Sam Altman's World Tour, in 16 Moments

00:00:00.000 | There have been 16 surprising and or fascinating moments from Sam Altman's world tour.

00:00:06.880 | I could have done a video on each of them, but after watching over 10 hours of interviews,

00:00:11.780 | I decided, you know what, let's just show you everything in one video.

00:00:15.940 | From AIs designing new AIs, to fresh ChatGPT leaks, shooting railguns, to open source,

00:00:23.740 | here's all 16 things I learnt in no particular order.

00:00:27.480 | Let's start with Sam Altman's warning about AIs designing their own architecture.

00:00:32.940 | Seems like a good idea, but Satsukawa could see one of their models designing the next model.

00:00:54.260 | We are definitely very concerned about superintelligence.

00:00:57.460 | It will be possible to build a computer, a computer cluster, a GPU farm, that is just smarter than any person,

00:01:04.820 | that can do science and engineering much, much faster than even a large team of really experienced scientists and engineers.

00:01:11.700 | And that is crazy. That is going to be unbelievably, extremely impactful.

00:01:17.060 | It could engineer the next version of the system.

00:01:20.780 | AI building AI. It's just crazy.

00:01:23.540 | Let's return to Abu Dhabi where Sam Altman said he enjoys the power that being CEO of OpenAI brings,

00:01:32.180 | but also mentioned strange decisions he might have to make.

00:01:36.020 | I mean, I have like lots of selfish reasons for doing this.

00:01:38.220 | And as you said, I get like all of the power of running OpenAI, but I can't think of like anything more fulfilling to work on.

00:01:44.900 | I don't think it's like particularly altruistic because it would be if I like didn't already have a bunch of money.

00:01:49.300 | Yeah, the money is going to like pile up faster than I can spend it anyway.

00:01:52.460 | I like being non-confident.

00:01:53.460 | I don't think I can flick that on OpenAI because I think the chance that we have to make a very strange decision someday is non-trivial.

00:02:01.260 | Speaking of big decisions, Sam Altman hinted twice, once in Jordan and once in India,

00:02:06.100 | of possible regrets he might have had over firing the starting gun in the AI race.

00:02:11.420 | We're definitely going to have some huge regrets 20 years from now.

00:02:15.180 | I hope what we can say is that we did far, far, far more good than bad.

00:02:19.620 | And I think we will. I think that's true.

00:02:21.740 | But the downside here is pretty big.

00:02:23.460 | And I think we feel that weight every day.

00:02:25.620 | Honestly, I think if we're going to regret something, it may be that we already pushed the button.

00:02:32.100 | Like we've already launched this revolution.

00:02:34.340 | It's somewhat out of our hands.

00:02:36.220 | I think it's going to be great.

00:02:37.660 | But like this is going to happen now, right?

00:02:40.300 | Like this, we're out of the world is like out of the gates.

00:02:43.820 | I guess the thing that I lose the most sleep over is that we already have done something really bad.

00:02:49.900 | I don't think we have.

00:02:51.220 | But the hypothetical that we're going to have to do something really bad is that we're going to have to do something really bad.

00:02:52.220 | I don't think we have.

00:02:52.300 | But the hypothetical that we're going to have to do something really bad is that we're going to have to do something really bad.

00:02:53.300 | that we, by launching Chachi PT into the world,

00:02:57.260 | shot the industry out of a rail gun

00:02:59.540 | and we now don't get to have much impact anymore.

00:03:02.460 | And there's gonna be an acceleration

00:03:05.100 | towards making these systems which again,

00:03:08.200 | I think will be used for tremendous good

00:03:10.240 | and I think we're gonna address all the problems.

00:03:12.840 | But maybe there's something in there

00:03:14.220 | that was really hard and complicated

00:03:15.940 | in a way we didn't understand,

00:03:17.780 | and we've now already kicked this off.

00:03:20.240 | - But back to Tel Aviv where both Sam Altman

00:03:22.700 | and OpenAI's chief scientist Ilya Satskova

00:03:25.640 | agreed that the risks from superintelligence

00:03:28.420 | were not science fiction.

00:03:30.020 | - To the last question, the superintelligent AI

00:03:32.000 | that's out of control, yeah, that'd be pretty bad.

