Latest / Faster, Please! — The Podcast / ✨ The Age of AI, an update: My chat with policy analyst Dean Ball
Transcript
- 0:06SPEAKER_00: Welcome to Faster Please, the podcast.
- 0:09SPEAKER_00: I'm your host, Jim Pethikoukis.
- 0:12SPEAKER_00: Several times a month,
- 0:13I'll feature a lively conversation with a fascinating and
- 0:16provocative guest about
- 0:18SPEAKER_00: how to make the world a better place through scientific discovery,
- 0:21SPEAKER_00: technological innovation,
- 0:23SPEAKER_00: and economic growth.
- 0:24SPEAKER_00: You're also going to want to check out my Faster Please newsletter.
- 0:27SPEAKER_00: You're on Substack throughout the week for fresh essays,
- 0:30SPEAKER_00: Q&As,
- 0:31SPEAKER_00: and stories from around the internet and around the world.
- 0:37Anxiety is running rampant about the future of AI and its place
- 0:41in our society.
- 0:43SPEAKER_00: When tech CEOs warn of an impending white-collar jobpocalypse,
- 0:48SPEAKER_00: it's no wonder public pessimism is so widespread.
- 0:51SPEAKER_00: Today on Faster Please,
- 0:52SPEAKER_00: the podcast,
- 0:54SPEAKER_00: I chat with Dean Ball to help us sift through some of the uncertainty.
- 0:59We talk about recursive self-improvement,
- 1:02the role of AI in everything from medicine to defense,
- 1:06and what to think about the possible growing risk of AI company
- 1:10nationalization.
- 1:12SPEAKER_00: Ball is a senior fellow at FAI, the Foundation for American Innovation.
- 1:17He recently served as senior policy advisor for Artificial
- 1:20Intelligence and
- 1:21Emerging Technology at the White House Office of Science and
- 1:25Technology Policy,
- 1:27SPEAKER_00: as well as a strategic advisor for AI at the National Science Foundation.
- 1:32He is also the author of the excellent Hyperdimensional Substack
- 1:35newsletter.
- 1:37There was this recent poll from a company called Blue Rose
- 1:41Research that said the
- 1:43arrival of AI is happening in a society that's already really
- 1:47anxious about its
- 1:48SPEAKER_00: financial future.
- 1:50SPEAKER_00: About half of people think AI is advancing too fast.
- 1:53SPEAKER_00: Only about a fifth are optimistic about its long-term social impact.
- 1:58SPEAKER_00: And like eight in 10 worry the government has no plan to protect workers.
- 2:04SPEAKER_00: You and I are both, I think, very optimistic about AI.
- 2:08SPEAKER_00: So is there any remedy for all this pessimism?
- 2:12SPEAKER_00: And does it really matter?
- 2:14SPEAKER_01: In terms of remedies for the pessimism,
- 2:18I think that some of these concerns are speculative concerns
- 2:26that there's really
- 2:27SPEAKER_01: not that much you can do about because of like,
- 2:30broader deterioration of trust in institutions and things like
- 2:34that that have
- 2:35SPEAKER_01: happened in America that there's not that much you can do about.
- 2:37There are other things that are more like immediate problems
- 2:42that people actually
- 2:43SPEAKER_01: do see in their lives that in principle you could pass laws about or that
- 2:48policymakers could talk about in a way that makes voters feel
- 2:51as though the
- 2:52SPEAKER_01: concerns are heard.
- 2:54But in general,
- 2:55I think that AI is kind of almost like a literary device for
- 2:59people.
- 2:59SPEAKER_01: And it's just generally like the future,
- 3:02SPEAKER_01: the big scary future in which lots of stuff happens that we don't like.
- 3:06SPEAKER_01: And we attribute a lot of it to AI.
- 3:08SPEAKER_01: So I think in some ways asking people,
- 3:10are you optimistic about AI is actually not that different of
- 3:13asking them,
- 3:13SPEAKER_01: like,
- 3:13SPEAKER_01: are you optimistic about sort of like the broad future?
- 3:16SPEAKER_01: Not the future of your particular life, but like the future of the world.
- 3:21SPEAKER_01: And I don't know how much there is you can do about that.
- 3:23The question of does it matter is a really interesting one
- 3:26because on the one hand...
- 3:28SPEAKER_01: Of course it matters,
- 3:29SPEAKER_01: you know,
- 3:30the enthusiasm with which our people adopt and perceive AI
- 3:36vis-a-vis other
- 3:37SPEAKER_01: countries where,
- 3:38SPEAKER_01: by the way,
- 3:38SPEAKER_01: people are much more optimistic in other countries.
- 3:41SPEAKER_01: That totally does matter, like a great deal.
- 3:43SPEAKER_01: It's like really important.
- 3:44SPEAKER_01: Yeah.
- 3:46SPEAKER_01: On the other hand,
- 3:48SPEAKER_01: if the technology is going to be as powerful as some people think,
- 3:53SPEAKER_01: and I increasingly count myself among those people,
- 3:56SPEAKER_01: well,
- 3:57SPEAKER_01: it's probably going to change the world a lot,
- 3:59SPEAKER_01: no matter what.
- 4:00SPEAKER_01: You know what I mean?
- 4:01SPEAKER_01: Probably people's general pessimism about it doesn't matter that much.
- 4:04SPEAKER_01: But again, I think that pessimism has a way of being self-fulfilling.
- 4:13And so to the extent we think it's going to go poorly,
- 4:16we might actually make it go worse.
- 4:19SPEAKER_01: So I do think it ultimately matters.
- 4:21SPEAKER_00: I have several concerns, but maybe the main concern.
- 4:25is that I am far more concerned about bad regulation with AI
- 4:29than I ever was with social media.
- 4:32I think it might've been Noah Smith who said that the message
- 4:36coming out about AI
- 4:39SPEAKER_00: from American businesses,
- 4:42you will soon be obsolete and you may be killed or enslaved by
- 4:47our product.
- 4:49SPEAKER_00: Not a great message, not a great value proposition to the public,
- 4:54SPEAKER_00: On the regulatory front,
- 4:55SPEAKER_00: does that that message,
- 4:57SPEAKER_00: granted,
- 4:57SPEAKER_00: it's exaggerated,
- 4:59SPEAKER_00: does that does that perception matter as far as regulation and potential
- 5:04SPEAKER_00: regulation?
- 5:06SPEAKER_01: I actually don't know how much it does, because like.
- 5:12It's not clear to me how many regulations have really been
- 5:16passed that attempt to
- 5:17SPEAKER_01: grapple with those kinds of issues.
- 5:20If you look at the regulations that I think are most pernicious
- 5:25in American law
- 5:27SPEAKER_01: that have been passed in the recent months and years,
- 5:32SPEAKER_01: For the most part,
- 5:33SPEAKER_01: they're about things like algorithmic bias and essentially protecting
- 5:38SPEAKER_01: occupationally licensed professions from AI competition.
- 5:42SPEAKER_01: So like mental health therapists and whatnot.
- 5:44SPEAKER_00: You cannot get legal advice from your chatbot.
- 5:48SPEAKER_01: Yeah, you can't get legal advice.
- 5:49SPEAKER_01: You can't get mental health advice, stuff like that.
