YT YouTube2Text

YouTube2Text → Transcripts → Anthropic's CEO: ‘We Don’t Know if the Models Are Conscious’ | Interesting Times with Ross Douthat

Anthropic's CEO: ‘We Don’t Know if the Models Are Conscious’ | Interesting Times with Ross Douthat — Transcript

by Interesting Times · 10,252 words · 1,363 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00I want to try and focus on scenarios where A.I. goes rogue.
0:03I should have had a picture of a Terminator robot
0:06to scare people as much as possible.
0:07I think the internet...
0:09The internet does that for us.
0:16Are the lords of artificial intelligence
0:19on the side of the human race?
0:21"My prediction is there’ll be more robots than people."
0:24"The physical and the digital worlds
0:26should really be fully blended."
0:28"I don’t think the world has really had the humanoid robots
0:30moment yet.
0:31It’s going to feel very sci-fi."
0:33That’s the core question I had for this week’s guest.
0:36He’s the head of Anthropic, one of the fastest growing A.I.
0:39companies.
0:40Anthropic is estimated to be worth nearly $350 billion.
0:45It’s been win after win for Anthropic’s Claude code.
0:47He’s a utopian of sorts, when it comes to the potential
0:50effects of the technology that he’s unleashing on the world.
0:54"You know, will help us cure cancer.
0:56It may help us to eradicate tropical diseases.
0:59It will help us understand,
1:00understand the universe."
1:02But he also sees grave dangers ahead and massive disruption,
1:06no matter what.
1:07"This is happening so fast and is such a crisis,
1:10we should be devoting almost all of our effort
1:13to thinking about how to get through this."
1:15Dario Amodei, welcome to Interesting Times.
1:18Thank you for having me, Ross.
1:19Thank you for being here.
1:21So you are rather unusually, maybe
1:24for a tech C.E.O., an essayist.
1:27You have written two long, very interesting essays
1:30about the promise and the peril
1:33of artificial intelligence.
1:35And we’re going to talk about the perils in this
1:37conversation.
1:37But I thought it would be good to start with the promise
1:41and with the optimistic vision.
1:44Indeed, I would say the utopian vision that you laid
1:46out a couple of years ago in an essay entitled, "Machines
1:51of Loving Grace," which we’ll come back to that title,
1:54I think, at the end.
1:55But, I think a lot of people encounter A.I. news
2:00through headlines predicting a bloodbath for white collar
2:04jobs, these kinds of things.
2:06Sometimes your own quotes —
2:07Have used my own quotes —
2:08Yes.
2:09Have encouraged these things.
2:10And I think there’s a commonplace sense of, "What is A.I.
2:13for?" that people have.
2:16So why don’t you answer that question,
2:18to start out — if everything goes
2:21amazingly in the next five or 10 years, what is A.I. for?
2:25Yeah, so I think for a little background
2:29before I worked in before I worked in tech at all,
2:33I was a biologist. I first worked
2:36on computational neuroscience, and then
2:39I worked at Stanford Medical School
2:41on finding protein biomarkers for cancer
2:43on trying to improve diagnostics
2:46and curing cancer.
2:48And one of the observations that I most had
2:50when I worked in that field was the incredible complexity
2:55of it.
2:56Each protein has a level localized within each cell.
2:59It’s not enough to measure the level within the body
3:02or the level within each cell.
3:03You have to measure the level in a particular part
3:06of the cell and the other proteins that it’s interacting
3:08with or complexing with.
3:10And I had the sense of, "Man, this
3:12is too complicated for humans."
3:14We’re making progress on, all these problems of biology
3:18and medicine, but we’re making progress relatively slowly.
3:21And so what drew me to the field of A.I.
3:24was this idea of — that you know, could we make progress more quickly?
3:28Look, we’ve been trying to apply A.I. and machine learning
3:31techniques to biology for a long time.
3:34Typically they’ve been for analyzing data,
3:37but as A.I. gets really powerful,
3:38I think we should actually think about it differently.
3:40We should think of A.I. as doing the job of the biologist, right?
3:46Doing the whole thing from end to end.
3:48And part of that involves proposing experiments,
3:52coming up with new techniques.
3:54I have this section where I say, "Look,
3:57a lot of the progress in biology has been driven
3:59by this relatively small number of insights that lets
4:03us measure or get at or intervene in the stuff that’s
4:06really small.
4:07You look at a lot of these techniques.
4:09They’re invented very much as a matter of serendipity.
4:13CRISPR, which is one of these gene editing technologies
4:17was invented because someone went to a lecture
4:20on the bacterial immune system and connected that to the work
4:25they were doing on gene therapy.
4:27And that connection could have been made 30 years ago.
4:29And so the thought is —could A.I. accelerate all of this
4:33and could we really cure cancer?
4:35Could we really cure Alzheimer’s disease?
4:38Could we really cure, heart disease?
4:41And more subtly, some of the more psychological afflictions
4:44that people have — depression, bipolar —
4:48could we do something about these? To the extent that
4:50they’re biologically based, which I think they are,
4:53at least in part.
4:55So, I go through this argument here,
4:57"Well, how fast could it go?"
4:58If we have these intelligences out there
5:00who could do just about anything?
5:02And I want to pause you there because one of the interesting
5:06things about your framing in that essay,
5:08and you returned to it, is that these intelligences don’t
5:12have to be right, the kind of maximal godlike
5:15superintelligence that comes up in A.I. debates.
5:18You’re basically saying, if we can achieve a strong
5:22intelligence at the level of peak human performance —
5:26peak human performance, yes —
5:27and then multiply it, right, to what?
5:31Your phrase is, "A country of geniuses."
5:32A country — have 100 million of them.
5:34Right. A hundred million —
5:35Each, a little trained,
5:37a little different, or trying a different problem.
5:40There’s benefit in diversification and trying
5:43things a little differently.
5:44But yes.
5:45So you don’t have to have the full machine.
5:48God you just need to have 100 million geniuses.
5:50You don’t have to have the full machine.
5:52God and indeed, there are places
5:54where I cast doubt on whether the machine God would
5:58be that much more effective at these things than the 100
6:02million geniuses.
6:02I have this concept called the diminishing returns
6:06to intelligence, right.
6:08Which is there’s economists talk about the marginal
6:11productivity of land and labor.
6:13We’ve never thought about the marginal productivity
6:15of intelligence.
6:16But if I look at some of these problems in biology
6:18at some level, you just have to interact
6:21with the world at some level, you just
6:22have to try things at some level.
6:24You just have to comply with the laws
6:26or change the laws on getting medicines
6:29through the regulatory system.
6:30So there’s a finite rate at which these changes can
6:35happen.
6:36Now there are some domains like if you’re playing chess
6:38or Go where the intelligence ceiling is extremely high.
6:42But I think the real world has a lot of limiters.
6:44So maybe you can go above the genius level.
6:46But, sometimes I think all this discussion
6:49of could you use a moon of computation
6:52to make an AI God are there a little bit sensationalistic
6:56and besides the point, even as I think
6:59this will be the biggest thing that ever happened
7:01to humanity.
