WEBVTT - Dario & Daniela Amodei: "This Is Gonna Determine the Fate of Nations and the Future of Humanity”

0:00:10.010 --> 0:00:14.350
<v Speaker 1>The Builders is a production of Kaleidoscope and iHeart, presented

0:00:14.510 --> 0:00:19.090
<v Speaker 1>by Ramp. Ramp is a financial operations platform that helps

0:00:19.210 --> 0:00:23.349
<v Speaker 1>ambitious companies manage their finances so their teams can spend

0:00:23.410 --> 0:00:27.230
<v Speaker 1>less time chasing receipts and approvals and more time bringing

0:00:27.380 --> 0:00:32.019
<v Speaker 1>extraordinary ideas to life. I'm Walter Isaacs, and welcome to

0:00:32.040 --> 0:00:35.920
<v Speaker 1>The Builders, my new series of interviews featuring the most

0:00:36.060 --> 0:00:42.210
<v Speaker 1>consequential minds shaping our future. A person's biography tells you

0:00:42.250 --> 0:00:47.229
<v Speaker 1>what values, what drives, sometimes what demons are bred into them.

0:00:47.870 --> 0:00:51.820
<v Speaker 1>It's particularly interesting in an era of AI because we're

0:00:51.860 --> 0:00:55.620
<v Speaker 1>watching the biographies of AI systems happen. We have to

0:00:55.660 --> 0:00:59.220
<v Speaker 1>make sure that we raise them carefully so that they

0:00:59.280 --> 0:01:02.220
<v Speaker 1>don't reflect the dark sides of humanity.

0:01:02.930 --> 0:01:07.289
<v Speaker 2>Alarming new details about robots from open AI hacking into

0:01:07.370 --> 0:01:10.270
<v Speaker 2>another company. Anthropic, the maker of Claude, says that it

0:01:10.290 --> 0:01:13.110
<v Speaker 2>blocked several potential plots this year by scientists who were

0:01:13.230 --> 0:01:16.030
<v Speaker 2>using its AI models to conduct research that could have

0:01:16.069 --> 0:01:17.850
<v Speaker 2>led to the creation of bioweapons.

0:01:18.950 --> 0:01:22.300
<v Speaker 1>And so I've begun to feel that the moral compass

0:01:22.330 --> 0:01:25.640
<v Speaker 1>of these builders is the most important thing you can

0:01:25.840 --> 0:01:30.339
<v Speaker 1>know about them, to understand what values we're imbuing into

0:01:30.400 --> 0:01:35.919
<v Speaker 1>these artificial agents. So today I'm speaking with Dario and

0:01:35.959 --> 0:01:37.459
<v Speaker 1>Daniela Amadei.

0:01:37.620 --> 0:01:40.260
<v Speaker 3>CEO and co-founder of AI giant Anthropic.

0:01:40.560 --> 0:01:44.140
<v Speaker 4>Dario Amadei with an urgent warning about artificial intelligence.

0:01:44.380 --> 0:01:48.320
<v Speaker 1>Dario and Daniela are the brother and sister duo who

0:01:48.380 --> 0:01:52.520
<v Speaker 1>run Anthropic, the maker of Claude. And for a while,

0:01:52.540 --> 0:01:56.160
<v Speaker 1>Anthropic has been seen as the good guys in the

0:01:56.240 --> 0:01:59.620
<v Speaker 1>AI race. But in the past few months, we've seen

0:02:00.000 --> 0:02:04.920
<v Speaker 1>worrying developments. Anthropic researcher Jacob Coxon resigned from his post,

0:02:05.000 --> 0:02:07.240
<v Speaker 1>sounding the alarm of what he called out of control.

0:02:07.260 --> 0:02:10.180
<v Speaker 2>In a now viral post, Coxon said the companies are, quote,

0:02:10.280 --> 0:02:14.260
<v Speaker 2>gambling with our lives. The people building AI earnestly believe

0:02:14.300 --> 0:02:16.440
<v Speaker 2>that it could kill us all by the end of

0:02:16.480 --> 0:02:17.040
<v Speaker 2>the decade.

0:02:17.840 --> 0:02:21.320
<v Speaker 1>All of this, along with reports of a possible IPO,

0:02:21.660 --> 0:02:25.840
<v Speaker 1>has begun to make me wonder, how deep does Anthropic's

0:02:25.940 --> 0:02:30.720
<v Speaker 1>commitment to humanity really run? And what happens to those principles?

0:02:31.500 --> 0:02:32.860
<v Speaker 1>when they get in the way of winning.