00:03:35.080 | Yeah, so it's like,

00:03:40.800 | it would be a big mistake to build a superintelligence AI

00:03:47.820 | that we don't know how to control.

00:03:49.240 | - I think the world should treat that not as a,

00:03:52.100 | you know, ha ha, never gonna come sci-fi risk,

00:03:55.160 | but something that we may have to confront

00:03:56.800 | in the next decade, which is not very long.

00:03:59.220 | - On a lighter note, Sam Altman didn't seem that perturbed,

00:04:02.120 | not just about a deep fake of himself,

00:04:04.640 | but also on society getting used to misinformation.

00:04:08.240 | - I wanna play a clip, maybe you guys can put on a clip

00:04:10.960 | of something I recently heard Sam speak somewhere

00:04:14.040 | and we can talk about it a bit.

00:04:15.500 | Could you play the clip please?

00:04:17.400 | - Hi, my name is Sam and I'm happy to be here today.

00:04:20.480 | Thank you all for joining.

00:04:21.780 | I also wanted to say that the gentleman on stage with me

00:04:24.920 | is incredibly good looking.

00:04:27.060 | And I also want to say that you should be very careful

00:04:29.360 | with videos generated with artificial intelligence technology.

00:04:32.700 | - Okay, so you didn't say that recently,

00:04:36.100 | but nonetheless, I think it raises a real question, right?

00:04:38.980 | When, you know, this video, if you look closely,

00:04:41.880 | you can see the lips aren't perfectly synced,

00:04:43.440 | but like you said, this stuff is only gonna get better

00:04:45.480 | and exponentially better.

00:04:46.900 | - Yeah, so that was like deeply in the uncanny valley.

00:04:48.940 | It's very strange to watch, but we're not that far away

00:04:51.780 | from something that looks perfect.

00:04:53.400 | There's a lot of fear right now about the impact

00:04:55.720 | this is gonna have on elections and on our society

00:04:58.900 | and how we ever trust media that we see.

00:05:01.860 | I have some fear there, but I think when it comes

00:05:04.160 | to like a video like that, I think as a society,

00:05:06.340 | we're gonna rise to the occasion.

00:05:07.520 | We're gonna learn very quickly that we don't trust videos

00:05:10.660 | unless we trust the sort of provenance.

00:05:13.240 | If people are saying something really important,

00:05:15.480 | they'll cryptographically sign it.

00:05:16.880 | - Indeed, throughout the world tour,

00:05:18.680 | Sam Altman repeatedly stated that he didn't believe there should be a video that looks perfect.

00:05:20.360 | Sam Altman repeatedly stated that he didn't believe there should be a video that looks perfect.

00:05:20.360 | Indeed, throughout the world tour, Sam Altman repeatedly stated that he didn't believe there should be a video that looks perfect.

00:05:20.360 | Sam Altman repeatedly stated that he didn't believe there should be a video that looks perfect.

00:05:21.580 | Sam Altman repeatedly stated that he didn't believe there should be any regulation of current models.

00:05:24.520 | - Everybody wants great education, productivity gains, discovery of new science,

00:05:29.340 | all of this stuff that's gonna happen, and no one wants to destroy the world.

00:05:32.580 | No one wants to do things, not even that bad, but still bad.

00:05:35.460 | I totally believe it is possible to not stifle innovation and to address the big risks.

00:05:42.020 | I think it would be a mistake to go regulate the current models of today.

00:05:45.880 | - And in Poland, his co-founder Wozzech Zaremba agreed,

00:05:49.460 | saying the risks of a video that looks perfect would be a mistake.

00:05:49.480 | saying the risks of a video that looks perfect would be a mistake.

00:05:50.360 | But the risks of superintelligence were 10 years away.

00:05:52.800 | - Also, I would say that the fear is the fear of AI of the future, not the AI of today.

00:06:01.100 | If the trajectory that we are on will continue, then in the decade or so, there will be built systems which are as powerful as today corporations.

00:06:13.420 | But if I could speak to Sam Altman, I would bring his attention to this paper published this week.

00:06:18.920 | This is a study out of Harvard University, which is a university that is very well-versed in AI.

00:06:20.360 | This study was published in the Harvard University and MIT, and it involved some non-scientist students working for one hour.