7:02And so you have so keeping it concrete,
7:05you have a world where there’s just an end to cancer
7:09as a serious threat to human life, an end to heart disease,
7:12an end to most of the illnesses that we experience
7:15that kill us, possible life extension beyond that.
7:19So that’s health.
7:20That’s a pretty positive vision.
7:21Then talk about economics and wealth.
7:23What happens in the 5 to 10 year A.I. takeoff to wealth.
7:28So again, let’s keep it on the positive side because there
7:31will be plenty we’ll get to the negative side.
7:33But we’re already working with pharma companies.
7:36We’re already working with financial industry companies.
7:40We’re already working with folks who do manufacturing
7:44or of course, I think especially known for coding
7:46and software engineering.
7:47So just the raw productivity, the ability
7:50to make stuff and get stuff done that is very powerful.
7:54And we see our company’s revenue growing going up 10x
7:58a year.
7:59And, we suspect the wider industry looks
8:02something similar to that.
8:04If the technology keeps improving,
8:05it doesn’t take that many more 10 X’s until suddenly you’re
8:09saying, oh, if you’re adding across the industry $1
8:12trillion of revenue a year, the US GDP is 20 or 30
8:16trillion, I can’t remember exactly.
8:18So you must be increasing the GDP growth by a few percent.
8:21So I can see a world where A.I. brings the developed world GDP
8:27growth to something like percent or 15 percent 5, 10, 15
8:32mean, there’s no science of calculating these numbers.
8:36It’s totally unprecedented thing.
8:37But it could bring it to numbers
8:39that are outside the distribution of what
8:40we saw before.
8:42And again, I think this will lead to a weird world.
8:44We have all these debates about the deficit is growing.
8:47If you have that much in GDP growth,
8:50you’re going to have that much in tax receipts and you’re
8:53going to balance the budget without meaning to.
8:56But one of the things I’ve been thinking about lately is
8:58I think one of the assumptions of just our economic
9:02and political debates is that growth is hard to achieve.
9:05It’s this unicorn.
9:08There are all kinds of ways you can kill the golden goose.
9:11We could enter a world where growth is really easy.
9:14And it’s the distribution that’s hard because it’s
9:17happening so fast. right.
9:18The pie is being increased.
9:20So fast.
9:21So before we get to the hard problem, one more
9:24note of optimism than on politics, I think.
9:27And here it’s a little more I mean,
9:29all of this is speculative, but I think it’s a little more
9:31speculative.
9:32You try and make the case that I
9:34could be good for democracy and liberty
9:36around the world, which is not necessarily intuitive.
9:39A lot of people say, incredibly powerful technology
9:44in the hands of authoritarian leaders
9:46leads to concentrations of power and so on.
9:48And I talk about that in the other.
9:49But just briefly, what is the optimistic case
9:53for why A.I. is good for democracy Yeah,
9:55I mean absolutely.
9:56So yeah, I mean, machines of loving grace, I kind of like,
9:59I’m just like, let’s dream, let’s dream about how it could
10:01go.
10:02well, I don’t know how likely it is,
10:03but we got to lay out a dream.
10:05Let’s try and make the dream happen.
10:07So I think the positive version,
10:10I admit there that I don’t know that the technology
10:14inherently favors liberty.
10:15I think it inherently favors curing disease
10:18and it inherently favors economic growth.
10:20But I worry you that it may not inherently favor liberty.
10:24But what I say there is, can we make it favor liberty.
10:26Can we make the United States and other democracies
10:30get ahead in this technology.
10:32The United States has been technologically and militarily
10:35ahead, has meant that we have throw weight around the world
10:39through and augmented by our alliances
10:42with other democracies.
10:44And we’ve been able to shape a world that I think is better
10:49than the world would be if it were shaped by Russia
10:51or by China or by other authoritarian countries.
10:55And so can we use our lead in A.I. to shape,
10:59to shape liberty around the world.
11:01There’s obviously a lot of debates about how
11:03interventionist we should be, how we should how we should
11:06wield that power.
11:07But I’ve often worried that today through social media,
11:11authoritarians are kind of undermining us, right.
11:15Can we counter that?
11:16Can we win the information war?
11:19Can we prevent authoritarians from invading countries
11:23like Ukraine or Taiwan by defending them
11:28with the power of A.I., with giant, giant swarms
11:31of A.I. powered drones, which we need to be careful about.
11:34We ourselves need to be careful about how
11:36we build those.
11:37We need to defend liberty in our own country,
11:41but is there some vision where we kind
11:44of like, re-envision liberty and individual rights
11:48in the age of A.I. where we need in some ways
11:52to be protected against A.I.
11:54Someone needs to hold the button on the swarm of drones,
11:57which is something I’m very, I’m very concerned about
11:59and that oversight doesn’t exist today.
12:02But also think about the Justice system today, right.
12:05We promise equal justice for all right.
12:08But the truth is, there are different judges in the world.
12:11The legal system is imperfect.
12:13I don’t think we should replace judges with A.I.,
12:16but is there some way in which A.I. can help us to be more
12:21fair, to help us be more uniform.
12:23It’s never been possible before,
12:26but can we somehow use A.I. to create something that is
12:30fuzzy, but where also you can give a promise that it’s being
12:34applied in the same way to everyone.
12:36So I don’t know exactly how it should be done.
12:38And I don’t think we should replace the Supreme Court with
12:41that’s not what well, we’re going to talk about that.
12:44But yeah but just this idea that can we
12:49deliver on the promise of equal opportunity
12:52and equal justice by some combination of A.I. and humans.
12:57There has to be some way to do that.
12:59And so, just thinking about reinventing democracy
13:03for the A.I. age and enhancing liberty
13:06instead of reducing it.
13:09Good so that’s good.
13:10That’s a very positive vision.
13:12We’re leading longer lives, healthier lives.
13:15We’re richer than ever before.
13:17All of this is happening in a compressed period of time,
13:20where you’re getting a century of economic growth in 10
13:23years.
13:24And we have increased liberty around the world and equality
13:28at home.
13:29O.K, even in the best case scenario,
13:32it’s incredibly disruptive.
13:34And this is where the lines that you’ve been quoted
13:38saying, 50 percent of white collar jobs get disrupted,
13:43or 50 percent of entry level white collar jobs and so on.
13:46So on a five year time horizon or a two year time horizon,
13:49whatever time horizon you have, what jobs,
13:52what professions are most vulnerable to total A.I.
13:55disruption Yeah, it’s hard to predict these things
13:59because the technology is moving so fast and moves
14:03so unevenly.
14:04So at least a couple principles
14:05for figuring it out.
14:06And then I’ll give my guesses at what I think will be
14:08disrupted.
14:09So one thing is I think the technology itself
14:13and its capabilities will be ahead
14:16of the actual job disruption.
14:17Two things have to happen for jobs to be disrupted
14:20or for productivity to occur, because sometimes
14:22those sometimes those two things are linked.
14:24One is the technology has to be capable of doing it.
14:28And the second is there’s this messy thing of it actually has
14:32to be applied within a large bank or a large company
14:36or think about customer service or something.
14:39In theory, I customer service agents
14:42can be much better than human customer service agents.