- 5:52SPEAKER_01: And so like,
- 5:52I don't know how much of that is actually motivated or in any
- 5:56way like when people
- 5:58SPEAKER_01: go and say like,
- 5:58SPEAKER_01: yeah,
- 5:59SPEAKER_01: AI has these like existential implications for the future.
- 6:03SPEAKER_01: I don't really see the connection between that and,
- 6:05SPEAKER_01: oh,
- 6:05SPEAKER_01: we have to protect occupational licensed positions,
- 6:09SPEAKER_01: interest groups.
- 6:10SPEAKER_01: Look, does it create a climate that justifies all sorts of things?
- 6:14SPEAKER_01: For sure, it does.
- 6:16SPEAKER_01: And I think it's reasonable to argue that that's not good.
- 6:20But I don't actually know that we have that many laws being
- 6:22passed that are truly
- 6:23SPEAKER_01: motivated by those kinds of concerns.
- 6:27I think the people that earnestly have those concerns don't
- 6:29feel as though they
- 6:31don't feel as though any laws have been passed that deal with
- 6:34those things.
- 6:36You know,
- 6:36there were all these concerns during the late 2010s about
- 6:39social media.
- 6:40And,
- 6:41of course,
- 6:42Washington failed to pass any sort of big privacy or content
- 6:47moderation laws.
- 6:49SPEAKER_00: And at the time,
- 6:50SPEAKER_00: you had social media CEOs talking about the advantages of their platform,
- 6:56SPEAKER_00: bringing us together,
- 6:58SPEAKER_00: the great value people derive from them.
- 7:01SPEAKER_00: What you did not have,
- 7:03SPEAKER_00: which is what I think CEOs are doing now,
- 7:06SPEAKER_00: was people saying things like,
- 7:08SPEAKER_00: in 18 months,
- 7:09SPEAKER_00: basically every white-collar job will be able to be done by AI.
- 7:13SPEAKER_00: Now, I don't know if when CEOs say things like that, what they mean is,
- 7:19theoretically every job can be done or every job task behind a
- 7:23computer can be done
- 7:24SPEAKER_00: or whether it will actually be done in practice.
- 7:28SPEAKER_00: And I think that lack of clarity itself is a problem.
- 7:31SPEAKER_00: But why do CEOs keep doing that?
- 7:33I should see how that's helpful for their company or these
- 7:36products or what I hope
- 7:38SPEAKER_00: is an emerging AI revolution.
- 7:41SPEAKER_01: Look,
- 7:41I think that technologists should be careful about predicting
- 7:48too much of economic
- 7:52effects that are beyond the object level capabilities of the
- 7:56technology itself.
- 7:57Because as you said,
- 7:58there's a complex difference between can automate and does
- 8:01automate.
- 8:03SPEAKER_01: There's a really big difference between those two things, right?
- 8:05SPEAKER_01: There's a...
- 8:07SPEAKER_01: Usually when AI people talk about these things,
- 8:09SPEAKER_01: by the way,
- 8:10SPEAKER_01: what they mean,
- 8:11SPEAKER_01: not always,
- 8:11SPEAKER_01: but usually what they mean is like,
- 8:14SPEAKER_01: can automate.
- 8:15They mean a system that is capable of automating huge amounts
- 8:18of knowledge work,
- 8:19SPEAKER_01: not necessarily one that does automate huge amounts of knowledge work.
- 8:22But the difference between those things is enormous and
- 8:25involves all kinds of
- 8:26SPEAKER_01: Hayekian complexity that I think we shouldn't just wave off.
- 8:33SPEAKER_01: I do always hesitate when CEOs make these observations about,
- 8:38SPEAKER_01: like,
- 8:40SPEAKER_01: oh,
- 8:40SPEAKER_01: you know,
- 8:41SPEAKER_01: yeah,
- 8:42SPEAKER_01: it's going to be,
- 8:43SPEAKER_01: you know,
- 8:43SPEAKER_01: 80% of the jobs are going to get automated or whatever.
- 8:45SPEAKER_01: Like, you don't know that.
- 8:46SPEAKER_01: You don't know what that means.
- 8:48SPEAKER_01: And, like, you should make more bracketed, you know, types of...
- 8:57SPEAKER_01: However,
- 8:58SPEAKER_01: one thing I will say is like,
- 8:59I think this ultimately comes down to do you think AI is
- 9:04different from other kinds
- 9:07SPEAKER_01: of technologies,
- 9:08SPEAKER_01: right?
- 9:09Is it akin to the PC and the internet or is it something
- 9:12fundamentally different?
- 9:15SPEAKER_01: Yeah.
- 9:15Like,
- 9:16is this a new kind,
- 9:17is this like the next big like internet platform business,
- 9:20right?
- 9:20SPEAKER_01: Is it like mobile or cloud or social media or whatever?
- 9:23SPEAKER_01: Or is it like something qualitatively different and bigger and more,
- 9:28SPEAKER_01: you know,
- 9:29SPEAKER_01: more,
- 9:30SPEAKER_01: more,
- 9:31SPEAKER_01: you know,
- 9:31SPEAKER_01: impactful?
- 9:32SPEAKER_01: Yeah.
- 9:34SPEAKER_01: And I lean toward probably the latter.
- 9:36SPEAKER_01: I lean toward probably it is different than those other things.
- 9:38And so it's worth thinking about in somewhat different terms
- 9:42and also thinking
- 9:43SPEAKER_01: about the stakes in somewhat different terms.
- 9:44SPEAKER_01: And so,
- 9:45SPEAKER_01: you know,
- 9:46SPEAKER_01: one of the personal frustrations that I experienced,
- 9:49while I definitely don't think you should make claims about all
- 9:52the jobs are going
- 9:52SPEAKER_01: to be automated in 18 months,
- 9:55SPEAKER_01: I do think that we tend to bounce between these extremes.
- 9:59On the one hand,
- 10:00we say all the jobs are going to be automated in 18 months.
- 10:03SPEAKER_01: And on the other hand,
- 10:04SPEAKER_01: you have people saying,
- 10:06SPEAKER_01: no,
- 10:06it's purely going to be augmentative and there's nothing to
- 10:10worry about at all when
- 10:12SPEAKER_01: it comes to the jobs.
- 10:13SPEAKER_01: I don't think that resonates with people.
- 10:15SPEAKER_01: I don't think that's what people feel.
- 10:18SPEAKER_01: It could be true.
- 10:19SPEAKER_01: I think the odds of that view being true,
- 10:21SPEAKER_01: the,
- 10:22SPEAKER_01: like,
- 10:22truly there is nothing to worry about here view being true are
- 10:29higher than the odds
- 10:31SPEAKER_01: of it's going to automate all the jobs being true,
- 10:34SPEAKER_01: right?
- 10:35SPEAKER_01: But still, I don't know that that's, like, my modal...
- 10:38SPEAKER_01: projection, right?
- 10:40I think it's quite possible that there will be meaningful labor
- 10:42dislocation and
- 10:44that that will cause political problems and that it could cause
- 10:47economic problems.