00:06:27.100 | In that hour, they were able to get chatbots to suggest four potential pandemic pathogens,

00:06:32.740 | explain how they can be generated from synthetic DNA using reverse genetics,

00:06:37.680 | supplied the names of DNA synthesis companies unlikely to screen orders,

00:06:42.320 | and identify detailed protocols and how to troubleshoot them.

00:06:46.320 | And they say that collectively, these results suggest that LLMs,

00:06:50.360 | will make pandemic class agents widely accessible, even to people they say with little or no lab training.

00:06:58.040 | And then there's this, these results strongly suggest that the existing evaluation and training process for large language models

00:07:05.680 | is inadequate to prevent them from providing malicious actors with accessible expertise relevant to inflicting mass death.

00:07:13.880 | And that more immediately, if unmitigated LLM chatbots render pandemic class agents more accessible,

00:07:20.360 | even to people without training in the life sciences,

00:07:22.920 | the number of individuals capable of killing tens of millions of people will dramatically increase.

00:07:28.200 | They recommend that, at a minimum, new LLMs larger than GPT-3 should undergo evaluation by third parties,

00:07:35.800 | skilled in assessing catastrophic biological risks before controlled access is given to the general public.

00:07:41.840 | Notice they said "larger than GPT-3",

00:07:44.720 | so that strongly contradicts Sam Altman's assertion that current models like GPT-4 shouldn't have

00:07:50.360 | any regulation.

00:07:51.400 | They say that even open source communities should welcome safeguards because a single instance of misuse and mass death

00:07:58.280 | would trigger a backlash including the imposition of extremely harsh regulations.

00:08:03.480 | One specific recommendation was that if biotech and information security experts

00:08:08.440 | were able to identify the set of publications most relevant to causing mass death,

00:08:13.400 | and companies like OpenAI and Google curated their training datasets to remove those publications,

00:08:20.360 | then future models trained on the curated data would be far less capable of providing anyone intent on harm

00:08:26.760 | with the "recipes for the creation or enhancement of pathogens".

00:08:30.920 | This seems like an absolutely obvious move to me and I think Ilya Satskova would agree.

00:08:35.560 | We are talking about as time goes by and the capability keeps increasing,

00:08:39.960 | you know, and eventually it goes all the way to here, right?

00:08:43.000 | Right now we are here. Today, that's where we are.

00:08:45.960 | That's where we're going to get to.

00:08:47.720 | When we get to this point, then yeah,

00:08:50.360 | it's very powerful technology. It can be used for amazing applications.

00:08:54.360 | You can say cure all disease. On the flip side, you can say create a disease.

00:08:59.560 | Much more worse than anything that existed before. That'd be bad.

00:09:03.240 | Moving on to the ChatGPT leak, it seems like we're going to get a new workspace

00:09:09.160 | where we can customize our interaction with ChatGPT, giving it files and a profile with

00:09:15.000 | any information that you'd like ChatGPT to remember about you and your preferences.

00:09:19.640 | This was hinted at in the chat, but I'm not sure if this is the right way to do it.

00:09:20.360 | We were hinted at on the world tour when one of Sam Altman's guests,

00:09:23.000 | Johannes Heidecker from OpenAI Research, talked about customizing models.

00:09:27.320 | We are trying to make our models both better at following certain guardrails that should

00:09:31.320 | never be overwritten, not with jailbreaks, not if you ask nicely, not if you threaten it.

00:09:35.560 | And we're also trying to make our models better at being customizable,

00:09:39.960 | making them listen more to additional instructions of what kind of behavior the user or the developer

00:09:45.320 | wants.

00:09:45.720 | On a lighter note, the leaders of OpenAI were asked in Seoul, the capital of

00:09:50.360 | South Korea, about the mixing of AI and religion.

00:09:54.120 | Do you expect AI to replace the role of religious organizations like church?

00:09:58.520 | I think that it's a good question how all human societies will integrate AI.

00:10:07.080 | And we've already seen people building AI pastors, for example, and so the constituents

00:10:11.880 | can ask questions to this pastor that can cite Bible verses and it can give advice.

00:10:15.880 | But now back to Poland, where Sam Altman called open source "unstable"

00:10:20.360 | "Realizing that open source is unstoppable and shouldn't be stopped,

00:10:24.280 | and so this stuff is going to be out there and as a society we have to adapt."