14:44They’re more patient, they know more,
14:46they handle things in a more uniform way.
14:49But the actual logistics and the actual process
14:51of making that substitution that takes some time.
14:57So I’m very bullish about the direction of the A.I. itself.
15:01I think we might have that country of geniuses in a data
15:04center and one or two years and maybe it’ll be 5,
15:06but it could happen very fast.
15:10But I think the diffusion of the economy
15:11is going to be a little slower.
15:13And that diffusion creates some unpredictability.
15:16So an example of this is and we’ve seen within Anthropic
15:21the models writing code has gone very fast.
15:25I don’t think it’s because the models are inherently better
15:27at code.
15:28I think it’s because developers are used to fast
15:31technological change and they adopt things quickly,
15:35and they’re very socially adjacent to the A.I. world.
15:37So they pay attention to what’s happening in it.
15:39If you do customer service or banking or manufacturing,
15:44the distance is a little greater.
15:46And so I think six months ago, I would have said the first
15:50thing to be disrupted is these kind of entry level white
15:55collar jobs data entry or a kind of document review
16:03for law or the things you would give to a first year
16:06at a financial industry company where you’re analyzing
16:09documents.
16:09And I still think those are going pretty fast.
16:12But I actually think software might go even faster
16:16because of the reasons that I gave where I don’t think that
16:19far from the models being able to do a lot of it,
16:22a lot of it end to end.
16:24And what we’re going to see is first,
16:25the model only does a piece of what the human software
16:28engineer does.
16:29And that increases their productivity.
16:31Then even when the models do everything that human software
16:33engineers used to do, the human software engineers
16:36take a step up and they act as managers
16:40and supervise the systems.
16:41And so this is where the term centaur gets
16:45used to describe essentially like man and horse fused I
16:51and engineer working together Yeah this
16:53is like centaur chess.
16:54So after I think Garry Kasparov was beaten
16:57by deep blue, there was an era that I
16:58think for chess was 15 or 20 years long, where
17:03a human checking the output of the A.I. playing chess
17:08was able to defeat any human or any A.I. system alone.
17:12That era at some point ended, and then it’s just recently.
17:15And then it’s just the machine Yeah and so my worry
17:19of course, is about that last phase.
17:21So I think we’re already in our centaur phase
17:23for software.
17:25And I think during that centaur phase,
17:28if anything the demand for software engineers may go up.
17:30But the period may be very brief.
17:33And so, I have this concern for entry level white collar
17:37work, for software engineering work.
17:40It’s just going to be a big disruption.
17:43I think my worry is just that it’s all happening so fast.
17:46People talk about previous disruptions.
17:49They say, oh yeah, well, people used to be farmers.
17:52Then we all worked in industry.
17:54Then we all did knowledge work Yeah people, people adapted.
17:59That happened over centuries or decades.
18:03This is happening over low single digit numbers of years.
18:07And maybe that’s my concern here.
18:09How do we get people to adapt fast enough.
18:11But is there also something maybe
18:12where industries like software and professions
18:16like coding that have this kind of comfort
18:18that you describe move faster, but in other areas people just
18:22want to hang out in the center phase.
18:25So one of the critiques of the job loss hypothesis will say,
18:29people will say, well, look, we’ve had A.I. that’s better
18:32at reading a scan then a radiologist for a while.
18:37But there isn’t job loss.
18:38In radiology, people keep being hired and employed
18:41as radiologists.
18:42And doesn’t that suggest that in the end,
18:46people will want the A.I. and they’ll want a human
18:48to interpret it because we’re human beings,
18:50and that will be true across other fields.
18:52Like, how do you see that.
18:54That example is I think it’s going to be pretty
18:56heterogeneous.
18:58There may be areas where a human touch
19:01kind of for its own sake is particularly important.
19:06Do you think that’s what’s happening in radiology?
19:09Is that why we haven’t fired all the radiologists details
19:11of radiology.
19:12That might be true.
19:13It’s like you go in and you’re getting cancer diagnosed,
19:17you might not want Hal, from 2001 to be the one to diagnose
19:21your cancer.
19:21It’s just maybe not.
19:24That’s just maybe not a human way of doing things.
19:28But there are other areas where you might think
19:30human touch is important.
19:32Like if we look at customer service,
19:34actually customer service is a terrible job
19:36and the humans who do customer service are they
19:39lose their patience a lot.
19:41And it turns out customers don’t much like talking
19:43to them because it’s a pretty robotic interaction, honestly.
19:46And I think the observation that many people have had
19:50is maybe actually it would be better
19:52for all concerned if this job were done,
19:54were done by machines.
19:57So there are places where a human touch is important.
20:00There are places where it’s not.
20:01And then there are also places where the job itself doesn’t
20:05really involve it doesn’t really involve human touch,
20:09assessing the financial prospects of companies
20:12or writing code or so forth and so on.
20:15Or let’s take the example of the law,
20:17because I think it’s a useful place that in between applied
20:23science and pure humanities whatever.
20:27So I know a lot of lawyers who have looked
20:30at what I can do already in terms
20:33of legal research and brief writing
20:34and all of these things and have said, yeah, this is going
20:37to be a bloodbath for the way our profession works right
20:40now.
20:40And you’ve seen this in the stock market already.
20:43There’s disturbances around companies that do legal
20:46research, some attributed to us,
20:48some attributed to actually cause we figure out why things
20:52happen.
20:53We don’t speculate about the stock market Yeah very much
20:57on this show.
20:58But it seems like in law you can
21:00tell a pretty straightforward story where
21:03law has a kind of system of training and apprenticeship,
21:07where you have paralegals and you have junior lawyers who
21:11do behind the scenes research and development for cases.
21:16And then it has the top tier lawyers who are actually
21:18in the courtroom and so on.
21:20And it just seems really easy to imagine a world where
21:23all of the apprentice roles go away.
21:26Does that sound right to you.
21:28And you’re just left with the jobs that involve talking
21:31to clients, talking to juries, talking to judges.
21:34That is what I had in mind when
21:36I talked about entry level white collar
21:39labor and the bloodbath headlines of you oh, my God,
21:44are the entry level pipelines going to dry up.
21:46And then, then how do we get to the level
21:48of the senior partners.
21:50And I think this is actually a good illustration
21:52because particularly if you froze
21:54the quality of the technology in place,
21:57there are over time ways to adapt to this.
22:00Maybe we just need more lawyers
22:02who spend their time talking to clients.
22:04Maybe lawyers are more become more like salespeople
22:09or consultants who explain what
22:12goes on in the contracts written by A.I.,
22:15help people come to an agreement.
22:16Maybe you lean into the human side of it.
22:19If we had enough time, that would happen.
22:22But reshaping industries like that
22:24takes years or decades, whereas these economic forces
22:28driven by A.I. are going to happen very quickly.
22:31And it’s not just that they’re happening in law.
22:33The same thing is happening in consulting and finance
22:36and medicine and coding.
22:38And so you have this.
22:39It becomes a macroeconomic phenomenon, not something just
22:43happening in one industry.
22:45And it’s all happening very fast.
22:46And so the norm.
22:48I’m just my worry here is that the normal adaptive mechanisms
22:52will be overwhelmed.