- 10:48SPEAKER_01: And particularly in the near term during a kind of transitional period,
- 10:52SPEAKER_01: because you and I both know that it's possible to imagine
- 10:56automating things and then just like letting your labor force
- 11:00turn out by attrition
- 11:01or doing layoffs that like it's easier to imagine that
- 11:04happening quite quickly than
- 11:06like all of the new jobs that ai ultimately creates like that
- 11:09takes time to form
- 11:10because you need like new types of businesses to form new types
- 11:13of products new
- 11:14types of consumer desires to materialize and like these things
- 11:18take time and so
- 11:20SPEAKER_01: it's very possible to imagine that the negative
- 11:23SPEAKER_01: effects are front loaded and the positive are back loaded,
- 11:27SPEAKER_01: similar to how the early innings of the Industrial Revolution went.
- 11:30SPEAKER_00: Clearly,
- 11:31there is a difference between how the kinds of economists who
- 11:37might work at a Wall
- 11:38SPEAKER_00: Street Bank or the Fed or at lots of kind of Washington think tanks,
- 11:43how they think about AI as being a powerful new technology and
- 11:48the message that
- 11:49SPEAKER_00: you're getting from lots of technologists
- 11:53SPEAKER_00: economists and founders,
- 11:55SPEAKER_00: and I guess even some,
- 11:56and I think there are some economists if they're sort of
- 11:58affiliated with that
- 11:59SPEAKER_00: world.
- 12:00SPEAKER_00: They really do speak about the technology in two very different ways.
- 12:05SPEAKER_00: The best example,
- 12:06SPEAKER_00: as I experienced,
- 12:07SPEAKER_00: was a panel where you had an economist,
- 12:11SPEAKER_00: when he was asked about what AI could do,
- 12:14SPEAKER_00: he said,
- 12:15SPEAKER_00: well,
- 12:15SPEAKER_00: I can imagine a future where,
- 12:17SPEAKER_00: you know,
- 12:17SPEAKER_00: when you're having a doctor's appointment,
- 12:19SPEAKER_00: the AI is there transcribing
- 12:23transcribing the appointment and coming up with interesting
- 12:28things the doctor might
- 12:29SPEAKER_00: think about,
- 12:30SPEAKER_00: but the doctor can just focus on the patient.
- 12:32SPEAKER_00: Like that was the vision from the economist.
- 12:35And the vision from the AI company CEO,
- 12:38it was in 10 years,
- 12:41death will be optional.
- 12:43SPEAKER_00: So two very different views.
- 12:45SPEAKER_00: Where do you think the disconnect comes from?
- 12:48SPEAKER_01: Yeah, I think both people are missing things here.
- 12:51SPEAKER_01: So first of all, like
- 12:54The world where the doctor...
- 12:58Let's set regulatory obstacles aside for just a minute.
- 13:01SPEAKER_01: Let's pretend like we live in a completely unregulated world.
- 13:06The world where the AI purely just provides advice and
- 13:12transcription services and
- 13:13whatnot for the doctor is a world where there is some reason
- 13:17you want the human
- 13:19SPEAKER_01: doctor to be there with input.
- 13:23And what if it turns out that the AI is just smarter than the
- 13:27human doctor in every
- 13:28SPEAKER_01: appreciable way?
- 13:30SPEAKER_01: and you just don't want one, right?
- 13:32SPEAKER_01: So you can imagine there being a future where there's,
- 13:35SPEAKER_01: like,
- 13:38SPEAKER_01: imagine there is,
- 13:40SPEAKER_01: like,
- 13:41SPEAKER_01: a little,
- 13:41SPEAKER_01: you know,
- 13:42SPEAKER_01: stand on a street corner that you walk into that has,
- 13:45SPEAKER_01: like,
- 13:45SPEAKER_01: a sort of,
- 13:46SPEAKER_01: like,
- 13:46SPEAKER_01: a medical chair in it,
- 13:48SPEAKER_01: like the kind you would sit in in a physical.
- 13:50And you go in,
- 13:51and it's just,
- 13:52like,
- 13:52a totally automated system that,
- 13:54like,
- 13:55draws your blood,
- 13:57SPEAKER_01: And, you know, does an eye exam and does all these things.
- 14:01SPEAKER_01: And it costs like $5.
- 14:03And those results are analyzed by an AI system whose opinions
- 14:09and advice is
- 14:11SPEAKER_01: basically always better than the human doctors we have today.
- 14:16SPEAKER_01: Well, okay, like I can totally imagine that being a thing, right?
- 14:20SPEAKER_01: Obviously there's like regulatory obstacles to that,
- 14:23SPEAKER_01: but I can totally imagine that being a thing.
- 14:25SPEAKER_01: When it comes to like,
- 14:26SPEAKER_01: well,
- 14:26SPEAKER_01: okay,
- 14:27SPEAKER_01: like longevity is,
- 14:28SPEAKER_01: you know,
- 14:29SPEAKER_01: we're gonna,
- 14:30SPEAKER_01: death will be optional.
- 14:33SPEAKER_01: You know, I think that assumes a level of scientific progress.
- 14:38SPEAKER_01: It assumes unbounded scientific progress, right?
- 14:43One thing that I'm so fond of that Tyler Cowen always says is
- 14:46that people in DC
- 14:47think on the margins and people in Silicon Valley think in
- 14:51terms of infinities.
- 14:53So when people in Silicon Valley hear,
- 14:56oh,
- 14:57well,
- 14:58we will have automated biological research.
- 15:01SPEAKER_01: We'll have automated biologists.
- 15:03So there'll be millions and millions of biologists who are all
- 15:06genius level.
- 15:09who can be brought to bear on arbitrary problems and will have
- 15:11automated labs and
- 15:12SPEAKER_01: will be able to test everything,
- 15:14SPEAKER_01: then yeah,
- 15:15SPEAKER_01: it's like,
- 15:15SPEAKER_01: okay,
- 15:15SPEAKER_01: we cure every disease,
- 15:17SPEAKER_01: sure,
- 15:17SPEAKER_01: sure.
- 15:17SPEAKER_01: But like,
- 15:19what if there are some things that are just orders of magnitude
- 15:25more complex than
- 15:26SPEAKER_01: we think?
- 15:28SPEAKER_01: And it actually is just much harder.
- 15:30SPEAKER_01: And it's harder for some other reason,
- 15:33SPEAKER_01: or it's harder for regulatory reasons,
- 15:34SPEAKER_01: or it's just not possible.
- 15:36SPEAKER_01: Curing death, quote unquote, just might not be possible.
- 15:39And so I find myself somewhat in the middle where I look at
- 15:43where the systems are.
- 15:44And I probably think that like today's systems are better than
- 15:48most human doctors
- 15:49SPEAKER_01: most of the time.
- 15:50SPEAKER_01: Mm-hmm.
- 15:51So I'm not entirely persuaded that it's just going to be like
- 15:56pure augmentation and
- 15:58SPEAKER_01: like,
- 15:58SPEAKER_01: oh,
- 15:58SPEAKER_01: we're actually,
- 15:59SPEAKER_01: you know,
- 15:59SPEAKER_01: the demand for human doctors is going to go up.
- 16:01SPEAKER_01: Like, a little like, I don't know about that.
- 16:06SPEAKER_01: But at the same time,
- 16:08SPEAKER_01: like this idea that it just scales to infinity,
- 16:11SPEAKER_01: I kind of have some skepticism of that too.
- 16:15SPEAKER_00: I remember a couple of years ago,
- 16:16SPEAKER_00: there was that essay,
- 16:17SPEAKER_00: Situational Awareness by Leopold Ashenbrenner.