0:02:33.639 --> 0:02:37.520
<v Speaker 2>Anthropic has reportedly warned investors that its technology may pose

0:02:37.740 --> 0:02:41.750
<v Speaker 2>existential risk to humanity in its IPO prospectus. And you

0:02:41.810 --> 0:02:45.429
<v Speaker 2>might then wonder, with a company warning of this, why

0:02:45.650 --> 0:02:47.450
<v Speaker 2>it's still planning to go ahead with the.

0:02:47.450 --> 0:02:48.169
<v Speaker 5>IPO at all.

0:02:48.190 --> 0:02:52.230
<v Speaker 2>The public sale could value Anthropic at $ 2 trillion.

0:02:54.610 --> 0:02:57.890
<v Speaker 1>I chatted with Daniela and Dario in July of 2026,

0:02:57.870 --> 0:03:03.130
<v Speaker 1>at the time I didn't know just how prescient this

0:03:03.210 --> 0:03:09.410
<v Speaker 1>conversation would be. Here's my interview with Dario and Daniela Amadei.

0:03:12.120 --> 0:03:13.139
<v Speaker 1>Welcome to The Builders.

0:03:13.600 --> 0:03:14.760
<v Speaker 3>Thank you so much for having us.

0:03:15.520 --> 0:03:18.260
<v Speaker 1>You know, Steve Jobs, whenever he launched a product, always

0:03:18.280 --> 0:03:20.660
<v Speaker 1>had that street sign of, you know, the humanities and

0:03:20.720 --> 0:03:24.970
<v Speaker 1>technology standing at that intersection. And when you were growing up, Daniela,

0:03:25.270 --> 0:03:26.290
<v Speaker 1>you were a flautist.

0:03:26.990 --> 0:03:27.970
<v Speaker 5>You loved the arts.

0:03:28.450 --> 0:03:31.740
<v Speaker 1>And Dario, you always a scientist. Did that help you gel?

0:03:31.800 --> 0:03:34.800
<v Speaker 1>And do you think that combination is important?

0:03:35.640 --> 0:03:39.240
<v Speaker 2>Yeah, I think that combination is a lot of what's

0:03:39.320 --> 0:03:43.710
<v Speaker 2>behind the company, right? One theory I have is that

0:03:44.290 --> 0:03:47.290
<v Speaker 2>running a company is too big a task for any

0:03:47.410 --> 0:03:50.090
<v Speaker 2>one person. The range of skills that it involves, it

0:03:50.150 --> 0:03:52.810
<v Speaker 2>is not possible for any single person to have. So

0:03:52.850 --> 0:03:55.680
<v Speaker 2>you kind of have to assemble a full complement of

0:03:55.750 --> 0:03:58.900
<v Speaker 2>skills across the top, like 10 or 15 people in

0:03:58.940 --> 0:04:02.030
<v Speaker 2>the company or something like that. And every very successful

0:04:02.070 --> 0:04:04.870
<v Speaker 2>company splits it up in different ways and solves the

0:04:04.910 --> 0:04:08.590
<v Speaker 2>problem in different ways. And I think with the sciences

0:04:08.650 --> 0:04:11.380
<v Speaker 2>and the humanities thing, you know, Daniela and I kind

0:04:11.410 --> 0:04:12.800
<v Speaker 2>of embody that and split that.

0:04:13.080 --> 0:04:13.320
<v Speaker 1>Yeah.

0:04:13.840 --> 0:04:15.560
<v Speaker 4>And I would say, you know, it's interesting, like, there

0:04:15.580 --> 0:04:19.740
<v Speaker 4>was an off-site we did in maybe 2023 where, like,

0:04:20.320 --> 0:04:22.210
<v Speaker 4>we sort of were talking about just, like, what does

0:04:22.250 --> 0:04:23.490
<v Speaker 4>it feel like to work at Anthropic?

0:04:23.510 --> 0:04:25.710
<v Speaker 3>Like, how do we kind of embody what our culture is?

0:04:26.150 --> 0:04:29.489
<v Speaker 4>And we actually said, you know, it's an unusual blend

0:04:29.529 --> 0:04:32.529
<v Speaker 4>of science and humanism. And it was only later that

0:04:32.550 --> 0:04:34.330
<v Speaker 4>I'm like, that also is how you would describe me

0:04:34.390 --> 0:04:34.890
<v Speaker 4>and Dario.

0:04:35.000 --> 0:04:36.920
<v Speaker 1>Well, when you were kids, did you know you wanted

0:04:36.960 --> 0:04:38.560
<v Speaker 1>to work together? Yes.

