Datasets:
Languages:
English
Multilinguality:
monolingual
Size Categories:
n<1K
Language Creators:
found
Source Datasets:
original
Tags:
karpathy,whisper,openai
| WEBVTT | |
| 00:00.000 --> 00:02.920 | |
| The following is a conversation with Greg Brockman. | |
| 00:02.920 --> 00:05.400 | |
| He's the cofounder and CTO of OpenAI, | |
| 00:05.400 --> 00:07.440 | |
| a world class research organization | |
| 00:07.440 --> 00:10.840 | |
| developing ideas in AI with a goal of eventually | |
| 00:10.840 --> 00:14.200 | |
| creating a safe and friendly artificial general | |
| 00:14.200 --> 00:18.840 | |
| intelligence, one that benefits and empowers humanity. | |
| 00:18.840 --> 00:22.520 | |
| OpenAI is not only a source of publications, algorithms, | |
| 00:22.520 --> 00:24.480 | |
| tools, and data sets. | |
| 00:24.480 --> 00:28.120 | |
| Their mission is a catalyst for an important public discourse | |
| 00:28.120 --> 00:32.720 | |
| about our future with both narrow and general intelligence | |
| 00:32.720 --> 00:34.000 | |
| systems. | |
| 00:34.000 --> 00:36.640 | |
| This conversation is part of the Artificial Intelligence | |
| 00:36.640 --> 00:39.480 | |
| Podcast at MIT and beyond. | |
| 00:39.480 --> 00:42.720 | |
| If you enjoy it, subscribe on YouTube, iTunes, | |
| 00:42.720 --> 00:44.520 | |
| or simply connect with me on Twitter | |
| 00:44.520 --> 00:48.040 | |
| at Lex Friedman, spelled F R I D. | |
| 00:48.040 --> 00:52.760 | |
| And now here's my conversation with Greg Brockman. | |
| 00:52.760 --> 00:54.640 | |
| So in high school and right after you wrote | |
| 00:54.640 --> 00:56.640 | |
| a draft of a chemistry textbook, | |
| 00:56.640 --> 00:58.120 | |
| I saw that that covers everything | |
| 00:58.120 --> 01:01.400 | |
| from basic structure of the atom to quantum mechanics. | |
| 01:01.400 --> 01:04.400 | |
| So it's clear you have an intuition and a passion | |
| 01:04.400 --> 01:09.880 | |
| for both the physical world with chemistry and non robotics | |
| 01:09.880 --> 01:13.680 | |
| to the digital world with AI, deep learning, | |
| 01:13.680 --> 01:15.360 | |
| reinforcement learning, so on. | |
| 01:15.360 --> 01:17.360 | |
| Do you see the physical world and the digital world | |
| 01:17.360 --> 01:18.600 | |
| as different? | |
| 01:18.600 --> 01:20.480 | |
| And what do you think is the gap? | |
| 01:20.480 --> 01:23.040 | |
| A lot of it actually boils down to iteration speed, | |
| 01:23.040 --> 01:25.240 | |
| that I think that a lot of what really motivates me | |
| 01:25.240 --> 01:27.720 | |
| is building things, right? | |
| 01:27.720 --> 01:29.320 | |
| Think about mathematics, for example, | |
| 01:29.320 --> 01:30.920 | |
| where you think really hard about a problem. | |
| 01:30.920 --> 01:31.680 | |
| You understand it. | |
| 01:31.680 --> 01:33.320 | |
| You're right down in this very obscure form | |
| 01:33.320 --> 01:34.560 | |
| that we call a proof. | |
| 01:34.560 --> 01:37.600 | |
| But then this is in humanity's library, right? | |
| 01:37.600 --> 01:38.400 | |
| It's there forever. | |
| 01:38.400 --> 01:40.400 | |
| This is some truth that we've discovered. | |
| 01:40.400 --> 01:43.040 | |
| And maybe only five people in your field will ever read it. | |
| 01:43.040 --> 01:45.440 | |
| But somehow you've kind of moved humanity forward. | |
| 01:45.440 --> 01:46.880 | |
| And so I actually used to really think | |
| 01:46.880 --> 01:48.680 | |
| that I was going to be a mathematician. | |
| 01:48.680 --> 01:51.400 | |
| And then I actually started writing this chemistry textbook. | |
| 01:51.400 --> 01:53.360 | |
| One of my friends told me, you'll never publish it | |
| 01:53.360 --> 01:54.880 | |
| because you don't have a PhD. | |
| 01:54.880 --> 01:58.000 | |
| So instead, I decided to build a website | |
| 01:58.000 --> 01:59.960 | |
| and try to promote my ideas that way. | |
| 01:59.960 --> 02:01.480 | |
| And then I discovered programming. | |
| 02:01.480 --> 02:05.320 | |
| And in programming, you think hard about a problem. | |
| 02:05.320 --> 02:06.080 | |
| You understand it. | |
| 02:06.080 --> 02:08.040 | |
| You're right down in a very obscure form | |
| 02:08.040 --> 02:09.720 | |
| that we call a program. | |
| 02:09.720 --> 02:12.240 | |
| But then once again, it's in humanity's library, right? | |
| 02:12.240 --> 02:14.120 | |
| And anyone can get the benefit from it. | |
| 02:14.120 --> 02:15.720 | |
| And the scalability is massive. | |
| 02:15.720 --> 02:17.720 | |
| And so I think that the thing that really appeals to me | |
| 02:17.720 --> 02:19.440 | |
| about the digital world is that you | |
| 02:19.440 --> 02:21.960 | |
| can have this insane leverage, right? | |
| 02:21.960 --> 02:24.960 | |
| A single individual with an idea is able to affect | |
| 02:24.960 --> 02:26.120 | |
| the entire planet. | |
| 02:26.120 --> 02:28.240 | |
| And that's something I think is really hard to do | |
| 02:28.240 --> 02:30.240 | |
| if you're moving around physical atoms. | |
| 02:30.240 --> 02:32.440 | |
| But you said mathematics. | |
| 02:32.440 --> 02:36.880 | |
| So if you look at the wet thing over here, our mind, | |
| 02:36.880 --> 02:39.800 | |
| do you ultimately see it as just math, | |
| 02:39.800 --> 02:41.800 | |
| as just information processing? | |
| 02:41.800 --> 02:44.880 | |
| Or is there some other magic if you've | |
| 02:44.880 --> 02:46.960 | |
| seen through biology and chemistry and so on? | |
| 02:46.960 --> 02:49.000 | |
| I think it's really interesting to think about humans | |
| 02:49.000 --> 02:51.000 | |
| as just information processing systems. | |
| 02:51.000 --> 02:54.080 | |
| And it seems like it's actually a pretty good way | |
| 02:54.080 --> 02:57.560 | |
| of describing a lot of how the world works or a lot of what | |
| 02:57.560 --> 03:01.000 | |
| we're capable of to think that, again, if you just | |
| 03:01.000 --> 03:03.640 | |
| look at technological innovations over time, | |
| 03:03.640 --> 03:05.920 | |
| that in some ways, the most transformative innovation | |
| 03:05.920 --> 03:07.760 | |
| that we've had has been the computer, right? | |
| 03:07.760 --> 03:10.560 | |
| In some ways, the internet, what has the internet done? | |
| 03:10.560 --> 03:12.760 | |
| The internet is not about these physical cables. | |
| 03:12.760 --> 03:14.560 | |
| It's about the fact that I am suddenly | |
| 03:14.560 --> 03:16.560 | |
| able to instantly communicate with any other human | |
| 03:16.560 --> 03:17.680 | |
| on the planet. | |
| 03:17.680 --> 03:19.680 | |
| I'm able to retrieve any piece of knowledge | |
| 03:19.680 --> 03:22.640 | |
| that, in some ways, the human race has ever had, | |
| 03:22.640 --> 03:26.080 | |
| and that those are these insane transformations. | |
| 03:26.080 --> 03:29.320 | |
| Do you see our society as a whole the collective | |
| 03:29.320 --> 03:31.240 | |
| as another extension of the intelligence | |
| 03:31.240 --> 03:32.280 | |
| of the human being? | |
| 03:32.280 --> 03:34.440 | |
| So if you look at the human being as an information processing | |
| 03:34.440 --> 03:36.920 | |
| system, you mentioned the internet, the networking. | |
| 03:36.920 --> 03:39.360 | |
| Do you see us all together as a civilization | |
| 03:39.360 --> 03:41.680 | |
| as a kind of intelligent system? | |
| 03:41.680 --> 03:43.560 | |
| Yeah, I think this is actually a really interesting | |
| 03:43.560 --> 03:45.840 | |
| perspective to take and to think about | |
| 03:45.840 --> 03:48.080 | |
| that you sort of have this collective intelligence | |
| 03:48.080 --> 03:49.520 | |
| of all of society. | |
| 03:49.520 --> 03:51.680 | |
| The economy itself is this superhuman machine | |
| 03:51.680 --> 03:54.440 | |
| that is optimizing something, right? | |
| 03:54.440 --> 03:56.400 | |
| And it's almost, in some ways, a company | |
| 03:56.400 --> 03:57.960 | |
| has a will of its own, right? | |
| 03:57.960 --> 03:59.400 | |
| That you have all these individuals who are all | |
| 03:59.400 --> 04:00.800 | |
| pursuing their own individual goals | |
| 04:00.800 --> 04:02.400 | |
| and thinking really hard and thinking | |
| 04:02.400 --> 04:04.640 | |
| about the right things to do, but somehow the company does | |
| 04:04.640 --> 04:07.880 | |
| something that is this emergent thing | |
| 04:07.880 --> 04:10.640 | |
| and that it's a really useful abstraction. | |
| 04:10.640 --> 04:12.440 | |
| And so I think that in some ways, | |
| 04:12.440 --> 04:14.880 | |
| we think of ourselves as the most intelligent things | |
| 04:14.880 --> 04:17.480 | |
| on the planet and the most powerful things on the planet. | |
| 04:17.480 --> 04:19.320 | |
| But there are things that are bigger than us, | |
| 04:19.320 --> 04:21.480 | |
| that are these systems that we all contribute to. | |
| 04:21.480 --> 04:25.000 | |
| And so I think actually, it's interesting to think about, | |
| 04:25.000 --> 04:27.440 | |
| if you've read Asa Geismov's foundation, right, | |
| 04:27.440 --> 04:30.160 | |
| that there's this concept of psycho history in there, | |
| 04:30.160 --> 04:31.920 | |
| which is effectively this, that if you have trillions | |
| 04:31.920 --> 04:35.200 | |
| or quadrillions of beings, then maybe you could actually | |
| 04:35.200 --> 04:39.080 | |
| predict what that huge macro being will do | |
| 04:39.080 --> 04:42.400 | |
| and almost independent of what the individuals want. | |
| 04:42.400 --> 04:44.240 | |
| And I actually have a second angle on this | |
| 04:44.240 --> 04:46.760 | |
| that I think is interesting, which is thinking about | |
| 04:46.760 --> 04:48.400 | |
| technological determinism. | |
| 04:48.400 --> 04:51.480 | |
| One thing that I actually think a lot about with OpenAI | |
| 04:51.480 --> 04:54.720 | |
| is that we're kind of coming onto this insanely | |
| 04:54.720 --> 04:57.400 | |
| transformational technology of general intelligence | |
| 04:57.400 --> 04:58.760 | |
| that will happen at some point. | |
| 04:58.760 --> 05:01.560 | |
| And there's a question of how can you take actions | |
| 05:01.560 --> 05:04.880 | |
| that will actually steer it to go better rather than worse? | |
| 05:04.880 --> 05:06.720 | |
| And that I think one question you need to ask is, | |
| 05:06.720 --> 05:09.320 | |
| as a scientist, as an event or as a creator, | |
| 05:09.320 --> 05:11.720 | |
| what impact can you have in general? | |
| 05:11.720 --> 05:12.880 | |
| You look at things like the telephone | |
| 05:12.880 --> 05:14.840 | |
| invented by two people on the same day. | |
| 05:14.840 --> 05:16.600 | |
| Like what does that mean, like what does that mean | |
| 05:16.600 --> 05:18.080 | |
| about the shape of innovation? | |
| 05:18.080 --> 05:20.160 | |
| And I think that what's going on is everyone's building | |
| 05:20.160 --> 05:21.720 | |
| on the shoulders of the same giants. | |
| 05:21.720 --> 05:23.840 | |
| And so you can kind of, you can't really hope | |
| 05:23.840 --> 05:25.720 | |
| to create something no one else ever would. | |
| 05:25.720 --> 05:27.040 | |
| You know, if Einstein wasn't born, | |
| 05:27.040 --> 05:29.200 | |
| someone else would have come up with relativity. | |
| 05:29.200 --> 05:31.000 | |
| You know, he changed the timeline a bit, right? | |
| 05:31.000 --> 05:33.000 | |
| That maybe it would have taken another 20 years, | |
| 05:33.000 --> 05:34.560 | |
| but it wouldn't be that fundamentally humanity | |
| 05:34.560 --> 05:37.360 | |
| would never discover these fundamental truths. | |
| 05:37.360 --> 05:40.440 | |
| So there's some kind of invisible momentum | |
| 05:40.440 --> 05:45.400 | |
| that some people like Einstein or OpenAI is plugging into | |
| 05:45.400 --> 05:47.800 | |
| that anybody else can also plug into. | |
| 05:47.800 --> 05:50.800 | |
| And ultimately, that wave takes us into a certain direction. | |
| 05:50.800 --> 05:51.840 | |
| That's what you mean by digitalism? | |
| 05:51.840 --> 05:52.840 | |
| That's right, that's right. | |
| 05:52.840 --> 05:54.240 | |
| And you know, this kind of seems to play out | |
| 05:54.240 --> 05:55.720 | |
| in a bunch of different ways. | |
| 05:55.720 --> 05:58.040 | |
| That there's some exponential that is being written | |
| 05:58.040 --> 05:59.960 | |
| and that the exponential itself, which one it is, | |
| 05:59.960 --> 06:01.520 | |
| changes, think about Moore's Law, | |
| 06:01.520 --> 06:04.800 | |
| an entire industry set, it's clocked to it for 50 years. | |
| 06:04.800 --> 06:06.200 | |
| Like how can that be, right? | |
| 06:06.200 --> 06:07.360 | |
| How is that possible? | |
| 06:07.360 --> 06:09.320 | |
| And yet somehow it happened. | |
| 06:09.320 --> 06:12.200 | |
| And so I think you can't hope to ever invent something | |
| 06:12.200 --> 06:13.360 | |
| that no one else will. | |
| 06:13.360 --> 06:15.360 | |
| Maybe you can change the timeline a little bit. | |
| 06:15.360 --> 06:17.400 | |
| But if you really want to make a difference, | |
| 06:17.400 --> 06:19.440 | |
| I think that the thing that you really have to do, | |
| 06:19.440 --> 06:21.320 | |
| the only real degree of freedom you have | |
| 06:21.320 --> 06:23.040 | |
| is to set the initial conditions | |
| 06:23.040 --> 06:24.960 | |
| under which a technology is born. | |
| 06:24.960 --> 06:26.680 | |
| And so you think about the internet, right? | |
| 06:26.680 --> 06:27.840 | |
| That there are lots of other competitors | |
| 06:27.840 --> 06:29.400 | |
| trying to build similar things. | |
| 06:29.400 --> 06:33.240 | |
| And the internet one, and that the initial conditions | |
| 06:33.240 --> 06:34.680 | |
| where it was created by this group | |
| 06:34.680 --> 06:37.760 | |
| that really valued people being able to be, | |
| 06:37.760 --> 06:39.120 | |
| you know, anyone being able to plug in | |
| 06:39.120 --> 06:42.480 | |
| this very academic mindset of being open and connected. | |
| 06:42.480 --> 06:44.400 | |
| And I think that the internet for the next 40 years | |
| 06:44.400 --> 06:46.360 | |
| really played out that way. | |
| 06:46.360 --> 06:47.680 | |
| You know, maybe today, | |
| 06:47.680 --> 06:49.840 | |
| things are starting to shift in a different direction, | |
| 06:49.840 --> 06:51.120 | |
| but I think that those initial conditions | |
| 06:51.120 --> 06:52.720 | |
| were really important to determine | |
| 06:52.720 --> 06:55.080 | |
| the next 40 years worth of progress. | |
| 06:55.080 --> 06:56.440 | |
| That's really beautifully put. | |
| 06:56.440 --> 06:58.800 | |
| So another example of that I think about, | |
| 06:58.800 --> 07:00.800 | |
| you know, I recently looked at it. | |
| 07:00.800 --> 07:03.800 | |
| I looked at Wikipedia, the formation of Wikipedia. | |
| 07:03.800 --> 07:05.520 | |
| And I wonder what the internet would be like | |
| 07:05.520 --> 07:07.760 | |
| if Wikipedia had ads. | |
| 07:07.760 --> 07:09.640 | |
| You know, there's an interesting argument | |
| 07:09.640 --> 07:14.280 | |
| that why they chose not to put advertisement on Wikipedia. | |
| 07:14.280 --> 07:17.800 | |
| I think Wikipedia is one of the greatest resources | |
| 07:17.800 --> 07:18.920 | |
| we have on the internet. | |
| 07:18.920 --> 07:21.280 | |
| It's extremely surprising how well it works | |
| 07:21.280 --> 07:22.960 | |
| and how well it was able to aggregate | |
| 07:22.960 --> 07:25.000 | |
| all this kind of good information. | |
| 07:25.000 --> 07:27.320 | |
| And essentially the creator of Wikipedia, | |
| 07:27.320 --> 07:29.360 | |
| I don't know, there's probably some debates there, | |
| 07:29.360 --> 07:31.200 | |
| but set the initial conditions | |
| 07:31.200 --> 07:33.240 | |
| and now it carried itself forward. | |
| 07:33.240 --> 07:34.080 | |
| That's really interesting. | |
| 07:34.080 --> 07:36.520 | |
| So the way you're thinking about AGI | |
| 07:36.520 --> 07:38.640 | |
| or artificial intelligence is you're focused on | |
| 07:38.640 --> 07:41.200 | |
| setting the initial conditions for the progress. | |
| 07:41.200 --> 07:42.320 | |
| That's right. | |
| 07:42.320 --> 07:43.160 | |
| That's powerful. | |
| 07:43.160 --> 07:45.560 | |
| Okay, so look into the future. | |
| 07:45.560 --> 07:48.160 | |
| If you create an AGI system, | |
| 07:48.160 --> 07:51.560 | |
| like one that can ace the Turing test, natural language, | |
| 07:51.560 --> 07:54.800 | |
| what do you think would be the interactions | |
| 07:54.800 --> 07:55.840 | |
| you would have with it? | |
| 07:55.840 --> 07:57.720 | |
| What do you think are the questions you would ask? | |
| 07:57.720 --> 08:00.560 | |
| Like what would be the first question you would ask? | |
| 08:00.560 --> 08:01.840 | |
| It, her, him. | |
| 08:01.840 --> 08:02.680 | |
| That's right. | |
| 08:02.680 --> 08:03.920 | |
| I think that at that point, | |
| 08:03.920 --> 08:05.960 | |
| if you've really built a powerful system | |
| 08:05.960 --> 08:08.480 | |
| that is capable of shaping the future of humanity, | |
| 08:08.480 --> 08:10.240 | |
| the first question that you really should ask | |
| 08:10.240 --> 08:12.280 | |
| is how do we make sure that this plays out well? | |
| 08:12.280 --> 08:13.960 | |
| And so that's actually the first question | |
| 08:13.960 --> 08:17.600 | |
| that I would ask a powerful AGI system is. | |
| 08:17.600 --> 08:19.160 | |
| So you wouldn't ask your colleague, | |
| 08:19.160 --> 08:22.280 | |
| you wouldn't ask like Ilya, you would ask the AGI system. | |
| 08:22.280 --> 08:24.640 | |
| Oh, we've already had the conversation with Ilya, right? | |
| 08:24.640 --> 08:25.720 | |
| And everyone here. | |
| 08:25.720 --> 08:27.480 | |
| And so you want as many perspectives | |
| 08:27.480 --> 08:29.720 | |
| and a piece of wisdom as you can | |
| 08:29.720 --> 08:31.200 | |
| for answering this question. | |
| 08:31.200 --> 08:33.120 | |
| So I don't think you necessarily defer to | |
| 08:33.120 --> 08:35.480 | |
| whatever your powerful system tells you, | |
| 08:35.480 --> 08:37.120 | |
| but you use it as one input | |
| 08:37.120 --> 08:39.280 | |
| to try to figure out what to do. | |
| 08:39.280 --> 08:40.920 | |
| But, and I guess fundamentally, | |
| 08:40.920 --> 08:42.160 | |
| what it really comes down to is | |
| 08:42.160 --> 08:43.960 | |
| if you built something really powerful | |
| 08:43.960 --> 08:45.280 | |
| and you think about, think about, for example, | |
| 08:45.280 --> 08:47.640 | |
| the creation of, of shortly after | |
| 08:47.640 --> 08:48.880 | |
| the creation of nuclear weapons, right? | |
| 08:48.880 --> 08:50.400 | |
| The most important question in the world | |
| 08:50.400 --> 08:52.800 | |
| was what's the world we're going to be like? | |
| 08:52.800 --> 08:54.880 | |
| How do we set ourselves up in a place | |
| 08:54.880 --> 08:58.320 | |
| where we're going to be able to survive as a species? | |
| 08:58.320 --> 09:00.640 | |
| With AGI, I think the question is slightly different, right? | |
| 09:00.640 --> 09:02.720 | |
| That there is a question of how do we make sure | |
| 09:02.720 --> 09:04.440 | |
| that we don't get the negative effects? | |
| 09:04.440 --> 09:06.240 | |
| But there's also the positive side, right? | |
| 09:06.240 --> 09:08.040 | |
| You imagine that, you know, like, | |
| 09:08.040 --> 09:09.720 | |
| like what will AGI be like? | |
| 09:09.720 --> 09:11.280 | |
| Like what will it be capable of? | |
| 09:11.280 --> 09:13.520 | |
| And I think that one of the core reasons | |
| 09:13.520 --> 09:15.760 | |
| that an AGI can be powerful and transformative | |
| 09:15.760 --> 09:18.920 | |
| is actually due to technological development, right? | |
| 09:18.920 --> 09:20.560 | |
| If you have something that's capable, | |
| 09:20.560 --> 09:23.880 | |
| that's capable as a human and that it's much more scalable, | |
| 09:23.880 --> 09:25.880 | |
| that you absolutely want that thing | |
| 09:25.880 --> 09:27.640 | |
| to go read the whole scientific literature | |
| 09:27.640 --> 09:30.000 | |
| and think about how to create cures for all the diseases, right? | |
| 09:30.000 --> 09:31.480 | |
| You want it to think about how to go | |
| 09:31.480 --> 09:33.360 | |
| and build technologies to help us | |
| 09:33.360 --> 09:37.320 | |
| create material abundance and to figure out societal problems | |
| 09:37.320 --> 09:38.160 | |
| that we have trouble with, | |
| 09:38.160 --> 09:40.000 | |
| like how are we supposed to clean up the environment? | |
| 09:40.000 --> 09:42.200 | |
| And, you know, maybe you want this | |
| 09:42.200 --> 09:44.120 | |
| to go and invent a bunch of little robots that will go out | |
| 09:44.120 --> 09:47.280 | |
| and be biodegradable and turn ocean debris | |
| 09:47.280 --> 09:49.640 | |
| into harmless molecules. | |
| 09:49.640 --> 09:54.040 | |
| And I think that that positive side | |
| 09:54.040 --> 09:55.720 | |