- 16:22And one of the things that was in that essay was that AI would
- 16:26become so advanced,
- 16:27SPEAKER_00: I guess you might at that point call it artificial general intelligence,
- 16:31SPEAKER_00: all the labs are going to get nationalized by the government.
- 16:34SPEAKER_00: Seemed a little science fictional,
- 16:35SPEAKER_00: but then we have this anthropic thing with the Defense Department,
- 16:39which some people could interpret as the first step toward a
- 16:42path where these labs
- 16:43SPEAKER_00: are effectively nationalized.
- 16:46SPEAKER_00: Is that the path we're on?
- 16:48SPEAKER_00: Quite possibly, yeah.
- 16:49SPEAKER_00: Quite critical of that decision.
- 16:51SPEAKER_01: I am extremely critical of that, right?
- 16:53SPEAKER_01: I'm extremely critical of that.
- 16:55SPEAKER_01: But if I were to just observe like the direction of travel,
- 17:00SPEAKER_01: it's a very scary direction of travel.
- 17:02SPEAKER_01: The little bit of background here without,
- 17:03SPEAKER_01: you know,
- 17:04I won't belabor the point on the whole anthropic Department of
- 17:09War controversy,
- 17:10SPEAKER_01: but basically the story is that- Thank you for saying Department of War,
- 17:15SPEAKER_00: your access to the Pentagon is maintained.
- 17:21SPEAKER_01: I actually,
- 17:22SPEAKER_01: it's funny,
- 17:22SPEAKER_01: when I first started covering this topic,
- 17:24SPEAKER_01: I would mess it up all the time.
- 17:26SPEAKER_01: And I would say DOD because I was so used to it,
- 17:28SPEAKER_01: but I've now gotten used to saying Department of War.
- 17:30SPEAKER_01: So I'm pretty consistent nowadays.
- 17:34But yeah,
- 17:35no,
- 17:35by the way,
- 17:36I think my access to the Department of War is very deeply
- 17:39constrained.
- 17:40SPEAKER_00: Not great at the moment.
- 17:40SPEAKER_00: Maybe not great at the moment.
- 17:42SPEAKER_01: Yeah, yeah, yeah.
- 17:43SPEAKER_00: It may take more than a phrase here or two to help that, but continue.
- 17:48SPEAKER_00: Yeah.
- 17:48SPEAKER_01: I think that's probably so, yes.
- 17:49SPEAKER_01: Basically, the Department of War decided that they...
- 17:55They had a contract with Anthropic to use Anthropic models in
- 17:58classified context,
- 17:59which might include like active combat and intelligence
- 18:02analysis and things like
- 18:03SPEAKER_01: this.
- 18:04SPEAKER_01: And that contract said that there are some usage restrictions,
- 18:08SPEAKER_01: including on mass domestic surveillance and autonomous lethal weapons.
- 18:12The Department of War had agreed to those terms and then
- 18:15decided later on that
- 18:16SPEAKER_01: actually no,
- 18:17SPEAKER_01: they wanted a contract that allowed for just all lawful use.
- 18:22SPEAKER_01: So no usage restrictions imposed by Anthropic.
- 18:25Anthropic ultimately said,
- 18:26no,
- 18:27these usage restrictions are important to us.
- 18:29SPEAKER_01: And the Department of War said, great.
- 18:32SPEAKER_01: If that's true...
- 18:34SPEAKER_01: then we are going to designate your company a supply chain risk.
- 18:38SPEAKER_01: And we're going to ban both all of the Department of War itself,
- 18:41but also all Department of War contractors from using anthropic
- 18:45technology.
- 18:46SPEAKER_01: This is a very severe punishment.
- 18:48SPEAKER_01: It's the equivalent of basically debarment from federal contracting.
- 18:52SPEAKER_01: It's actually in some ways more severe than that.
- 18:55SPEAKER_01: And this is a very severe penalty to pay.
- 19:01SPEAKER_01: And...
- 19:03SPEAKER_01: So in some ways,
- 19:05it seems to me as though what you're looking at here is kind of
- 19:11equivalent to...
- 19:14It's like pushing in the direction of nationalization because
- 19:17the government is
- 19:18SPEAKER_01: basically saying,
- 19:19SPEAKER_01: well...
- 19:20SPEAKER_01: you know,
- 19:20SPEAKER_01: you can't set terms of usage of your product,
- 19:25SPEAKER_01: not just with us,
- 19:26SPEAKER_01: but also with any of our contractors.
- 19:28SPEAKER_01: And of course there are,
- 19:29SPEAKER_01: I don't know how many Department of War contractors there are,
- 19:32SPEAKER_01: but like,
- 19:33SPEAKER_01: tens of thousands,
- 19:34SPEAKER_01: hundreds of thousands,
- 19:36probably like most Fortune 500 companies in America in some way
- 19:40or another.
- 19:41SPEAKER_01: Apple and Microsoft,
- 19:42SPEAKER_01: Google,
- 19:43SPEAKER_01: these are all,
- 19:44SPEAKER_01: Amazon,
- 19:44SPEAKER_01: these are all Department of War contractors,
- 19:46SPEAKER_01: right?
- 19:46SPEAKER_01: A huge number of people are contractors for the government here.
- 19:50SPEAKER_01: And like,
- 19:51SPEAKER_01: if that's the case,
- 19:54that's a quite broad regulation and it's kind of an assertion
- 19:57of public authority
- 19:58SPEAKER_01: over what we typically think of as private property.
- 20:02SPEAKER_01: And so, yeah, I think this is just the beginning of it.
- 20:05SPEAKER_01: During these negotiations between DOW and Anthropic,
- 20:09the Department of War has threatened to invoke the Defense
- 20:13Production Act,
- 20:14SPEAKER_01: which is basically saying,
- 20:16SPEAKER_01: well,
- 20:16SPEAKER_01: if you don't want to provide us with the model that we want,
- 20:18SPEAKER_01: we'll make you do it.
- 20:20SPEAKER_01: will make you provide us with this model.
- 20:22I mean,
- 20:23that is way more explicit about,
- 20:24like,
- 20:25this is actually just nationalization.
- 20:28SPEAKER_01: And so, yeah, all that feels quite concerning to me.
- 20:35SPEAKER_01: I don't think it's inevitable.
- 20:37But I do think these kinds of conversations and debates are
- 20:41going to play a really
- 20:42SPEAKER_01: big role over the coming years and probably a decade.
- 20:46SPEAKER_01: And I think it's certainly in the cards.
- 20:48SPEAKER_01: I don't think it's crazy to think about nationalization.
- 20:51I think it's a terrible idea,
- 20:53but I don't think it's outside the Overton window anymore.
- 20:56I've been reading more and more wargaming kinds of scenarios
- 21:01involving AI.
- 21:02SPEAKER_00: And you might want to even call AI at that point
- 21:06SPEAKER_00: artificial general intelligence or super intelligence.
- 21:08And these war games really show that this technology could be
- 21:11amazingly decisive in
- 21:14SPEAKER_00: a military conflict.
- 21:16SPEAKER_00: And with a potential weapon that powerful,
- 21:20SPEAKER_00: it makes me think these nationalization scenarios are realistic.