0:04:38.620 --> 0:04:41.730
<v Speaker 2>Yes. I don't think either of us imagined... that we

0:04:41.750 --> 0:04:46.140
<v Speaker 2>would start a company. That's not really how either of

0:04:46.240 --> 0:04:48.659
<v Speaker 2>us thought of our lives or about what we wanted

0:04:48.740 --> 0:04:52.080
<v Speaker 2>to do. I had always wanted to be a scientist.

0:04:52.100 --> 0:04:56.460
<v Speaker 2>Daniela can speak to her aspirations. But I actually found

0:04:56.940 --> 0:04:59.289
<v Speaker 2>companies and business to be of very little interest to me.

0:05:00.470 --> 0:05:03.909
<v Speaker 4>I would say Dario and I were very unusual in

0:05:03.970 --> 0:05:07.130
<v Speaker 4>that from a really early age, we just genuinely really

0:05:07.190 --> 0:05:07.729
<v Speaker 4>liked each other.

0:05:07.750 --> 0:05:09.029
<v Speaker 3>And I agree with Dario.

0:05:09.089 --> 0:05:11.260
<v Speaker 4>I don't think starting a company was ever on our

0:05:11.279 --> 0:05:15.760
<v Speaker 4>bingo card. Neither of us were particularly company-oriented. I was

0:05:15.800 --> 0:05:17.980
<v Speaker 4>always really interested in, I think I would call myself

0:05:18.020 --> 0:05:22.010
<v Speaker 4>a classic do-gooder. So I worked in international development. I

0:05:22.050 --> 0:05:25.330
<v Speaker 4>felt really strongly about helping people everywhere in the world

0:05:25.390 --> 0:05:30.110
<v Speaker 4>have equal access and opportunity. And I became a vegetarian

0:05:30.130 --> 0:05:31.730
<v Speaker 4>when I was like quite young. It's like the only

0:05:31.770 --> 0:05:34.570
<v Speaker 4>thing that I did first. Dario did everything first because

0:05:34.589 --> 0:05:36.589
<v Speaker 4>he's older, but I became a vegetarian and then he

0:05:36.610 --> 0:05:37.470
<v Speaker 4>became a vegetarian.

0:05:38.070 --> 0:05:41.300
<v Speaker 1>And Dario, tell me how you and Daniela divide up

0:05:41.360 --> 0:05:43.180
<v Speaker 1>the roles and then how you fill it in with

0:05:43.260 --> 0:05:44.020
<v Speaker 1>others in the team.

0:05:45.120 --> 0:05:48.839
<v Speaker 2>Yeah, so I think I can function best when I

0:05:48.880 --> 0:05:52.600
<v Speaker 2>think about the big picture strategy and the research side

0:05:52.640 --> 0:05:56.170
<v Speaker 2>of things. I'm best at laying out a vision, seeing

0:05:56.250 --> 0:05:59.310
<v Speaker 2>what comes next. I have found that when I'm freed

0:05:59.350 --> 0:06:02.229
<v Speaker 2>up to focus entirely on that, that is the thing

0:06:02.250 --> 0:06:03.929
<v Speaker 2>that has the most value for the company.

0:06:05.540 --> 0:06:06.820
<v Speaker 5>Running the company day to.

0:06:06.779 --> 0:06:11.670
<v Speaker 2>Day, recruiting the people, just making sure that the machine doesn't,

0:06:11.990 --> 0:06:14.570
<v Speaker 2>you know, the wheels don't fall off the machine while

0:06:14.610 --> 0:06:16.549
<v Speaker 2>I'm trying to figure out what comes next.

0:06:17.029 --> 0:06:18.010
<v Speaker 5>Daniela does this.

0:06:18.150 --> 0:06:22.370
<v Speaker 2>And I think this division of labor allows both things

0:06:22.450 --> 0:06:26.170
<v Speaker 2>to be done to a much higher standard than either

0:06:26.230 --> 0:06:29.300
<v Speaker 2>of them would be done if we were distracted by

0:06:29.339 --> 0:06:30.240
<v Speaker 2>having to do the other thing.

0:06:30.580 --> 0:06:32.360
<v Speaker 4>Our chief people officer who knows both of us really

0:06:32.400 --> 0:06:34.820
<v Speaker 4>well is in a lot of conversations with us. The

0:06:34.860 --> 0:06:36.880
<v Speaker 4>way she put it is that Dario makes us ambitious

0:06:36.920 --> 0:06:39.360
<v Speaker 4>and I make us successful. And I thought that was

0:06:39.380 --> 0:06:43.180
<v Speaker 4>a really good encapsulation, which is, I think Anthropic could

0:06:43.220 --> 0:06:46.970
<v Speaker 4>be like a good successful business, but not change the

0:06:47.010 --> 0:06:49.530
<v Speaker 4>world if we didn't have Dario. And I think with Dario,

0:06:49.550 --> 0:06:51.290
<v Speaker 4>we would have the perfect vision, but our ability to

0:06:51.370 --> 0:06:52.270
<v Speaker 4>execute would.