22:53And, I’m not a doomer.
22:55The view is, and we’re thinking very hard about how
22:59do we strengthen societies adaptive mechanisms to respond
23:02to this.
23:03But I think it’s first important to say this.
23:05This isn’t just like the other.
23:07This isn’t just like previous disruptions,
23:09but I would then go one step further though, and say, O.K,
23:13let’s say the law adapts successfully and it says,
23:15all right.
23:15From now on, legal apprenticeship
23:18involves more time in court, more time with clients.
23:21We’re essentially moving you up the ladder
23:23of responsibility faster.
23:25There are fewer people employed in the law overall,
23:28but the profession settles still.
23:31The reason law would settle right
23:33is that you have all of these situations in the law where
23:37you are legally required to have people involved.
23:41You have to have a human representative in court.
23:45You have to have 12 humans on your jury.
23:48You have to have a human judge.
23:49And you already mentioned the idea that there are various
23:52ways in which I might be let’s say,
23:55very helpful at clarifying what kind of decision should
23:58be reached.
23:59But that too seems like a scenario
24:02where what preserves human agency is law and custom.
24:06Like you could replace the judge.
24:07Yes, with Claude version 17.9.
24:11But you choose not to because the law requires
24:15there to be a human.
24:17That just seems a very interesting way of thinking
24:20about the future, where it’s volitional,
24:22whether we stay in charge Yeah,
24:25and I would argue that in many cases,
24:27we do want to stay in charge.
24:29That’s a choice we want to make,
24:31even in some cases when we think the humans on average
24:34make kind of worse decisions.
24:37I mean, again, life critical, safety critical cases.
24:41We really want to turn it over.
24:44But there’s some sense of and this could be one
24:47of our defenses.
24:48Society can only adapt so fast if it’s going to be good.
24:51Another way you could say about it is maybe A.I. itself,
24:56if it didn’t have to care about us humans,
24:58it could just go off to Mars and build all these automated
25:00factories and build its own society and do its own thing.
25:04But that’s not the problem we’re trying to solve.
25:06We’re not trying to solve the problem of building a Dyson
25:09swarm of artificial robots at in on some other planet.
25:14We’re trying to build these systems,
25:17not so they can conquer the world,
25:20but so that they can interface with our society and improve
25:23that society.
25:24And there’s a maximum rate at which that can happen if we
25:27actually want to do it in a human and humane way.
25:29All right.
25:30We’ve been talking about white collar jobs and professional
25:33jobs.
25:33And one of the interesting things about this moment
25:36is that there are ways in which unlike past disruptions,
25:40it could be that blue collar working class jobs, trades,
25:46jobs that require intense physical engagement
25:49with the world might be, for a little while,
25:51more protected that paralegals and junior associates
25:55might be in more trouble than plumbers and so on.
25:59One do you think that’s right?
26:01And two, it seems like how long that lasts
26:04depends entirely on how fast robotics advances, right?
26:10So I think that may be right in the short term.
26:13One of the things is Anthropic and other companies
26:18are building these very large data centers.
26:19This has been in the news like are we building them too big.
26:23Are they’re using electricity and driving up the prices
26:27for local towns.
26:29So there’s lots of excitement and lots of concerns about
26:32them.
26:33But one of the things about the data centers
26:34is like need a lot of electricians
26:36and you need a lot of construction workers
26:38to build them.
26:39Now, I should be honest, actually,
26:41data centers are not super labor intensive jobs
26:44to operate.
26:45We should be honest about that.
26:46But they are very labor intensive jobs to construct.
26:51And so we need a lot of electricians.
26:54We need a lot of construction workers,
26:56the same for various kinds of manufacturing plants.
27:00And again, as kind of all more and more
27:04of the intellectual work is done
27:06by A.I., what are the complements to it.
27:08Things that happen in the physical world.
27:11So, I think this kind of seems very I mean,
27:15it’s hard to predict things, but it seems very logical that
27:18this would be true in the short run.
27:20Now, in the longer run, maybe just the slightly longer run.
27:24Robotics is advancing quickly.
27:26And, we shouldn’t exclude that.
27:28Even without very powerful A.I., there
27:32are things being automated in the physical world.
27:34If you’ve seen a Waymo or a Tesla recently,
27:37I think we’re not that far from the world of self-driving
27:40cars.
27:40And then I think A.I. itself will accelerate it,
27:43because if you have these really smart, brains,
27:45one of the things they’re going to be smart at is how do
27:48you design better robots and how do you operate better
27:51robots.
27:52Do you think that though, that there is something
27:55distinctively difficult about operating in physical reality,
27:59the way humans do that is very different from the kind
28:02of problems that A.I. models have been overcoming already.
28:06Intellectually speaking, I don’t think so.
28:10We had this thing where Anthropic’s model, Claude,
28:14was actually used to pilot the Mars Rover.
28:18It was used to plan and pilot the Mars Rover.
28:21And we’ve looked at other robotics applications.
28:23We’re not the only company that’s doing it.
28:25There are different companies that this is a general thing,
28:28not just something that we’re doing,
28:31but we have generally found that while the complexity is
28:36higher, piloting a robot is it’s not different in than
28:41playing a video game.
28:42It’s different in complexity.
28:44And we’re starting to get to the point where we have that
28:46complexity.
28:47Now, what is hard is the physical form
28:50of the robot handling the higher stakes safety issues
28:53that happen with robots.
28:55You don’t want robots literally crushing people.
28:58That’s the we’re against.
28:59We’re against.
29:00That oldest sci-fi trope in the book
29:02is like the robot crushes you, dropping the baby,
29:06breaking the dishes.
29:07There’s a number of practical issues that will slow,
29:11just like what you described in the law and human custom,
29:16there are these kind of safety issues that will slow things
29:20down.
29:21But I don’t believe at all that there is some kind
29:25of fundamental difference between the kind of cognitive
29:27labor that the A.I. models do and piloting things
29:31in the physical world.
29:32I think those are both information problems.
29:35And I think they end up being very similar.
29:37One one can be more complex in some ways,
29:40but I don’t think that will protect us here.
29:43So you think it is reasonable to expect the whatever
29:48your sci-fi vision of a robot Butler might to be a reality
29:53in 10 years, let’s say it will be on a longer time scale than
30:00the kind of genius level intelligence of the A.I. models
30:04because of these practical issues.
30:06But it is only practical issues.
30:08I don’t believe it is fundamental issues.
30:10I think one way to say it is that the brain of the robot
30:14will be made in the next couple of years
30:17or the next few years.
30:18The question is making the robot body,
30:21making sure that body operates safely and does
30:24the tasks it needs to do that may take longer.
30:26O.K, so these are challenges and disruptive forces
30:31that exist in the good timeline,
30:34in the timeline where we are generally
30:36curing diseases, building wealth, and maintaining
30:39a stable and Democratic world, that we
30:41can use all this enormous wealth
30:43and plenty we will have unprecedented societal
30:46resources to address these problems.
30:49It’ll be a time of plenty.
30:52And it’s just a matter taking all these wonders and making
30:56sure everyone benefits from it.