- 21:24SPEAKER_00: They don't seem very science fictional to me at all.
- 21:26SPEAKER_00: Now, maybe you have to assume that the technology is that possible.
- 21:29SPEAKER_00: I'm not sure if you think how realistic you think that is,
- 21:32SPEAKER_00: if that's where we're at or if that's where we'll be in two years.
- 21:35No,
- 21:35I think like very dramatic capability enhancements are totally,
- 21:41totally possible.
- 21:42SPEAKER_01: I think I'd say a couple of things to that.
- 21:44I think the first thing I would say is think about the
- 21:50following technologies,
- 21:51SPEAKER_01: the printing press,
- 21:53SPEAKER_01: banking,
- 21:54SPEAKER_01: or financial services in general,
- 22:00SPEAKER_01: the computer,
- 22:02SPEAKER_01: electricity,
- 22:03SPEAKER_01: internal combustion.
- 22:04SPEAKER_01: These are all things that
- 22:07SPEAKER_01: it's essentially impossible to imagine the modern military without.
- 22:13If you took electricity away from the military,
- 22:16we wouldn't really have a military anymore.
- 22:18So it's possible for a technology to be utterly foundational
- 22:24for all military
- 22:26SPEAKER_01: capability while also not really being a weapon first and foremost.
- 22:31SPEAKER_01: So that's an interesting paradox, right?
- 22:33SPEAKER_01: Like, in a way...
- 22:36SPEAKER_01: or microelectronics and computers are another good example.
- 22:39SPEAKER_01: It's almost impossible to imagine the military without computers,
- 22:41SPEAKER_01: but we don't think of computers as weapons,
- 22:44SPEAKER_01: yet in some sense,
- 22:44SPEAKER_01: they're some of the most dangerous weapons we've ever built.
- 22:50SPEAKER_01: I think AI is like that.
- 22:52SPEAKER_01: Broadly speaking, I think AI is like that.
- 22:55SPEAKER_01: It's not literally a weapon.
- 22:56It doesn't literally kill people,
- 22:58but it can enable capabilities when it is adopted well.
- 23:03It can enable organizations to do things that were simply
- 23:05impossible before.
- 23:07And those things are going to be table stakes for participating
- 23:10in the future.
- 23:11SPEAKER_01: And that'll be true of corporations.
- 23:13SPEAKER_01: It'll be true of individuals.
- 23:15SPEAKER_01: It'll be true of think tanks.
- 23:17SPEAKER_01: And it'll be true of militaries.
- 23:19SPEAKER_01: So that's how I would see that.
- 23:21SPEAKER_01: You know, the other thing I think...
- 23:24SPEAKER_01: that follows from that in many ways is,
- 23:26SPEAKER_01: um,
- 23:29one big difference between like the atomic bomb or a missile or
- 23:32a fighter jet is
- 23:34SPEAKER_01: that,
- 23:34SPEAKER_01: um,
- 23:34SPEAKER_01: those technologies are not indelible parts of how you and me,
- 23:40SPEAKER_01: or you and I,
- 23:41SPEAKER_01: um,
- 23:45SPEAKER_01: express our liberty and communicate and think.
- 23:48SPEAKER_01: Whereas AI is this like really important cognitive tool that already,
- 23:54SPEAKER_01: like already,
- 23:54SPEAKER_01: I just can't imagine how I would do my job,
- 23:58SPEAKER_01: which is to literally,
- 23:59SPEAKER_01: my job is to exercise my First Amendment rights.
- 24:02SPEAKER_01: That's my job, that's what I get paid to do.
- 24:04SPEAKER_01: I can't see how I would do that without the benefit of AI systems.
- 24:10SPEAKER_01: And already that's true for me.
- 24:12SPEAKER_01: I think it'll be like even like just profoundly more true in the future.
- 24:16And so the notion that that system should be controlled by the
- 24:22government,
- 24:23SPEAKER_01: I think we want to just be extremely careful about that.
- 24:26Even if the control is regulatory in nature as opposed to like
- 24:30nationalization.
- 24:31I think even this is why we need to be so cautious on the
- 24:34regulatory front.
- 24:37And then I think the final thing is that if you step back one
- 24:40more step and think
- 24:41SPEAKER_01: about the AI system itself and what is it,
- 24:47ultimately what it's going to be is an agent of mine that
- 24:52represents me in
- 24:54SPEAKER_01: interactions with the world,
- 24:55SPEAKER_01: that makes decisions on my behalf,
- 24:57SPEAKER_01: and that helps inform my own decisions.
- 25:00SPEAKER_01: And...
- 25:01I think the alignment of those things to particular sets of
- 25:04values and philosophies
- 25:06SPEAKER_01: and morals is deeply important.
- 25:08SPEAKER_01: But then we get into the question of,
- 25:10SPEAKER_01: oh,
- 25:10which set of ethics and morals and philosophical principles
- 25:15should it be aligned
- 25:16SPEAKER_01: to?
- 25:17And the second you're asking that question,
- 25:20you are deeply inside of First Amendment territory.
- 25:24And it starts to become clear that the act of making an AI
- 25:27system is the First Amendment act.
- 25:30And so one thing that follows from that is like the government
- 25:32totally has First
- 25:33SPEAKER_01: Amendment rights.
- 25:35SPEAKER_01: So the government can maybe make its own AI system,
- 25:38but I don't think that they can tell other people what sorts of
- 25:42values AI systems
- 25:45SPEAKER_01: can and can't have.
- 25:46I think basically we answered that question 250 years ago with
- 25:51a decisive no in the
- 25:53SPEAKER_01: form of the First Amendment.
- 25:55If I have this right,
- 25:55you would like to see these labs have a mandatory audit.
- 25:59SPEAKER_00: You would like to see some sort of independent,
- 26:03mandatory private auditors making sure that whatever safety
- 26:08plans they have,
- 26:09either is it whatever safety plans they have that they follow
- 26:13them or that they're
- 26:14SPEAKER_00: following government safety plans?
- 26:15SPEAKER_01: You know, I have recently spoken out in favor of
- 26:21Basically requiring frontier labs to undergo mandatory audits
- 26:29for specifically
- 26:31SPEAKER_01: whether or not they follow their safety,
- 26:35SPEAKER_01: their stated safety plans.
- 26:37SPEAKER_01: So what's the reasoning for this?
- 26:38SPEAKER_01: So...
- 26:39um the reasoning for this is that there are a handful of risks
- 26:45that again we have a
- 26:46plausible threat model and we also have like pretty substantial
- 26:48uncertainty about
- 26:50like what to do about them so like two of them two of those
- 26:53risks are things like
- 26:55cyber attacks right like autonomous cyber attacks which we're
- 26:58starting to see
- 26:59happen we're starting to see people using ai systems for for
- 27:03these catastrophic
- 27:07SPEAKER_01: On the other hand,
- 27:08there's this thing that sounds like science fiction that's
- 27:11called recursive
- 27:12SPEAKER_01: self-improvement,
- 27:14SPEAKER_01: which is where the AI models get good enough to train future versions of
- 27:18SPEAKER_01: themselves.