0:06:52.130 --> 0:06:53.990
<v Speaker 3>Be, it would just be more complicated.

0:06:54.630 --> 0:06:57.609
<v Speaker 4>And so I think there's something special about, A, the

0:06:57.650 --> 0:06:59.820
<v Speaker 4>fact that we know each other really well, B, it's

0:06:59.880 --> 0:07:04.250
<v Speaker 4>extremely convenient that our skills are like 100% complimentary, or

0:07:04.270 --> 0:07:06.489
<v Speaker 4>maybe not 100%, but I would say close to 100%

0:07:06.490 --> 0:07:09.610
<v Speaker 4>complimentary in this fashion. And this is not even to

0:07:09.650 --> 0:07:12.310
<v Speaker 4>mention the fact that we have five co-founders along with

0:07:12.330 --> 0:07:16.190
<v Speaker 4>the two of us who also own like incredibly important

0:07:16.330 --> 0:07:19.890
<v Speaker 4>parts of the business, the culture, like the organization. It's

0:07:19.930 --> 0:07:21.510
<v Speaker 4>hard for me to imagine how you would do this

0:07:21.550 --> 0:07:22.330
<v Speaker 4>with like just the.

0:07:22.270 --> 0:07:22.690
<v Speaker 3>Two of us.

0:07:23.250 --> 0:07:26.120
<v Speaker 1>So go back to growing up and you have these

0:07:26.220 --> 0:07:30.780
<v Speaker 1>extraordinary parents and they help teach you right from wrong.

0:07:30.880 --> 0:07:32.360
<v Speaker 1>How do they do that?

0:07:33.440 --> 0:07:36.280
<v Speaker 4>I don't know that it's something that was ever... You know,

0:07:36.320 --> 0:07:38.440
<v Speaker 4>sometimes I hear people say, like, you know, we covered

0:07:39.060 --> 0:07:42.480
<v Speaker 4>different forms of philosophical ethics at the dinner table.

0:07:42.500 --> 0:07:43.760
<v Speaker 3>That was not how it was done.

0:07:44.200 --> 0:07:46.110
<v Speaker 4>My mom does have a philosophy degree, but she didn't

0:07:46.150 --> 0:07:48.010
<v Speaker 4>subject us to talking about Kant or anything like that.

0:07:48.530 --> 0:07:51.050
<v Speaker 4>I think it was much more, like, inherent in terms

0:07:51.090 --> 0:07:55.410
<v Speaker 4>of values, like, just the belief that people everywhere in

0:07:55.450 --> 0:07:58.960
<v Speaker 4>the world are, like... deserving of opportunity and dignity and

0:07:59.280 --> 0:08:01.740
<v Speaker 4>rights and respect. I think there was something special about

0:08:01.760 --> 0:08:04.360
<v Speaker 4>the feeling that like we could be whoever we wanted

0:08:04.420 --> 0:08:06.300
<v Speaker 4>to be as long as there was this kind of

0:08:06.740 --> 0:08:09.390
<v Speaker 4>faith and belief that like, hey, part of what it

0:08:09.470 --> 0:08:11.950
<v Speaker 4>means to be a successful human is to be a

0:08:11.990 --> 0:08:13.930
<v Speaker 4>good person and to be kind to other people.

0:08:14.070 --> 0:08:15.390
<v Speaker 5>Yeah, it felt implicit.

0:08:15.510 --> 0:08:17.610
<v Speaker 2>There wasn't any kind of, you know, this is what

0:08:17.630 --> 0:08:18.510
<v Speaker 2>you should do or this is.

0:08:18.490 --> 0:08:19.440
<v Speaker 5>How you live in the world.

0:08:20.420 --> 0:08:22.800
<v Speaker 1>Well, one reason I ask, as you might suspect, is

0:08:22.860 --> 0:08:26.570
<v Speaker 1>now you've got a really important kid named Claude And

0:08:26.610 --> 0:08:29.790
<v Speaker 1>to what extent does that way that you grew up

0:08:30.230 --> 0:08:33.240
<v Speaker 1>help you, inform you about how you teach Claude to

0:08:33.250 --> 0:08:34.240
<v Speaker 1>know right from wrong?