| is something that I think people miss | |
| 09:55.720 --> 09:58.160 | |
| sometimes when thinking about what an AGI will be like. | |
| 09:58.160 --> 10:00.280 | |
| And so I think that if you have a system | |
| 10:00.280 --> 10:01.640 | |
| that's capable of all of that, | |
| 10:01.640 --> 10:03.960 | |
| you absolutely want its advice about how do I make sure | |
| 10:03.960 --> 10:07.600 | |
| that we're using your capabilities | |
| 10:07.600 --> 10:09.200 | |
| in a positive way for humanity. | |
| 10:09.200 --> 10:11.400 | |
| So what do you think about that psychology | |
| 10:11.400 --> 10:14.800 | |
| that looks at all the different possible trajectories | |
| 10:14.800 --> 10:17.520 | |
| of an AGI system, many of which, | |
| 10:17.520 --> 10:19.960 | |
| perhaps the majority of which are positive | |
| 10:19.960 --> 10:23.320 | |
| and nevertheless focuses on the negative trajectories? | |
| 10:23.320 --> 10:24.720 | |
| I mean, you get to interact with folks, | |
| 10:24.720 --> 10:28.840 | |
| you get to think about this maybe within yourself as well. | |
| 10:28.840 --> 10:30.560 | |
| You look at Sam Harris and so on. | |
| 10:30.560 --> 10:32.720 | |
| It seems to be, sorry to put it this way, | |
| 10:32.720 --> 10:36.720 | |
| but almost more fun to think about the negative possibilities. | |
| 10:37.800 --> 10:39.560 | |
| Whatever that's deep in our psychology, | |
| 10:39.560 --> 10:40.760 | |
| what do you think about that? | |
| 10:40.760 --> 10:41.920 | |
| And how do we deal with it? | |
| 10:41.920 --> 10:44.400 | |
| Because we want AI to help us. | |
| 10:44.400 --> 10:47.880 | |
| So I think there's kind of two problems | |
| 10:47.880 --> 10:49.960 | |
| entailed in that question. | |
| 10:49.960 --> 10:52.360 | |
| The first is more of the question of, | |
| 10:52.360 --> 10:54.600 | |
| how can you even picture what a world | |
| 10:54.600 --> 10:56.600 | |
| with a new technology will be like? | |
| 10:56.600 --> 10:57.880 | |
| Now imagine we're in 1950 | |
| 10:57.880 --> 11:01.040 | |
| and I'm trying to describe Uber to someone. | |
| 11:01.040 --> 11:05.360 | |
| Aps and the internet. | |
| 11:05.360 --> 11:08.920 | |
| Yeah, I mean, that's going to be extremely complicated, | |
| 11:08.920 --> 11:10.160 | |
| but it's imaginable. | |
| 11:10.160 --> 11:11.400 | |
| It's imaginable, right? | |
| 11:11.400 --> 11:14.000 | |
| But, and now imagine being in 1950 | |
| 11:14.000 --> 11:15.280 | |
| and predicting Uber, right? | |
| 11:15.280 --> 11:17.680 | |
| And you need to describe the internet, | |
| 11:17.680 --> 11:18.720 | |
| you need to describe GPS, | |
| 11:18.720 --> 11:20.280 | |
| you need to describe the fact | |
| 11:20.280 --> 11:23.920 | |
| that everyone's going to have this phone in their pocket. | |
| 11:23.920 --> 11:26.160 | |
| And so I think that just the first truth | |
| 11:26.160 --> 11:28.040 | |
| is that it is hard to picture | |
| 11:28.040 --> 11:31.160 | |
| how a transformative technology will play out in the world. | |
| 11:31.160 --> 11:32.760 | |
| We've seen that before with technologies | |
| 11:32.760 --> 11:35.560 | |
| that are far less transformative than AGI will be. | |
| 11:35.560 --> 11:37.480 | |
| And so I think that one piece | |
| 11:37.480 --> 11:39.560 | |
| is that it's just even hard to imagine | |
| 11:39.560 --> 11:41.640 | |
| and to really put yourself in a world | |
| 11:41.640 --> 11:44.600 | |
| where you can predict what that positive vision | |
| 11:44.600 --> 11:45.760 | |
| would be like. | |
| 11:46.920 --> 11:49.520 | |
| And I think the second thing is that it is, | |
| 11:49.520 --> 11:53.280 | |
| I think it is always easier to support | |
| 11:53.280 --> 11:55.080 | |
| the negative side than the positive side. | |
| 11:55.080 --> 11:57.120 | |
| It's always easier to destroy than create. | |
| 11:58.200 --> 12:00.800 | |
| And, you know, less in a physical sense | |
| 12:00.800 --> 12:03.080 | |
| and more just in an intellectual sense, right? | |
| 12:03.080 --> 12:05.680 | |
| Because, you know, I think that with creating something, | |
| 12:05.680 --> 12:07.440 | |
| you need to just get a bunch of things right | |
| 12:07.440 --> 12:10.280 | |
| and to destroy, you just need to get one thing wrong. | |
| 12:10.280 --> 12:12.080 | |
| And so I think that what that means | |
| 12:12.080 --> 12:14.240 | |
| is that I think a lot of people's thinking dead ends | |
| 12:14.240 --> 12:16.880 | |
| as soon as they see the negative story. | |
| 12:16.880 --> 12:20.360 | |
| But that being said, I actually have some hope, right? | |
| 12:20.360 --> 12:23.160 | |
| I think that the positive vision | |
| 12:23.160 --> 12:26.000 | |
| is something that I think can be, | |
| 12:26.000 --> 12:27.600 | |
| is something that we can talk about. | |
| 12:27.600 --> 12:30.200 | |
| I think that just simply saying this fact of, | |
| 12:30.200 --> 12:32.000 | |
| yeah, like there's positive, there's negatives, | |
| 12:32.000 --> 12:33.600 | |
| everyone likes to dwell on the negative, | |
| 12:33.600 --> 12:35.360 | |
| people actually respond well to that message and say, | |
| 12:35.360 --> 12:37.040 | |
| huh, you're right, there's a part of this | |
| 12:37.040 --> 12:39.640 | |
| that we're not talking about, not thinking about. | |
| 12:39.640 --> 12:41.240 | |
| And that's actually something that's, | |
| 12:41.240 --> 12:43.800 | |
| I think really been a key part | |
| 12:43.800 --> 12:46.640 | |
| of how we think about AGI at OpenAI, right? | |
| 12:46.640 --> 12:48.160 | |
| You can kind of look at it as like, okay, | |
| 12:48.160 --> 12:51.000 | |
| like OpenAI talks about the fact that there are risks | |
| 12:51.000 --> 12:53.160 | |
| and yet they're trying to build this system. | |
| 12:53.160 --> 12:56.080 | |
| Like how do you square those two facts? | |
| 12:56.080 --> 12:59.120 | |
| So do you share the intuition that some people have, | |
| 12:59.120 --> 13:02.680 | |
| I mean, from Sam Harris to even Elon Musk himself, | |
| 13:02.680 --> 13:06.600 | |
| that it's tricky as you develop AGI | |
| 13:06.600 --> 13:10.400 | |
| to keep it from slipping into the existential threats, | |
| 13:10.400 --> 13:11.760 | |
| into the negative. | |
| 13:11.760 --> 13:13.640 | |
| What's your intuition about, | |
| 13:13.640 --> 13:17.720 | |
| how hard is it to keep AI development | |
| 13:17.720 --> 13:19.640 | |
| on the positive track? | |
| 13:19.640 --> 13:20.680 | |
| What's your intuition there? | |
| 13:20.680 --> 13:21.560 | |
| To answer that question, | |
| 13:21.560 --> 13:23.960 | |
| you can really look at how we structure OpenAI. | |
| 13:23.960 --> 13:25.840 | |
| So we really have three main arms. | |
| 13:25.840 --> 13:26.960 | |
| So we have capabilities, | |
| 13:26.960 --> 13:29.040 | |
| which is actually doing the technical work | |
| 13:29.040 --> 13:31.160 | |
| and pushing forward what these systems can do. | |
| 13:31.160 --> 13:35.120 | |
| There's safety, which is working on technical mechanisms | |
| 13:35.120 --> 13:36.920 | |
| to ensure that the systems we build | |
| 13:36.920 --> 13:38.480 | |
| are aligned with human values. | |
| 13:38.480 --> 13:39.640 | |
| And then there's policy, | |
| 13:39.640 --> 13:42.040 | |
| which is making sure that we have governance mechanisms, | |
| 13:42.040 --> 13:45.280 | |
| answering that question of, well, whose values? | |
| 13:45.280 --> 13:47.360 | |
| And so I think that the technical safety one | |
| 13:47.360 --> 13:50.480 | |
| is the one that people kind of talk about the most, right? | |
| 13:50.480 --> 13:52.080 | |
| You talk about, like think about, | |
| 13:52.080 --> 13:54.200 | |
| you know, all of the dystopic AI movies, | |
| 13:54.200 --> 13:55.960 | |
| a lot of that is about not having good | |
| 13:55.960 --> 13:57.520 | |
| technical safety in place. | |
| 13:57.520 --> 13:59.960 | |
| And what we've been finding is that, you know, | |
| 13:59.960 --> 14:01.360 | |
| I think that actually a lot of people | |
| 14:01.360 --> 14:02.680 | |
| look at the technical safety problem | |
| 14:02.680 --> 14:05.400 | |
| and think it's just intractable, right? | |
| 14:05.400 --> 14:07.840 | |
| This question of what do humans want? | |
| 14:07.840 --> 14:09.160 | |
| How am I supposed to write that down? | |
| 14:09.160 --> 14:11.240 | |
| Can I even write down what I want? | |
| 14:11.240 --> 14:12.080 | |
| No way. | |
| 14:13.040 --> 14:14.800 | |
| And then they stop there. | |
| 14:14.800 --> 14:16.880 | |
| But the thing is we've already built systems | |
| 14:16.880 --> 14:20.920 | |
| that are able to learn things that humans can't specify. | |
| 14:20.920 --> 14:22.920 | |
| You know, even the rules for how to recognize | |
| 14:22.920 --> 14:25.000 | |
| if there's a cat or a dog in an image. | |
| 14:25.000 --> 14:26.520 | |
| Turns out it's intractable to write that down | |
| 14:26.520 --> 14:28.400 | |
| and yet we're able to learn it. | |
| 14:28.400 --> 14:31.040 | |
| And that what we're seeing with systems we build at OpenAI | |
| 14:31.040 --> 14:33.800 | |
| and they're still in early proof of concept stage | |
| 14:33.800 --> 14:36.320 | |
| is that you are able to learn human preferences. | |
| 14:36.320 --> 14:38.920 | |
| You're able to learn what humans want from data. | |
| 14:38.920 --> 14:40.400 | |
| And so that's kind of the core focus | |
| 14:40.400 --> 14:41.760 | |
| for our technical safety team. | |
| 14:41.760 --> 14:43.800 | |
| And I think that they're actually, | |
| 14:43.800 --> 14:45.640 | |
| we've had some pretty encouraging updates | |
| 14:45.640 --> 14:48.040 | |
| in terms of what we've been able to make work. | |
| 14:48.040 --> 14:51.680 | |
| So you have an intuition and a hope that from data, | |
| 14:51.680 --> 14:53.640 | |
| you know, looking at the value alignment problem, | |
| 14:53.640 --> 14:57.040 | |
| from data we can build systems that align | |
| 14:57.040 --> 15:00.600 | |
| with the collective better angels of our nature. | |
| 15:00.600 --> 15:04.600 | |
| So align with the ethics and the morals of human beings. | |
| 15:04.600 --> 15:05.880 | |
| To even say this in a different way, | |
| 15:05.880 --> 15:08.560 | |
| I mean, think about how do we align humans, right? | |
| 15:08.560 --> 15:10.400 | |
| Think about like a human baby can grow up | |
| 15:10.400 --> 15:12.880 | |
| to be an evil person or a great person. | |
| 15:12.880 --> 15:15.200 | |
| And a lot of that is from learning from data, right? | |
| 15:15.200 --> 15:17.720 | |
| That you have some feedback as a child is growing up. | |
| 15:17.720 --> 15:19.160 | |
| They get to see positive examples. | |
| 15:19.160 --> 15:23.120 | |
| And so I think that just like the only example | |
| 15:23.120 --> 15:25.400 | |
| we have of a general intelligence | |
| 15:25.400 --> 15:28.040 | |
| that is able to learn from data | |
| 15:28.040 --> 15:31.440 | |
| to align with human values and to learn values, | |
| 15:31.440 --> 15:32.880 | |
| I think we shouldn't be surprised | |
| 15:32.880 --> 15:36.040 | |
| that we can do the same sorts of techniques | |
| 15:36.040 --> 15:37.440 | |
| or whether the same sort of techniques | |
| 15:37.440 --> 15:41.080 | |
| end up being how we solve value alignment for AGI's. | |
| 15:41.080 --> 15:42.680 | |
| So let's go even higher. | |
| 15:42.680 --> 15:44.800 | |
| I don't know if you've read the book, Sapiens. | |
| 15:44.800 --> 15:48.320 | |
| But there's an idea that, you know, | |
| 15:48.320 --> 15:50.000 | |
| that as a collective, as us human beings, | |
| 15:50.000 --> 15:54.720 | |
| we kind of develop together ideas that we hold. | |
| 15:54.720 --> 15:57.920 | |
| There's no, in that context, objective truth. | |
| 15:57.920 --> 16:00.000 | |
| We just kind of all agree to certain ideas | |
| 16:00.000 --> 16:01.440 | |
| and hold them as a collective. | |
| 16:01.440 --> 16:03.480 | |
| Did you have a sense that there is | |
| 16:03.480 --> 16:05.360 | |
| in the world of good and evil, | |
| 16:05.360 --> 16:07.560 | |
| do you have a sense that to the first approximation, | |
| 16:07.560 --> 16:10.280 | |
| there are some things that are good | |
| 16:10.280 --> 16:14.520 | |
| and that you could teach systems to behave to be good? | |
| 16:14.520 --> 16:18.440 | |
| So I think that this actually blends into our third team, | |
| 16:18.440 --> 16:19.880 | |
| which is the policy team. | |
| 16:19.880 --> 16:22.320 | |
| And this is the one, the aspect that I think people | |
| 16:22.320 --> 16:25.280 | |
| really talk about way less than they should. | |
| 16:25.280 --> 16:27.640 | |
| Because imagine that we build super powerful systems | |
| 16:27.640 --> 16:29.720 | |
| that we've managed to figure out all the mechanisms | |
| 16:29.720 --> 16:32.800 | |
| for these things to do whatever the operator wants. | |
| 16:32.800 --> 16:34.480 | |
| The most important question becomes, | |
| 16:34.480 --> 16:36.720 | |
| who's the operator, what do they want, | |
| 16:36.720 --> 16:39.400 | |
| and how is that going to affect everyone else? | |
| 16:39.400 --> 16:43.080 | |
| And I think that this question of what is good, | |
| 16:43.080 --> 16:44.720 | |
| what are those values, I mean, | |
| 16:44.720 --> 16:45.960 | |
| I think you don't even have to go | |
| 16:45.960 --> 16:48.400 | |
| to those very grand existential places | |
| 16:48.400 --> 16:50.920 | |
| to realize how hard this problem is. | |
| 16:50.920 --> 16:52.880 | |
| You just look at different countries | |
| 16:52.880 --> 16:54.520 | |
| and cultures across the world. | |
| 16:54.520 --> 16:57.120 | |
| And that there's a very different conception | |
| 16:57.120 --> 17:01.920 | |
| of how the world works and what kinds of ways | |
| 17:01.920 --> 17:03.400 | |
| that society wants to operate. | |
| 17:03.400 --> 17:07.000 | |
| And so I think that the really core question | |
| 17:07.000 --> 17:09.560 | |
| is actually very concrete. | |
| 17:09.560 --> 17:10.960 | |
| And I think it's not a question | |
| 17:10.960 --> 17:12.880 | |
| that we have ready answers to, | |
| 17:12.880 --> 17:16.560 | |
| how do you have a world where all the different countries | |
| 17:16.560 --> 17:19.720 | |
| that we have, United States, China, Russia, | |
| 17:19.720 --> 17:22.720 | |
| and the hundreds of other countries out there | |
| 17:22.720 --> 17:26.600 | |
| are able to continue to not just operate | |
| 17:26.600 --> 17:28.440 | |
| in the way that they see fit, | |
| 17:28.440 --> 17:32.520 | |
| but in the world that emerges in these, | |
| 17:32.520 --> 17:34.680 | |
| where you have these very powerful systems, | |
| 17:36.040 --> 17:37.800 | |
| operating alongside humans, | |
| 17:37.800 --> 17:39.800 | |
| ends up being something that empowers humans more, | |
| 17:39.800 --> 17:44.120 | |
| that makes human existence be a more meaningful thing | |
| 17:44.120 --> 17:46.400 | |
| and that people are happier and wealthier | |
| 17:46.400 --> 17:48.960 | |
| and able to live more fulfilling lives. | |
| 17:48.960 --> 17:51.560 | |
| It's not an obvious thing for how to design that world | |
| 17:51.560 --> 17:53.600 | |
| once you have that very powerful system. | |
| 17:53.600 --> 17:55.800 | |
| So if we take a little step back, | |
| 17:55.800 --> 17:58.200 | |
| and we're having like a fascinating conversation | |
| 17:58.200 --> 18:01.880 | |
| and open as in many ways a tech leader in the world, | |
| 18:01.880 --> 18:05.440 | |
| and yet we're thinking about these big existential questions | |
| 18:05.440 --> 18:07.000 | |
| which is fascinating, really important. | |
| 18:07.000 --> 18:09.160 | |
| I think you're a leader in that space | |
| 18:09.160 --> 18:10.840 | |
| and that's a really important space | |
| 18:10.840 --> 18:13.080 | |
| of just thinking how AI affects society | |
| 18:13.080 --> 18:14.360 | |
| in a big picture view. | |
| 18:14.360 --> 18:17.320 | |
| So Oscar Wilde said, we're all in the gutter, | |
| 18:17.320 --> 18:19.000 | |
| but some of us are looking at the stars | |
| 18:19.000 --> 18:22.320 | |
| and I think OpenAI has a charter | |
| 18:22.320 --> 18:24.600 | |
| that looks to the stars, I would say, | |
| 18:24.600 --> 18:26.880 | |
| to create intelligence, to create general intelligence, | |
| 18:26.880 --> 18:29.440 | |
| make it beneficial, safe, and collaborative. | |
| 18:29.440 --> 18:33.680 | |
| So can you tell me how that came about? | |
| 18:33.680 --> 18:36.320 | |
| How a mission like that and the path | |
| 18:36.320 --> 18:39.120 | |
| to creating a mission like that at OpenAI was founded? | |
| 18:39.120 --> 18:41.640 | |
| Yeah, so I think that in some ways | |
| 18:41.640 --> 18:45.040 | |
| it really boils down to taking a look at the landscape. | |
| 18:45.040 --> 18:47.040 | |
| So if you think about the history of AI | |
| 18:47.040 --> 18:49.920 | |
| that basically for the past 60 or 70 years, | |
| 18:49.920 --> 18:51.640 | |
| people have thought about this goal | |
| 18:51.640 --> 18:53.960 | |
| of what could happen if you could automate | |
| 18:53.960 --> 18:55.640 | |
| human intellectual labor. | |
| 18:56.680 --> 18:58.280 | |
| Imagine you can build a computer system | |
| 18:58.280 --> 19:00.560 | |
| that could do that, what becomes possible? | |
| 19:00.560 --> 19:02.400 | |
| We have a lot of sci fi that tells stories | |
| 19:02.400 --> 19:04.920 | |
| of various dystopias and increasingly you have movies | |
| 19:04.920 --> 19:06.480 | |
| like Her that tell you a little bit about | |
| 19:06.480 --> 19:09.440 | |
| maybe more of a little bit utopic vision. | |
| 19:09.440 --> 19:12.560 | |
| You think about the impacts that we've seen | |
| 19:12.560 --> 19:16.280 | |
| from being able to have bicycles for our minds | |
| 19:16.280 --> 19:20.360 | |
| and computers and that I think that the impact | |
| 19:20.360 --> 19:23.480 | |
| of computers and the internet has just far outstripped | |
| 19:23.480 --> 19:26.200 | |
| what anyone really could have predicted. | |
| 19:26.200 --> 19:27.400 | |
| And so I think that it's very clear | |
| 19:27.400 --> 19:29.360 | |
| that if you can build an AGI, | |
| 19:29.360 --> 19:31.600 | |
| it will be the most transformative technology | |
| 19:31.600 --> 19:33.040 | |
| that humans will ever create. | |
| 19:33.040 --> 19:36.840 | |
| And so what it boils down to then is a question of, | |
| 19:36.840 --> 19:38.680 | |
| well, is there a path? | |
| 19:38.680 --> 19:39.520 | |
| Is there hope? | |
| 19:39.520 --> 19:41.680 | |
| Is there a way to build such a system? | |
| 19:41.680 --> 19:43.640 | |
| And I think that for 60 or 70 years | |
| 19:43.640 --> 19:48.040 | |
| that people got excited and that ended up not being able | |
| 19:48.040 --> 19:51.480 | |
| to deliver on the hopes that people had pinned on them. | |
| 19:51.480 --> 19:54.880 | |
| And I think that then, that after two winters | |
| 19:54.880 --> 19:57.600 | |
| of AI development, that people, | |
| 19:57.600 --> 20:00.560 | |
| I think kind of almost stopped daring to dream, right? | |
| 20:00.560 --> 20:03.280 | |
| That really talking about AGI or thinking about AGI | |
| 20:03.280 --> 20:05.640 | |
| became almost this taboo in the community. | |
| 20:06.640 --> 20:08.720 | |
| But I actually think that people took the wrong lesson | |
| 20:08.720 --> 20:10.080 | |
| from AI history. | |
| 20:10.080 --> 20:12.400 | |
| And if you look back, starting in 1959 | |
| 20:12.400 --> 20:14.240 | |
| is when the Perceptron was released. | |
| 20:14.240 --> 20:17.720 | |
| And this is basically one of the earliest neural networks. | |
| 20:17.720 --> 20:19.280 | |
| It was released to what was perceived | |
| 20:19.280 --> 20:20.840 | |
| as this massive overhype. | |
| 20:20.840 --> 20:22.360 | |
| So in the New York Times in 1959, | |
| 20:22.360 --> 20:26.400 | |
| you have this article saying that the Perceptron | |
| 20:26.400 --> 20:29.160 | |
| will one day recognize people, call out their names, | |
| 20:29.160 --> 20:31.480 | |
| instantly translate speech between languages. | |
| 20:31.480 --> 20:33.800 | |
| And people at the time looked at this and said, | |
| 20:33.800 --> 20:36.120 | |
| this is, your system can't do any of that. | |
| 20:36.120 --> 20:38.080 | |
| And basically spent 10 years trying to discredit | |
| 20:38.080 --> 20:40.640 | |
| the whole Perceptron direction and succeeded. | |
| 20:40.640 --> 20:41.840 | |
| And all the funding dried up. | |
| 20:41.840 --> 20:44.960 | |
| And people kind of went in other directions. | |
| 20:44.960 --> 20:46.920 | |
| And in the 80s, there was this resurgence. | |
| 20:46.920 --> 20:49.320 | |
| And I'd always heard that the resurgence in the 80s | |
| 20:49.320 --> 20:51.520 | |
| was due to the invention of back propagation | |
| 20:51.520 --> 20:53.720 | |
| and these algorithms that got people excited. | |