- 27:20SPEAKER_01: And there has been,
- 27:21SPEAKER_01: I think,
- 27:23SPEAKER_01: one of the things,
- 27:24SPEAKER_01: you know,
- 27:24kind of what I was alluding to earlier about like a frustration
- 27:27of mine has been
- 27:28SPEAKER_01: that maybe some of these things don't get taken,
- 27:31SPEAKER_01: people don't get taken at face value when they say things like this.
- 27:35SPEAKER_01: It's like,
- 27:37Recursive self-improvement has been treated as a science
- 27:39fiction matter for a long time.
- 27:41SPEAKER_01: And my view has been like, this is really not science fiction.
- 27:43SPEAKER_01: This is like really quite seriously going to be a thing.
- 27:46SPEAKER_01: And now it's starting to become a thing.
- 27:48SPEAKER_01: And we have...
- 27:50SPEAKER_01: Frontier Labs that have said,
- 27:51SPEAKER_01: and OpenAI has said that,
- 27:54SPEAKER_01: you know,
- 27:55SPEAKER_01: GPT 5.3 and 5.4 were sort of created by their own models and,
- 28:01SPEAKER_01: you know,
- 28:01SPEAKER_01: helped not exclusively,
- 28:03SPEAKER_01: but that their models were like helping in really,
- 28:05SPEAKER_01: really big ways.
- 28:06SPEAKER_01: And I think that's just going to,
- 28:07SPEAKER_01: every time the model gets better,
- 28:09SPEAKER_01: it's able to help more and more.
- 28:10SPEAKER_01: And so that's why it's called recursive self-improvement.
- 28:14SPEAKER_01: We don't know where that goes.
- 28:16We don't know where it goes,
- 28:17and there's a couple different things that we don't understand.
- 28:20SPEAKER_01: One of them is like, okay, so...
- 28:25do we trust the ai systems that are making the future ai
- 28:28systems because they could
- 28:29get quite capable quite quickly because this is one of those
- 28:32areas where like the
- 28:34the issue of infinites and on the margin it actually might be
- 28:37more appropriate to
- 28:38think in something closer to infinities right because this
- 28:41isn't something that
- 28:42SPEAKER_01: gets bottlenecked by that many real world processes yes there are compute
- 28:45limitations um which could end up being a really big deal but
- 28:49in the end like
- 28:52OpenAI's workforce goes from single-digit thousands to tens of
- 28:57thousands to
- 28:57SPEAKER_01: hundreds of thousands to millions.
- 28:59SPEAKER_01: And we've never seen something like that happen before.
- 29:03SPEAKER_01: And the research that they do is purely digital.
- 29:06SPEAKER_01: So that's one thing.
- 29:08The other thing is that there's all these interesting
- 29:10geostrategic implications that get raised.
- 29:12SPEAKER_01: And even almost like antitrust and FTC-style questions.
- 29:18if all of a sudden AI systems are able to make better versions
- 29:21of themselves,
- 29:24SPEAKER_01: does that mean that like someone
- 29:27SPEAKER_01: can start from nothing.
- 29:30SPEAKER_01: You can start from scratch with one of these systems and all of a sudden,
- 29:33like build a really a much more powerful system just with
- 29:36access to computing
- 29:37SPEAKER_01: power.
- 29:38SPEAKER_01: Because if that's true,
- 29:40SPEAKER_01: then all of a sudden,
- 29:41SPEAKER_01: like you don't need all the talent that America has,
- 29:44SPEAKER_01: right?
- 29:45SPEAKER_01: Like any country on earth,
- 29:47SPEAKER_01: including France or North Korea in principle,
- 29:51SPEAKER_01: if they have access to the computing power,
- 29:53SPEAKER_01: they can,
- 29:55SPEAKER_01: They can run millions of agents and have a model that's,
- 29:58SPEAKER_01: you know,
- 29:58SPEAKER_01: maybe not quite as good,
- 29:59SPEAKER_01: but close to as good as open AI is the best stuff we can produce.
- 30:03That would really,
- 30:04really change the contours of like AI competition,
- 30:07right?
- 30:08SPEAKER_01: So there's all these questions we have,
- 30:10SPEAKER_01: and you just can't know without an inside look at the labs.
- 30:16SPEAKER_01: And so I have proposed this notion of like,
- 30:19SPEAKER_01: you know,
- 30:19SPEAKER_01: when I auditing is like a bad word for it almost,
- 30:22but it's like some combination of like an expert think tank
- 30:26that has that like
- 30:29works in collaboration with the government and formulates
- 30:32research questions in
- 30:33SPEAKER_01: collaboration with the government,
- 30:35SPEAKER_01: but that is independent of the government.
- 30:37SPEAKER_01: And it basically goes in and kind of like, um,
- 30:41SPEAKER_01: you know,
- 30:41it basically works in such a way that like the frontier labs
- 30:46are not grading their
- 30:47SPEAKER_01: own work.
- 30:48And we're also asking some questions that are salient to the
- 30:50public interest.
- 30:52SPEAKER_01: And that's like the idea behind this.
- 30:54I think that this is not something that's intended to be a
- 30:57massive regulatory regime.
- 30:59SPEAKER_01: Certainly it's a requirement that I wouldn't
- 31:04SPEAKER_01: put on social media companies.
- 31:05SPEAKER_01: It's a requirement I wouldn't put on smartphone companies, right?
- 31:08SPEAKER_01: Like I've largely argued against,
- 31:10SPEAKER_01: I've been a more or less total libertarian on like the vast majority of
- 31:14SPEAKER_01: technologies.
- 31:16But this is one where I think the implications here are serious
- 31:21enough that it's
- 31:21SPEAKER_01: worth sort of minimum viable regulations,
- 31:24SPEAKER_01: I guess is how I would put it.
- 31:25SPEAKER_00: What kind of feedback have you gotten on this idea?
- 31:30SPEAKER_01: Among people that
- 31:32SPEAKER_01: think seriously about this type of stuff,
- 31:36SPEAKER_01: in truth,
- 31:37SPEAKER_01: I think it's kind of where everyone has landed.
- 31:41SPEAKER_01: So I first proposed this notion a little more than a year ago,
- 31:45SPEAKER_01: and I called it a framework for private governance.
- 31:49I called this private governance as opposed to auditing or
- 31:52something like that.
- 31:53SPEAKER_01: I think private governance is a little bit closer to what
- 31:58SPEAKER_01: to what we're looking for.
- 31:58SPEAKER_01: And the idea is like,
- 31:59these would be entities that would provide sort of regulatory
- 32:04services to the
- 32:06SPEAKER_01: government,
- 32:06SPEAKER_01: but who would ultimately be licensed by the government,
- 32:11SPEAKER_01: but independent of the government.
- 32:14SPEAKER_01: And the reason that I came to this conclusion is that I realized,
- 32:19SPEAKER_01: okay,
- 32:19SPEAKER_01: well,
- 32:19SPEAKER_01: there's going to need to be some kind of a regulatory apparatus here.
- 32:23But the problem is that there are these structural incentives
- 32:27that in particular,
- 32:28SPEAKER_01: the national security state has that lean in the direction of very,
- 32:31SPEAKER_01: very heavy regulatory regimes,
- 32:34SPEAKER_01: either like arms control style regulations,
- 32:36SPEAKER_01: which is how like missile producers are regulated or like straight up
- 32:39SPEAKER_01: nationalization,
- 32:40SPEAKER_01: things like this.