0:08:34.760 --> 0:08:38.120
<v Speaker 2>Yeah, you know, we put out a constitution for Claude,

0:08:38.300 --> 0:08:40.260
<v Speaker 2>I think it was a few months ago. And I

0:08:40.300 --> 0:08:43.600
<v Speaker 2>think there are some interesting lessons in it, which we've

0:08:43.700 --> 0:08:48.370
<v Speaker 2>learned from several years of trying to train Claude in

0:08:48.429 --> 0:08:52.410
<v Speaker 2>successive versions. I think early on, we had been imagining

0:08:52.470 --> 0:08:54.290
<v Speaker 2>that there would be various rules that, you know, you

0:08:54.309 --> 0:08:57.280
<v Speaker 2>would tell Claude, you know, don't, Don't help someone commit

0:08:57.300 --> 0:08:59.960
<v Speaker 2>a crime or don't make biological weapons or whatever.

0:09:00.440 --> 0:09:02.720
<v Speaker 5>But it turns out that when you make this kind.

0:09:02.600 --> 0:09:05.360
<v Speaker 2>Of discrete list of rules, you end up with this

0:09:05.420 --> 0:09:07.580
<v Speaker 2>kind of spiky shape where you're trying to teach the

0:09:07.620 --> 0:09:11.880
<v Speaker 2>model like 170 different rules that are kind of arbitrary.

0:09:12.320 --> 0:09:15.660
<v Speaker 2>We found it's better to teach the model some basic

0:09:16.160 --> 0:09:20.200
<v Speaker 2>principles like what it should prioritize, how it should reason

0:09:20.280 --> 0:09:23.040
<v Speaker 2>about what the right thing to do is. Which, of course,

0:09:23.080 --> 0:09:25.660
<v Speaker 2>I think that's the right way to teach people as well, right? Like,

0:09:25.679 --> 0:09:29.000
<v Speaker 2>you know, if I just give someone a list of rules,

0:09:29.040 --> 0:09:31.500
<v Speaker 2>thou shalt do this, thou shalt not do this. You know,

0:09:31.520 --> 0:09:35.570
<v Speaker 2>I don't think that produces as robust a morality as

0:09:35.630 --> 0:09:37.230
<v Speaker 2>I say, this is how you should think about things.

0:09:37.250 --> 0:09:40.130
<v Speaker 2>You should learn how to think for yourself. And these

0:09:40.170 --> 0:09:42.410
<v Speaker 2>are the principles on which you should think, right? Principles

0:09:42.450 --> 0:09:47.690
<v Speaker 2>of open inquiry, principles of care for other people, principles

0:09:47.929 --> 0:09:51.209
<v Speaker 2>of fairness and justice. And then everything is derived from that.

0:09:51.620 --> 0:09:55.630
<v Speaker 2>And that is, frankly, how Claude learns. We find that

0:09:55.650 --> 0:09:59.429
<v Speaker 2>to be surprisingly effective because that allows Claude to generalize

0:09:59.470 --> 0:10:01.810
<v Speaker 2>to new situations that it might never have imagined.

0:10:02.830 --> 0:10:05.220
<v Speaker 1>Danielle, I can't help myself when you said your mom

0:10:05.240 --> 0:10:10.800
<v Speaker 1>was an expert on Kant. Did you actually wrestle with philosophy?

0:10:11.160 --> 0:10:16.000
<v Speaker 4>You know, what's interesting is the woman who is responsible

0:10:16.040 --> 0:10:19.840
<v Speaker 4>for training Claude's character has a PhD in philosophy.

0:10:20.730 --> 0:10:22.350
<v Speaker 3>It's just, it's unusual, right?

0:10:22.370 --> 0:10:24.569
<v Speaker 4>You would think, oh, the person that's training the sort

0:10:24.590 --> 0:10:27.530
<v Speaker 4>of character of these models must be a computer scientist

0:10:27.590 --> 0:10:30.729
<v Speaker 4>or somebody who's studied ML or AI for many decades.

0:10:31.100 --> 0:10:34.280
<v Speaker 4>Amanda's technical and she like technically trains the model, but

0:10:34.300 --> 0:10:38.340
<v Speaker 4>I think it is very felt in Claude's character when

0:10:38.380 --> 0:10:41.980
<v Speaker 4>you communicate with it, that there's this kind of experience

0:10:42.040 --> 0:10:45.700
<v Speaker 4>of something that's very wise, kind of the better angel

0:10:45.720 --> 0:10:48.860
<v Speaker 4>of your nature trying to answer your questions, But there

0:10:49.100 --> 0:10:51.239
<v Speaker 4>are like if you kind of poke Claude in these

0:10:51.300 --> 0:10:56.740
<v Speaker 4>like difficult areas where maybe different philosophical traditions might conflict.