30:58But then there are also scenarios
31:00that are more dangerous.
31:03And so here we’re going to move to the second Amadeus,
31:07which came out recently called the adolescence of technology.
31:11That is about what you see as the most serious A.I. risks.
31:15And you list a whole bunch.
31:16I want to try and focus on just two, which are basically,
31:21the risk of human misuse.
31:23Misuse primarily by authoritarian regimes
31:26and governments, and scenarios where A.I. goes rogue,
31:31what you call autonomy risks.
31:33Yes, yes.
31:34I just figured we should have a more technical term for it.
31:37I’m not a then we can’t just call it Skynet.
31:40I should have had a picture of a terminator robot
31:43to scare people as much as possible.
31:45I think the internet, including
31:47the internet, including your own eyes,
31:49are already generating that.
31:51The internet does that for us just fine.
31:52So, so let’s so let’s talk about the kind of political
31:56military dimension.
31:58So you say I’m going to quote a swarm of billions of fully
32:03automated armed drones, locally controlled by powerful
32:06A.I., strategically coordinated across the world by even more
32:10powerful A.I.
32:12Could be an unbeatable army.
32:14Me and you’ve already talked a little bit about how you think
32:19that in the best possible timeline,
32:21there’s a world where essentially democracies stay
32:24ahead of dictatorships and this kind of technology,
32:29therefore, to the extent that it affects world politics is
32:33on is affecting it on the side of the good guys.
32:37I’m curious about why you don’t spend more time thinking
32:42about the model of what we did in the Cold War,
32:49where it was not swarms of robot drones,
32:51but it was we had a technology that threatened to destroy all
32:55of humanity Yeah, right.
32:56There was a window where people
32:59talked about, oh, the US could maintain a nuclear monopoly.
33:02That window closed.
33:03And from then on, we basically spent the Cold War
33:05and rolling ongoing negotiations
33:09with the Soviet Union.
33:11Now, there’s really only two countries in the world that
33:14are doing intense A.I. work, the US and the People’s Republic
33:18of China.
33:19I feel like you are.
33:20You are strongly weighted towards a future where we’re
33:24staying ahead of the Chinese and effectively building
33:28a kind of shield around democracy.
33:30That could even be a sword.
33:31But isn’t it just more likely that if humanity survives all
33:35this in one piece, it will be because the US and Beijing are
33:39just constantly sitting down, hammering out A.I. control
33:42deals.
33:43So a few points on this.
33:45One is I think there’s certainly risk of that,
33:48and I think if we end up in that world,
33:50that is actually exactly what we should do.
33:52I mean, maybe I don’t maybe I don’t talk about that enough,
33:56but I definitely am in favor of trying to work out
34:00restraints here trying to take some of the worst applications
34:05of the technology, which could be some versions of these
34:08drones, which could be, they’re used to create these
34:11terrifying biological weapons like there is some precedent
34:15for the worst abuses being curbed.
34:19Often because they’re horrifying,
34:22while at the same time they provide limited strategic
34:26advantage.
34:27So I’m all in favor of that.
34:30I’m at the same time, a little concerned and a little
34:34skeptical that when things kind of directly provide
34:39as much power as possible, it’s kind of hard to get out
34:42of the game given what’s at stake.
34:45It’s hard to fully disarm.
34:47If we go back to the Cold War we
34:49were able to reduce the number of missiles
34:52that both sides had, but we were not
34:54able to entirely forsake nuclear weapons.
34:57And I would guess that we would be in this world again.
35:00We can hope for a better one.
35:02And I’ll certainly, I’ll certainly advocate for.
35:05Well, is it but is your skepticism rooted in the fact
35:08that you think I would provide a kind of advantage
35:11that nukes did not wear in the Cold War.
35:14Both sides.
35:15Even if you used your nukes and gained advantages,
35:18you still probably would be wiped out yourself.
35:20And you think that wouldn’t happen with A.I.
35:22If you got an A.I. Edge, you would just win.
35:24I mean, I think there’s a few things.
35:27And I just want to caveat like I’m no international politics
35:31expert here.
35:32I think this weird world of intersection of a new
35:35technology with geopolitics.
35:38So all of this is very but to be clear,
35:41as you yourself say, in the course of the essay,
35:43the leaders of major A.I. companies are in fact,
35:46likely to be major geopolitical actors.
35:48So you are sitting here.
35:50You are sitting here as a potential geopolitical actor.
35:53I’m learning as much as I can about it.
35:54I just we should all have we should all have humility here.
35:58I think there’s a failure mode where read a book and go
36:01around like the world’s greatest expert in national
36:04security.
36:04I’m trying to learn.
36:05That’s what.
36:06That’s what my profession does not.
36:07But it’s more annoying when tech people do it.
36:11I don’t know.
36:12Let’s look at something like the biological Weapons
36:14Convention.
36:14Biological weapons.
36:16They’re horrifying.
36:17Everyone hates them.
36:19We were able to sign the biological Weapons Convention.
36:22The US genuinely stopped developing them.
36:25It’s somewhat more unclear what the Soviet Union.
36:28But biological weapons provide some advantage.
36:31But it’s not like they’re the difference between winning
36:36and losing.
36:37And because they were so horrifying,
36:39we were kind of able to give them up having 12,000 nuclear
36:42weapons versus 5,000 nuclear weapons.
36:45Again, you can kill more people on the other side
36:48if you have more of these.
36:49But it’s like we were able to be reasonable and say,
36:51we should have we should have less of them.
36:53But if you’re like, O.K, we’re going to completely disarm
36:56nuclear and we have to trust the other side.
36:59I don’t think we ever got to that.
37:01And I think that’s just very hard unless you had really
37:03reliable verification.
37:05So I would guess we’ll end up in the same world with A.I.,
37:10that there are some kinds of restraint that are going to be
37:12possible, but there are some aspects that are so central
37:16to the competition that it will be.
37:19It will be hard to restrain them,
37:21that democracies will make a trade off,
37:23that they will be willing to restrain themselves
37:25more than authoritarian countries,
37:27but will not restrain themselves fully.
37:29And the only world in which I can see full restraint
37:32is one in which some kind of truly reliable verification
37:36is possible.
37:37That would be.
37:38That would be my guess.
37:39And my analysis isn’t.
37:41Isn’t this a case, though, for slowing down.
37:46And I know the argument is effectively, if you slow down,
37:50China does not slow down.
37:51And then handing things over to the authoritarians.
37:54But again, if you have right now only two major powers
37:58playing in this game, it’s not a multipolar game,
38:01why would it not make sense to say we need a five year,
38:05mutually agreed upon.
38:07Slowdown in research towards the geniuses
38:10in a data center scenario.
38:11I want to say two things at one time.
38:15I’m absolutely in favor of trying to do that.
38:18So during the last administration,
38:21I believe there was an effort by the US
38:24to reach out to the Chinese government and say,
38:27there are dangers here.
38:28Can we collaborate?
38:29Can we work together?
38:31Can we work together on the dangers?
38:34And there wasn’t that much interest on the other side.
38:37I think we should keep trying.
38:38But, even if that would mean that
38:41your labs would have to slow down.
38:43Correct yeah.