00:10:27.400 | But speaking of stopping AI, Sam Altman was asked about his own loved ones,

00:10:31.880 | and in response he gave a utopic vision of the future and called the current world "barbaric".

00:10:37.400 | If you truly believe that AI imposes a danger to humankind, why keep developing it?

00:10:43.720 | Aren't you afraid for your own dear ones and family?

00:10:47.880 | I think it's a super fair

00:10:50.360 | and good question.

00:10:51.720 | And the most troublesome part of our jobs is that we have to balance this incredible

00:10:59.400 | promise and this technology that I think humans really need, and we can talk about

00:11:05.240 | why in a second, with confronting these very serious risks.

00:11:09.240 | Why to build it?

00:11:10.040 | Number one, I do think that when we look back at the standard of living and what we tolerate

00:11:15.480 | for people today, it will look even worse than

00:11:20.360 | when we look back at how people lived 500 or 1000 years ago.

00:11:24.120 | And we'll say like, man, can you imagine that people lived in poverty?

00:11:27.880 | Can you imagine people suffered from disease?

00:11:30.440 | Can you imagine that everyone didn't have a phenomenal education and were able to live

00:11:33.800 | their lives however they wanted?

00:11:35.080 | It's going to look barbaric.

00:11:36.360 | I think everyone in the future is going to have better lives than the best people of today.

00:11:40.120 | I think there's like a moral duty to figure out how to do that.

00:11:43.160 | I also think this is like unstoppable, like this is the progress of technology.

00:11:47.240 | It won't work to some degree.

00:11:50.360 | It's going to stop it.

00:11:50.920 | And so we have to figure out how to manage the risk.

00:11:53.320 | He doesn't seem to be 100% sure on this front though.

00:11:56.520 | And here is an interview he gave with The Guardian when he was in London for his world

00:12:00.920 | tour.

00:12:01.320 | Speaking of super intelligence, he said, "It's not that it's not stoppable.

00:12:05.400 | If governments around the world decided to act in concert to limit AI development, as

00:12:09.800 | they have in other fields such as human cloning or bioweapon research, they may be able to."

00:12:14.520 | But then he repeated, "But that would be to give up all that is possible.

00:12:17.800 | I think that this will be the most tremendous leap forward,

00:12:20.360 | in terms of the quality of life for people that we've ever had."

00:12:22.600 | I did try to get tickets for the London leg of his world tour,

00:12:25.800 | but they were sold out within half an hour.

00:12:27.480 | Oh well.

00:12:28.040 | Sam Oldman does think that behaviour will change, however,

00:12:30.920 | when these AGI labs stare existential risk in the face.

00:12:34.760 | Sam Oldman: "One of the things we talked about is what's a structure that would let us

00:12:38.920 | warmly embrace regulation that would hurt us the most.

00:12:42.760 | And now that the time has come for that, we're out here advocating around the world for regulation

00:12:47.480 | that will impact us the most. So, of course, we'll comply with it.

00:12:50.360 | But I think it's more easy to get good behaviour out of people when they are staring existential

00:12:56.520 | risk in the face. And so I think all of the people at the leading edge here, these different

00:13:01.480 | companies, now feel this, and you will see a different collective response than you saw from

00:13:06.280 | the social media companies."

00:13:07.480 | And in terms of opportunities, both Sam Oldman and Ilya Sutskova talked about solving climate change.

00:13:13.240 | Sam Oldman: "I don't want to say this because climate change is so serious and so hard of a

00:13:16.360 | problem, but I think once we have a really powerful super intelligence,

00:13:20.360 | addressing climate change will not be particularly difficult for a system like that."

00:13:24.040 | Ilya Sutskova: "We can even explain how. Here's how you solve climate change. You need a very large

00:13:28.360 | amount of efficient carbon capture. You need the energy for the carbon capture, you need the

00:13:33.320 | technology to build it, and you need to build a lot of it. If you can accelerate the scientific

00:13:38.120 | progress, which is something that a powerful AI could do, we could get to a very advanced carbon

00:13:43.400 | capture much faster. We could get to a very cheap power much faster. We could get to cheaper

00:13:48.680 | manufacturing much faster. Ilya Sutskova: "I think that's a very important point.

00:13:50.360 | Combine those three: cheap power, cheap manufacturing, advanced carbon capture.

00:13:54.840 | Now you build lots of them. And now you sucked out all the excess CO2 from the atmosphere."