0:10:57.100 --> 0:10:58.900
<v Speaker 3>I think that's where you especially feel, Amanda.

0:10:59.179 --> 0:11:02.449
<v Speaker 4>Because you'll sort of see like, oh, well, you know,

0:11:02.530 --> 0:11:04.410
<v Speaker 4>you might want to consider things from like this point

0:11:04.450 --> 0:11:06.370
<v Speaker 4>of view, but you could also consider it from this

0:11:06.450 --> 0:11:09.170
<v Speaker 4>point of view. And I do think it's it's important

0:11:09.210 --> 0:11:13.090
<v Speaker 4>because as the world evolves. begins to rely on these

0:11:13.150 --> 0:11:15.250
<v Speaker 4>models more and more, you don't want to be locked

0:11:15.270 --> 0:11:19.800
<v Speaker 4>into sort of one philosophical tradition because ethics and morality

0:11:19.840 --> 0:11:23.580
<v Speaker 4>are quite complicated. And different people and different cultures might

0:11:23.620 --> 0:11:26.140
<v Speaker 4>have different answers about how they would like you to behave.

0:11:28.350 --> 0:11:43.199
<v Speaker 1>Don't go anywhere. There's a lot more right after this break. Dario,

0:11:43.260 --> 0:11:46.500
<v Speaker 1>you were there at one of the seminal moments of

0:11:46.580 --> 0:11:52.090
<v Speaker 1>modern times, which is when GPT-1 and 2 suddenly burst

0:11:52.190 --> 0:11:55.689
<v Speaker 1>on the scene, surprising, perhaps even surprising you. You were

0:11:56.130 --> 0:11:59.929
<v Speaker 1>head of research at OpenAI. What was that moment like?

0:12:01.130 --> 0:12:04.809
<v Speaker 2>Yeah, so I wasn't very involved in GPT-1, but I

0:12:04.850 --> 0:12:07.030
<v Speaker 2>was very involved in the idea that we should run

0:12:07.050 --> 0:12:13.790
<v Speaker 2>a larger scale, more expensive, more ambitious experiment. on the

0:12:13.890 --> 0:12:16.329
<v Speaker 2>idea that if we ran these language models at a

0:12:16.370 --> 0:12:18.350
<v Speaker 2>larger scale, they would be better at language.

0:12:18.390 --> 0:12:19.630
<v Speaker 5>They would be more coherent.

0:12:20.050 --> 0:12:22.719
<v Speaker 2>So it's almost like one of these moments in scientific

0:12:22.750 --> 0:12:26.179
<v Speaker 2>history where you're like, you propose a definitive experiment to

0:12:26.220 --> 0:12:28.780
<v Speaker 2>just make a very strange prediction. And then if the

0:12:28.820 --> 0:12:31.660
<v Speaker 2>prediction turns out to be correct, you're like, God, we're really,

0:12:32.080 --> 0:12:36.540
<v Speaker 2>we're really onto something. So I see GPT-2 as the

0:12:36.610 --> 0:12:40.850
<v Speaker 2>first real test of the scaling laws hypothesis. And What

0:12:40.910 --> 0:12:45.220
<v Speaker 2>I found super interesting was the enormous divergence in reactions

0:12:45.280 --> 0:12:45.530
<v Speaker 2>to it.

0:12:46.420 --> 0:12:48.040
<v Speaker 1>And were you kind of surprised?

0:12:48.920 --> 0:12:52.300
<v Speaker 5>I personally was super freaked out by it.

0:12:52.380 --> 0:12:55.079
<v Speaker 2>Like, I looked at it and I said, my God,

0:12:55.120 --> 0:12:57.339
<v Speaker 2>you just scale this thing up another factor of a

0:12:57.380 --> 0:13:00.340
<v Speaker 2>million and it'll be better than humans at everything. Like,

0:13:00.720 --> 0:13:04.690
<v Speaker 2>this is the defining thing. economic and scientific moment of

0:13:04.770 --> 0:13:06.689
<v Speaker 2>the modern age. We just don't know it yet, right?