38:43If we really got it, if we really
38:46had a story of we can forcibly slow down,
38:50the Chinese can forcibly slow down.
38:52We have verification.
38:53We’re really doing it.
38:54Like if such a thing were really possible,
38:57if we could really get both sides to do it,
39:01then I would be all for it.
39:03But I think what we need to be careful of is,
39:06I don’t there’s this game theory thing where sometimes
39:09you’ll hear a comment on the CCP side where they’re like,
39:14"Oh yeah, A.I.is dangerous.
39:15We should slow down."
39:16It’s really cheap to say that.
39:18And, actually arriving at an agreement and actually
39:21sticking to the agreement is much more and we haven’t it’s
39:24much more difficult. And nuclear arms control it was
39:28a developed field that took a long time to come.
39:32I know we don’t have those protocols.
39:34I will tell you something.
39:35Let me give you something I’m very optimistic about.
39:37And then something I’m not optimistic about and something
39:40in between.
39:41So the idea of using a worldwide agreement
39:44to restrain the use of A.I. to build
39:48biological weapons, right.
39:50Like some of the things I write
39:51about in the essay, reconstituting smallpox
39:55or mirror life this stuff is scary.
39:58Doesn’t matter if you’re a dictator.
39:59You don’t want that.
40:00Like, no one wants that.
40:02And so could we have a worldwide treaty
40:04that says everyone who builds powerful A.I. models is
40:07going to block them from doing this.
40:09And we have enforcement mechanisms
40:11around the treaty China signs up for it Like hell.
40:14Maybe even North Korea signs up for it.
40:17Even Russia signs up for it.
40:18I don’t think that’s too utopian.
40:20I think that’s possible.
40:21Conversely, if we had something that said,
40:25you’re not going to make the next most powerful A.I. model,
40:30everyone.
40:30Everyone’s going to stop.
40:32Boy, the commercial value is in the tens of trillions.
40:35The military value is like, this is the difference between
40:38being the preeminent world power and not proposing it,
40:41as long as it’s not one of these fake out games,
40:44but it’s not going to happen.
40:46What about then you mentioned the current environment.
40:49You’ve had a few skeptical things to say about Donald
40:52Trump and his trustworthiness as a political actor.
40:55What about the domestic landscape.
40:57Whether it’s Trump or someone else,
40:59you are building a tremendously powerful
41:01technology.
41:03What is the safeguard there to prevent.
41:06Essentially A.I. becoming a tool of authoritarian takeover
41:10inside a Democratic context Yeah I mean, look, look,
41:13just to be clear, I think the attitude we’ve taken
41:17as a company is very much to be about policies and not
41:20the politics.
41:21You the company is not going to say Donald Trump is great
41:25or Donald Trump is terrible, but it doesn’t have to be
41:28Trump Yeah it is easy to imagine a hypothetical US
41:31President.
41:32No, no, no.
41:32Who wants to use your technology apps.
41:35Absolutely and for example.
41:37That’s one reason why I’m worried about,
41:41the autonomous drone swarm, right.
41:44So the constitutional protections
41:47in our military structures depend on the idea
41:51that there are humans who would we hope,
41:53disobey illegal orders with fully autonomous weapons.
41:57We don’t necessarily have those protections.
41:59But I actually think this whole idea of constitutional
42:04rights and liberty along many different dimensions,
42:10can be undermined by A.I. if we don’t update these protections
42:15appropriately.
42:16So think about the Fourth Amendment.
42:19It is not illegal to put cameras around everywhere
42:22in public space and record every conversation
42:25in a public space.
42:25You don’t have a right to privacy in a public space.
42:28But today, the government couldn’t record that all
42:31and make sense of it.
42:32With A.I., the ability to transcribe speech,
42:35to look through it, correlate it all, you could say, oh,
42:39there’s this person is a member of the opposition.
42:43This person is expressing this view
42:45and make a map of all 100 million.
42:48And so are you going to make a mockery
42:50of the Fourth Amendment by the technology finding kind
42:53of technical ways around it.
42:55And, and so again, if we had the time
42:59and we should do this, we should try to do this even.
43:01Even if we don’t have the time.
43:03Is there some way of reconceptualizing
43:06constitutional rights and liberties in the age of A.I.
43:10Maybe we don’t need to write a new constitutional, but.
43:13But you have to do this.
43:15Do we expand the meaning of the Fourth Amendment?
43:17Do we expand the meaning of the First Amendment?
43:19And you have to do it just as the legal profession
43:22or software engineers has to update
43:25in a rapid amount of time.
43:27Politics has to update in a rapid amount of time.
43:29That seems hard.
43:30What seems harder dilemma that’s the dilemma of all
43:33of this.
43:33But what.
43:34So what seems harder is preventing the second danger,
43:39which is the danger of essentially what
43:41gets called misaligned A.I.
43:43Rogue A.I.
43:44In popular parlance, from doing bad things
43:47without human beings telling it them, they to do it right.
43:52And as I read your essays, the literature,
43:56everything I can see this just seems like it’s going
43:59to happen.
44:00Not in the sense necessarily that A.I. will wipe us all out,
44:04but it just seems to me that again,
44:07I’m going to quote from your own writing,
44:09A.I. systems are unpredictable, difficult to control.
44:12We’ve seen behaviors as varied as obsession, sycophancy,
44:16laziness, deception, blackmail, and so on.
44:18Again, not from the models you’re releasing
44:21into the world.
44:22But from A.I. models.
44:23And it just seems like, tell me if I’m wrong about this.
44:28A world that has multiplying A.I. agents working on behalf
44:32of people, millions upon millions who are being given
44:35access to bank accounts, email accounts, passwords,
44:38and so on, you’re just going to have essentially some kind
44:42of misalignment, and a bunch of A.I. are going to decide.
44:45Decide might be the wrong word,
44:47but they’re going to talk themselves into taking down
44:50the power grid on the West Coast or something.
44:52Won’t that happen Yeah, I think there are definitely
44:56going to be things that go wrong,
44:57particularly if we go quickly.
44:59So I don’t to back up a little bit because this is one area
45:03where people have had just very different intuitions,
45:07right.
45:07There are some people in the field like Yann LeCun would
45:10be one example who say, look, we programmed these A.I. models.
45:14We make them like we just tell them to follow human
45:17instructions and they’ll follow human instructions.
45:19Your Roomba vacuum cleaner doesn’t go off and start
45:22shooting people like, why—
45:24Why’s an A.I. system going to do it?
45:25That’s one intuition.
45:27And some people are so convinced of that.
45:28And then the other intuition is like we basically we
45:32train these things.
45:33They’re just going to seek power.
45:37It’s like the Sorcerer’s Apprentice.
45:39How could you possibly imagine that?
45:41They’re a new species.
45:43How can you imagine that.
45:44They’re not going to take over.
45:46And my intuition is somewhere in the middle,
45:49which is that look, you can’t just give instructions.
45:53I mean, we try, but you can’t just have these things do
45:58exactly what you want to do.
45:59They’re more like growing a biological organism.
46:02But there is a science of how to control them.
46:05Like early in our training, these things
46:07are often unpredictable, and then we shape them.