00:13:59.880 | Sam Oldman: "You know, if you think about a system where you can say,

00:14:02.120 | 'Tell me how to make a lot of clean energy cheaply. Tell me how to efficiently capture carbon.

00:14:07.960 | And then tell me how to build a factory to do this at planetary scale.' If you can do that,

00:14:11.880 | you can do a lot of other things too." Ilya Sutskova: "Yeah. With one addition

00:14:15.240 | that not only you ask it to tell it, you ask it to do it." Sam Oldman: "That would indeed be

00:14:20.360 | amazing. But think of the power we would be giving to an AI if it was able to just do it,

00:14:25.960 | just create those carbon capture factories." Ilya Sutskova: "If we did make that decision,

00:14:30.120 | one thing that would help would be reducing hallucinations." Sam Oldman: "I think we will

00:14:34.280 | get the hallucination problem to a much, much better place. It will take us,

00:14:38.040 | my colleagues weigh in, I think it'll take us a year and a half, two years, something like that.

00:14:43.400 | But at that point, we won't still talk about these." Ilya Sutskova:

00:14:45.560 | Sam Oldman talked about that in New Delhi. That timeframe of 18 months to two years is

00:14:50.360 | ambitious and surprising. But now onto jobs, which Sam Oldman was asked about on every leg of the

00:14:56.440 | tour. On this front though, I do think it was Ilya Sutskova who gave the more honest answer.

00:15:01.320 | Sam Oldman: "Economic dislocation indeed, like we already know that there are jobs that are

00:15:06.600 | being impacted or they're being affected. In other words, some chunks of the jobs can be done. You

00:15:11.880 | know, if you're a programmer, you don't write functions anymore, copilot writes them for you.

00:15:15.800 | If you're an artist though, it's a bit different because a big chunk of the artists' economic

00:15:20.360 | part of them activity has been taken by some of the image generators. And while new jobs will be

00:15:25.240 | created, it's going to be a long period of economic uncertainty. There is an argument to be

00:15:29.720 | made that even when they have full human level AI, full AGI, people will still have economic activity

00:15:35.800 | to do. I don't know whether that's the case, but in either event, we will need to have something

00:15:42.600 | that will soften the blow to allow for a smoother transition either to the totally new professions

00:15:50.360 | or even if not, then we want government, the social systems will need to keep keen."

00:15:55.320 | I do think the changes in the job market will be dramatic and we'll be following the story closely.

00:15:59.960 | One thing I definitely agree with Sam Oldman on though, is the deep,

00:16:03.560 | almost philosophical change that this solving of intelligence has brought to humanity.

00:16:08.680 | Sam Oldman: "I grew up implicitly thinking that intelligence was this like really special

00:16:16.120 | human thing and kind of somewhat magical. And I now think that it's sort of a fundamental

00:16:20.360 | property of matter. And that's definitely a change to my worldview. The history of scientific

00:16:28.440 | discovery is that humans are less and less at the center. We used to think that sun rotated around

00:16:33.880 | us and then maybe at least we were, if not that, we were going to be the center of the galaxy and

00:16:38.280 | there wasn't this big universe. And then multiverse really is kind of weird and depressing. And

00:16:42.520 | if intelligence is a special, again, we're just further and further

00:16:45.560 | away from main character energy. But that's all right. That's sort of like a

00:16:50.360 | nice thing to realize actually." It's a bit like a Copernican and

00:16:53.880 | Darwinian revolution all rolled in one. But I'll give the final word to Greg Brockman in

00:16:59.320 | Seoul who talked about the unpredictability of scaling up models 10 times.

00:17:03.960 | Greg Brockman: "That is the biggest theme in the history of AI is that it's full of surprises.

00:17:07.480 | Every time you think you know something, you scale it up 10x, turns out you knew nothing.

00:17:11.080 | And so I think that we as a humanity, as a species are really exploring this together."

00:17:15.240 | Being all in it together and knowing nothing sounds about right. But thank you

00:17:20.360 | for watching to the end. I know that Sam Altman has a couple more stops. I think it's Jakarta

00:17:25.720 | and Melbourne on the world tour and I'll be watching those of course. But for now,

00:17:29.880 | thank you and have a wonderful day.

Sam Altman's World Tour, in 16 Moments

Chapters