0:13:06.710 --> 0:13:09.650
<v Speaker 2>This is going to determine the fate of nations and

0:13:09.730 --> 0:13:12.870
<v Speaker 2>economies and the future of humanity. Like, you know, you

0:13:12.890 --> 0:13:16.410
<v Speaker 2>peer into the crystal ball, you see the whole future unfurled. And,

0:13:16.530 --> 0:13:18.589
<v Speaker 2>you know, it really scared me. I said, oh my God,

0:13:18.630 --> 0:13:20.429
<v Speaker 2>what are the implications of that? Is this the right

0:13:20.500 --> 0:13:23.280
<v Speaker 2>place to be developing it? Is, you know, what's going

0:13:23.300 --> 0:13:23.940
<v Speaker 2>to happen next?

0:13:23.980 --> 0:13:25.440
<v Speaker 5>What do we do with this? Who do we tell

0:13:25.500 --> 0:13:26.020
<v Speaker 5>about this?

0:13:26.640 --> 0:13:29.260
<v Speaker 2>And maybe 10% of the people around me had exactly

0:13:29.280 --> 0:13:30.160
<v Speaker 2>the same reaction.

0:13:30.179 --> 0:13:31.580
<v Speaker 5>And it was often the smartest ones.

0:13:31.580 --> 0:13:37.689
<v Speaker 2>90% of the people around me, because GBD2' s capabilities in

0:13:37.750 --> 0:13:41.230
<v Speaker 2>any particular area were not that impressive, right?

0:13:41.610 --> 0:13:42.090
<v Speaker 5>It was the.

0:13:42.050 --> 0:13:45.740
<v Speaker 2>Difference between GBD1 and GBD2, the generality, the fact that

0:13:45.760 --> 0:13:51.560
<v Speaker 2>it could translate languages and summarize a document and write

0:13:51.600 --> 0:13:53.390
<v Speaker 2>a few lines of code here and there.

0:13:53.880 --> 0:13:55.080
<v Speaker 5>But people would look at it and they would say,

0:13:55.100 --> 0:13:57.500
<v Speaker 5>in any particular area, this doesn't seem so smart. And

0:13:57.540 --> 0:14:00.359
<v Speaker 5>so many people just shrugged and they just dismissed it.

0:14:00.730 --> 0:14:02.830
<v Speaker 2>And you get the usual comments from kind of the

0:14:02.910 --> 0:14:05.890
<v Speaker 2>jaded people, ah, tech people have done this before. Here

0:14:06.170 --> 0:14:08.670
<v Speaker 2>are all the limitations of the models. And I just

0:14:08.710 --> 0:14:10.250
<v Speaker 2>look at it and I say, it's the same thing

0:14:10.370 --> 0:14:13.530
<v Speaker 2>over and over again. The limitations of the models now,

0:14:13.870 --> 0:14:17.089
<v Speaker 2>six months from now, the things that seemed hard, they're

0:14:17.110 --> 0:14:18.969
<v Speaker 2>going to be easy. 12 months from now, the things

0:14:18.990 --> 0:14:21.130
<v Speaker 2>that seemed impossible, the models are going to be doing

0:14:21.150 --> 0:14:25.739
<v Speaker 2>them regularly. Everything we saw with GPT-2, we're seeing it

0:14:25.960 --> 0:14:28.900
<v Speaker 2>over and over again in wider and wider swaths.

0:14:28.590 --> 0:14:29.120
<v Speaker 1>Of the world.

0:14:29.590 --> 0:14:33.070
<v Speaker 2>And the world is gradually finally being convinced. You know,

0:14:33.310 --> 0:14:36.690
<v Speaker 2>we just announced that we're going to be developing drugs for,

0:14:36.870 --> 0:14:38.450
<v Speaker 2>you know, neglected, rare diseases.

0:14:39.390 --> 0:14:40.590
<v Speaker 3>I think you were quite polite about it.

0:14:40.590 --> 0:14:42.510
<v Speaker 4>But I mean, I remember at the time there was

0:14:42.550 --> 0:14:47.160
<v Speaker 4>some like very well-respected, you know, computer scientists and like

0:14:47.300 --> 0:14:50.500
<v Speaker 4>AI and ML leaders who were like basically like bullied

0:14:50.520 --> 0:14:52.560
<v Speaker 4>some of the researchers who worked on it on Twitter

0:14:52.600 --> 0:14:54.220
<v Speaker 4>and were like, this is stupid.

0:14:54.280 --> 0:14:55.840
<v Speaker 3>This model is so incapable.

0:14:55.940 --> 0:14:57.800
<v Speaker 4>And it was like you were sort of a quack

0:14:57.870 --> 0:15:01.090
<v Speaker 4>if you believed this was like a really significant moment.

0:15:01.110 --> 0:15:02.750
<v Speaker 5>And that's all been memory hold.