| 20:53.720 --> 20:55.760 | |
| But actually the causality was due to people | |
| 20:55.760 --> 20:57.200 | |
| building larger computers. | |
| 20:57.200 --> 20:59.280 | |
| That you can find these articles from the 80s saying | |
| 20:59.280 --> 21:01.760 | |
| that the democratization of computing power | |
| 21:01.760 --> 21:04.040 | |
| suddenly meant that you could run these larger neural networks. | |
| 21:04.040 --> 21:06.280 | |
| And then people started to do all these amazing things, | |
| 21:06.280 --> 21:08.000 | |
| back propagation algorithm was invented. | |
| 21:08.000 --> 21:10.120 | |
| And the neural nets people were running | |
| 21:10.120 --> 21:13.000 | |
| were these tiny little like 20 neuron neural nets. | |
| 21:13.000 --> 21:15.160 | |
| What are you supposed to learn with 20 neurons? | |
| 21:15.160 --> 21:18.640 | |
| And so of course they weren't able to get great results. | |
| 21:18.640 --> 21:21.960 | |
| And it really wasn't until 2012 that this approach, | |
| 21:21.960 --> 21:24.680 | |
| that's almost the most simple, natural approach | |
| 21:24.680 --> 21:27.720 | |
| that people had come up with in the 50s, right? | |
| 21:27.720 --> 21:30.360 | |
| In some ways, even in the 40s before there were computers | |
| 21:30.360 --> 21:32.000 | |
| with the Pits McCullin neuron, | |
| 21:33.040 --> 21:37.480 | |
| suddenly this became the best way of solving problems, right? | |
| 21:37.480 --> 21:39.280 | |
| And I think there are three core properties | |
| 21:39.280 --> 21:42.120 | |
| that deep learning has that I think | |
| 21:42.120 --> 21:44.120 | |
| are very worth paying attention to. | |
| 21:44.120 --> 21:45.920 | |
| The first is generality. | |
| 21:45.920 --> 21:48.760 | |
| We have a very small number of deep learning tools, | |
| 21:48.760 --> 21:52.360 | |
| SGD, deep neural net, maybe some, you know, RL. | |
| 21:52.360 --> 21:55.600 | |
| And it solves this huge variety of problems, | |
| 21:55.600 --> 21:57.240 | |
| speech recognition, machine translation, | |
| 21:57.240 --> 22:00.200 | |
| game playing, all of these problems, | |
| 22:00.200 --> 22:01.040 | |
| small set of tools. | |
| 22:01.040 --> 22:02.760 | |
| So there's the generality. | |
| 22:02.760 --> 22:05.000 | |
| There's a second piece, which is the competence. | |
| 22:05.000 --> 22:07.040 | |
| You wanna solve any of those problems? | |
| 22:07.040 --> 22:10.640 | |
| Throughout 40 years worth of normal computer vision research | |
| 22:10.640 --> 22:13.640 | |
| replaced with a deep neural net, it's gonna work better. | |
| 22:13.640 --> 22:16.320 | |
| And there's a third piece, which is the scalability, right? | |
| 22:16.320 --> 22:18.720 | |
| That one thing that has been shown time and time again | |
| 22:18.720 --> 22:21.760 | |
| is that you, if you have a larger neural network, | |
| 22:21.760 --> 22:25.120 | |
| throw more compute, more data at it, it will work better. | |
| 22:25.120 --> 22:28.880 | |
| Those three properties together feel like essential parts | |
| 22:28.880 --> 22:30.800 | |
| of building a general intelligence. | |
| 22:30.800 --> 22:33.000 | |
| Now, it doesn't just mean that if we scale up | |
| 22:33.000 --> 22:35.200 | |
| what we have, that we will have an AGI, right? | |
| 22:35.200 --> 22:36.800 | |
| There are clearly missing pieces. | |
| 22:36.800 --> 22:38.000 | |
| There are missing ideas. | |
| 22:38.000 --> 22:40.000 | |
| We need to have answers for reasoning. | |
| 22:40.000 --> 22:44.800 | |
| But I think that the core here is that for the first time, | |
| 22:44.800 --> 22:46.880 | |
| it feels that we have a paradigm | |
| 22:46.880 --> 22:48.960 | |
| that gives us hope that general intelligence | |
| 22:48.960 --> 22:50.560 | |
| can be achievable. | |
| 22:50.560 --> 22:52.160 | |
| And so as soon as you believe that, | |
| 22:52.160 --> 22:54.480 | |
| everything else becomes into focus, right? | |
| 22:54.480 --> 22:56.560 | |
| If you imagine that you may be able to, | |
| 22:56.560 --> 22:59.920 | |
| and that the timeline I think remains uncertain, | |
| 22:59.920 --> 23:02.200 | |
| but I think that certainly within our lifetimes | |
| 23:02.200 --> 23:04.640 | |
| and possibly within a much shorter period of time | |
| 23:04.640 --> 23:06.560 | |
| than people would expect, | |
| 23:06.560 --> 23:09.360 | |
| if you can really build the most transformative technology | |
| 23:09.360 --> 23:11.720 | |
| that will ever exist, you stop thinking about yourself | |
| 23:11.720 --> 23:12.560 | |
| so much, right? | |
| 23:12.560 --> 23:14.240 | |
| And you start thinking about just like, | |
| 23:14.240 --> 23:16.440 | |
| how do you have a world where this goes well? | |
| 23:16.440 --> 23:18.160 | |
| And that you need to think about the practicalities | |
| 23:18.160 --> 23:19.560 | |
| of how do you build an organization | |
| 23:19.560 --> 23:22.000 | |
| and get together a bunch of people and resources | |
| 23:22.000 --> 23:25.160 | |
| and to make sure that people feel motivated | |
| 23:25.160 --> 23:26.800 | |
| and ready to do it. | |
| 23:28.080 --> 23:30.720 | |
| But I think that then you start thinking about, | |
| 23:30.720 --> 23:32.080 | |
| well, what if we succeed? | |
| 23:32.080 --> 23:34.280 | |
| And how do we make sure that when we succeed, | |
| 23:34.280 --> 23:35.600 | |
| that the world is actually the place | |
| 23:35.600 --> 23:38.200 | |
| that we want ourselves to exist in? | |
| 23:38.200 --> 23:41.080 | |
| And almost in the Rawlsian Vale sense of the word. | |
| 23:41.080 --> 23:43.880 | |
| And so that's kind of the broader landscape. | |
| 23:43.880 --> 23:46.680 | |
| And Open AI was really formed in 2015 | |
| 23:46.680 --> 23:51.480 | |
| with that high level picture of AGI might be possible | |
| 23:51.480 --> 23:52.880 | |
| sooner than people think | |
| 23:52.880 --> 23:55.840 | |
| and that we need to try to do our best | |
| 23:55.840 --> 23:57.480 | |
| to make sure it's going to go well. | |
| 23:57.480 --> 23:59.360 | |
| And then we spent the next couple of years | |
| 23:59.360 --> 24:00.840 | |
| really trying to figure out what does that mean? | |
| 24:00.840 --> 24:01.960 | |
| How do we do it? | |
| 24:01.960 --> 24:04.800 | |
| And I think that typically with a company, | |
| 24:04.800 --> 24:07.320 | |
| you start out very small. | |
| 24:07.320 --> 24:09.000 | |
| So you want a cofounder and you build a product, | |
| 24:09.000 --> 24:11.360 | |
| you get some users, you get a product market fit, | |
| 24:11.360 --> 24:13.320 | |
| then at some point you raise some money, | |
| 24:13.320 --> 24:14.840 | |
| you hire people, you scale, | |
| 24:14.840 --> 24:17.440 | |
| and then down the road, then the big companies | |
| 24:17.440 --> 24:19.080 | |
| realize you exist and try to kill you. | |
| 24:19.080 --> 24:21.520 | |
| And for Open AI, it was basically everything | |
| 24:21.520 --> 24:22.960 | |
| in exactly the opposite order. | |
| 24:25.480 --> 24:26.760 | |
| Let me just pause for a second. | |
| 24:26.760 --> 24:27.520 | |
| He said a lot of things. | |
| 24:27.520 --> 24:31.240 | |
| And let me just admire the jarring aspect | |
| 24:31.240 --> 24:35.160 | |
| of what Open AI stands for, which is daring to dream. | |
| 24:35.160 --> 24:37.120 | |
| I mean, you said it's pretty powerful. | |
| 24:37.120 --> 24:40.080 | |
| You caught me off guard because I think that's very true. | |
| 24:40.080 --> 24:44.040 | |
| The step of just daring to dream | |
| 24:44.040 --> 24:46.720 | |
| about the possibilities of creating intelligence | |
| 24:46.720 --> 24:48.760 | |
| in a positive and a safe way, | |
| 24:48.760 --> 24:50.640 | |
| but just even creating intelligence | |
| 24:50.640 --> 24:55.640 | |
| is a much needed, refreshing catalyst | |
| 24:56.280 --> 24:57.360 | |
| for the AI community. | |
| 24:57.360 --> 24:58.800 | |
| So that's the starting point. | |
| 24:58.800 --> 25:02.840 | |
| Okay, so then formation of Open AI, what's your point? | |
| 25:02.840 --> 25:05.640 | |
| I would just say that when we were starting Open AI, | |
| 25:05.640 --> 25:07.760 | |
| that kind of the first question that we had is, | |
| 25:07.760 --> 25:12.000 | |
| is it too late to start a lab with a bunch of the best people? | |
| 25:12.000 --> 25:13.160 | |
| Right, is that even possible? | |
| 25:13.160 --> 25:14.320 | |
| That was an actual question. | |
| 25:14.320 --> 25:17.280 | |
| That was the core question of, | |
| 25:17.280 --> 25:19.320 | |
| we had this dinner in July of 2015, | |
| 25:19.320 --> 25:21.240 | |
| and that was really what we spent the whole time | |
| 25:21.240 --> 25:22.320 | |
| talking about. | |
| 25:22.320 --> 25:26.800 | |
| And because you think about kind of where AI was, | |
| 25:26.800 --> 25:30.200 | |
| is that it transitioned from being an academic pursuit | |
| 25:30.200 --> 25:32.240 | |
| to an industrial pursuit. | |
| 25:32.240 --> 25:34.240 | |
| And so a lot of the best people were in these big | |
| 25:34.240 --> 25:37.000 | |
| research labs and that we wanted to start our own one | |
| 25:37.000 --> 25:40.560 | |
| that no matter how much resources we could accumulate | |
| 25:40.560 --> 25:43.520 | |
| would be pale in comparison to the big tech companies. | |
| 25:43.520 --> 25:44.720 | |
| And we knew that. | |
| 25:44.720 --> 25:45.800 | |
| And there's a question of, | |
| 25:45.800 --> 25:47.720 | |
| are we going to be actually able to get this thing | |
| 25:47.720 --> 25:48.720 | |
| off the ground? | |
| 25:48.720 --> 25:49.760 | |
| You need critical mass. | |
| 25:49.760 --> 25:52.120 | |
| You can't just do you and a cofounder build a product, right? | |
| 25:52.120 --> 25:55.600 | |
| You really need to have a group of five to 10 people. | |
| 25:55.600 --> 25:59.480 | |
| And we kind of concluded it wasn't obviously impossible. | |
| 25:59.480 --> 26:00.840 | |
| So it seemed worth trying. | |
| 26:02.240 --> 26:04.800 | |
| Well, you're also a dreamer, so who knows, right? | |
| 26:04.800 --> 26:05.640 | |
| That's right. | |
| 26:05.640 --> 26:07.720 | |
| Okay, so speaking of that, | |
| 26:07.720 --> 26:10.520 | |
| competing with the big players, | |
| 26:11.520 --> 26:14.080 | |
| let's talk about some of the tricky things | |
| 26:14.080 --> 26:17.480 | |
| as you think through this process of growing, | |
| 26:17.480 --> 26:20.080 | |
| of seeing how you can develop these systems | |
| 26:20.080 --> 26:22.640 | |
| at a scale that competes. | |
| 26:22.640 --> 26:25.720 | |
| So you recently formed OpenAI LP, | |
| 26:26.560 --> 26:30.800 | |
| a new cap profit company that now carries the name OpenAI. | |
| 26:30.800 --> 26:33.280 | |
| So OpenAI is now this official company. | |
| 26:33.280 --> 26:36.520 | |
| The original nonprofit company still exists | |
| 26:36.520 --> 26:39.800 | |
| and carries the OpenAI nonprofit name. | |
| 26:39.800 --> 26:42.000 | |
| So can you explain what this company is, | |
| 26:42.000 --> 26:44.280 | |
| what the purpose of its creation is, | |
| 26:44.280 --> 26:48.800 | |
| and how did you arrive at the decision to create it? | |
| 26:48.800 --> 26:53.280 | |
| OpenAI, the whole entity and OpenAI LP as a vehicle | |
| 26:53.280 --> 26:55.560 | |
| is trying to accomplish the mission | |
| 26:55.560 --> 26:57.520 | |
| of ensuring that artificial general intelligence | |
| 26:57.520 --> 26:58.800 | |
| benefits everyone. | |
| 26:58.800 --> 27:00.240 | |
| And the main way that we're trying to do that | |
| 27:00.240 --> 27:01.840 | |
| is by actually trying to build | |
| 27:01.840 --> 27:03.240 | |
| general intelligence to ourselves | |
| 27:03.240 --> 27:05.920 | |
| and make sure the benefits are distributed to the world. | |
| 27:05.920 --> 27:07.200 | |
| That's the primary way. | |
| 27:07.200 --> 27:09.600 | |
| We're also fine if someone else does this, right? | |
| 27:09.600 --> 27:10.640 | |
| It doesn't have to be us. | |
| 27:10.640 --> 27:12.640 | |
| If someone else is going to build an AGI | |
| 27:12.640 --> 27:14.840 | |
| and make sure that the benefits don't get locked up | |
| 27:14.840 --> 27:18.160 | |
| in one company or with one set of people, | |
| 27:19.280 --> 27:21.160 | |
| like we're actually fine with that. | |
| 27:21.160 --> 27:25.400 | |
| And so those ideas are baked into our charter, | |
| 27:25.400 --> 27:28.400 | |
| which is kind of the foundational document | |
| 27:28.400 --> 27:31.920 | |
| that describes kind of our values and how we operate. | |
| 27:31.920 --> 27:36.360 | |
| And it's also really baked into the structure of OpenAI LP. | |
| 27:36.360 --> 27:37.960 | |
| And so the way that we've set up OpenAI LP | |
| 27:37.960 --> 27:42.160 | |
| is that in the case where we succeed, right? | |
| 27:42.160 --> 27:45.320 | |
| If we actually build what we're trying to build, | |
| 27:45.320 --> 27:47.800 | |
| then investors are able to get a return, | |
| 27:47.800 --> 27:50.400 | |
| and but that return is something that is capped. | |
| 27:50.400 --> 27:53.000 | |
| And so if you think of AGI in terms of the value | |
| 27:53.000 --> 27:54.160 | |
| that you could really create, | |
| 27:54.160 --> 27:56.320 | |
| you're talking about the most transformative technology | |
| 27:56.320 --> 27:58.000 | |
| ever created, it's gonna create, | |
| 27:58.000 --> 28:01.880 | |
| or does the magnitude more value than any existing company? | |
| 28:01.880 --> 28:05.960 | |
| And that all of that value will be owned by the world, | |
| 28:05.960 --> 28:07.880 | |
| like legally titled to the nonprofit | |
| 28:07.880 --> 28:09.560 | |
| to fulfill that mission. | |
| 28:09.560 --> 28:12.800 | |
| And so that's the structure. | |
| 28:12.800 --> 28:15.200 | |
| So the mission is a powerful one, | |
| 28:15.200 --> 28:18.920 | |
| and it's one that I think most people would agree with. | |
| 28:18.920 --> 28:22.960 | |
| It's how we would hope AI progresses. | |
| 28:22.960 --> 28:25.440 | |
| And so how do you tie yourself to that mission? | |
| 28:25.440 --> 28:29.240 | |
| How do you make sure you do not deviate from that mission | |
| 28:29.240 --> 28:34.240 | |
| that other incentives that are profit driven | |
| 28:34.560 --> 28:36.800 | |
| wouldn't don't interfere with the mission? | |
| 28:36.800 --> 28:39.560 | |
| So this was actually a really core question for us | |
| 28:39.560 --> 28:40.920 | |
| for the past couple of years, | |
| 28:40.920 --> 28:43.560 | |
| because I'd say that the way that our history went | |
| 28:43.560 --> 28:44.960 | |
| was that for the first year, | |
| 28:44.960 --> 28:46.240 | |
| we were getting off the ground, right? | |
| 28:46.240 --> 28:47.960 | |
| We had this high level picture, | |
| 28:47.960 --> 28:51.880 | |
| but we didn't know exactly how we wanted to accomplish it. | |
| 28:51.880 --> 28:53.440 | |
| And really two years ago, | |
| 28:53.440 --> 28:55.040 | |
| it's when we first started realizing | |
| 28:55.040 --> 28:56.160 | |
| in order to build AGI, | |
| 28:56.160 --> 28:58.720 | |
| we're just gonna need to raise way more money | |
| 28:58.720 --> 29:00.680 | |
| than we can as a nonprofit. | |
| 29:00.680 --> 29:02.800 | |
| We're talking many billions of dollars. | |
| 29:02.800 --> 29:05.440 | |
| And so the first question is, | |
| 29:05.440 --> 29:06.840 | |
| how are you supposed to do that | |
| 29:06.840 --> 29:08.680 | |
| and stay true to this mission? | |
| 29:08.680 --> 29:10.560 | |
| And we looked at every legal structure out there | |
| 29:10.560 --> 29:11.960 | |
| and included none of them were quite right | |
| 29:11.960 --> 29:13.400 | |
| for what we wanted to do. | |
| 29:13.400 --> 29:14.600 | |
| And I guess it shouldn't be too surprising | |
| 29:14.600 --> 29:16.920 | |
| if you're gonna do some crazy unprecedented technology | |
| 29:16.920 --> 29:17.920 | |
| that you're gonna have to come | |
| 29:17.920 --> 29:20.320 | |
| with some crazy unprecedented structure to do it in. | |
| 29:20.320 --> 29:25.320 | |
| And a lot of our conversation was with people at OpenAI, | |
| 29:26.080 --> 29:27.240 | |
| the people who really joined | |
| 29:27.240 --> 29:29.160 | |
| because they believe so much in this mission | |
| 29:29.160 --> 29:32.120 | |
| and thinking about how do we actually raise the resources | |
| 29:32.120 --> 29:35.920 | |
| to do it and also stay true to what we stand for. | |
| 29:35.920 --> 29:38.000 | |
| And the place you gotta start is to really align | |
| 29:38.000 --> 29:39.560 | |
| on what is it that we stand for, right? | |
| 29:39.560 --> 29:40.560 | |
| What are those values? | |
| 29:40.560 --> 29:41.840 | |
| What's really important to us? | |
| 29:41.840 --> 29:43.760 | |
| And so I'd say that we spent about a year | |
| 29:43.760 --> 29:46.240 | |
| really compiling the OpenAI charter. | |
| 29:46.240 --> 29:47.560 | |
| And that determines, | |
| 29:47.560 --> 29:50.240 | |
| and if you even look at the first line item in there, | |
| 29:50.240 --> 29:52.360 | |
| it says that, look, we expect we're gonna have to marshal | |
| 29:52.360 --> 29:53.760 | |
| huge amounts of resources, | |
| 29:53.760 --> 29:55.160 | |
| but we're going to make sure | |
| 29:55.160 --> 29:57.920 | |
| that we minimize conflict of interest with the mission. | |
| 29:57.920 --> 30:00.720 | |
| And that kind of aligning on all of those pieces | |
| 30:00.720 --> 30:04.240 | |
| was the most important step towards figuring out | |
| 30:04.240 --> 30:06.040 | |
| how do we structure a company | |
| 30:06.040 --> 30:08.240 | |
| that can actually raise the resources | |
| 30:08.240 --> 30:10.360 | |
| to do what we need to do. | |
| 30:10.360 --> 30:14.760 | |
| I imagine OpenAI, the decision to create OpenAI LP | |
| 30:14.760 --> 30:16.360 | |
| was a really difficult one. | |
| 30:16.360 --> 30:17.920 | |
| And there was a lot of discussions | |
| 30:17.920 --> 30:19.640 | |
| as you mentioned for a year. | |
| 30:19.640 --> 30:22.760 | |
| And there was different ideas, | |
| 30:22.760 --> 30:25.120 | |
| perhaps detractors within OpenAI, | |
| 30:26.120 --> 30:28.920 | |
| sort of different paths that you could have taken. | |
| 30:28.920 --> 30:30.240 | |
| What were those concerns? | |
| 30:30.240 --> 30:32.040 | |
| What were the different paths considered? | |
| 30:32.040 --> 30:34.080 | |
| What was that process of making that decision like? | |
| 30:34.080 --> 30:35.000 | |
| Yep. | |
| 30:35.000 --> 30:37.200 | |
| But so if you look actually at the OpenAI charter, | |
| 30:37.200 --> 30:40.880 | |
| that there's almost two paths embedded within it. | |
| 30:40.880 --> 30:44.880 | |
| There is, we are primarily trying to build AGI ourselves, | |
| 30:44.880 --> 30:47.360 | |
| but we're also okay if someone else does it. | |
| 30:47.360 --> 30:49.040 | |
| And this is a weird thing for a company. | |
| 30:49.040 --> 30:50.480 | |
| It's really interesting, actually. | |
| 30:50.480 --> 30:51.320 | |
| Yeah. | |
| 30:51.320 --> 30:53.280 | |
| But there is an element of competition | |
| 30:53.280 --> 30:56.680 | |
| that you do want to be the one that does it, | |
| 30:56.680 --> 30:59.040 | |
| but at the same time, you're okay if somebody else doesn't. | |
| 30:59.040 --> 31:01.000 | |
| We'll talk about that a little bit, that trade off, | |
| 31:01.000 --> 31:02.960 | |
| that dance that's really interesting. | |
| 31:02.960 --> 31:04.600 | |
| And I think this was the core tension | |
| 31:04.600 --> 31:06.360 | |
| as we were designing OpenAI LP | |
| 31:06.360 --> 31:08.240 | |
| and really the OpenAI strategy, | |
| 31:08.240 --> 31:11.080 | |
| is how do you make sure that both you have a shot | |
| 31:11.080 --> 31:12.640 | |
| at being a primary actor, | |
| 31:12.640 --> 31:15.840 | |
| which really requires building an organization, | |
| 31:15.840 --> 31:17.720 | |
| raising massive resources, | |
| 31:17.720 --> 31:19.440 | |
| and really having the will to go | |
| 31:19.440 --> 31:22.000 | |
| and execute on some really, really hard vision, right? | |
| 31:22.000 --> 31:23.760 | |
| You need to really sign up for a long period | |
| 31:23.760 --> 31:27.120 | |
| to go and take on a lot of pain and a lot of risk. | |
| 31:27.120 --> 31:29.000 | |
| And to do that, | |
| 31:29.000 --> 31:31.720 | |
| normally you just import the startup mindset, right? | |
| 31:31.720 --> 31:32.760 | |
| And that you think about, okay, | |
| 31:32.760 --> 31:34.240 | |
| like how do we out execute everyone? | |
| 31:34.240 --> 31:36.160 | |
| You have this very competitive angle. | |
| 31:36.160 --> 31:38.120 | |
| But you also have the second angle of saying that, | |
| 31:38.120 --> 31:41.600 | |
| well, the true mission isn't for OpenAI to build AGI. | |
| 31:41.600 --> 31:45.080 | |
| The true mission is for AGI to go well for humanity. | |
| 31:45.080 --> 31:48.080 | |
