're going to see next. like this (00:06:24) maybe why we see ghosts but either way (00:06:26) your brain's constantly looking at what (00:06:27) it thinks it expects to see and then if (00:06:29) something is not unexpected it kind of (00:06:30) jumps out at you is there anything like (00:06:32) that with transformers or there's (00:06:33) nothing there's nothing like that where (00:06:34) it's trying to predict things (00:06:35) >> well that's exactly how you train a (00:06:36) model like GPT4 so um you know if you (00:06:40) take it to to train a model like GB4 (00:06:42) basically we take all of the text (00:06:43) everything we can scrape off the (00:06:44) internet um and (00:06:48) what the model is trying to do is it (00:06:49) goes character by character and it's (00:06:51) trying to predict the next character So, (00:06:54) you know, uh if if it sees, you know, (00:06:56) the rain in Spain falls mainly on the (00:06:58) it's going to guess, oh, that's a P and (00:06:59) it's going to be plain. And it does this (00:07:01) over and over again for huge amounts of (00:07:04) documents, like trillions of characters. (00:07:06) And over time, it seems as though that (00:07:10) yields something that looks like (00:07:11) intelligence. (00:07:11) >> The prediction seems very very similar (00:07:12) to intelligence, which which may which (00:07:14) may be what we're doing, too. (00:07:15) >> Yeah. I I at this point, I think we (00:07:18) understand neural networks a lot better (00:07:19) than we understand the brain. So (00:07:20) whenever someone talks to me about the (00:07:21) brain I always think well what does a (00:07:23) neural network do and then probably (00:07:24) actually the brain does that. (00:07:26) >> Do you think of yourself as a neural (00:07:27) network now? (00:07:28) >> I do actually. Um I actually think of my (00:07:30) kids as neural networks. (00:07:31) >> How has that changed your interactions (00:07:33) with them? (00:07:35) >> You know as I was starting at OpenAI (00:07:36) first I worked on robotics and then I (00:07:38) worked on I we did some problems with (00:07:39) math and then programming and uh early (00:07:42) on you know the kids would always get (00:07:44) the problems before the robots were and (00:07:46) now it's the other way around. Now the (00:07:48) neural networks are better are you know (00:07:50) better at solving problems than the (00:07:51) kids. (00:07:52) >> Has it given you any intuition on how to (00:07:53) train young minds? (00:07:55) >> Uh yeah, you just show them lots and (00:07:57) lots of things and you just have them (00:07:58) read. (00:07:59) >> Show them three trillion things. (00:08:00) >> I mean that's that's probably how you (00:08:02) and I learned, right? (00:08:03) >> Just just read a lot of books when we (00:08:05) were kids. (00:08:05) >> I guess there's a thing about attention, (00:08:07) too. I've always found that kids do (00:08:08) better when they're confident. I don't (00:08:10) know if the computers need to be given (00:08:11) confidence, although maybe that's a (00:08:12) signal for something else like (00:08:13) attention. Well, that's actually one of (00:08:14) the cool things that we've discovered um (00:08:17) with GBD3 and GBD4 is that uh the model (00:08:21) so the models are trained on the (00:08:22) internet and um they're reading (00:08:26) everything, right? Then they try to act (00:08:28) like what they've seen, right? They're (00:08:29) predicting and so they don't know (00:08:31) whether to predict being someone who's (00:08:33) dumb or someone who's really smart. (00:08:35) >> That's why you have to tell them they're (00:08:36) really smart when you're talking to (00:08:37) them. It's so funny. (00:08:38) >> Exactly. Like you should you're a (00:08:39) confident, you know, powerful physicist (00:08:41) who knows everything. the best physicist (00:08:43) in the world. Now answer my question. (00:08:45) >> And you sound like Shakespeare because (00:08:46) it's fun if you make them sound cool. (00:08:47) >> That's right. Or or you you have to do (00:08:49) everything in in a limick.