- 32:41SPEAKER_01: And I wanna avoid those outcomes,
- 32:42SPEAKER_01: but I also think that A,
- 32:44I think the purely decentralized libertarian model is quite
- 32:47plausibly bad on the
- 32:48SPEAKER_01: merits,
- 32:49SPEAKER_01: but regardless of whether or not it's bad,
- 32:51SPEAKER_01: I think it's not possible.
- 32:52So there is this hybrid that we're going to have to walk is the
- 32:55conclusion that I've come to.
- 32:57SPEAKER_01: And I've just, I've decided that this is the minimum viable way to do it.
- 33:01SPEAKER_01: And what's funny is that when I came up with this proposal a year ago,
- 33:05SPEAKER_01: libertarians were like,
- 33:06SPEAKER_01: this is too regulatory.
- 33:08SPEAKER_01: And AI safety people were like, this isn't regulatory enough.
- 33:13And in the intervening year,
- 33:15a lot of them have come to me and been like,
- 33:17you know,
- 33:18um,
- 33:21This actually does kind of seem like the sort of natural point
- 33:25of convergence.
- 33:26And so I feel as though the momentum for ideas like this has
- 33:29grown considerably on
- 33:30SPEAKER_01: both the right and the left in the last year.
- 33:33SPEAKER_01: Again, among people that are paying attention specifically to AI policy.
- 33:37SPEAKER_01: I don't know that this is like...
- 33:38SPEAKER_00: I had the analyst Miles Brundage on the podcast a while ago,
- 33:41SPEAKER_00: and he has an idea.
- 33:42It seems some...
- 33:43There might be differences I'm not picking up right now between
- 33:47your two plans.
- 33:47SPEAKER_00: He also wants...
- 33:49SPEAKER_00: mandatory, you know, third party audits.
- 33:52And it seems so you might be right that this is where a lot of
- 33:55people are,
- 33:56SPEAKER_00: are sort of landing,
- 33:57SPEAKER_00: especially,
- 33:57SPEAKER_00: you know,
- 33:57SPEAKER_00: whether you worry about no,
- 33:59SPEAKER_00: whether you worry that there's no regulation or what you worry that,
- 34:02SPEAKER_00: will have something like nationalization.
- 34:03SPEAKER_00: It seems like a middle ground worth talking about.
- 34:06And by the way,
- 34:07full disclosure,
- 34:08I'm actually on the board of advisors of Miles' organization.
- 34:11SPEAKER_01: Oh, well, there you go.
- 34:13SPEAKER_01: Yeah, we do have some disagreements, but largely substantively agree.
- 34:18SPEAKER_00: Thank you for mentioning that.
- 34:19SPEAKER_00: One thing I wanted to ask,
- 34:21SPEAKER_00: Jason Furman,
- 34:22SPEAKER_00: the economist,
- 34:22SPEAKER_00: he did this thing on Twitter or X where he had people sort of
- 34:28SPEAKER_00: give their probabilities for what's the productivity impact for AI.
- 34:33SPEAKER_00: Is it going to be sort of stagnation?
- 34:34SPEAKER_00: So it's not much impact.
- 34:36SPEAKER_00: Is it everywhere?
- 34:37SPEAKER_00: Is this going to be super intelligence and we're going to,
- 34:40SPEAKER_00: it's going to be,
- 34:40SPEAKER_00: you know,
- 34:41SPEAKER_00: 30% a year or something like that.
- 34:43SPEAKER_00: I didn't want to,
- 34:44I'm not going to ask you to give me all these percentages and
- 34:45different scenarios,
- 34:46SPEAKER_00: but I did want to ask you this.
- 34:49SPEAKER_00: Is there a pace of productivity growth that's so fast,
- 34:54SPEAKER_00: and maybe this would be a nice problem to have,
- 34:56SPEAKER_00: that would make you uncomfortable?
- 34:58SPEAKER_00: That would be too fast.
- 35:00SPEAKER_00: That would cause too much social chaos.
- 35:02SPEAKER_00: Again,
- 35:03SPEAKER_00: not a problem we've had lately,
- 35:05SPEAKER_00: but certainly when you listen to some technologists,
- 35:08SPEAKER_00: they talk about very,
- 35:09SPEAKER_00: very fast economic growth.
- 35:11SPEAKER_00: And is there a...
- 35:13SPEAKER_00: Is there a too fast number for you?
- 35:15Is 3% okay,
- 35:17but 8% would be,
- 35:18you know,
- 35:19crazy go nuts,
- 35:20something in between?
- 35:21SPEAKER_00: Have you thought about that?
- 35:22SPEAKER_01: Yeah.
- 35:22SPEAKER_01: Oh, yeah.
- 35:23SPEAKER_01: So, I mean, 8% is great still, right?
- 35:25SPEAKER_01: 8% is still like where, you know, the...
- 35:29uh we haven't reached scary zone yet but when you start to get
- 35:32into the really
- 35:33crazy numbers about productivity or GDP growth um it seems to
- 35:37me that in practice
- 35:39it's impossible to imagine that kind of growth without
- 35:42basically private property
- 35:43seizure that sort of seems like a world where um we are uh
- 35:49eminent domaining
- 35:50SPEAKER_01: people's houses and like
- 35:53SPEAKER_01: building solar farms and data centers,
- 35:55SPEAKER_01: you know,
- 35:55SPEAKER_01: or nuclear fusion plants and data centers or whatever.
- 35:59SPEAKER_01: But yeah,
- 36:00SPEAKER_01: no,
- 36:00SPEAKER_01: I mean,
- 36:01SPEAKER_01: generally speaking,
- 36:01SPEAKER_01: like,
- 36:02SPEAKER_01: I do think it's possible to have growth that's too fast.
- 36:05SPEAKER_01: I don't expect that.
- 36:06SPEAKER_01: I'm not in the camp of people that expects 30% economic growth.
- 36:11I also am not in the sort of...
- 36:15I'm not in the camp that expects very modest growth.
- 36:20SPEAKER_01: I would say I'm in the, like...
- 36:23SPEAKER_01: I think annualized,
- 36:26SPEAKER_01: looking back 10 or 15 years from now,
- 36:29SPEAKER_01: we will say that AI,
- 36:31SPEAKER_01: including the data centers,
- 36:33SPEAKER_01: but also the things that...
- 36:35SPEAKER_01: So there's the GDP growth that comes from enabling AI,
- 36:39SPEAKER_01: and then there's the GDP growth that comes from what AI enables.
- 36:42SPEAKER_01: And thus far, all we've seen really is the former.
- 36:45SPEAKER_01: We haven't really seen the latter.
- 36:46SPEAKER_01: But I think when you add those things up,
- 36:49SPEAKER_01: Looking back 10 or 15 years,
- 36:51we'll look back and we'll say the annual additional GDP growth
- 36:55upon sort of
- 36:57SPEAKER_01: over-the-counter factual without AI is about two and a half percent.
- 37:01So I expect AI to add about two to two and a half percent of
- 37:04GDP growth per year
- 37:06SPEAKER_01: over the next 10 to 15 years.
- 37:09SPEAKER_01: That's really aggressive, by the way.
- 37:10SPEAKER_00: That's a very aggressive.