| And so how do you take all of those first actions | |
| 31:48.080 --> 31:51.320 | |
| and make sure you don't close the door on outcomes | |
| 31:51.320 --> 31:54.480 | |
| that would actually be positive and fulfill the mission? | |
| 31:54.480 --> 31:56.680 | |
| And so I think it's a very delicate balance, right? | |
| 31:56.680 --> 31:59.560 | |
| And I think that going 100% one direction or the other | |
| 31:59.560 --> 32:01.320 | |
| is clearly not the correct answer. | |
| 32:01.320 --> 32:03.920 | |
| And so I think that even in terms of just how we talk about | |
| 32:03.920 --> 32:05.400 | |
| OpenAI and think about it, | |
| 32:05.400 --> 32:07.600 | |
| there's just like one thing that's always | |
| 32:07.600 --> 32:09.680 | |
| in the back of my mind is to make sure | |
| 32:09.680 --> 32:12.120 | |
| that we're not just saying OpenAI's goal | |
| 32:12.120 --> 32:14.000 | |
| is to build AGI, right? | |
| 32:14.000 --> 32:15.560 | |
| That it's actually much broader than that, right? | |
| 32:15.560 --> 32:19.360 | |
| That first of all, it's not just AGI, it's safe AGI | |
| 32:19.360 --> 32:20.320 | |
| that's very important. | |
| 32:20.320 --> 32:23.120 | |
| But secondly, our goal isn't to be the ones to build it, | |
| 32:23.120 --> 32:24.720 | |
| our goal is to make sure it goes well for the world. | |
| 32:24.720 --> 32:26.120 | |
| And so I think that figuring out, | |
| 32:26.120 --> 32:27.960 | |
| how do you balance all of those | |
| 32:27.960 --> 32:30.280 | |
| and to get people to really come to the table | |
| 32:30.280 --> 32:35.280 | |
| and compile a single document that encompasses all of that | |
| 32:36.360 --> 32:37.560 | |
| wasn't trivial. | |
| 32:37.560 --> 32:41.680 | |
| So part of the challenge here is your mission is, | |
| 32:41.680 --> 32:44.240 | |
| I would say, beautiful, empowering, | |
| 32:44.240 --> 32:47.520 | |
| and a beacon of hope for people in the research community | |
| 32:47.520 --> 32:49.200 | |
| and just people thinking about AI. | |
| 32:49.200 --> 32:51.880 | |
| So your decisions are scrutinized | |
| 32:51.880 --> 32:55.920 | |
| more than, I think, a regular profit driven company. | |
| 32:55.920 --> 32:57.400 | |
| Do you feel the burden of this | |
| 32:57.400 --> 32:58.560 | |
| in the creation of the charter | |
| 32:58.560 --> 33:00.200 | |
| and just in the way you operate? | |
| 33:00.200 --> 33:01.040 | |
| Yes. | |
| 33:03.040 --> 33:05.920 | |
| So why do you lean into the burden | |
| 33:07.040 --> 33:08.640 | |
| by creating such a charter? | |
| 33:08.640 --> 33:10.440 | |
| Why not keep it quiet? | |
| 33:10.440 --> 33:12.920 | |
| I mean, it just boils down to the mission, right? | |
| 33:12.920 --> 33:15.200 | |
| Like, I'm here and everyone else is here | |
| 33:15.200 --> 33:17.880 | |
| because we think this is the most important mission, right? | |
| 33:17.880 --> 33:19.000 | |
| Dare to dream. | |
| 33:19.000 --> 33:23.360 | |
| All right, so do you think you can be good for the world | |
| 33:23.360 --> 33:26.000 | |
| or create an AGI system that's good | |
| 33:26.000 --> 33:28.320 | |
| when you're a for profit company? | |
| 33:28.320 --> 33:32.920 | |
| From my perspective, I don't understand why profit | |
| 33:32.920 --> 33:37.640 | |
| interferes with positive impact on society. | |
| 33:37.640 --> 33:40.760 | |
| I don't understand why Google | |
| 33:40.760 --> 33:42.920 | |
| that makes most of its money from ads | |
| 33:42.920 --> 33:45.040 | |
| can't also do good for the world | |
| 33:45.040 --> 33:47.520 | |
| or other companies, Facebook, anything. | |
| 33:47.520 --> 33:50.240 | |
| I don't understand why those have to interfere. | |
| 33:50.240 --> 33:55.120 | |
| You know, you can, profit isn't the thing in my view | |
| 33:55.120 --> 33:57.240 | |
| that affects the impact of a company. | |
| 33:57.240 --> 34:00.360 | |
| What affects the impact of the company is the charter, | |
| 34:00.360 --> 34:04.160 | |
| is the culture, is the people inside | |
| 34:04.160 --> 34:07.360 | |
| and profit is the thing that just fuels those people. | |
| 34:07.360 --> 34:08.760 | |
| What are your views there? | |
| 34:08.760 --> 34:10.920 | |
| Yeah, so I think that's a really good question | |
| 34:10.920 --> 34:14.200 | |
| and there's some real like longstanding debates | |
| 34:14.200 --> 34:16.520 | |
| in human society that are wrapped up in it. | |
| 34:16.520 --> 34:18.680 | |
| The way that I think about it is just think about | |
| 34:18.680 --> 34:21.520 | |
| what are the most impactful nonprofits in the world? | |
| 34:24.000 --> 34:26.760 | |
| What are the most impactful for profits in the world? | |
| 34:26.760 --> 34:29.280 | |
| Right, it's much easier to list the for profits. | |
| 34:29.280 --> 34:30.120 | |
| That's right. | |
| 34:30.120 --> 34:32.400 | |
| And I think that there's some real truth here | |
| 34:32.400 --> 34:34.600 | |
| that the system that we set up, | |
| 34:34.600 --> 34:38.320 | |
| the system for kind of how today's world is organized | |
| 34:38.320 --> 34:41.760 | |
| is one that really allows for huge impact | |
| 34:41.760 --> 34:45.400 | |
| and that kind of part of that is that you need to be, | |
| 34:45.400 --> 34:48.080 | |
| that for profits are self sustaining | |
| 34:48.080 --> 34:51.200 | |
| and able to kind of build on their own momentum. | |
| 34:51.200 --> 34:53.080 | |
| And I think that's a really powerful thing. | |
| 34:53.080 --> 34:55.880 | |
| It's something that when it turns out | |
| 34:55.880 --> 34:57.920 | |
| that we haven't set the guardrails correctly, | |
| 34:57.920 --> 34:58.840 | |
| causes problems, right? | |
| 34:58.840 --> 35:02.720 | |
| Think about logging companies that go into the rainforest, | |
| 35:02.720 --> 35:04.680 | |
| that's really bad, we don't want that. | |
| 35:04.680 --> 35:06.520 | |
| And it's actually really interesting to me | |
| 35:06.520 --> 35:08.480 | |
| that kind of this question of | |
| 35:08.480 --> 35:11.400 | |
| how do you get positive benefits out of a for profit company? | |
| 35:11.400 --> 35:12.600 | |
| It's actually very similar to | |
| 35:12.600 --> 35:15.800 | |
| how do you get positive benefits out of an AGI, right? | |
| 35:15.800 --> 35:18.000 | |
| That you have this like very powerful system, | |
| 35:18.000 --> 35:19.680 | |
| it's more powerful than any human | |
| 35:19.680 --> 35:21.760 | |
| and it's kind of autonomous in some ways. | |
| 35:21.760 --> 35:23.800 | |
| You know, it's super human in a lot of axes | |
| 35:23.800 --> 35:25.400 | |
| and somehow you have to set the guardrails | |
| 35:25.400 --> 35:26.800 | |
| to get good things to happen. | |
| 35:26.800 --> 35:29.360 | |
| But when you do, the benefits are massive. | |
| 35:29.360 --> 35:32.920 | |
| And so I think that when I think about nonprofit | |
| 35:32.920 --> 35:36.120 | |
| versus for profit, I think just not enough happens | |
| 35:36.120 --> 35:37.800 | |
| in nonprofits, they're very pure, | |
| 35:37.800 --> 35:39.200 | |
| but it's just kind of, you know, | |
| 35:39.200 --> 35:40.840 | |
| it's just hard to do things there. | |
| 35:40.840 --> 35:44.000 | |
| And for profits in some ways, like too much happens, | |
| 35:44.000 --> 35:46.440 | |
| but if kind of shaped in the right way, | |
| 35:46.440 --> 35:47.840 | |
| it can actually be very positive. | |
| 35:47.840 --> 35:52.160 | |
| And so with OpenILP, we're picking a road in between. | |
| 35:52.160 --> 35:54.880 | |
| Now, the thing that I think is really important to recognize | |
| 35:54.880 --> 35:57.160 | |
| is that the way that we think about OpenILP | |
| 35:57.160 --> 36:00.440 | |
| is that in the world where AGI actually happens, right? | |
| 36:00.440 --> 36:01.720 | |
| In a world where we are successful, | |
| 36:01.720 --> 36:03.800 | |
| we build the most transformative technology ever, | |
| 36:03.800 --> 36:06.600 | |
| the amount of value we're going to create will be astronomical. | |
| 36:07.600 --> 36:12.600 | |
| And so then in that case, that the cap that we have | |
| 36:12.760 --> 36:15.520 | |
| will be a small fraction of the value we create. | |
| 36:15.520 --> 36:17.800 | |
| And the amount of value that goes back to investors | |
| 36:17.800 --> 36:20.000 | |
| and employees looks pretty similar to what would happen | |
| 36:20.000 --> 36:21.680 | |
| in a pretty successful startup. | |
| 36:23.760 --> 36:26.520 | |
| And that's really the case that we're optimizing for, right? | |
| 36:26.520 --> 36:28.560 | |
| That we're thinking about in the success case, | |
| 36:28.560 --> 36:32.120 | |
| making sure that the value we create doesn't get locked up. | |
| 36:32.120 --> 36:34.920 | |
| And I expect that in other for profit companies | |
| 36:34.920 --> 36:37.800 | |
| that it's possible to do something like that. | |
| 36:37.800 --> 36:39.720 | |
| I think it's not obvious how to do it, right? | |
| 36:39.720 --> 36:41.440 | |
| And I think that as a for profit company, | |
| 36:41.440 --> 36:44.240 | |
| you have a lot of fiduciary duty to your shareholders | |
| 36:44.240 --> 36:45.640 | |
| and that there are certain decisions | |
| 36:45.640 --> 36:47.520 | |
| that you just cannot make. | |
| 36:47.520 --> 36:49.080 | |
| In our structure, we've set it up | |
| 36:49.080 --> 36:52.440 | |
| so that we have a fiduciary duty to the charter, | |
| 36:52.440 --> 36:54.400 | |
| that we always get to make the decision | |
| 36:54.400 --> 36:56.720 | |
| that is right for the charter, | |
| 36:56.720 --> 36:58.800 | |
| rather than even if it comes at the expense | |
| 36:58.800 --> 37:00.680 | |
| of our own stakeholders. | |
| 37:00.680 --> 37:03.400 | |
| And so I think that when I think about | |
| 37:03.400 --> 37:04.360 | |
| what's really important, | |
| 37:04.360 --> 37:06.280 | |
| it's not really about nonprofit versus for profit. | |
| 37:06.280 --> 37:09.600 | |
| It's really a question of if you build a GI | |
| 37:09.600 --> 37:10.600 | |
| and you kind of, you know, | |
| 37:10.600 --> 37:13.080 | |
| humanity is now at this new age, | |
| 37:13.080 --> 37:15.760 | |
| who benefits, whose lives are better? | |
| 37:15.760 --> 37:17.120 | |
| And I think that what's really important | |
| 37:17.120 --> 37:20.320 | |
| is to have an answer that is everyone. | |
| 37:20.320 --> 37:23.400 | |
| Yeah, which is one of the core aspects of the charter. | |
| 37:23.400 --> 37:26.520 | |
| So one concern people have, not just with OpenAI, | |
| 37:26.520 --> 37:28.400 | |
| but with Google, Facebook, Amazon, | |
| 37:28.400 --> 37:33.400 | |
| anybody really that's creating impact at scale | |
| 37:35.000 --> 37:37.680 | |
| is how do we avoid, as your charter says, | |
| 37:37.680 --> 37:40.080 | |
| avoid enabling the use of AI or AGI | |
| 37:40.080 --> 37:43.640 | |
| to unduly concentrate power? | |
| 37:43.640 --> 37:45.920 | |
| Why would not a company like OpenAI | |
| 37:45.920 --> 37:48.640 | |
| keep all the power of an AGI system to itself? | |
| 37:48.640 --> 37:49.520 | |
| The charter. | |
| 37:49.520 --> 37:50.360 | |
| The charter. | |
| 37:50.360 --> 37:51.960 | |
| So, you know, how does the charter | |
| 37:53.120 --> 37:57.240 | |
| actualize itself in day to day? | |
| 37:57.240 --> 38:00.480 | |
| So I think that first to zoom out, right, | |
| 38:00.480 --> 38:01.880 | |
| that the way that we structure the company | |
| 38:01.880 --> 38:04.560 | |
| is so that the power for sort of, you know, | |
| 38:04.560 --> 38:06.720 | |
| dictating the actions that OpenAI takes | |
| 38:06.720 --> 38:08.600 | |
| ultimately rests with the board, right? | |
| 38:08.600 --> 38:11.720 | |
| The board of the nonprofit and the board is set up | |
| 38:11.720 --> 38:13.480 | |
| in certain ways, with certain restrictions | |
| 38:13.480 --> 38:16.280 | |
| that you can read about in the OpenAI LP blog post. | |
| 38:16.280 --> 38:19.200 | |
| But effectively the board is the governing body | |
| 38:19.200 --> 38:21.200 | |
| for OpenAI LP. | |
| 38:21.200 --> 38:24.400 | |
| And the board has a duty to fulfill the mission | |
| 38:24.400 --> 38:26.360 | |
| of the nonprofit. | |
| 38:26.360 --> 38:28.800 | |
| And so that's kind of how we tie, | |
| 38:28.800 --> 38:30.960 | |
| how we thread all these things together. | |
| 38:30.960 --> 38:32.880 | |
| Now there's a question of so day to day, | |
| 38:32.880 --> 38:34.800 | |
| how do people, the individuals, | |
| 38:34.800 --> 38:36.960 | |
| who in some ways are the most empowered ones, right? | |
| 38:36.960 --> 38:38.800 | |
| You know, the board sort of gets to call the shots | |
| 38:38.800 --> 38:41.920 | |
| at the high level, but the people who are actually executing | |
| 38:41.920 --> 38:43.120 | |
| are the employees, right? | |
| 38:43.120 --> 38:45.480 | |
| The people here on a day to day basis who have the, | |
| 38:45.480 --> 38:47.720 | |
| you know, the keys to the technical kingdom. | |
| 38:48.960 --> 38:51.720 | |
| And there I think that the answer looks a lot like, | |
| 38:51.720 --> 38:55.120 | |
| well, how does any company's values get actualized, right? | |
| 38:55.120 --> 38:56.720 | |
| And I think that a lot of that comes down to | |
| 38:56.720 --> 38:58.160 | |
| that you need people who are here | |
| 38:58.160 --> 39:01.320 | |
| because they really believe in that mission | |
| 39:01.320 --> 39:02.800 | |
| and they believe in the charter | |
| 39:02.800 --> 39:05.440 | |
| and that they are willing to take actions | |
| 39:05.440 --> 39:08.600 | |
| that maybe are worse for them, but are better for the charter. | |
| 39:08.600 --> 39:11.440 | |
| And that's something that's really baked into the culture. | |
| 39:11.440 --> 39:13.200 | |
| And honestly, I think it's, you know, | |
| 39:13.200 --> 39:14.560 | |
| I think that that's one of the things | |
| 39:14.560 --> 39:18.200 | |
| that we really have to work to preserve as time goes on. | |
| 39:18.200 --> 39:20.760 | |
| And that's a really important part of how we think | |
| 39:20.760 --> 39:23.040 | |
| about hiring people and bringing people into OpenAI. | |
| 39:23.040 --> 39:25.320 | |
| So there's people here, there's people here | |
| 39:25.320 --> 39:30.320 | |
| who could speak up and say, like, hold on a second, | |
| 39:30.840 --> 39:34.600 | |
| this is totally against what we stand for, culture wise. | |
| 39:34.600 --> 39:35.440 | |
| Yeah, yeah, for sure. | |
| 39:35.440 --> 39:37.120 | |
| I mean, I think that we actually have, | |
| 39:37.120 --> 39:38.760 | |
| I think that's like a pretty important part | |
| 39:38.760 --> 39:41.920 | |
| of how we operate and how we have, | |
| 39:41.920 --> 39:44.160 | |
| even again with designing the charter | |
| 39:44.160 --> 39:46.680 | |
| and designing OpenAI in the first place, | |
| 39:46.680 --> 39:48.760 | |
| that there has been a lot of conversation | |
| 39:48.760 --> 39:50.480 | |
| with employees here and a lot of times | |
| 39:50.480 --> 39:52.400 | |
| where employees said, wait a second, | |
| 39:52.400 --> 39:53.920 | |
| this seems like it's going in the wrong direction | |
| 39:53.920 --> 39:55.120 | |
| and let's talk about it. | |
| 39:55.120 --> 39:57.360 | |
| And so I think one thing that's, I think are really, | |
| 39:57.360 --> 39:58.880 | |
| and you know, here's actually one thing | |
| 39:58.880 --> 40:02.080 | |
| that I think is very unique about us as a small company, | |
| 40:02.080 --> 40:04.360 | |
| is that if you're at a massive tech giant, | |
| 40:04.360 --> 40:05.680 | |
| that's a little bit hard for someone | |
| 40:05.680 --> 40:08.120 | |
| who's a line employee to go and talk to the CEO | |
| 40:08.120 --> 40:10.520 | |
| and say, I think that we're doing this wrong. | |
| 40:10.520 --> 40:13.040 | |
| And you know, you'll get companies like Google | |
| 40:13.040 --> 40:15.720 | |
| that have had some collective action from employees | |
| 40:15.720 --> 40:19.400 | |
| to make ethical change around things like Maven. | |
| 40:19.400 --> 40:20.680 | |
| And so maybe there are mechanisms | |
| 40:20.680 --> 40:22.240 | |
| that other companies that work, | |
| 40:22.240 --> 40:24.480 | |
| but here, super easy for anyone to pull me aside, | |
| 40:24.480 --> 40:26.320 | |
| to pull Sam aside, to pull Eli aside, | |
| 40:26.320 --> 40:27.800 | |
| and people do it all the time. | |
| 40:27.800 --> 40:29.800 | |
| One of the interesting things in the charter | |
| 40:29.800 --> 40:31.640 | |
| is this idea that it'd be great | |
| 40:31.640 --> 40:34.240 | |
| if you could try to describe or untangle | |
| 40:34.240 --> 40:36.440 | |
| switching from competition to collaboration | |
| 40:36.440 --> 40:38.920 | |
| and late stage AGI development. | |
| 40:38.920 --> 40:39.760 | |
| It's really interesting, | |
| 40:39.760 --> 40:42.160 | |
| this dance between competition and collaboration, | |
| 40:42.160 --> 40:43.400 | |
| how do you think about that? | |
| 40:43.400 --> 40:45.000 | |
| Yeah, assuming that you can actually do | |
| 40:45.000 --> 40:47.040 | |
| the technical side of AGI development, | |
| 40:47.040 --> 40:48.960 | |
| I think there's going to be two key problems | |
| 40:48.960 --> 40:50.400 | |
| with figuring out how do you actually deploy it | |
| 40:50.400 --> 40:51.520 | |
| and make it go well. | |
| 40:51.520 --> 40:53.160 | |
| The first one of these is the run up | |
| 40:53.160 --> 40:56.360 | |
| to building the first AGI. | |
| 40:56.360 --> 40:58.920 | |
| You look at how self driving cars are being developed, | |
| 40:58.920 --> 41:00.680 | |
| and it's a competitive race. | |
| 41:00.680 --> 41:02.560 | |
| And the thing that always happens in competitive race | |
| 41:02.560 --> 41:04.160 | |
| is that you have huge amounts of pressure | |
| 41:04.160 --> 41:05.600 | |
| to get rid of safety. | |
| 41:06.800 --> 41:08.920 | |
| And so that's one thing we're very concerned about, right? | |
| 41:08.920 --> 41:12.000 | |
| Is that people, multiple teams figuring out, | |
| 41:12.000 --> 41:13.600 | |
| we can actually get there, | |
| 41:13.600 --> 41:16.680 | |
| but you know, if we took the slower path | |
| 41:16.680 --> 41:20.240 | |
| that is more guaranteed to be safe, we will lose. | |
| 41:20.240 --> 41:22.360 | |
| And so we're going to take the fast path. | |
| 41:22.360 --> 41:25.480 | |
| And so the more that we can, both ourselves, | |
| 41:25.480 --> 41:27.280 | |
| be in a position where we don't generate | |
| 41:27.280 --> 41:29.000 | |
| that competitive race, where we say, | |
| 41:29.000 --> 41:31.520 | |
| if the race is being run and that someone else | |
| 41:31.520 --> 41:33.280 | |
| is further ahead than we are, | |
| 41:33.280 --> 41:35.600 | |
| we're not going to try to leapfrog. | |
| 41:35.600 --> 41:37.200 | |
| We're going to actually work with them, right? | |
| 41:37.200 --> 41:38.800 | |
| We will help them succeed. | |
| 41:38.800 --> 41:40.440 | |
| As long as what they're trying to do | |
| 41:40.440 --> 41:42.920 | |
| is to fulfill our mission, then we're good. | |
| 41:42.920 --> 41:44.800 | |
| We don't have to build AGI ourselves. | |
| 41:44.800 --> 41:47.080 | |
| And I think that's a really important commitment from us, | |
| 41:47.080 --> 41:49.080 | |
| but it can't just be unilateral, right? | |
| 41:49.080 --> 41:50.400 | |
| I think that it's really important | |
| 41:50.400 --> 41:53.120 | |
| that other players who are serious about building AGI | |
| 41:53.120 --> 41:54.680 | |
| make similar commitments, right? | |
| 41:54.680 --> 41:56.640 | |
| And I think that, you know, again, | |
| 41:56.640 --> 41:57.840 | |
| to the extent that everyone believes | |
| 41:57.840 --> 42:00.080 | |
| that AGI should be something to benefit everyone, | |
| 42:00.080 --> 42:01.240 | |
| then it actually really shouldn't matter | |
| 42:01.240 --> 42:02.440 | |
| which company builds it. | |
| 42:02.440 --> 42:04.160 | |
| And we should all be concerned about the case | |
| 42:04.160 --> 42:06.080 | |
| where we just race so hard to get there | |
| 42:06.080 --> 42:07.640 | |
| that something goes wrong. | |
| 42:07.640 --> 42:09.600 | |
| So what role do you think government, | |
| 42:10.560 --> 42:13.840 | |
| our favorite entity has in setting policy and rules | |
| 42:13.840 --> 42:18.320 | |
| about this domain, from research to the development | |
| 42:18.320 --> 42:22.880 | |
| to early stage, to late stage AI and AGI development? | |
| 42:22.880 --> 42:25.640 | |
| So I think that, first of all, | |
| 42:25.640 --> 42:28.080 | |
| it's really important that government's in there, right? | |
| 42:28.080 --> 42:29.800 | |
| In some way, shape, or form, you know, | |
| 42:29.800 --> 42:30.920 | |
| at the end of the day, we're talking about | |
| 42:30.920 --> 42:35.080 | |
| building technology that will shape how the world operates | |