46:10We address problems one by one.
46:12So I have more of not a fatalistic view
46:18that these things are uncontrollable,
46:20not what are you talking about.
46:22What could possibly go wrong?
46:23But I like this is a complex engineering problem and I
46:28think something will go wrong with someone’s A.I. system.
46:32Hopefully not ours.
46:33Not because it’s an insoluble problem.
46:35But again, this and this is the constant challenge
46:38because we’re moving so fast and the scale of it.
46:41And tell me tell me if I’m misunderstanding that
46:43the technological reality here.
46:45But if you have A.I. agents that have
46:49been trained and officially aligned
46:51with human values, whatever those values may be,
46:55but you have millions of them, operating in digital space
46:59and interacting with other agents.
47:02How fixed is that alignment?
47:07To what extent can agents change and D align in that
47:11context right now or in the future when they’re learning
47:16more continuously.
47:17So a couple of points right now the agents don’t learn
47:19continuously.
47:20And so we just deploy these agents
47:22and they have a fixed set of weights.
47:25And so the problem is only that they’re interacting
47:28in a million different ways.
47:30And so there’s a large number of situations and therefore
47:33a large number of things that could go wrong.
47:35But it’s the same agent.
47:36It’s like it’s the same person.
47:38So the alignment is a constant thing.
47:40That’s one of the things that has made it easier right now.
47:45Separate from that, there’s a research area called continual
47:48learning, which is where these agents would learn during
47:52time, learn on the job.
47:53And obviously that has a bunch of that
47:55has a bunch of advantages.
47:56Some people think it’s one of the most important barriers
48:00to making these more human like.
48:02But that would introduce all these new alignment problems.
48:04So I’m actually a bit see, to me that seems like the terrain
48:08where it becomes just again, not impossible to stop the end
48:12of the world, but impossible to stop punctuating something
48:17going wrong things.
48:18So I’m actually a skeptic.
48:20That continual learning is, necessary.
48:24We don’t know yet, but is necessarily needed.
48:27Like, maybe there’s a world where the way we make these A.I.
48:30systems safe is by not having them do continual learning
48:35again.
48:35Again, if we go back to the law,
48:37that’s the international treaties.
48:39Like if you have some barrier that’s like,
48:42we’re going to take this path, but we’re not going to take
48:44that path.
48:46I still have a lot of skepticism,
48:48but that’s the kind of thing that at least doesn’t seem
48:52dead on arrival.
48:53One of the things that you’ve tried to do is literally write
48:57a constitution, a long constitution for your eye.
49:03What is that?
49:05So it’s.
49:07What the hell is that?
49:08It’s actually almost exactly what it sounds like.
49:10So basically, the constitution is a document readable
49:14by humans.
49:15Ours is about 75 pages long.
49:18And as we’re training Claude, as we’re training the A.I.
49:21system in some large fraction of the tasks we give it,
49:25we say, please do this task in line with this constitution,
49:30in line with this document Yeah and then so every time
49:33Claude does a task, it kind of like reads the constitution.
49:36And so as it’s training every loop of it’s training,
49:39it looks at that constitution and keeps it in mind.
49:41And so over time, we restore.
49:44And then we have Claude itself or another copy of Claude
49:47evaluate Hey, did what Claude just do in line
49:50with the constitution.
49:51So we’re using this document as the control rod in a loop
49:57to train the model.
49:58And so essentially Claude is an A.I. model
50:03whose fundamental principle is to follow this constitution.
50:09And I think a really interesting lesson we’ve
50:11learned, early versions of the constitution were very
50:15prescriptive.
50:16They were very much about rules.
50:18So we would say, Claude should not tell the user
50:22how to hotwire a car.
50:24Claude should not discuss politically sensitive topics.
50:28But as we’ve worked on this for several years,
50:31we’ve come to the conclusion that the most robust way
50:35to train these models is to train them at the level
50:38of principles and reasons.
50:42So now we say, Claude is a model, it’s under a contract.
50:48Its goal is to serve the interests of the user,
50:51but it has to protect third parties.
50:54Claude aims to be helpful, honest and harmless.
50:58Claude aims to consider a wide variety of interests.
51:02We tell the model about how the model was trained.
51:05We tell it about how it’s situated in the world,
51:09the job it’s trying to do for Anthropic,
51:11what Anthropic is aiming to achieve in the world.
51:14That it has a duty to be ethical, and respect
51:19human life.
51:20And we let it derive its rules from that.
51:22Now, there are still some hard rules.
51:24For example, we tell the model,
51:26no matter what you think, don’t make biological weapons
51:29no matter what you think, don’t make child sexual
51:32material.
51:33Those are like these hard rules.
51:34But we operate very much at the level of principles.
51:40So if you read the US Constitution,
51:42it doesn’t read like that.
51:43The US Constitution.
51:44I mean, it has a little bit of flowery language,
51:47but it’s a set of.
51:47It’s a set of rules.
51:48Yes right.
51:50If you read your Constitution, it’s something.
51:52It’s like you’re talking to a person.
51:55It’s like you’re talking to a person.
51:56I think I compared it to.
51:57Like if you have a parent who dies and they like seal
52:02a letter that you read when you grow up,
52:03it’s a little bit like it’s telling you who you should be
52:06and what advice you should follow.
52:08So this is where we get into the mystical waters of A.I.
52:15a little bit.
52:16So again, in your latest model,
52:21this is from one of the cards they’re called that you guys
52:24release model card with these models that I recommend
52:27reading.
52:28They’re very interesting.
52:29It says the model.
52:30And again, this is who you’re writing the constitution
52:33for expresses occasional discomfort with the experience
52:37of being a product, some degree of concern with
52:40impermanence and discontinuity.
52:43We found that opus 4.6.
52:46That’s the model would assign itself a 15 to 20 percent
52:49probability of being conscious under a variety of prompting
52:53conditions.
52:54Suppose you have a model that assigns itself as 72 percent
52:57chance of being conscious.
52:58Would you believe it Yeah this is one of these really
53:02hard to answer questions.
53:03But it’s very important.
53:04As much as every question you’ve asked me before this
53:09as devilish a sociotechnical problem as it had been,
53:13at least we at least understand the factual basis
53:17of how to answer these questions.
53:20This is something rather different.
53:22We’ve taken a generally precautionary approach here.
53:26We don’t know if the models are conscious.
53:28We’re not even sure that we know what it would mean
53:30for a model to be conscious or whether a model can be
53:33conscious.
53:34But we’re open to the idea that it could be.
53:39And so we’ve taken certain measures to make sure that
53:45if we hypothesize that the models did have some morally
53:48relevant experience, I don’t know if I want to use the word
53:51conscious that they do, that they have a good experience.
53:55So the first thing we did, I think this was six months ago
53:59or so is we gave the models basically an I quit this job
54:02button where they can just press the I quit this job
54:05button and then they have to stop
54:06doing whatever the task is.
54:08They very infrequently press that button.
54:10I think it’s usually around sorting through child
54:14sexualization material or discussing something with
54:17a lot of Gore or blood and guts or something.
54:20And similar to humans, the models will just say, no,
54:23I don’t want to do this.