- 37:12That would be quite,
- 37:14you know,
- 37:14when you're talking about,
- 37:15you know,
- 37:16most long-term forecasts,
- 37:19SPEAKER_00: Pre-AI, Gen AI forecasts were like 1.8% economic growth overall.
- 37:28SPEAKER_00: It's quite an acceleration.
- 37:31SPEAKER_01: I'm forecasting an annualized GDP growth regime in the 4% to 6% range.
- 37:34SPEAKER_01: At what point?
- 37:39Will we look at the world around us and there will be no debate
- 37:45that this is a
- 37:46SPEAKER_00: powerful technology with a sweeping economic impact?
- 37:51SPEAKER_00: So I'm not asking you like, when is AGI?
- 37:55SPEAKER_00: But like, when will, when you look around, like that debate will be over?
- 37:58SPEAKER_00: Because clearly some people think it's,
- 38:01They do wonder because they're not seeing the impacts and
- 38:03there's like pilot
- 38:04SPEAKER_00: programs and there's,
- 38:05SPEAKER_00: yeah,
- 38:06SPEAKER_00: we're spending a lot on infrastructure,
- 38:08SPEAKER_00: but,
- 38:08SPEAKER_00: you know,
- 38:08SPEAKER_00: is it paying off?
- 38:09SPEAKER_00: When will sort of the payoff question be over?
- 38:12SPEAKER_00: And we'll know like,
- 38:13SPEAKER_00: yes,
- 38:14SPEAKER_00: this is having a big,
- 38:16SPEAKER_00: big impact,
- 38:17SPEAKER_00: not just on productivity growth,
- 38:18SPEAKER_00: but on the bottom lines of companies everywhere.
- 38:20SPEAKER_01: It's a much better question to ask because my view is that,
- 38:26SPEAKER_01: and I've said this publicly,
- 38:27SPEAKER_01: is that I actually think that current systems satisfy many commonly used
- 38:34SPEAKER_01: definitions of artificial general intelligence.
- 38:37SPEAKER_01: No one's built super intelligence yet,
- 38:38but I think we have systems that satisfy at least some
- 38:41definitions of artificial
- 38:42SPEAKER_01: general intelligence that are commonly used in the field,
- 38:45SPEAKER_01: OpenAI's definition in particular.
- 38:46SPEAKER_01: I think we've effectively satisfied that definition.
- 38:49And some people would say we hit AGI,
- 38:51by the way,
- 38:52three years ago,
- 38:53right?
- 38:54SPEAKER_01: So the question of like,
- 38:59when is it going to be a social fact as opposed to a
- 39:02technological fact is a really
- 39:05SPEAKER_01: interesting one.
- 39:07SPEAKER_01: My guess would be that it will be in about...
- 39:12would say two to three years from now feels right to me um that
- 39:16it will be like
- 39:17unquestionable it could be sooner i think it's already
- 39:20happening to some extent um
- 39:23like i do think that the department of war thing has changed
- 39:27the tenor in a pretty
- 39:29dramatic way and so like that's one good example but yeah um uh
- 39:34uh i think
- 39:35SPEAKER_01: basically two to two to three years
- 39:38SPEAKER_00: And where do you think we ultimately land on regulation?
- 39:41SPEAKER_00: Is the technology moving much sort of within Congress's decision circle?
- 39:46SPEAKER_00: It's moving much faster than the debate?
- 39:49SPEAKER_00: Or do you think ultimately there is some...
- 39:51SPEAKER_00: Washington does something,
- 39:52SPEAKER_00: and what is that something?
- 39:54SPEAKER_01: Washington will ultimately do something.
- 39:56I think we're sufficiently far out from that happening that
- 39:59it's hard for me to
- 39:59SPEAKER_01: predict the substance of it.
- 40:01SPEAKER_01: But I think in broad strokes, what I would say is like...
- 40:09I predict that it will involve some relationship between the
- 40:13government and the
- 40:14SPEAKER_01: Frontier Labs and some sort of structured relationship,
- 40:17SPEAKER_01: some sort of new institutional mechanisms that get built to structure the
- 40:24SPEAKER_01: relationship between the government and the Frontier Labs.
- 40:27SPEAKER_01: I think that's basically what we will build, and I would hope...
- 40:33SPEAKER_01: that as that happens,
- 40:35we realize that like institutions themselves can be structured
- 40:44in different ways
- 40:45SPEAKER_01: based on technology.
- 40:47SPEAKER_01: And so we can use technology to build different kinds of institutions.
- 40:52And that would kind of be my hope is that we start to work in
- 40:55that direction.
- 40:56SPEAKER_00: I led off with that gloomy poll.
- 40:58SPEAKER_00: And again, let me just remind you of one of the results.
- 41:0018% of people said they're optimistic about the long-term
- 41:03social impact of AI.
- 41:05If you're optimistic,
- 41:06and I'm going to assume you're optimistic,
- 41:08why are you optimistic?
- 41:10SPEAKER_01: I am a techno-optimist and a human optimist,
- 41:13which means that I think that humans are adaptable and more
- 41:16ingenious than we give
- 41:17SPEAKER_01: ourselves credit for.
- 41:19I think that technology itself is going to be this wonderful
- 41:22enabler of all kinds of people.
- 41:24SPEAKER_01: In the fullness of time,
- 41:25I think we are going to be able to realize an entirely
- 41:30different kind of quality of
- 41:33SPEAKER_01: life,
- 41:33SPEAKER_01: a whole new level of quality of life,
- 41:35SPEAKER_01: a whole new level of fulfillment.
- 41:37SPEAKER_01: We're going to be able to do things like explore the solar system,
- 41:41SPEAKER_01: and that will become much more common.
- 41:44SPEAKER_01: We'll all become richer and smarter and healthier.
- 41:50SPEAKER_01: And I think that that will feel qualitatively.
- 41:53It will be an incredible world that we're going to build that
- 41:57will make us all today
- 42:00SPEAKER_01: We will look back and we will feel poor.
- 42:02SPEAKER_01: We'll look back at the way we live right now.
- 42:04SPEAKER_01: And right now,
- 42:04SPEAKER_01: I think I live,
- 42:06SPEAKER_01: compared to the vast majority of humans that have ever lived,
- 42:08SPEAKER_01: I live like a king,
- 42:10SPEAKER_01: even though I'm a middle-class person,
- 42:12SPEAKER_01: right?
- 42:13SPEAKER_01: I'm not some extraordinary wealth, but I live like a king.
- 42:18SPEAKER_01: compared to almost everyone who's ever lived.
- 42:20SPEAKER_01: And we'll look back and we'll feel that way about ourselves,
- 42:23SPEAKER_01: the way we feel about someone who lived a hundred years ago.
- 42:26SPEAKER_01: And not so long from now,
- 42:27SPEAKER_01: I think we'll feel about ourselves,
- 42:29SPEAKER_01: you know,
- 42:29SPEAKER_01: it's a couple of decades from now.
- 42:31SPEAKER_01: And so I think it's going to be fantastic,
- 42:34SPEAKER_01: but I don't dispute the notion that,
- 42:37SPEAKER_01: you know,
- 42:38SPEAKER_01: we have,
- 42:41SPEAKER_01: I'm an institutional pessimist.
- 42:43And so I think we have some real political and sort of
- 42:48governance bumpiness ahead of us.