| 42:35.080 --> 42:39.040 | |
| and that there needs to be government as part of that answer. | |
| 42:39.040 --> 42:42.160 | |
| And so that's why we've done a number | |
| 42:42.160 --> 42:43.600 | |
| of different congressional testimonies. | |
| 42:43.600 --> 42:46.440 | |
| We interact with a number of different lawmakers | |
| 42:46.440 --> 42:50.040 | |
| and that right now, a lot of our message to them | |
| 42:50.040 --> 42:54.360 | |
| is that it's not the time for regulation, | |
| 42:54.360 --> 42:56.400 | |
| it is the time for measurement, right? | |
| 42:56.400 --> 42:59.080 | |
| That our main policy recommendation is that people, | |
| 42:59.080 --> 43:00.680 | |
| and you know, the government does this all the time | |
| 43:00.680 --> 43:04.880 | |
| with bodies like NIST, spend time trying to figure out | |
| 43:04.880 --> 43:07.920 | |
| just where the technology is, how fast it's moving, | |
| 43:07.920 --> 43:11.200 | |
| and can really become literate and up to speed | |
| 43:11.200 --> 43:13.520 | |
| with respect to what to expect. | |
| 43:13.520 --> 43:15.240 | |
| So I think that today, the answer really | |
| 43:15.240 --> 43:17.320 | |
| is about measurement. | |
| 43:17.320 --> 43:20.160 | |
| And I think that there will be a time and place | |
| 43:20.160 --> 43:21.720 | |
| where that will change. | |
| 43:21.720 --> 43:24.840 | |
| And I think it's a little bit hard to predict exactly | |
| 43:24.840 --> 43:27.120 | |
| what exactly that trajectory should look like. | |
| 43:27.120 --> 43:31.080 | |
| So there will be a point at which regulation, | |
| 43:31.080 --> 43:34.200 | |
| federal in the United States, the government steps in | |
| 43:34.200 --> 43:39.200 | |
| and helps be the, I don't wanna say the adult in the room, | |
| 43:39.520 --> 43:42.400 | |
| to make sure that there is strict rules, | |
| 43:42.400 --> 43:45.200 | |
| maybe conservative rules that nobody can cross. | |
| 43:45.200 --> 43:47.400 | |
| Well, I think there's kind of maybe two angles to it. | |
| 43:47.400 --> 43:49.800 | |
| So today with narrow AI applications, | |
| 43:49.800 --> 43:51.960 | |
| that I think there are already existing bodies | |
| 43:51.960 --> 43:54.880 | |
| that are responsible and should be responsible for regulation. | |
| 43:54.880 --> 43:57.040 | |
| You think about, for example, with self driving cars, | |
| 43:57.040 --> 43:59.440 | |
| that you want the national highway. | |
| 44:00.720 --> 44:02.920 | |
| Yeah, exactly to be regulated in that. | |
| 44:02.920 --> 44:04.040 | |
| That makes sense, right? | |
| 44:04.040 --> 44:04.960 | |
| That basically what we're saying | |
| 44:04.960 --> 44:08.120 | |
| is that we're going to have these technological systems | |
| 44:08.120 --> 44:10.600 | |
| that are going to be performing applications | |
| 44:10.600 --> 44:12.280 | |
| that humans already do. | |
| 44:12.280 --> 44:14.800 | |
| Great, we already have ways of thinking about standards | |
| 44:14.800 --> 44:16.160 | |
| and safety for those. | |
| 44:16.160 --> 44:18.880 | |
| So I think actually empowering those regulators today | |
| 44:18.880 --> 44:20.040 | |
| is also pretty important. | |
| 44:20.040 --> 44:24.760 | |
| And then I think for AGI, that there's going to be a point | |
| 44:24.760 --> 44:26.040 | |
| where we'll have better answers. | |
| 44:26.040 --> 44:27.640 | |
| And I think that maybe a similar approach | |
| 44:27.640 --> 44:30.520 | |
| of first measurement and start thinking about | |
| 44:30.520 --> 44:31.640 | |
| what the rules should be. | |
| 44:31.640 --> 44:32.640 | |
| I think it's really important | |
| 44:32.640 --> 44:36.280 | |
| that we don't prematurely squash progress. | |
| 44:36.280 --> 44:40.160 | |
| I think it's very easy to kind of smother a budding field. | |
| 44:40.160 --> 44:42.160 | |
| And I think that's something to really avoid. | |
| 44:42.160 --> 44:43.760 | |
| But I don't think that the right way of doing it | |
| 44:43.760 --> 44:46.920 | |
| is to say, let's just try to blaze ahead | |
| 44:46.920 --> 44:50.280 | |
| and not involve all these other stakeholders. | |
| 44:51.480 --> 44:56.240 | |
| So you've recently released a paper on GPT2 | |
| 44:56.240 --> 45:01.240 | |
| language modeling, but did not release the full model | |
| 45:02.040 --> 45:05.280 | |
| because you had concerns about the possible negative effects | |
| 45:05.280 --> 45:07.480 | |
| of the availability of such model. | |
| 45:07.480 --> 45:10.680 | |
| It's outside of just that decision, | |
| 45:10.680 --> 45:14.360 | |
| and it's super interesting because of the discussion | |
| 45:14.360 --> 45:17.000 | |
| at a societal level, the discourse it creates. | |
| 45:17.000 --> 45:19.320 | |
| So it's fascinating in that aspect. | |
| 45:19.320 --> 45:22.880 | |
| But if you think that's the specifics here at first, | |
| 45:22.880 --> 45:25.920 | |
| what are some negative effects that you envisioned? | |
| 45:25.920 --> 45:28.600 | |
| And of course, what are some of the positive effects? | |
| 45:28.600 --> 45:30.640 | |
| Yeah, so again, I think to zoom out, | |
| 45:30.640 --> 45:34.040 | |
| like the way that we thought about GPT2 | |
| 45:34.040 --> 45:35.800 | |
| is that with language modeling, | |
| 45:35.800 --> 45:38.560 | |
| we are clearly on a trajectory right now | |
| 45:38.560 --> 45:40.880 | |
| where we scale up our models | |
| 45:40.880 --> 45:44.480 | |
| and we get qualitatively better performance, right? | |
| 45:44.480 --> 45:47.360 | |
| GPT2 itself was actually just a scale up | |
| 45:47.360 --> 45:50.680 | |
| of a model that we've released in the previous June, right? | |
| 45:50.680 --> 45:52.880 | |
| And we just ran it at much larger scale | |
| 45:52.880 --> 45:53.880 | |
| and we got these results | |
| 45:53.880 --> 45:57.240 | |
| where suddenly starting to write coherent pros, | |
| 45:57.240 --> 46:00.040 | |
| which was not something we'd seen previously. | |
| 46:00.040 --> 46:01.320 | |
| And what are we doing now? | |
| 46:01.320 --> 46:05.760 | |
| Well, we're gonna scale up GPT2 by 10x by 100x by 1000x | |
| 46:05.760 --> 46:07.840 | |
| and we don't know what we're gonna get. | |
| 46:07.840 --> 46:10.120 | |
| And so it's very clear that the model | |
| 46:10.120 --> 46:12.840 | |
| that we released last June, | |
| 46:12.840 --> 46:16.440 | |
| I think it's kind of like, it's a good academic toy. | |
| 46:16.440 --> 46:18.920 | |
| It's not something that we think is something | |
| 46:18.920 --> 46:20.440 | |
| that can really have negative applications | |
| 46:20.440 --> 46:21.680 | |
| or to the extent that it can, | |
| 46:21.680 --> 46:24.360 | |
| that the positive of people being able to play with it | |
| 46:24.360 --> 46:28.280 | |
| is far outweighs the possible harms. | |
| 46:28.280 --> 46:32.600 | |
| You fast forward to not GPT2, but GPT20, | |
| 46:32.600 --> 46:34.720 | |
| and you think about what that's gonna be like. | |
| 46:34.720 --> 46:38.200 | |
| And I think that the capabilities are going to be substantive. | |
| 46:38.200 --> 46:41.120 | |
| And so there needs to be a point in between the two | |
| 46:41.120 --> 46:43.480 | |
| where you say, this is something | |
| 46:43.480 --> 46:45.200 | |
| where we are drawing the line | |
| 46:45.200 --> 46:48.000 | |
| and that we need to start thinking about the safety aspects. | |
| 46:48.000 --> 46:50.160 | |
| And I think for GPT2, we could have gone either way. | |
| 46:50.160 --> 46:52.720 | |
| And in fact, when we had conversations internally | |
| 46:52.720 --> 46:54.760 | |
| that we had a bunch of pros and cons | |
| 46:54.760 --> 46:58.160 | |
| and it wasn't clear which one outweighed the other. | |
| 46:58.160 --> 46:59.840 | |
| And I think that when we announced | |
| 46:59.840 --> 47:02.160 | |
| that, hey, we decide not to release this model, | |
| 47:02.160 --> 47:03.600 | |
| then there was a bunch of conversation | |
| 47:03.600 --> 47:05.200 | |
| where various people said it's so obvious | |
| 47:05.200 --> 47:06.360 | |
| that you should have just released it. | |
| 47:06.360 --> 47:07.520 | |
| There are other people that said it's so obvious | |
| 47:07.520 --> 47:08.840 | |
| you should not have released it. | |
| 47:08.840 --> 47:10.960 | |
| And I think that that almost definitionally means | |
| 47:10.960 --> 47:13.800 | |
| that holding it back was the correct decision. | |
| 47:13.800 --> 47:17.000 | |
| If it's not obvious whether something is beneficial | |
| 47:17.000 --> 47:19.720 | |
| or not, you should probably default to caution. | |
| 47:19.720 --> 47:22.440 | |
| And so I think that the overall landscape | |
| 47:22.440 --> 47:23.760 | |
| for how we think about it | |
| 47:23.760 --> 47:25.920 | |
| is that this decision could have gone either way. | |
| 47:25.920 --> 47:27.960 | |
| There are great arguments in both directions. | |
| 47:27.960 --> 47:30.080 | |
| But for future models down the road, | |
| 47:30.080 --> 47:32.320 | |
| and possibly sooner than you'd expect, | |
| 47:32.320 --> 47:33.880 | |
| because scaling these things up doesn't actually | |
| 47:33.880 --> 47:36.800 | |
| take that long, those ones, | |
| 47:36.800 --> 47:39.600 | |
| you're definitely not going to want to release into the wild. | |
| 47:39.600 --> 47:42.640 | |
| And so I think that we almost view this as a test case | |
| 47:42.640 --> 47:45.360 | |
| and to see, can we even design, | |
| 47:45.360 --> 47:47.960 | |
| how do you have a society or how do you have a system | |
| 47:47.960 --> 47:50.520 | |
| that goes from having no concept of responsible disclosure | |
| 47:50.520 --> 47:53.440 | |
| where the mere idea of not releasing something | |
| 47:53.440 --> 47:55.960 | |
| for safety reasons is unfamiliar | |
| 47:55.960 --> 47:57.440 | |
| to a world where you say, okay, | |
| 47:57.440 --> 47:58.720 | |
| we have a powerful model. | |
| 47:58.720 --> 47:59.720 | |
| Let's at least think about it. | |
| 47:59.720 --> 48:01.280 | |
| Let's go through some process. | |
| 48:01.280 --> 48:02.680 | |
| And you think about the security community. | |
| 48:02.680 --> 48:03.880 | |
| It took them a long time | |
| 48:03.880 --> 48:05.960 | |
| to design responsible disclosure. | |
| 48:05.960 --> 48:07.200 | |
| You think about this question of, | |
| 48:07.200 --> 48:08.800 | |
| well, I have a security exploit. | |
| 48:08.800 --> 48:09.760 | |
| I send it to the company. | |
| 48:09.760 --> 48:12.000 | |
| The company is like, tries to prosecute me | |
| 48:12.000 --> 48:14.760 | |
| or just ignores it. | |
| 48:14.760 --> 48:16.080 | |
| What do I do? | |
| 48:16.080 --> 48:17.320 | |
| And so the alternatives of, | |
| 48:17.320 --> 48:19.120 | |
| oh, I just always publish your exploits. | |
| 48:19.120 --> 48:20.200 | |
| That doesn't seem good either. | |
| 48:20.200 --> 48:21.600 | |
| And so it really took a long time | |
| 48:21.600 --> 48:25.320 | |
| and it was bigger than any individual. | |
| 48:25.320 --> 48:27.080 | |
| It's really about building a whole community | |
| 48:27.080 --> 48:28.760 | |
| that believe that, okay, we'll have this process | |
| 48:28.760 --> 48:30.160 | |
| where you send it to the company | |
| 48:30.160 --> 48:31.680 | |
| if they don't act at a certain time, | |
| 48:31.680 --> 48:33.120 | |
| then you can go public | |
| 48:33.120 --> 48:34.440 | |
| and you're not a bad person. | |
| 48:34.440 --> 48:36.240 | |
| You've done the right thing. | |
| 48:36.240 --> 48:38.680 | |
| And I think that in AI, | |
| 48:38.680 --> 48:41.400 | |
| part of the response to GPT2 just proves | |
| 48:41.400 --> 48:44.200 | |
| that we don't have any concept of this. | |
| 48:44.200 --> 48:47.080 | |
| So that's the high level picture. | |
| 48:47.080 --> 48:48.720 | |
| And so I think that, | |
| 48:48.720 --> 48:51.240 | |
| I think this was a really important move to make. | |
| 48:51.240 --> 48:54.000 | |
| And we could have maybe delayed it for GPT3, | |
| 48:54.000 --> 48:56.080 | |
| but I'm really glad we did it for GPT2. | |
| 48:56.080 --> 48:57.760 | |
| And so now you look at GPT2 itself | |
| 48:57.760 --> 48:59.440 | |
| and you think about the substance of, okay, | |
| 48:59.440 --> 49:01.320 | |
| what are potential negative applications? | |
| 49:01.320 --> 49:04.120 | |
| So you have this model that's been trained on the internet, | |
| 49:04.120 --> 49:06.520 | |
| which is also going to be a bunch of very biased data, | |
| 49:06.520 --> 49:09.600 | |
| a bunch of very offensive content in there. | |
| 49:09.600 --> 49:13.240 | |
| And you can ask it to generate content for you | |
| 49:13.240 --> 49:14.600 | |
| on basically any topic, right? | |
| 49:14.600 --> 49:15.440 | |
| You just give it a prompt | |
| 49:15.440 --> 49:16.800 | |
| and it'll just start writing | |
| 49:16.800 --> 49:19.120 | |
| and it writes content like you see on the internet, | |
| 49:19.120 --> 49:21.960 | |
| you know, even down to like saying advertisement | |
| 49:21.960 --> 49:24.200 | |
| in the middle of some of its generations. | |
| 49:24.200 --> 49:26.200 | |
| And you think about the possibilities | |
| 49:26.200 --> 49:29.280 | |
| for generating fake news or abusive content. | |
| 49:29.280 --> 49:30.120 | |
| And, you know, it's interesting | |
| 49:30.120 --> 49:31.880 | |
| seeing what people have done with, you know, | |
| 49:31.880 --> 49:34.400 | |
| we released a smaller version of GPT2 | |
| 49:34.400 --> 49:37.480 | |
| and the people have done things like try to generate, | |
| 49:37.480 --> 49:40.760 | |
| you know, take my own Facebook message history | |
| 49:40.760 --> 49:43.360 | |
| and generate more Facebook messages like me | |
| 49:43.360 --> 49:47.360 | |
| and people generating fake politician content | |
| 49:47.360 --> 49:49.520 | |
| or, you know, there's a bunch of things there | |
| 49:49.520 --> 49:51.920 | |
| where you at least have to think, | |
| 49:51.920 --> 49:54.720 | |
| is this going to be good for the world? | |
| 49:54.720 --> 49:56.320 | |
| There's the flip side, which is I think | |
| 49:56.320 --> 49:57.840 | |
| that there's a lot of awesome applications | |
| 49:57.840 --> 50:01.640 | |
| that we really want to see like creative applications | |
| 50:01.640 --> 50:04.000 | |
| in terms of if you have sci fi authors | |
| 50:04.000 --> 50:06.760 | |
| that can work with this tool and come with cool ideas, | |
| 50:06.760 --> 50:09.720 | |
| like that seems awesome if we can write better sci fi | |
| 50:09.720 --> 50:11.360 | |
| through the use of these tools. | |
| 50:11.360 --> 50:13.080 | |
| And we've actually had a bunch of people right into us | |
| 50:13.080 --> 50:16.160 | |
| asking, hey, can we use it for, you know, | |
| 50:16.160 --> 50:18.360 | |
| a variety of different creative applications? | |
| 50:18.360 --> 50:21.880 | |
| So the positive are actually pretty easy to imagine. | |
| 50:21.880 --> 50:26.880 | |
| There are, you know, the usual NLP applications | |
| 50:26.880 --> 50:30.960 | |
| that are really interesting, but let's go there. | |
| 50:30.960 --> 50:32.960 | |
| It's kind of interesting to think about a world | |
| 50:32.960 --> 50:37.960 | |
| where, look at Twitter, where not just fake news | |
| 50:37.960 --> 50:42.960 | |
| but smarter and smarter bots being able to spread | |
| 50:43.040 --> 50:47.400 | |
| in an interesting complex networking way in information | |
| 50:47.400 --> 50:50.800 | |
| that just floods out us regular human beings | |
| 50:50.800 --> 50:52.880 | |
| with our original thoughts. | |
| 50:52.880 --> 50:57.880 | |
| So what are your views of this world with GPT 20? | |
| 50:58.760 --> 51:01.600 | |
| Right, how do we think about, again, | |
| 51:01.600 --> 51:03.560 | |
| it's like one of those things about in the 50s | |
| 51:03.560 --> 51:08.560 | |
| trying to describe the internet or the smartphone. | |
| 51:08.720 --> 51:09.960 | |
| What do you think about that world, | |
| 51:09.960 --> 51:11.400 | |
| the nature of information? | |
| 51:12.920 --> 51:16.760 | |
| One possibility is that we'll always try to design systems | |
| 51:16.760 --> 51:19.680 | |
| that identify a robot versus human | |
| 51:19.680 --> 51:21.280 | |
| and we'll do so successfully. | |
| 51:21.280 --> 51:24.600 | |
| And so we'll authenticate that we're still human. | |
| 51:24.600 --> 51:27.520 | |
| And the other world is that we just accept the fact | |
| 51:27.520 --> 51:30.360 | |
| that we're swimming in a sea of fake news | |
| 51:30.360 --> 51:32.120 | |
| and just learn to swim there. | |
| 51:32.120 --> 51:34.800 | |
| Well, have you ever seen the, there's a, you know, | |
| 51:34.800 --> 51:39.800 | |
| popular meme of a robot with a physical arm and pen | |
| 51:41.520 --> 51:43.440 | |
| clicking the I'm not a robot button? | |
| 51:43.440 --> 51:44.280 | |
| Yeah. | |
| 51:44.280 --> 51:48.560 | |
| I think the truth is that really trying to distinguish | |
| 51:48.560 --> 51:52.160 | |
| between robot and human is a losing battle. | |
| 51:52.160 --> 51:53.800 | |
| Ultimately, you think it's a losing battle? | |
| 51:53.800 --> 51:55.520 | |
| I think it's a losing battle ultimately, right? | |
| 51:55.520 --> 51:57.800 | |
| I think that that is that in terms of the content, | |
| 51:57.800 --> 51:59.360 | |
| in terms of the actions that you can take. | |
| 51:59.360 --> 52:01.200 | |
| I mean, think about how captures have gone, right? | |
| 52:01.200 --> 52:02.920 | |
| The captures used to be a very nice, simple. | |
| 52:02.920 --> 52:06.320 | |
| You just have this image, all of our OCR is terrible. | |
| 52:06.320 --> 52:08.880 | |
| You put a couple of artifacts in it, you know, | |
| 52:08.880 --> 52:11.040 | |
| humans are gonna be able to tell what it is | |
| 52:11.040 --> 52:13.840 | |
| an AI system wouldn't be able to today. | |
| 52:13.840 --> 52:15.720 | |
| Like I could barely do captures. | |
| 52:15.720 --> 52:18.360 | |
| And I think that this is just kind of where we're going. | |
| 52:18.360 --> 52:20.400 | |
| I think captures where we're a moment in time thing. | |
| 52:20.400 --> 52:22.520 | |
| And as AI systems become more powerful, | |
| 52:22.520 --> 52:25.520 | |
| that there being human capabilities that can be measured | |
| 52:25.520 --> 52:29.360 | |
| in a very easy automated way that the AIs will not be | |
| 52:29.360 --> 52:31.120 | |
| capable of, I think that's just like, | |
| 52:31.120 --> 52:34.160 | |
| it's just an increasingly hard technical battle. | |
| 52:34.160 --> 52:36.240 | |
| But it's not that all hope is lost, right? | |
| 52:36.240 --> 52:39.760 | |
| And you think about how do we already authenticate | |
| 52:39.760 --> 52:40.600 | |
| ourselves, right? | |
| 52:40.600 --> 52:41.760 | |
| That, you know, we have systems. | |
| 52:41.760 --> 52:43.440 | |
| We have social security numbers. | |
| 52:43.440 --> 52:46.560 | |
| If you're in the U S or, you know, you have, you have, | |
| 52:46.560 --> 52:48.920 | |
| you know, ways of identifying individual people | |
| 52:48.920 --> 52:51.880 | |
| and having real world identity tied to digital identity | |
| 52:51.880 --> 52:54.880 | |
| seems like a step towards, you know, | |
| 52:54.880 --> 52:56.200 | |
| authenticating the source of content | |
| 52:56.200 --> 52:58.240 | |
| rather than the content itself. | |
| 52:58.240 --> 53:00.000 | |
| Now, there are problems with that. | |
| 53:00.000 --> 53:03.000 | |
| How can you have privacy and anonymity in a world | |
| 53:03.000 --> 53:05.440 | |
| where the only content you can really trust is, | |
| 53:05.440 --> 53:06.560 | |
| or the only way you can trust content | |
| 53:06.560 --> 53:08.560 | |
| is by looking at where it comes from. | |
| 53:08.560 --> 53:11.400 | |
| And so I think that building out good reputation networks | |
| 53:11.400 --> 53:14.080 | |
| maybe one possible solution. | |
| 53:14.080 --> 53:16.280 | |
| But yeah, I think that this question is not | |
| 53:16.280 --> 53:17.720 | |
| an obvious one. | |
| 53:17.720 --> 53:19.320 | |
| And I think that we, you know, | |
| 53:19.320 --> 53:20.880 | |
| maybe sooner than we think we'll be in a world | |
| 53:20.880 --> 53:23.800 | |
| where, you know, today I often will read a tweet | |
| 53:23.800 --> 53:25.960 | |
| and be like, do I feel like a real human wrote this? | |
| 53:25.960 --> 53:27.560 | |
| Or, you know, do I feel like this was like genuine? | |
| 53:27.560 --> 53:30.160 | |
| I feel like I can kind of judge the content a little bit. | |
| 53:30.160 --> 53:32.640 | |
| And I think in the future, it just won't be the case. | |
| 53:32.640 --> 53:36.880 | |
| You look at, for example, the FCC comments on net neutrality. | |
| 53:36.880 --> 53:39.880 | |
| It came out later that millions of those were auto generated | |
| 53:39.880 --> 53:41.960 | |
| and that the researchers were able to do various | |