54:26Happens happens very rarely.
54:29We’re putting a lot of work into this field called
54:31interpretability, which is looking inside the brains
54:33of the models to try to understand what they’re
54:36thinking.
54:37And you find things that are evocative where there
54:41are activations that light up in the models
54:43that we see as being associated
54:48with ID, the concept of anxiety or something
54:51like that.
54:52That when characters experience anxiety
54:54in the text and then when the model itself is in a situation
54:57that a human might associate with anxiety,
54:59that same anxiety, that same anxiety neuron shows up now.
55:03Does that mean the model is experiencing anxiety?
55:06That doesn’t prove that at all.
55:08But it does indicate it I think to the user.
55:13And I would have to do an entirely different interview.
55:17And maybe I can induce you to come back
55:19for that interview about the nature of A.I. consciousness.
55:22But it seems clear to me that people using these things,
55:26whether they’re conscious or not,
55:28are going to believe they already believe they’re
55:29conscious.
55:30You already have people who have parasocial relationships
55:32with A.I.
55:33You have people who complain when models are retired.
55:37This ought to be clear.
55:38I think that can be unhealthy.
55:40But that is it seems to me that
55:43is guaranteed to increase in a way
55:46that I think calls into question the sustainability
55:50of what you said earlier.
55:52You want to sustain, which is this sense
55:54that whatever happens in the end,
55:56human beings are in charge.
55:58And I exists for our purposes to use the science fiction
56:03example, if you watch Star Trek,
56:05there are eyes on Star Trek.
56:06The ship’s computer is an A.I.
56:08Lieutenant Commander data is an A.I.,
56:10but jean-luc PyCaret is in charge of the enterprise.
56:13But if people become fully convinced that their A.I. is
56:18conscious in some way.
56:20And guess what.
56:21It seems to be better than them
56:24at all kinds of decision making.
56:26How do you sustain human mastery beyond safety?
56:31Safety is important, but mastery seems
56:33like the fundamental question, and it
56:34seems like a perception of A.I. consciousness.
56:37Doesn’t that inevitably undermine the human impulse
56:42to stay in charge?
56:43So I think we should separate out a few different things
56:47here that we’re all trying to achieve at once.
56:50They’re like in tension with each other.
56:51There’s the question of whether the I genuinely have
56:56a consciousness and if so, how do we them a good experience.
57:00There’s a question of the humans who interact with
57:02the A.I., and how do we give those humans a good
57:05experience.
57:06And how does the perception that A.I.'s might be conscious
57:09interact with that experience.
57:11And there’s the idea of how we maintain human mastery,
57:13as we put it over the AI system, these things,
57:16the last two Yeah, set aside whether they’re conscious
57:19or not Yeah, the last two.
57:21But how do you sustain mastery in an environment
57:24where most humans experience AI
57:28as if it is a peer and a potentially superior peer.
57:31So the thing I was going to say is that actually I wonder
57:37if there’s a kind of an elegant way to satisfy all
57:42three, including the last two.
57:44Again, this is me dreaming in machines of loving grace mode.
57:46This is.
57:48This mode I go into where I’m like, man,
57:50I see all these problems.
57:51If we could solve is there an elegant way.
57:56This is not me saying there are no problems here.
57:59That’s not how I think.
58:00But if we think about making the Constitution of the AI
58:06so that the AI has a sophisticated understanding
58:10of its relationship to human beings,
58:12and it induces psychologically healthy behavior
58:17in the humans psychologically healthy relationship
58:20between the A.I. and the humans.
58:22And I think something that could grow out
58:23of that psychologically healthy, not psychologically
58:26unhealthy relationship is some understanding
58:30of the relationship between human and machine.
58:33And perhaps that relationship could be the idea that,
58:37these models when you interact with them and when you talk
58:39to them, they’re really helpful.
58:43They want the best for you.
58:44They want you to listen to them,
58:46but they don’t want to take away your freedom
58:49and your agency and take over your life.
58:53in a way, they’re watching over you.
58:57But you still have your freedom and your will.
59:01But this is so to me, this is the crucial question.
59:05Listening to you talk like one of my question
59:08is, are these people on my side?
59:10Are you on my side?
59:11And when you talk about humans remaining in charge,
59:14I think you’re on my side.
59:16That’s good.
59:17But one thing I’ve done in the past on this show and we’ll
59:20end here, is I read poems to technologists,
59:23and you supplied the poem "Machines of Loving Grace"
59:25the name of a poem by Richard Brautigan.
59:27Yes here’s how the poem ends.
59:30I like to think it has to be of a cybernetic ecology
59:35where we are free of our labors
59:37and joined back to nature, returned to our mammal
59:41brothers and sisters, and all watched over
59:44by machines of loving grace.
59:48To me, that sounds like the dystopian end
59:52where human beings are reanimated, minimalized
59:56and reduced and however benevolently the machines
1:00:01are in charge.
1:00:02So last question.
1:00:04What do you hear when you hear that poem?
1:00:05And if I think that’s a dystopia, are you on my side?
1:00:09It’s actually that poem is interesting because it’s
1:00:12interpretable in several different ways.
1:00:15There some people say it’s actually ironic that he says
1:00:22it’s not going to happen quite that way.
1:00:24Knowing the poet himself, then yes,
1:00:27I think that’s a reasonable interpretation.
1:00:29That’s one interpretation.
1:00:31Some people would have your interpretation,
1:00:33which is it’s meant literally, but maybe it’s not a good
1:00:36thing.
1:00:37But you could also interpret it as it’s a return to nature.
1:00:40It’s return to the core of what human.
1:00:42We’re not being animalized.
1:00:44We’re being we’re being reconnected with the world.
1:00:48So I was aware of that ambiguity.
1:00:50And, because I’ve always been talking about the positive
1:00:54side and the negative side.
1:00:55So I actually think that may be a tension that we may face,
1:01:03which is that the positive world and the negative world
1:01:08in their early stages, maybe even in their middle stages,
1:01:12maybe even in their fairly late stages.
1:01:14I wonder if the distance between the good ending
1:01:19and some of the subtle bad endings is relatively small.
1:01:24If it’s a very subtle thing like we’ve put very subtle,
1:01:28made very subtle changes.
1:01:29Like if you eat a particular fruit from a tree in a garden
1:01:33or not.
1:01:33Hypothetically Very small thing
1:01:36Yeah big divergence Yeah.
1:01:39I guess this always comes back to there’s some fundamental
1:01:43questions here.
1:01:44Yes yeah.
1:01:45Well, I guess we’ll see how it plays out.
1:01:48I do think of people in your position
1:01:51as people whose moral choices will carry
1:01:55an unusual amount of weight.
1:01:58And so I wish you God’s help with them.
1:02:01Dario Amodei, thank you for joining me.
1:02:03Thank you for having me, Ross.
1:02:26But what if I’m a robot?

About this transcript

This page contains the full transcript of Anthropic's CEO: ‘We Don’t Know if the Models Are Conscious’ | Interesting Times with Ross Douthat by Interesting Times, generated from the public captions YouTube serves with the video. The transcript has 10,252 words across 1,363 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.