| 53:41.960 --> 53:44.040 | |
| statistical techniques to do that. | |
| 53:44.040 --> 53:47.160 | |
| What do you do in a world where those statistical techniques | |
| 53:47.160 --> 53:48.000 | |
| don't exist? | |
| 53:48.000 --> 53:49.120 | |
| It's just impossible to tell the difference | |
| 53:49.120 --> 53:50.640 | |
| between humans and AI's. | |
| 53:50.640 --> 53:53.960 | |
| And in fact, the most persuasive arguments | |
| 53:53.960 --> 53:57.200 | |
| are written by AI, all that stuff. | |
| 53:57.200 --> 53:58.600 | |
| It's not sci fi anymore. | |
| 53:58.600 --> 54:01.320 | |
| You look at GPT2 making a great argument for why recycling | |
| 54:01.320 --> 54:02.560 | |
| is bad for the world. | |
| 54:02.560 --> 54:04.440 | |
| You got to read that and be like, huh, you're right. | |
| 54:04.440 --> 54:06.520 | |
| We are addressing just the symptoms. | |
| 54:06.520 --> 54:08.120 | |
| Yeah, that's quite interesting. | |
| 54:08.120 --> 54:11.320 | |
| I mean, ultimately it boils down to the physical world | |
| 54:11.320 --> 54:13.680 | |
| being the last frontier of proving. | |
| 54:13.680 --> 54:16.080 | |
| So you said like basically networks of people, | |
| 54:16.080 --> 54:19.400 | |
| humans vouching for humans in the physical world. | |
| 54:19.400 --> 54:22.960 | |
| And somehow the authentication ends there. | |
| 54:22.960 --> 54:24.560 | |
| I mean, if I had to ask you, | |
| 54:25.520 --> 54:28.160 | |
| I mean, you're way too eloquent for a human. | |
| 54:28.160 --> 54:31.240 | |
| So if I had to ask you to authenticate, | |
| 54:31.240 --> 54:33.120 | |
| like prove how do I know you're not a robot | |
| 54:33.120 --> 54:34.920 | |
| and how do you know I'm not a robot? | |
| 54:34.920 --> 54:35.760 | |
| Yeah. | |
| 54:35.760 --> 54:40.520 | |
| I think that's so far were in this space, | |
| 54:40.520 --> 54:42.120 | |
| this conversation we just had, | |
| 54:42.120 --> 54:44.000 | |
| the physical movements we did | |
| 54:44.000 --> 54:47.040 | |
| is the biggest gap between us and AI systems | |
| 54:47.040 --> 54:49.360 | |
| is the physical manipulation. | |
| 54:49.360 --> 54:51.280 | |
| So maybe that's the last frontier. | |
| 54:51.280 --> 54:53.040 | |
| Well, here's another question is, | |
| 54:53.040 --> 54:57.320 | |
| why is solving this problem important, right? | |
| 54:57.320 --> 54:59.080 | |
| Like what aspects are really important to us? | |
| 54:59.080 --> 55:01.200 | |
| And I think that probably where we'll end up | |
| 55:01.200 --> 55:03.600 | |
| is we'll hone in on what do we really want | |
| 55:03.600 --> 55:06.400 | |
| out of knowing if we're talking to a human. | |
| 55:06.400 --> 55:09.480 | |
| And I think that again, this comes down to identity. | |
| 55:09.480 --> 55:11.760 | |
| And so I think that the internet of the future, | |
| 55:11.760 --> 55:14.840 | |
| I expect to be one that will have lots of agents out there | |
| 55:14.840 --> 55:16.320 | |
| that will interact with you. | |
| 55:16.320 --> 55:17.880 | |
| But I think that the question of, | |
| 55:17.880 --> 55:21.520 | |
| is this real flesh and blood human | |
| 55:21.520 --> 55:23.800 | |
| or is this an automated system? | |
| 55:23.800 --> 55:25.800 | |
| May actually just be less important. | |
| 55:25.800 --> 55:27.360 | |
| Let's actually go there. | |
| 55:27.360 --> 55:32.360 | |
| It's GPT2 is impressive and let's look at GPT20. | |
| 55:32.440 --> 55:37.440 | |
| Why is it so bad that all my friends are GPT20? | |
| 55:37.440 --> 55:42.440 | |
| Why is it so important on the internet? | |
| 55:43.320 --> 55:47.360 | |
| Do you think to interact with only human beings? | |
| 55:47.360 --> 55:50.640 | |
| Why can't we live in a world where ideas can come | |
| 55:50.640 --> 55:52.960 | |
| from models trained on human data? | |
| 55:52.960 --> 55:55.720 | |
| Yeah, I think this is actually a really interesting question. | |
| 55:55.720 --> 55:56.560 | |
| This comes back to the, | |
| 55:56.560 --> 55:59.560 | |
| how do you even picture a world with some new technology? | |
| 55:59.560 --> 56:02.080 | |
| And I think that one thing that I think is important | |
| 56:02.080 --> 56:04.760 | |
| is, you know, let's say honesty. | |
| 56:04.760 --> 56:07.520 | |
| And I think that if you have, you know, almost in the | |
| 56:07.520 --> 56:11.120 | |
| Turing test style sense of technology, | |
| 56:11.120 --> 56:13.200 | |
| you have AIs that are pretending to be humans | |
| 56:13.200 --> 56:15.800 | |
| and deceiving you, I think that is, you know, | |
| 56:15.800 --> 56:17.560 | |
| that feels like a bad thing, right? | |
| 56:17.560 --> 56:19.720 | |
| I think that it's really important that we feel like | |
| 56:19.720 --> 56:21.280 | |
| we're in control of our environment, right? | |
| 56:21.280 --> 56:23.400 | |
| That we understand who we're interacting with. | |
| 56:23.400 --> 56:25.880 | |
| And if it's an AI or a human, | |
| 56:25.880 --> 56:28.680 | |
| that that's not something that we're being deceived about. | |
| 56:28.680 --> 56:30.240 | |
| But I think that the flip side of, | |
| 56:30.240 --> 56:32.680 | |
| can I have as meaningful of an interaction with an AI | |
| 56:32.680 --> 56:34.240 | |
| as I can with a human? | |
| 56:34.240 --> 56:36.880 | |
| Well, I actually think here you can turn to sci fi. | |
| 56:36.880 --> 56:40.040 | |
| And her, I think is a great example of asking this very | |
| 56:40.040 --> 56:40.880 | |
| question, right? | |
| 56:40.880 --> 56:42.800 | |
| And one thing I really love about her is it really starts | |
| 56:42.800 --> 56:45.800 | |
| out almost by asking how meaningful are human | |
| 56:45.800 --> 56:47.280 | |
| virtual relationships, right? | |
| 56:47.280 --> 56:51.200 | |
| And then you have a human who has a relationship with an AI | |
| 56:51.200 --> 56:54.320 | |
| and that you really start to be drawn into that, right? | |
| 56:54.320 --> 56:56.960 | |
| And that all of your emotional buttons get triggered | |
| 56:56.960 --> 56:59.000 | |
| in the same way as if there was a real human that was on | |
| 56:59.000 --> 57:00.400 | |
| the other side of that phone. | |
| 57:00.400 --> 57:03.800 | |
| And so I think that this is one way of thinking about it, | |
| 57:03.800 --> 57:07.160 | |
| is that I think that we can have meaningful interactions | |
| 57:07.160 --> 57:09.720 | |
| and that if there's a funny joke, | |
| 57:09.720 --> 57:11.320 | |
| some sense it doesn't really matter if it was written | |
| 57:11.320 --> 57:14.600 | |
| by a human or an AI, but what you don't want in a way | |
| 57:14.600 --> 57:17.360 | |
| where I think we should really draw hard lines is deception. | |
| 57:17.360 --> 57:20.200 | |
| And I think that as long as we're in a world where, | |
| 57:20.200 --> 57:22.640 | |
| you know, why do we build AI systems at all, right? | |
| 57:22.640 --> 57:25.000 | |
| The reason we want to build them is to enhance human lives, | |
| 57:25.000 --> 57:26.680 | |
| to make humans be able to do more things, | |
| 57:26.680 --> 57:29.040 | |
| to have humans feel more fulfilled. | |
| 57:29.040 --> 57:32.040 | |
| And if we can build AI systems that do that, | |
| 57:32.040 --> 57:33.200 | |
| you know, sign me up. | |
| 57:33.200 --> 57:35.160 | |
| So the process of language modeling, | |
| 57:37.120 --> 57:38.760 | |
| how far do you think it take us? | |
| 57:38.760 --> 57:40.680 | |
| Let's look at movie HER. | |
| 57:40.680 --> 57:45.040 | |
| Do you think a dialogue, natural language conversation | |
| 57:45.040 --> 57:47.840 | |
| is formulated by the Turing test, for example, | |
| 57:47.840 --> 57:50.760 | |
| do you think that process could be achieved through | |
| 57:50.760 --> 57:53.160 | |
| this kind of unsupervised language modeling? | |
| 57:53.160 --> 57:56.960 | |
| So I think the Turing test in its real form | |
| 57:56.960 --> 57:58.680 | |
| isn't just about language, right? | |
| 57:58.680 --> 58:00.560 | |
| It's really about reasoning too, right? | |
| 58:00.560 --> 58:01.920 | |
| That to really pass the Turing test, | |
| 58:01.920 --> 58:03.880 | |
| I should be able to teach calculus | |
| 58:03.880 --> 58:05.520 | |
| to whoever's on the other side | |
| 58:05.520 --> 58:07.480 | |
| and have it really understand calculus | |
| 58:07.480 --> 58:09.320 | |
| and be able to, you know, go and solve | |
| 58:09.320 --> 58:11.280 | |
| new calculus problems. | |
| 58:11.280 --> 58:13.960 | |
| And so I think that to really solve the Turing test, | |
| 58:13.960 --> 58:16.440 | |
| we need more than what we're seeing with language models. | |
| 58:16.440 --> 58:18.720 | |
| We need some way of plugging in reasoning. | |
| 58:18.720 --> 58:22.400 | |
| Now, how different will that be from what we already do? | |
| 58:22.400 --> 58:23.880 | |
| That's an open question, right? | |
| 58:23.880 --> 58:25.480 | |
| It might be that we need some sequence | |
| 58:25.480 --> 58:27.200 | |
| of totally radical new ideas, | |
| 58:27.200 --> 58:29.560 | |
| or it might be that we just need to kind of shape | |
| 58:29.560 --> 58:31.920 | |
| our existing systems in a slightly different way. | |
| 58:33.040 --> 58:34.640 | |
| But I think that in terms of how far | |
| 58:34.640 --> 58:35.920 | |
| language modeling will go, | |
| 58:35.920 --> 58:37.520 | |
| it's already gone way further | |
| 58:37.520 --> 58:39.760 | |
| than many people would have expected, right? | |
| 58:39.760 --> 58:40.960 | |
| I think that things like, | |
| 58:40.960 --> 58:42.720 | |
| and I think there's a lot of really interesting angles | |
| 58:42.720 --> 58:45.920 | |
| to poke in terms of how much does GPT2 | |
| 58:45.920 --> 58:47.880 | |
| understand physical world? | |
| 58:47.880 --> 58:49.360 | |
| Like, you know, you read a little bit | |
| 58:49.360 --> 58:52.360 | |
| about fire underwater in GPT2. | |
| 58:52.360 --> 58:54.200 | |
| So it's like, okay, maybe it doesn't quite understand | |
| 58:54.200 --> 58:55.680 | |
| what these things are. | |
| 58:55.680 --> 58:58.560 | |
| But at the same time, I think that you also see | |
| 58:58.560 --> 59:00.640 | |
| various things like smoke coming from flame, | |
| 59:00.640 --> 59:02.680 | |
| and you know, a bunch of these things that GPT2, | |
| 59:02.680 --> 59:04.880 | |
| it has no body, it has no physical experience, | |
| 59:04.880 --> 59:07.280 | |
| it's just statically read data. | |
| 59:07.280 --> 59:11.680 | |
| And I think that if the answer is like, | |
| 59:11.680 --> 59:14.600 | |
| we don't know yet, and these questions though, | |
| 59:14.600 --> 59:16.240 | |
| we're starting to be able to actually ask them | |
| 59:16.240 --> 59:18.720 | |
| to physical systems, to real systems that exist, | |
| 59:18.720 --> 59:19.880 | |
| and that's very exciting. | |
| 59:19.880 --> 59:21.160 | |
| Do you think, what's your intuition? | |
| 59:21.160 --> 59:24.040 | |
| Do you think if you just scale language modeling, | |
| 59:24.040 --> 59:29.040 | |
| like significantly scale, that reasoning can emerge | |
| 59:29.320 --> 59:31.320 | |
| from the same exact mechanisms? | |
| 59:31.320 --> 59:34.960 | |
| I think it's unlikely that if we just scale GPT2, | |
| 59:34.960 --> 59:38.600 | |
| that we'll have reasoning in the full fledged way. | |
| 59:38.600 --> 59:39.760 | |
| And I think that there's like, | |
| 59:39.760 --> 59:41.520 | |
| the type signature is a little bit wrong, right? | |
| 59:41.520 --> 59:44.560 | |
| That like, there's something we do with, | |
| 59:44.560 --> 59:45.800 | |
| that we call thinking, right? | |
| 59:45.800 --> 59:47.640 | |
| Where we spend a lot of compute, | |
| 59:47.640 --> 59:49.160 | |
| like a variable amount of compute | |
| 59:49.160 --> 59:50.680 | |
| to get to better answers, right? | |
| 59:50.680 --> 59:53.040 | |
| I think a little bit harder, I get a better answer. | |
| 59:53.040 --> 59:55.160 | |
| And that that kind of type signature | |
| 59:55.160 --> 59:58.880 | |
| isn't quite encoded in a GPT, right? | |
| 59:58.880 --> 1:00:01.880 | |
| GPT will kind of like, it's spent a long time | |
| 1:00:01.880 --> 1:00:03.640 | |
| in it's like evolutionary history, | |
| 1:00:03.640 --> 1:00:04.680 | |
| baking and all this information, | |
| 1:00:04.680 --> 1:00:07.000 | |
| getting very, very good at this predictive process. | |
| 1:00:07.000 --> 1:00:10.320 | |
| And then at runtime, I just kind of do one forward pass | |
| 1:00:10.320 --> 1:00:13.240 | |
| and am able to generate stuff. | |
| 1:00:13.240 --> 1:00:15.560 | |
| And so, there might be small tweaks | |
| 1:00:15.560 --> 1:00:18.040 | |
| to what we do in order to get the type signature, right? | |
| 1:00:18.040 --> 1:00:21.040 | |
| For example, well, it's not really one forward pass, right? | |
| 1:00:21.040 --> 1:00:22.640 | |
| You generate symbol by symbol. | |
| 1:00:22.640 --> 1:00:25.560 | |
| And so, maybe you generate like a whole sequence of thoughts | |
| 1:00:25.560 --> 1:00:28.200 | |
| and you only keep like the last bit or something. | |
| 1:00:28.200 --> 1:00:29.840 | |
| But I think that at the very least, | |
| 1:00:29.840 --> 1:00:32.160 | |
| I would expect you have to make changes like that. | |
| 1:00:32.160 --> 1:00:35.520 | |
| Yeah, just exactly how we, you said think | |
| 1:00:35.520 --> 1:00:38.400 | |
| is the process of generating thought by thought | |
| 1:00:38.400 --> 1:00:40.360 | |
| in the same kind of way, like you said, | |
| 1:00:40.360 --> 1:00:43.640 | |
| keep the last bit, the thing that we converge towards. | |
| 1:00:45.000 --> 1:00:47.280 | |
| And I think there's another piece which is interesting, | |
| 1:00:47.280 --> 1:00:50.240 | |
| which is this out of distribution generalization, right? | |
| 1:00:50.240 --> 1:00:52.600 | |
| That like thinking somehow lets us do that, right? | |
| 1:00:52.600 --> 1:00:54.400 | |
| That we have an experience of thing | |
| 1:00:54.400 --> 1:00:56.080 | |
| and yet somehow we just kind of keep refining | |
| 1:00:56.080 --> 1:00:58.040 | |
| our mental model of it. | |
| 1:00:58.040 --> 1:01:01.160 | |
| This is again, something that feels tied to | |
| 1:01:01.160 --> 1:01:03.360 | |
| whatever reasoning is. | |
| 1:01:03.360 --> 1:01:05.720 | |
| And maybe it's a small tweak to what we do. | |
| 1:01:05.720 --> 1:01:08.080 | |
| Maybe it's many ideas and we'll take as many decades. | |
| 1:01:08.080 --> 1:01:11.920 | |
| Yeah, so the assumption there, generalization | |
| 1:01:11.920 --> 1:01:14.160 | |
| out of distribution is that it's possible | |
| 1:01:14.160 --> 1:01:16.880 | |
| to create new ideas. | |
| 1:01:18.160 --> 1:01:20.840 | |
| It's possible that nobody's ever created any new ideas. | |
| 1:01:20.840 --> 1:01:25.360 | |
| And then with scaling GPT2 to GPT20, | |
| 1:01:25.360 --> 1:01:30.360 | |
| you would essentially generalize to all possible thoughts | |
| 1:01:30.520 --> 1:01:34.200 | |
| as humans can have, just to play devil's advocate. | |
| 1:01:34.200 --> 1:01:37.280 | |
| Right, I mean, how many new story ideas | |
| 1:01:37.280 --> 1:01:39.120 | |
| have we come up with since Shakespeare, right? | |
| 1:01:39.120 --> 1:01:40.160 | |
| Yeah, exactly. | |
| 1:01:41.600 --> 1:01:44.680 | |
| It's just all different forms of love and drama and so on. | |
| 1:01:44.680 --> 1:01:45.800 | |
| Okay. | |
| 1:01:45.800 --> 1:01:47.520 | |
| Not sure if you read Biddle Lesson, | |
| 1:01:47.520 --> 1:01:49.400 | |
| a recent blog post by Rich Sutton. | |
| 1:01:49.400 --> 1:01:50.880 | |
| Yep, I have. | |
| 1:01:50.880 --> 1:01:53.720 | |
| He basically says something that echoes | |
| 1:01:53.720 --> 1:01:55.480 | |
| some of the ideas that you've been talking about, | |
| 1:01:55.480 --> 1:01:58.320 | |
| which is, he says the biggest lesson | |
| 1:01:58.320 --> 1:02:00.680 | |
| that can be read from 70 years of AI research | |
| 1:02:00.680 --> 1:02:03.880 | |
| is that general methods that leverage computation | |
| 1:02:03.880 --> 1:02:07.920 | |
| are ultimately going to ultimately win out. | |
| 1:02:07.920 --> 1:02:08.960 | |
| Do you agree with this? | |
| 1:02:08.960 --> 1:02:13.520 | |
| So basically open AI in general about the ideas | |
| 1:02:13.520 --> 1:02:15.880 | |
| you're exploring about coming up with methods, | |
| 1:02:15.880 --> 1:02:20.120 | |
| whether it's GPT2 modeling or whether it's open AI5, | |
| 1:02:20.120 --> 1:02:23.160 | |
| playing Dota, where a general method | |
| 1:02:23.160 --> 1:02:27.160 | |
| is better than a more fine tuned, expert tuned method. | |
| 1:02:29.760 --> 1:02:32.200 | |
| Yeah, so I think that, well, one thing that I think | |
| 1:02:32.200 --> 1:02:33.800 | |
| was really interesting about the reaction | |
| 1:02:33.800 --> 1:02:36.480 | |
| to that blog post was that a lot of people have read this | |
| 1:02:36.480 --> 1:02:39.440 | |
| as saying that compute is all that matters. | |
| 1:02:39.440 --> 1:02:41.360 | |
| And that's a very threatening idea, right? | |
| 1:02:41.360 --> 1:02:43.720 | |
| And I don't think it's a true idea either, right? | |
| 1:02:43.720 --> 1:02:45.800 | |
| It's very clear that we have algorithmic ideas | |
| 1:02:45.800 --> 1:02:47.920 | |
| that have been very important for making progress. | |
| 1:02:47.920 --> 1:02:50.720 | |
| And to really build AI, you wanna push as far as you can | |
| 1:02:50.720 --> 1:02:52.760 | |
| on the computational scale, and you wanna push | |
| 1:02:52.760 --> 1:02:55.520 | |
| as far as you can on human ingenuity. | |
| 1:02:55.520 --> 1:02:57.040 | |
| And so I think you need both. | |
| 1:02:57.040 --> 1:02:58.320 | |
| But I think the way that you phrase the question | |
| 1:02:58.320 --> 1:02:59.640 | |
| is actually very good, right? | |
| 1:02:59.640 --> 1:03:02.200 | |
| That it's really about what kind of ideas | |
| 1:03:02.200 --> 1:03:04.040 | |
| should we be striving for? | |
| 1:03:04.040 --> 1:03:07.600 | |
| And absolutely, if you can find a scalable idea, | |
| 1:03:07.600 --> 1:03:08.640 | |
| you pour more compute into it, | |
| 1:03:08.640 --> 1:03:11.400 | |
| you pour more data into it, it gets better. | |
| 1:03:11.400 --> 1:03:13.800 | |
| Like that's the real Holy Grail. | |
| 1:03:13.800 --> 1:03:16.600 | |
| And so I think that the answer to the question, | |
| 1:03:16.600 --> 1:03:19.920 | |
| I think is yes, that's really how we think about it. | |
| 1:03:19.920 --> 1:03:22.760 | |
| And that part of why we're excited about the power | |
| 1:03:22.760 --> 1:03:25.320 | |
| of deep learning and the potential for building AGI | |
| 1:03:25.320 --> 1:03:27.600 | |
| is because we look at the systems that exist | |
| 1:03:27.600 --> 1:03:29.720 | |
| in the most successful AI systems, | |
| 1:03:29.720 --> 1:03:32.680 | |
| and we realize that you scale those up, | |
| 1:03:32.680 --> 1:03:34.000 | |
| they're gonna work better. | |
| 1:03:34.000 --> 1:03:36.320 | |
| And I think that that scalability is something | |
| 1:03:36.320 --> 1:03:37.160 | |
| that really gives us hope | |
| 1:03:37.160 --> 1:03:39.600 | |
| for being able to build transformative systems. | |
| 1:03:39.600 --> 1:03:43.240 | |
| So I'll tell you, this is partially an emotional, | |
| 1:03:43.240 --> 1:03:45.760 | |
| you know, a thing that a response that people often have, | |
| 1:03:45.760 --> 1:03:49.280 | |
| if compute is so important for state of the art performance, | |
| 1:03:49.280 --> 1:03:50.760 | |
| you know, individual developers, | |
| 1:03:50.760 --> 1:03:52.960 | |
| maybe a 13 year old sitting somewhere in Kansas | |
| 1:03:52.960 --> 1:03:55.040 | |
| or something like that, you know, they're sitting, | |
| 1:03:55.040 --> 1:03:56.760 | |
| they might not even have a GPU | |
| 1:03:56.760 --> 1:04:00.080 | |
| and or may have a single GPU, a 1080 or something like that. | |
| 1:04:00.080 --> 1:04:02.640 | |
| And there's this feeling like, well, | |
| 1:04:02.640 --> 1:04:07.280 | |
| how can I possibly compete or contribute to this world of AI | |
| 1:04:07.280 --> 1:04:09.840 | |
| if scale is so important? | |
| 1:04:09.840 --> 1:04:11.920 | |
| So if you can comment on that, | |
| 1:04:11.920 --> 1:04:14.320 | |
| and in general, do you think we need to also | |
| 1:04:14.320 --> 1:04:18.800 | |
| in the future focus on democratizing compute resources | |
| 1:04:18.800 --> 1:04:22.680 | |
| more or as much as we democratize the algorithms? | |
| 1:04:22.680 --> 1:04:23.960 | |
| Well, so the way that I think about it | |
| 1:04:23.960 --> 1:04:28.880 | |
| is that there's this space of possible progress, right? | |
| 1:04:28.880 --> 1:04:30.920 | |
| There's a space of ideas and sort of systems | |
| 1:04:30.920 --> 1:04:32.960 | |
| that will work, that will move us forward. | |
| 1:04:32.960 --> 1:04:34.840 | |
| And there's a portion of that space, | |
| 1:04:34.840 --> 1:04:35.760 | |
| and to some extent, | |
| 1:04:35.760 --> 1:04:37.960 | |
| an increasingly significant portion of that space | |
| 1:04:37.960 --> 1:04:41.080 | |
| that does just require massive compute resources. | |
| 1:04:41.080 --> 1:04:44.760 | |
| And for that, I think that the answer is kind of clear | |
| 1:04:44.760 --> 1:04:47.960 | |
| and that part of why we have the structure that we do | |
| 1:04:47.960 --> 1:04:49.640 | |
| is because we think it's really important | |
| 1:04:49.640 --> 1:04:50.600 | |
| to be pushing the scale | |
| 1:04:50.600 --> 1:04:53.840 | |
| and to be building these large clusters and systems. | |
| 1:04:53.840 --> 1:04:55.920 | |
| But there's another portion of the space | |
| 1:04:55.920 --> 1:04:57.880 | |
| that isn't about the large scale compute, | |
| 1:04:57.880 --> 1:04:59.960 | |
| that are these ideas that, and again, | |
| 1:04:59.960 --> 1:05:02.200 | |
| I think that for the ideas to really be impactful | |
| 1:05:02.200 --> 1:05:04.200 | |
| and really shine, that they should be ideas | |
| 1:05:04.200 --> 1:05:05.840 | |
| that if you scale them up, | |
| 1:05:05.840 --> 1:05:08.840 | |
| would work way better than they do at small scale. | |
| 1:05:08.840 --> 1:05:11.160 | |
| But you can discover them without massive | |
| 1:05:11.160 --> 1:05:12.760 | |
| computational resources. | |
| 1:05:12.760 --> 1:05:15.200 | |
| And if you look at the history of recent developments, | |
| 1:05:15.200 --> 1:05:17.680 | |
| you think about things like the GAN or the VAE, | |
| 1:05:17.680 --> 1:05:20.920 | |
| that these are ones that I think you could come up with them | |
| 1:05:20.920 --> 1:05:22.720 | |
| without having, and in practice, | |
| 1:05:22.720 --> 1:05:24.520 | |
| people did come up with them without having | |
| 1:05:24.520 --> 1:05:26.560 | |
| massive, massive computational resources. | |
| 1:05:26.560 --> 1:05:28.000 | |
| Right, I just talked to Ian Goodfellow, | |
| 1:05:28.000 --> 1:05:31.600 | |
| but the thing is the initial GAN | |
| 1:05:31.600 --> 1:05:34.200 | |
| produced pretty terrible results, right? | |
| 1:05:34.200 --> 1:05:36.880 | |
| So only because it was in a very specific, | |
| 1:05:36.880 --> 1:05:38.640 | |
| only because they're smart enough to know | |
| 1:05:38.640 --> 1:05:41.520 | |
| that this is quite surprising to generate anything | |
| 1:05:41.520 --> 1:05:43.160 | |
| that they know. | |
| 1:05:43.160 --> 1:05:46.040 | |
| Do you see a world, or is that too optimistic and dreamer, | |
| 1:05:46.040 --> 1:05:49.760 | |
| like, to imagine that the compute resources | |
| 1:05:49.760 --> 1:05:52.200 | |
| are something that's owned by governments | |
| 1:05:52.200 --> 1:05:55.040 | |
| and provided as a utility? | |
| 1:05:55.040 --> 1:05:57.120 | |
| Actually, to some extent, this question reminds me | |
| 1:05:57.120 --> 1:06:00.280 | |
| of a blog post from one of my former professors | |
| 1:06:00.280 --> 1:06:02.440 | |
| at Harvard, this guy, Matt Welch, | |
| 1:06:02.440 --> 1:06:03.760 | |
| who was a systems professor. | |
| 1:06:03.760 --> 1:06:05.280 | |
| I remember sitting in his tenure talk, right, | |
| 1:06:05.280 --> 1:06:08.800 | |
| and that he had literally just gotten tenure. | |
| 1:06:08.800 --> 1:06:10.960 | |
| He went to Google for the summer, | |
| 1:06:10.960 --> 1:06:15.680 | |
| and then decided he wasn't going back to academia, right? | |
| 1:06:15.680 --> 1:06:17.760 | |
| And kind of in his blog post, he makes this point | |
| 1:06:17.760 --> 1:06:20.800 | |
| that, look, as a systems researcher, | |
| 1:06:20.800 --> 1:06:23.040 | |
| that I come up with these cool system ideas, | |
| 1:06:23.040 --> 1:06:25.080 | |
| right, and kind of build a little proof of concept, | |
| 1:06:25.080 --> 1:06:27.080 | |
| and the best thing I could hope for | |
| 1:06:27.080 --> 1:06:30.120 | |
| is that the people at Google or Yahoo, | |
| 1:06:30.120 --> 1:06:32.600 | |
| which was around at the time, | |
| 1:06:32.600 --> 1:06:35.400 | |
| will implement it and actually make it work at scale, right? | |
| 1:06:35.400 --> 1:06:36.640 | |
| That's like the dream for me, right? | |
| 1:06:36.640 --> 1:06:38.000 | |
| I build the little thing, and they turn it into | |
| 1:06:38.000 --> 1:06:40.000 | |
| the big thing that's actually working. | |
| 1:06:40.000 --> 1:06:43.360 | |
| And for him, he said, I'm done with that. | |
| 1:06:43.360 --> 1:06:45.320 | |
| I want to be the person who's actually doing | |
| 1:06:45.320 --> 1:06:47.200 | |
| building and deploying. | |
| 1:06:47.200 --> 1:06:49.560 | |
| And I think that there's a similar dichotomy here, right? | |
| 1:06:49.560 --> 1:06:52.400 | |
| I think that there are people who really actually | |
| 1:06:52.400 --> 1:06:55.240 | |
| find value, and I think it is a valuable thing to do, | |
| 1:06:55.240 --> 1:06:57.440 | |
| to be the person who produces those ideas, right, | |
| 1:06:57.440 --> 1:06:58.840 | |
| who builds the proof of concept. | |
| 1:06:58.840 --> 1:07:00.600 | |
| And yeah, you don't get to generate | |
| 1:07:00.600 --> 1:07:02.760 | |
| the coolest possible GAN images, | |
| 1:07:02.760 --> 1:07:04.480 | |
| but you invented the GAN, right? | |
| 1:07:04.480 --> 1:07:07.560 | |
| And so there's a real trade off there. | |
| 1:07:07.560 --> 1:07:09.040 | |
| And I think that that's a very personal choice, | |
| 1:07:09.040 --> 1:07:10.840 | |
| but I think there's value in both sides. | |
| 1:07:10.840 --> 1:07:14.600 | |
| So do you think creating AGI, something, | |
| 1:07:14.600 --> 1:07:19.600 | |
| or some new models, we would see echoes of the brilliance | |
| 1:07:20.440 --> 1:07:22.240 | |
| even at the prototype level. | |
| 1:07:22.240 --> 1:07:24.080 | |
| So you would be able to develop those ideas | |
| 1:07:24.080 --> 1:07:27.240 | |
| without scale, the initial seeds. | |
| 1:07:27.240 --> 1:07:30.680 | |
| So take a look at, I always like to look at examples | |
| 1:07:30.680 --> 1:07:32.680 | |
| that exist, right, look at real precedent. | |
| 1:07:32.680 --> 1:07:36.240 | |
| And so take a look at the June 2018 model | |
| 1:07:36.240 --> 1:07:39.200 | |
| that we released that we scaled up to turn to GPT2. | |
| 1:07:39.200 --> 1:07:41.280 | |
| And you can see that at small scale, | |
| 1:07:41.280 --> 1:07:42.800 | |
| it set some records, right? | |
| 1:07:42.800 --> 1:07:44.800 | |
| This was the original GPT. | |
| 1:07:44.800 --> 1:07:46.840 | |
| We actually had some cool generations. | |
| 1:07:46.840 --> 1:07:49.840 | |
| They weren't nearly as amazing and really stunning | |
| 1:07:49.840 --> 1:07:52.000 | |
| as the GPT2 ones, but it was promising. | |
| 1:07:52.000 --> 1:07:53.040 | |
| It was interesting. | |
| 1:07:53.040 --> 1:07:55.280 | |
| And so I think it is the case that with a lot | |
| 1:07:55.280 --> 1:07:58.280 | |
| of these ideas that you see promise at small scale, | |
| 1:07:58.280 --> 1:08:00.800 | |
| but there isn't an asterisk here, a very big asterisk, | |
| 1:08:00.800 --> 1:08:05.240 | |
| which is sometimes we see behaviors that emerge | |
| 1:08:05.240 --> 1:08:07.280 | |
| that are qualitatively different | |
| 1:08:07.280 --> 1:08:09.080 | |
| from anything we saw at small scale. | |
| 1:08:09.080 --> 1:08:12.600 | |
| And that the original inventor of whatever algorithm | |
| 1:08:12.600 --> 1:08:15.520 | |
| looks at and says, I didn't think it could do that. | |
| 1:08:15.520 --> 1:08:17.400 | |
| This is what we saw in Dota, right? | |
| 1:08:17.400 --> 1:08:19.320 | |
| So PPO was created by John Shulman, | |
| 1:08:19.320 --> 1:08:20.560 | |
| who's a researcher here. | |
| 1:08:20.560 --> 1:08:24.680 | |
| And with Dota, we basically just ran PPO | |
| 1:08:24.680 --> 1:08:26.520 | |
| at massive, massive scale. | |
| 1:08:26.520 --> 1:08:29.120 | |
| And there's some tweaks in order to make it work, | |
| 1:08:29.120 --> 1:08:31.520 | |
| but fundamentally it's PPO at the core. | |
| 1:08:31.520 --> 1:08:35.280 | |
| And we were able to get this longterm planning, | |
| 1:08:35.280 --> 1:08:38.680 | |
| these behaviors to really play out on a time scale | |
| 1:08:38.680 --> 1:08:40.760 | |
| that we just thought was not possible. | |
| 1:08:40.760 --> 1:08:42.680 | |
| And John looked at that and was like, | |
| 1:08:42.680 --> 1:08:44.240 | |
| I didn't think it could do that. | |
| 1:08:44.240 --> 1:08:45.480 | |
| That's what happens when you're at three orders | |
| 1:08:45.480 --> 1:08:48.400 | |
| of magnitude more scale than you tested at. | |
| 1:08:48.400 --> 1:08:50.600 | |
| Yeah, but it still has the same flavors of, | |
| 1:08:50.600 --> 1:08:55.600 | |
| you know, at least echoes of the expected billions. | |
| 1:08:56.000 --> 1:08:57.880 | |
| Although I suspect with GPT, | |
| 1:08:57.880 --> 1:09:01.800 | |
| it's scaled more and more, you might get surprising things. | |
| 1:09:01.800 --> 1:09:03.200 | |
| So yeah, you're right. | |
| 1:09:03.200 --> 1:09:06.360 | |
| It's interesting that it's difficult to see | |
| 1:09:06.360 --> 1:09:09.320 | |
| how far an idea will go when it's scaled. | |
| 1:09:09.320 --> 1:09:11.080 | |
| It's an open question. | |
| 1:09:11.080 --> 1:09:13.080 | |
| Well, so to that point with Dota and PPO, | |
| 1:09:13.080 --> 1:09:15.040 | |
| like I mean, here's a very concrete one, right? | |
| 1:09:15.040 --> 1:09:16.680 | |
| It's like, it's actually one thing | |
| 1:09:16.680 --> 1:09:17.720 | |
| that's very surprising about Dota | |
| 1:09:17.720 --> 1:09:20.400 | |
| that I think people don't really pay that much attention to. | |
| 1:09:20.400 --> 1:09:22.360 | |
| Is the decree of generalization | |
| 1:09:22.360 --> 1:09:24.560 | |
| out of distribution that happens, right? | |
| 1:09:24.560 --> 1:09:26.320 | |
| That you have this AI that's trained | |
| 1:09:26.320 --> 1:09:28.880 | |
| against other bots for its entirety, | |
| 1:09:28.880 --> 1:09:30.360 | |
| the entirety of its existence. | |
| 1:09:30.360 --> 1:09:31.440 | |
| Sorry to take a step back. | |
| 1:09:31.440 --> 1:09:36.440 | |
| Can you talk through, you know, a story of Dota, | |
| 1:09:37.240 --> 1:09:42.040 | |
| a story of leading up to opening I5 and that past, | |
| 1:09:42.040 --> 1:09:43.920 | |
| and what was the process of self playing | |
| 1:09:43.920 --> 1:09:45.440 | |
| and so on of training on this? | |
| 1:09:45.440 --> 1:09:46.280 | |
| Yeah, yeah, yeah. | |
| 1:09:46.280 --> 1:09:47.120 | |
| So with Dota. | |
| 1:09:47.120 --> 1:09:47.960 | |
| What is Dota? | |
| 1:09:47.960 --> 1:09:50.000 | |
| Dota is a complex video game | |
| 1:09:50.000 --> 1:09:51.320 | |
| and we started training, | |
| 1:09:51.320 --> 1:09:52.720 | |
| we started trying to solve Dota | |
| 1:09:52.720 --> 1:09:55.680 | |
| because we felt like this was a step towards the real world | |
| 1:09:55.680 --> 1:09:58.040 | |
| relative to other games like Chess or Go, right? | |
| 1:09:58.040 --> 1:09:59.160 | |
| Those very cerebral games | |
| 1:09:59.160 --> 1:10:00.480 | |
| where you just kind of have this board | |
| 1:10:00.480 --> 1:10:01.880 | |
| of very discreet moves. | |
| 1:10:01.880 --> 1:10:04.040 | |
| Dota starts to be much more continuous time. | |
| 1:10:04.040 --> 1:10:06.200 | |
| So you have this huge variety of different actions | |
| 1:10:06.200 --> 1:10:07.680 | |
| that you have a 45 minute game | |
| 1:10:07.680 --> 1:10:09.360 | |
| with all these different units | |
| 1:10:09.360 --> 1:10:11.840 | |
| and it's got a lot of messiness to it | |
| 1:10:11.840 --> 1:10:14.480 | |
| that really hasn't been captured by previous games. | |
| 1:10:14.480 --> 1:10:17.320 | |
| And famously all of the hard coded bots for Dota | |
| 1:10:17.320 --> 1:10:18.400 | |
| were terrible, right? | |
| 1:10:18.400 --> 1:10:19.920 | |
| It's just impossible to write anything good for it | |
| 1:10:19.920 --> 1:10:21.240 | |
| because it's so complex. | |
| 1:10:21.240 --> 1:10:23.280 | |
| And so this seemed like a really good place | |
| 1:10:23.280 --> 1:10:25.240 | |
| to push what's the state of the art | |
| 1:10:25.240 --> 1:10:26.800 | |
| in reinforcement learning. | |
| 1:10:26.800 --> 1:10:29.000 | |
| And so we started by focusing on the one versus one | |
| 1:10:29.000 --> 1:10:32.360 | |
| version of the game and we're able to solve that. | |
| 1:10:32.360 --> 1:10:33.880 | |
| We're able to beat the world champions | |
| 1:10:33.880 --> 1:10:37.240 | |
| and the learning, the skill curve | |
| 1:10:37.240 --> 1:10:38.960 | |
| was this crazy exponential, right? | |
| 1:10:38.960 --> 1:10:41.000 | |
| It was like constantly we were just scaling up, | |
| 1:10:41.000 --> 1:10:43.240 | |
| that we were fixing bugs and that you look | |
| 1:10:43.240 --> 1:10:46.600 | |
| at the skill curve and it was really a very, very smooth one. | |
| 1:10:46.600 --> 1:10:47.440 | |
| So it's actually really interesting | |
| 1:10:47.440 --> 1:10:50.000 | |
| to see how that like human iteration loop | |
| 1:10:50.000 --> 1:10:52.680 | |
| yielded very steady exponential progress. | |
| 1:10:52.680 --> 1:10:55.160 | |
| And to one side note, first of all, | |
| 1:10:55.160 --> 1:10:57.080 | |
| it's an exceptionally popular video game. | |
| 1:10:57.080 --> 1:10:59.400 | |
| The side effect is that there's a lot | |
| 1:10:59.400 --> 1:11:01.920 | |
| of incredible human experts at that video game. | |
| 1:11:01.920 --> 1:11:05.200 | |
| So the benchmark that you're trying to reach is very high. | |
| 1:11:05.200 --> 1:11:07.840 | |
| And the other, can you talk about the approach | |
| 1:11:07.840 --> 1:11:10.600 | |
| that was used initially and throughout training | |
| 1:11:10.600 --> 1:11:12.040 | |
| these agents to play this game? | |
| 1:11:12.040 --> 1:11:12.880 | |
| Yep. | |
| 1:11:12.880 --> 1:11:14.400 | |
| And so the approach that we used is self play. | |
| 1:11:14.400 --> 1:11:17.320 | |
| And so you have two agents that don't know anything. | |
| 1:11:17.320 --> 1:11:18.640 | |
| They battle each other, | |
| 1:11:18.640 --> 1:11:20.760 | |
| they discover something a little bit good | |
| 1:11:20.760 --> 1:11:22.000 | |
| and now they both know it. | |
| 1:11:22.000 --> 1:11:24.520 | |
| And they just get better and better and better without bound. | |
| 1:11:24.520 --> 1:11:27.040 | |
| And that's a really powerful idea, right? | |
| 1:11:27.040 --> 1:11:30.160 | |
| That we then went from the one versus one version | |
| 1:11:30.160 --> 1:11:32.400 | |
| of the game and scaled up to five versus five, right? | |
| 1:11:32.400 --> 1:11:34.280 | |
| So you think about kind of like with basketball | |
| 1:11:34.280 --> 1:11:35.440 | |
| where you have this like team sport | |
| 1:11:35.440 --> 1:11:37.640 | |
| and you need to do all this coordination | |
| 1:11:37.640 --> 1:11:40.920 | |
| and we were able to push the same idea, | |
| 1:11:40.920 --> 1:11:45.920 | |
| the same self play to really get to the professional level | |
| 1:11:45.920 --> 1:11:48.880 | |
| at the full five versus five version of the game. | |
| 1:11:48.880 --> 1:11:52.400 | |
| And the things that I think are really interesting here | |
| 1:11:52.400 --> 1:11:54.760 | |
| is that these agents in some ways | |
| 1:11:54.760 --> 1:11:56.760 | |
| they're almost like an insect like intelligence, right? | |
| 1:11:56.760 --> 1:11:59.920 | |
| Where they have a lot in common with how an insect is trained, | |
| 1:11:59.920 --> 1:12:00.760 | |
| right? | |
| 1:12:00.760 --> 1:12:02.640 | |
| An insect kind of lives in this environment for a very long time | |
| 1:12:02.640 --> 1:12:05.280 | |
| or the ancestors of this insect have been around | |
| 1:12:05.280 --> 1:12:07.000 | |
| for a long time and had a lot of experience. | |
| 1:12:07.000 --> 1:12:09.680 | |
| I think it's baked into this agent. | |
| 1:12:09.680 --> 1:12:12.720 | |
| And it's not really smart in the sense of a human, right? | |
| 1:12:12.720 --> 1:12:14.560 | |
| It's not able to go and learn calculus, | |
| 1:12:14.560 --> 1:12:17.000 | |
| but it's able to navigate its environment extremely well. | |
| 1:12:17.000 --> 1:12:18.480 | |
| And it's able to handle unexpected things | |
| 1:12:18.480 --> 1:12:22.080 | |
| in the environment that's never seen before, pretty well. | |
| 1:12:22.080 --> 1:12:24.800 | |
| And we see the same sort of thing with our Dota bots, right? | |
| 1:12:24.800 --> 1:12:26.720 | |
| That they're able to, within this game, | |
| 1:12:26.720 --> 1:12:28.440 | |
| they're able to play against humans, | |
| 1:12:28.440 --> 1:12:30.000 | |
| which is something that never existed | |
| 1:12:30.000 --> 1:12:31.360 | |
| in its evolutionary environment. | |
| 1:12:31.360 --> 1:12:34.400 | |
| Totally different play styles from humans versus the bots. | |
| 1:12:34.400 --> 1:12:37.200 | |
| And yet it's able to handle it extremely well. | |
| 1:12:37.200 --> 1:12:40.400 | |
| And that's something that I think was very surprising to us | |
| 1:12:40.400 --> 1:12:43.440 | |
| was something that doesn't really emerge | |
| 1:12:43.440 --> 1:12:47.200 | |
| from what we've seen with PPO at smaller scale, right? | |
| 1:12:47.200 --> 1:12:48.560 | |
| And the kind of scale we're running this stuff at | |
| 1:12:48.560 --> 1:12:51.920 | |
| was I could take 100,000 CPU cores, | |
| 1:12:51.920 --> 1:12:54.040 | |
| running with like hundreds of GPUs. | |
| 1:12:54.040 --> 1:12:59.040 | |
| It was probably about something like hundreds of years | |
| 1:12:59.040 --> 1:13:03.800 | |
| of experience going into this bot every single real day. | |
| 1:13:03.800 --> 1:13:06.200 | |
| And so that scale is massive. | |
| 1:13:06.200 --> 1:13:08.400 | |
| And we start to see very different kinds of behaviors | |
| 1:13:08.400 --> 1:13:10.760 | |
| out of the algorithms that we all know and love. | |
| 1:13:10.760 --> 1:13:15.160 | |
| Dota, you mentioned, beat the world expert 1v1. | |
| 1:13:15.160 --> 1:13:21.160 | |
| And then you weren't able to win 5v5 this year | |
| 1:13:21.160 --> 1:13:24.080 | |
| at the best players in the world. | |
| 1:13:24.080 --> 1:13:26.640 | |
| So what's the comeback story? | |
| 1:13:26.640 --> 1:13:27.680 | |
| First of all, talk through that. | |
| 1:13:27.680 --> 1:13:29.480 | |
| That was an exceptionally exciting event. | |
| 1:13:29.480 --> 1:13:33.160 | |
| And what's the following months in this year look like? | |
| 1:13:33.160 --> 1:13:33.760 | |
| Yeah, yeah. | |
| 1:13:33.760 --> 1:13:38.640 | |
| So one thing that's interesting is that we lose all the time. | |
| 1:13:38.640 --> 1:13:40.040 | |
| Because we play here. | |
| 1:13:40.040 --> 1:13:42.840 | |
| So the Dota team at OpenAI, we play the bot | |
| 1:13:42.840 --> 1:13:45.800 | |
| against better players than our system all the time. | |
| 1:13:45.800 --> 1:13:47.400 | |
| Or at least we used to, right? | |
| 1:13:47.400 --> 1:13:50.680 | |
| Like the first time we lost publicly was we went up | |
| 1:13:50.680 --> 1:13:53.480 | |
| on stage at the international and we played against some | |
| 1:13:53.480 --> 1:13:54.800 | |
| of the best teams in the world. | |
| 1:13:54.800 --> 1:13:56.320 | |
| And we ended up losing both games. | |
| 1:13:56.320 --> 1:13:58.520 | |
| But we give them a run for their money, right? | |
| 1:13:58.520 --> 1:14:01.440 | |
| That both games were kind of 30 minutes, 25 minutes. | |
| 1:14:01.440 --> 1:14:04.200 | |
| And they went back and forth, back and forth, back and forth. | |
| 1:14:04.200 --> 1:14:06.360 | |
| And so I think that really shows that we're | |
| 1:14:06.360 --> 1:14:08.280 | |
| at the professional level. | |
| 1:14:08.280 --> 1:14:09.640 | |
| And that kind of looking at those games, | |
| 1:14:09.640 --> 1:14:12.280 | |
| we think that the coin could have gone a different direction | |
| 1:14:12.280 --> 1:14:13.560 | |
| and we could have had some wins. | |
| 1:14:13.560 --> 1:14:16.200 | |
| And so that was actually very encouraging for us. | |
| 1:14:16.200 --> 1:14:18.360 | |
| And you know, it's interesting because the international was | |
| 1:14:18.360 --> 1:14:19.720 | |
| at a fixed time, right? | |
| 1:14:19.720 --> 1:14:22.680 | |
| So we knew exactly what day we were going to be playing. | |
| 1:14:22.680 --> 1:14:25.480 | |
| And we pushed as far as we could, as fast as we could. | |
| 1:14:25.480 --> 1:14:28.040 | |
| Two weeks later, we had a bot that had an 80% win rate | |
| 1:14:28.040 --> 1:14:30.120 | |
| versus the one that played at TI. | |
| 1:14:30.120 --> 1:14:31.720 | |
| So the March of Progress, you know, | |
| 1:14:31.720 --> 1:14:33.480 | |
| that you should think of as a snapshot rather | |
| 1:14:33.480 --> 1:14:34.920 | |
| than as an end state. | |
| 1:14:34.920 --> 1:14:39.000 | |
| And so in fact, we'll be announcing our finals pretty soon. | |
| 1:14:39.000 --> 1:14:42.760 | |
| I actually think that we'll announce our final match | |
| 1:14:42.760 --> 1:14:45.240 | |
| prior to this podcast being released. | |
| 1:14:45.240 --> 1:14:49.240 | |
| So there should be, we'll be playing against the world | |
| 1:14:49.240 --> 1:14:49.720 | |
| champions. | |
| 1:14:49.720 --> 1:14:52.520 | |
| And you know, for us, it's really less about, | |
| 1:14:52.520 --> 1:14:55.400 | |
| like the way that we think about what's upcoming | |
| 1:14:55.400 --> 1:14:59.000 | |
| is the final milestone, the final competitive milestone | |
| 1:14:59.000 --> 1:15:00.280 | |
| for the project, right? | |
| 1:15:00.280 --> 1:15:02.760 | |
| That our goal in all of this isn't really | |
| 1:15:02.760 --> 1:15:05.160 | |
| about beating humans at Dota. | |
| 1:15:05.160 --> 1:15:06.760 | |
| Our goal is to push the state of the art | |
| 1:15:06.760 --> 1:15:07.800 | |
| in reinforcement learning. | |
| 1:15:07.800 --> 1:15:08.920 | |
| And we've done that, right? | |
| 1:15:08.920 --> 1:15:10.680 | |
| And we've actually learned a lot from our system | |
| 1:15:10.680 --> 1:15:13.320 | |
| and that we have, you know, I think a lot of exciting | |
| 1:15:13.320 --> 1:15:14.680 | |
| next steps that we want to take. | |
| 1:15:14.680 --> 1:15:16.440 | |
| And so, you know, kind of the final showcase | |
| 1:15:16.440 --> 1:15:18.760 | |
| of what we built, we're going to do this match. | |
| 1:15:18.760 --> 1:15:21.240 | |
| But for us, it's not really the success or failure | |
| 1:15:21.240 --> 1:15:23.800 | |
| to see, you know, do we have the coin flip go | |
| 1:15:23.800 --> 1:15:24.840 | |
| in our direction or against. | |
| 1:15:25.880 --> 1:15:28.680 | |
| Where do you see the field of deep learning | |
| 1:15:28.680 --> 1:15:30.680 | |
| heading in the next few years? | |
| 1:15:31.720 --> 1:15:35.480 | |
| Where do you see the work in reinforcement learning | |
| 1:15:35.480 --> 1:15:40.360 | |
| perhaps heading and more specifically with OpenAI, | |
| 1:15:41.160 --> 1:15:43.480 | |
| all the exciting projects that you're working on, | |
| 1:15:44.280 --> 1:15:46.360 | |
| what does 2019 hold for you? | |
| 1:15:46.360 --> 1:15:47.400 | |
| Massive scale. | |
| 1:15:47.400 --> 1:15:47.880 | |
| Scale. | |
| 1:15:47.880 --> 1:15:49.480 | |
| I will put an atrocious on that and just say, | |
| 1:15:49.480 --> 1:15:52.200 | |
| you know, I think that it's about ideas plus scale. | |
| 1:15:52.200 --> 1:15:52.840 | |
| You need both. | |
| 1:15:52.840 --> 1:15:54.920 | |
| So that's a really good point. | |
| 1:15:54.920 --> 1:15:57.720 | |
| So the question, in terms of ideas, | |
| 1:15:58.520 --> 1:16:02.200 | |
| you have a lot of projects that are exploring | |
| 1:16:02.200 --> 1:16:04.280 | |
| different areas of intelligence. | |
| 1:16:04.280 --> 1:16:07.480 | |
| And the question is, when you think of scale, | |
| 1:16:07.480 --> 1:16:09.560 | |
| do you think about growing the scale | |
| 1:16:09.560 --> 1:16:10.680 | |
| of those individual projects, | |
| 1:16:10.680 --> 1:16:12.600 | |
| or do you think about adding new projects? | |
| 1:16:13.160 --> 1:16:17.320 | |
| And sorry, if you were thinking about adding new projects, | |
| 1:16:17.320 --> 1:16:19.800 | |
| or if you look at the past, what's the process | |
| 1:16:19.800 --> 1:16:21.960 | |
| of coming up with new projects and new ideas? | |
| 1:16:21.960 --> 1:16:22.680 | |
| Yep. | |
| 1:16:22.680 --> 1:16:24.600 | |
| So we really have a life cycle of project here. | |
| 1:16:25.240 --> 1:16:27.320 | |
| So we start with a few people just working | |
| 1:16:27.320 --> 1:16:28.440 | |
| on a small scale idea. | |
| 1:16:28.440 --> 1:16:30.520 | |
| And language is actually a very good example of this, | |
| 1:16:30.520 --> 1:16:32.440 | |
| that it was really, you know, one person here | |
| 1:16:32.440 --> 1:16:34.840 | |
| who was pushing on language for a long time. | |
| 1:16:34.840 --> 1:16:36.680 | |
| I mean, then you get signs of life, right? | |
| 1:16:36.680 --> 1:16:38.440 | |
| And so this is like, let's say, you know, | |
| 1:16:38.440 --> 1:16:42.600 | |
| with the original GPT, we had something that was interesting. | |
| 1:16:42.600 --> 1:16:44.760 | |
| And we said, okay, it's time to scale this, right? | |
| 1:16:44.760 --> 1:16:45.960 | |
| It's time to put more people on it, | |
| 1:16:45.960 --> 1:16:48.120 | |
| put more computational resources behind it, | |
| 1:16:48.120 --> 1:16:51.560 | |
| and then we just kind of keep pushing and keep pushing. | |
| 1:16:51.560 --> 1:16:52.920 | |
| And the end state is something that looks like | |
| 1:16:52.920 --> 1:16:55.400 | |
| Dota or Robotics, where you have a large team of, | |
| 1:16:55.400 --> 1:16:57.800 | |
| you know, 10 or 15 people that are running things | |
| 1:16:57.800 --> 1:17:00.680 | |
| at very large scale, and that you're able to really have | |
| 1:17:00.680 --> 1:17:04.280 | |
| material engineering and, you know, | |
| 1:17:04.280 --> 1:17:06.520 | |
| sort of machine learning science coming together | |
| 1:17:06.520 --> 1:17:10.200 | |
| to make systems that work and get material results | |
| 1:17:10.200 --> 1:17:11.560 | |
| that just would have been impossible otherwise. | |
| 1:17:12.200 --> 1:17:13.560 | |
| So we do that whole life cycle. | |
| 1:17:13.560 --> 1:17:16.600 | |
| We've done it a number of times, you know, typically end to end. | |
| 1:17:16.600 --> 1:17:19.960 | |
| It's probably two years or so to do it. | |
| 1:17:19.960 --> 1:17:21.720 | |
| You know, the organization's been around for three years, | |
| 1:17:21.720 --> 1:17:23.000 | |
| so maybe we'll find that we also have | |
| 1:17:23.000 --> 1:17:24.760 | |
| longer life cycle projects. | |
| 1:17:24.760 --> 1:17:27.480 | |
| But, you know, we work up to those. | |
| 1:17:27.480 --> 1:17:30.280 | |
| So one team that we're actually just starting, | |
| 1:17:30.280 --> 1:17:32.200 | |
| Illy and I, are kicking off a new team | |
| 1:17:32.200 --> 1:17:35.080 | |
| called the Reasoning Team, and this is to really try to tackle | |
| 1:17:35.080 --> 1:17:37.400 | |
| how do you get neural networks to reason? | |
| 1:17:37.400 --> 1:17:41.400 | |
| And we think that this will be a long term project. | |
| 1:17:41.400 --> 1:17:42.840 | |
| It's one that we're very excited about. | |
| 1:17:42.840 --> 1:17:46.200 | |
| In terms of reasoning, super exciting topic, | |
| 1:17:47.400 --> 1:17:52.200 | |
| what kind of benchmarks, what kind of tests of reasoning | |
| 1:17:52.200 --> 1:17:53.800 | |
| do you envision? | |
| 1:17:53.800 --> 1:17:55.880 | |
| What would, if you set back, | |
| 1:17:55.880 --> 1:17:59.240 | |
| whatever drink, and you would be impressed | |
| 1:17:59.240 --> 1:18:01.640 | |
| that this system is able to do something, | |
| 1:18:01.640 --> 1:18:02.760 | |
| what would that look like? | |
| 1:18:02.760 --> 1:18:03.800 | |
| Theorem proving. | |
| 1:18:03.800 --> 1:18:04.840 | |
| Theorem proving. | |
| 1:18:04.840 --> 1:18:09.480 | |
| So some kind of logic, and especially mathematical logic. | |
| 1:18:09.480 --> 1:18:10.440 | |
| I think so, right? | |
| 1:18:10.440 --> 1:18:12.440 | |
| And I think that there's kind of other problems | |
| 1:18:12.440 --> 1:18:14.520 | |
| that are dual to theorem proving in particular. | |
| 1:18:14.520 --> 1:18:16.840 | |
| You know, you think about programming, | |
| 1:18:16.840 --> 1:18:19.960 | |
| you think about even like security analysis of code, | |
| 1:18:19.960 --> 1:18:24.200 | |
| that these all kind of capture the same sorts of core reasoning | |
| 1:18:24.200 --> 1:18:27.480 | |
| and being able to do some out of distribution generalization. | |
| 1:18:28.440 --> 1:18:31.880 | |
| It would be quite exciting if OpenAI Reasoning Team | |
| 1:18:31.880 --> 1:18:33.880 | |
| was able to prove that P equals NP. | |
| 1:18:33.880 --> 1:18:35.080 | |
| That would be very nice. | |
| 1:18:35.080 --> 1:18:37.720 | |
| It would be very, very exciting especially. | |
| 1:18:37.720 --> 1:18:39.080 | |
| If it turns out that P equals NP, | |
| 1:18:39.080 --> 1:18:40.120 | |
| that'll be interesting too. | |
| 1:18:40.120 --> 1:18:45.160 | |
| It would be ironic and humorous. | |
| 1:18:45.160 --> 1:18:51.800 | |
| So what problem stands out to you as the most exciting | |
| 1:18:51.800 --> 1:18:55.720 | |
| and challenging impactful to the work for us as a community | |
| 1:18:55.720 --> 1:18:58.440 | |
| in general and for OpenAI this year? | |
| 1:18:58.440 --> 1:18:59.480 | |
| You mentioned reasoning. | |
| 1:18:59.480 --> 1:19:01.320 | |
| I think that's a heck of a problem. | |
| 1:19:01.320 --> 1:19:01.480 | |
| Yeah. | |
| 1:19:01.480 --> 1:19:02.760 | |
| So I think reasoning is an important one. | |
| 1:19:02.760 --> 1:19:04.840 | |
| I think it's going to be hard to get good results in 2019. | |
| 1:19:05.480 --> 1:19:07.480 | |
| You know, again, just like we think about the lifecycle, | |
| 1:19:07.480 --> 1:19:07.960 | |
| takes time. | |
| 1:19:08.600 --> 1:19:11.320 | |
| I think for 2019, language modeling seems to be kind of | |
| 1:19:11.320 --> 1:19:12.520 | |
| on that ramp, right? | |
| 1:19:12.520 --> 1:19:14.760 | |
| It's at the point that we have a technique that works. | |
| 1:19:14.760 --> 1:19:17.000 | |
| We want to scale 100x, 1000x, see what happens. | |
| 1:19:18.040 --> 1:19:18.360 | |
| Awesome. | |
| 1:19:18.360 --> 1:19:21.800 | |
| Do you think we're living in a simulation? | |
| 1:19:21.800 --> 1:19:24.520 | |
| I think it's hard to have a real opinion about it. | |
| 1:19:25.560 --> 1:19:26.200 | |
| It's actually interesting. | |
| 1:19:26.200 --> 1:19:29.960 | |
| I separate out things that I think can have yield | |
| 1:19:29.960 --> 1:19:31.880 | |
| materially different predictions about the world | |
| 1:19:32.520 --> 1:19:35.640 | |
| from ones that are just kind of fun to speculate about. | |
| 1:19:35.640 --> 1:19:37.800 | |
| And I kind of view simulation as more like, | |
| 1:19:37.800 --> 1:19:40.200 | |
| is there a flying teapot between Mars and Jupiter? | |
| 1:19:40.200 --> 1:19:43.800 | |
| Like, maybe, but it's a little bit hard to know | |
| 1:19:43.800 --> 1:19:45.000 | |
| what that would mean for my life. | |
| 1:19:45.000 --> 1:19:46.360 | |
| So there is something actionable. | |
| 1:19:46.360 --> 1:19:50.680 | |
| So some of the best work opening as done | |
| 1:19:50.680 --> 1:19:52.200 | |
| is in the field of reinforcement learning. | |
| 1:19:52.760 --> 1:19:56.520 | |
| And some of the success of reinforcement learning | |
| 1:19:56.520 --> 1:19:59.080 | |
| come from being able to simulate the problem you're trying | |
| 1:19:59.080 --> 1:20:00.040 | |
| to solve. | |
| 1:20:00.040 --> 1:20:03.560 | |
| So do you have a hope for reinforcement, | |
| 1:20:03.560 --> 1:20:05.160 | |
| for the future of reinforcement learning | |
| 1:20:05.160 --> 1:20:06.920 | |
| and for the future of simulation? | |
| 1:20:06.920 --> 1:20:09.000 | |
| Like, whether we're talking about autonomous vehicles | |
| 1:20:09.000 --> 1:20:12.760 | |
| or any kind of system, do you see that scaling? | |
| 1:20:12.760 --> 1:20:16.280 | |
| So we'll be able to simulate systems and, hence, | |
| 1:20:16.280 --> 1:20:19.400 | |
| be able to create a simulator that echoes our real world | |
| 1:20:19.400 --> 1:20:22.520 | |
| and proving once and for all, even though you're denying it | |
| 1:20:22.520 --> 1:20:23.800 | |
| that we're living in a simulation. | |
| 1:20:24.840 --> 1:20:26.360 | |
| I feel like I've used that for questions, right? | |
| 1:20:26.360 --> 1:20:28.200 | |
| So, you know, kind of at the core there of, like, | |
| 1:20:28.200 --> 1:20:30.280 | |
| can we use simulation for self driving cars? | |
| 1:20:31.080 --> 1:20:33.720 | |
| Take a look at our robotic system, DACTL, right? | |
| 1:20:33.720 --> 1:20:37.720 | |
| That was trained in simulation using the Dota system, in fact. | |
| 1:20:37.720 --> 1:20:39.560 | |
| And it transfers to a physical robot. | |
| 1:20:40.280 --> 1:20:42.120 | |
| And I think everyone looks at our Dota system, | |
| 1:20:42.120 --> 1:20:43.400 | |
| they're like, okay, it's just a game. | |
| 1:20:43.400 --> 1:20:45.080 | |
| How are you ever going to escape to the real world? | |
| 1:20:45.080 --> 1:20:47.320 | |
| And the answer is, well, we did it with the physical robot, | |
| 1:20:47.320 --> 1:20:48.600 | |
| the no one could program. | |
| 1:20:48.600 --> 1:20:50.840 | |
| And so I think the answer is simulation goes a lot further | |
| 1:20:50.840 --> 1:20:53.400 | |
| than you think if you apply the right techniques to it. | |
| 1:20:54.040 --> 1:20:55.400 | |
| Now, there's a question of, you know, | |
| 1:20:55.400 --> 1:20:57.400 | |
| are the beings in that simulation going to wake up | |
| 1:20:57.400 --> 1:20:58.520 | |
| and have consciousness? | |
| 1:20:59.480 --> 1:21:02.840 | |
| I think that one seems a lot harder to, again, reason about. | |
| 1:21:02.840 --> 1:21:05.240 | |
| I think that, you know, you really should think about, like, | |
| 1:21:05.240 --> 1:21:07.800 | |
| where exactly does human consciousness come from | |
| 1:21:07.800 --> 1:21:09.000 | |
| in our own self awareness? | |
| 1:21:09.000 --> 1:21:10.600 | |
| And, you know, is it just that, like, | |
| 1:21:10.600 --> 1:21:12.280 | |
| once you have, like, a complicated enough neural net, | |
| 1:21:12.280 --> 1:21:14.440 | |
| do you have to worry about the agent's feeling pain? | |
| 1:21:15.720 --> 1:21:17.560 | |
| And, you know, I think there's, like, | |
| 1:21:17.560 --> 1:21:19.320 | |
| interesting speculation to do there. | |
| 1:21:19.320 --> 1:21:22.920 | |
| But, you know, again, I think it's a little bit hard to know for sure. | |
| 1:21:22.920 --> 1:21:24.840 | |
| Well, let me just keep with the speculation. | |
| 1:21:24.840 --> 1:21:28.040 | |
| Do you think to create intelligence, general intelligence, | |
| 1:21:28.600 --> 1:21:33.000 | |
| you need one consciousness and two a body? | |
| 1:21:33.000 --> 1:21:34.920 | |
| Do you think any of those elements are needed, | |
| 1:21:34.920 --> 1:21:38.360 | |
| or is intelligence something that's orthogonal to those? | |
| 1:21:38.360 --> 1:21:41.560 | |
| I'll stick to the kind of, like, the non grand answer first, | |
| 1:21:41.560 --> 1:21:41.720 | |
| right? | |
| 1:21:41.720 --> 1:21:43.960 | |
| So the non grand answer is just to look at, | |
| 1:21:43.960 --> 1:21:45.560 | |
| you know, what are we already making work? | |
| 1:21:45.560 --> 1:21:47.640 | |
| You look at GPT2, a lot of people would have said | |
| 1:21:47.640 --> 1:21:49.320 | |
| that to even get these kinds of results, | |
| 1:21:49.320 --> 1:21:50.920 | |
| you need real world experience. | |
| 1:21:50.920 --> 1:21:52.440 | |
| You need a body, you need grounding. | |
| 1:21:52.440 --> 1:21:54.920 | |
| How are you supposed to reason about any of these things? | |
| 1:21:54.920 --> 1:21:56.360 | |
| How are you supposed to, like, even kind of know | |
| 1:21:56.360 --> 1:21:57.960 | |
| about smoke and fire and those things | |
| 1:21:57.960 --> 1:21:59.560 | |
| if you've never experienced them? | |
| 1:21:59.560 --> 1:22:03.000 | |
| And GPT2 shows that you can actually go way further | |
| 1:22:03.000 --> 1:22:05.640 | |
| than that kind of reasoning would predict. | |
| 1:22:05.640 --> 1:22:09.240 | |
| So I think that in terms of, do we need consciousness? | |
| 1:22:09.240 --> 1:22:10.360 | |
| Do we need a body? | |
| 1:22:10.360 --> 1:22:11.880 | |
| It seems the answer is probably not, right? | |
| 1:22:11.880 --> 1:22:13.640 | |
| That we could probably just continue to push | |
| 1:22:13.640 --> 1:22:14.680 | |
| kind of the systems we have. | |
| 1:22:14.680 --> 1:22:16.520 | |
| They already feel general. | |
| 1:22:16.520 --> 1:22:19.080 | |
| They're not as competent or as general | |
| 1:22:19.080 --> 1:22:21.640 | |
| or able to learn as quickly as an AGI would, | |
| 1:22:21.640 --> 1:22:24.680 | |
| but, you know, they're at least like kind of proto AGI | |
| 1:22:24.680 --> 1:22:28.040 | |
| in some way, and they don't need any of those things. | |
| 1:22:28.040 --> 1:22:31.640 | |
| Now, let's move to the grand answer, which is, you know, | |
| 1:22:31.640 --> 1:22:34.840 | |
| if our neural nets consciousness, | |
| 1:22:34.840 --> 1:22:37.240 | |
| nets conscious already, would we ever know? | |
| 1:22:37.240 --> 1:22:38.680 | |
| How can we tell, right? | |
| 1:22:38.680 --> 1:22:40.920 | |
| And, you know, here's where the speculation starts | |
| 1:22:40.920 --> 1:22:44.760 | |
| to become, you know, at least interesting or fun | |
| 1:22:44.760 --> 1:22:46.200 | |
| and maybe a little bit disturbing, | |
| 1:22:46.200 --> 1:22:47.880 | |
| depending on where you take it. | |
| 1:22:47.880 --> 1:22:51.080 | |
| But it certainly seems that when we think about animals, | |
| 1:22:51.080 --> 1:22:53.080 | |
| that there's some continuum of consciousness. | |
| 1:22:53.080 --> 1:22:56.040 | |
| You know, my cat, I think, is conscious in some way, right? | |
| 1:22:56.040 --> 1:22:58.040 | |
| You know, not as conscious as a human. | |
| 1:22:58.040 --> 1:22:59.880 | |
| And you could imagine that you could build | |
| 1:22:59.880 --> 1:23:01.000 | |
| a little consciousness meter, right? | |
| 1:23:01.000 --> 1:23:02.840 | |
| You point at a cat, it gives you a little reading, | |
| 1:23:02.840 --> 1:23:06.200 | |
| you point at a human, it gives you much bigger reading. | |
| 1:23:06.200 --> 1:23:07.960 | |
| What would happen if you pointed one of those | |
| 1:23:07.960 --> 1:23:09.800 | |
| at a Dota neural net? | |
| 1:23:09.800 --> 1:23:11.960 | |
| And if you're training this massive simulation, | |
| 1:23:11.960 --> 1:23:14.600 | |
| do the neural nets feel pain? | |
| 1:23:14.600 --> 1:23:16.760 | |
| You know, it becomes pretty hard to know | |
| 1:23:16.760 --> 1:23:20.040 | |
| that the answer is no, and it becomes pretty hard | |
| 1:23:20.040 --> 1:23:22.360 | |
| to really think about what that would mean | |
| 1:23:22.360 --> 1:23:25.160 | |
| if the answer were yes. | |
| 1:23:25.160 --> 1:23:27.400 | |
| And it's very possible, you know, for example, | |
| 1:23:27.400 --> 1:23:29.400 | |
| you could imagine that maybe the reason | |
| 1:23:29.400 --> 1:23:31.400 | |
| that humans have consciousness | |
| 1:23:31.400 --> 1:23:35.000 | |
| is because it's a convenient computational shortcut, right? | |
| 1:23:35.000 --> 1:23:36.920 | |
| If you think about it, if you have a being | |
| 1:23:36.920 --> 1:23:39.320 | |
| that wants to avoid pain, which seems pretty important | |
| 1:23:39.320 --> 1:23:41.000 | |
| to survive in this environment | |
| 1:23:41.000 --> 1:23:43.640 | |
| and wants to, like, you know, eat food, | |
| 1:23:43.640 --> 1:23:45.400 | |
| then maybe the best way of doing it | |
| 1:23:45.400 --> 1:23:47.080 | |
| is to have a being that's conscious, right? | |
| 1:23:47.080 --> 1:23:49.480 | |
| That, you know, in order to succeed in the environment, | |
| 1:23:49.480 --> 1:23:51.080 | |
| you need to have those properties | |
| 1:23:51.080 --> 1:23:52.600 | |
| and how are you supposed to implement them? | |
| 1:23:52.600 --> 1:23:55.240 | |
| And maybe this consciousness is a way of doing that. | |
| 1:23:55.240 --> 1:23:57.720 | |
| If that's true, then actually maybe we should expect | |
| 1:23:57.720 --> 1:23:59.880 | |
| that really competent reinforcement learning agents | |
| 1:23:59.880 --> 1:24:01.960 | |
| will also have consciousness. | |
| 1:24:01.960 --> 1:24:03.240 | |
| But, you know, that's a big if. | |
| 1:24:03.240 --> 1:24:04.760 | |
| And I think there are a lot of other arguments | |
| 1:24:04.760 --> 1:24:05.880 | |
| that you can make in other directions. | |
| 1:24:06.680 --> 1:24:08.360 | |
| I think that's a really interesting idea | |
| 1:24:08.360 --> 1:24:11.400 | |
| that even GPT2 has some degree of consciousness. | |
| 1:24:11.400 --> 1:24:14.200 | |
| That's something that's actually not as crazy | |
| 1:24:14.200 --> 1:24:14.760 | |
| to think about. | |
| 1:24:14.760 --> 1:24:17.720 | |
| It's useful to think about as we think about | |
| 1:24:17.720 --> 1:24:19.800 | |
| what it means to create intelligence of a dog, | |
| 1:24:19.800 --> 1:24:24.360 | |
| intelligence of a cat, and the intelligence of a human. | |
| 1:24:24.360 --> 1:24:30.760 | |
| So, last question, do you think we will ever fall in love, | |
| 1:24:30.760 --> 1:24:33.560 | |
| like in the movie, Her, with an artificial intelligence system | |
| 1:24:34.360 --> 1:24:36.200 | |
| or an artificial intelligence system | |
| 1:24:36.200 --> 1:24:38.440 | |
| falling in love with a human? | |
| 1:24:38.440 --> 1:24:38.920 | |
| I hope so. | |
| 1:24:40.120 --> 1:24:43.640 | |
| If there's any better way to end it is on love. | |
| 1:24:43.640 --> 1:24:45.560 | |
| So, Greg, thanks so much for talking today. | |
| 1:24:45.560 --> 1:24:55.560 | |
| Thank you for having me. | |