0:15:02.810 --> 0:15:06.720
<v Speaker 2>Like all the researchers who like were like, oh, this is,

0:15:07.170 --> 0:15:08.430
<v Speaker 2>this is pseudoscience.

0:15:08.510 --> 0:15:09.450
<v Speaker 5>It doesn't make sense.

0:15:09.510 --> 0:15:12.190
<v Speaker 2>Like all of them are applying for jobs at like

0:15:12.610 --> 0:15:15.310
<v Speaker 2>the LLM companies. And like, you know, we never say

0:15:15.370 --> 0:15:18.410
<v Speaker 2>it out loud to any particular one, but like, you know,

0:15:18.460 --> 0:15:20.440
<v Speaker 2>we look at them and they're like, I, we remember

0:15:20.480 --> 0:15:23.780
<v Speaker 2>what you were saying in 2019, you have no credibility

0:15:23.840 --> 0:15:24.320
<v Speaker 2>on this matter.

0:15:25.860 --> 0:15:29.060
<v Speaker 1>So Daniela, you were vice president of safety back then

0:15:29.160 --> 0:15:33.510
<v Speaker 1>at OpenAI. And when you hit GPT-2, I assume you're

0:15:33.710 --> 0:15:37.290
<v Speaker 1>like Dario, somebody who's almost freaked out about it. Did

0:15:37.350 --> 0:15:39.250
<v Speaker 1>it freak you out about safety too?

0:15:40.150 --> 0:15:43.220
<v Speaker 4>I think, you know, there wasn't something about GBT2, the

0:15:43.260 --> 0:15:46.080
<v Speaker 4>model itself, that made us, you know, concerned, right? We

0:15:46.120 --> 0:15:48.360
<v Speaker 4>didn't look at that model and say, wow, we think

0:15:48.400 --> 0:15:51.720
<v Speaker 4>this model has like cybersecurity implications or something, right? The

0:15:51.760 --> 0:15:54.980
<v Speaker 4>model itself was in some ways kind of laughably wrong

0:15:55.040 --> 0:15:56.920
<v Speaker 4>about a lot of things, right? It was still a very,

0:15:56.940 --> 0:15:59.970
<v Speaker 4>I would say, harmless model. But I think the thing

0:16:00.330 --> 0:16:03.550
<v Speaker 4>that Dario mentioned that I also felt was a sense

0:16:03.610 --> 0:16:08.190
<v Speaker 4>of like, okay, the difference between GPT-1 and GPT-2, just

0:16:08.230 --> 0:16:11.770
<v Speaker 4>like extrapolate that out, right? Like there was something about seeing,

0:16:12.170 --> 0:16:14.240
<v Speaker 4>it's almost like if you'd never seen a child before

0:16:14.250 --> 0:16:16.680
<v Speaker 4>and you'd only seen adults and you were kind of like,

0:16:17.080 --> 0:16:18.240
<v Speaker 4>I don't know what this little thing is.

0:16:18.520 --> 0:16:19.740
<v Speaker 3>And then suddenly it starts to walk.

0:16:20.080 --> 0:16:23.300
<v Speaker 4>You're like, wow, huh, I bet this thing is going

0:16:23.320 --> 0:16:24.860
<v Speaker 4>to be capable of doing a lot of the things

0:16:24.900 --> 0:16:26.200
<v Speaker 4>that a grownup could do one day.

0:16:26.280 --> 0:16:27.780
<v Speaker 3>Like maybe this is where this is going.

0:16:28.540 --> 0:16:33.560
<v Speaker 1>Dario, you say your departure from OpenAI was partly due

0:16:33.600 --> 0:16:38.440
<v Speaker 1>to a breakdown of trust, an interesting word, with Sam Altman.

0:16:38.940 --> 0:16:42.320
<v Speaker 1>Explain how you felt there was a breakdown of trust

0:16:42.500 --> 0:16:44.120
<v Speaker 1>and honesty there.

0:16:45.040 --> 0:16:49.119
<v Speaker 2>What I will say is, look, leaving OpenAI and founding Anthropic,

0:16:49.200 --> 0:16:52.780
<v Speaker 2>I mean, it ultimately comes down to the principles and

0:16:52.820 --> 0:16:54.210
<v Speaker 2>how we're going to build the technology.

0:16:54.260 --> 0:16:56.540
<v Speaker 5>That is what matters at the end of the day.

0:16:57.160 --> 0:16:59.760
<v Speaker 2>And, you know, I think over the last five and

0:16:59.780 --> 0:17:03.420
<v Speaker 2>a half years, we have seen how those principles differ.

0:17:03.540 --> 0:17:06.300
<v Speaker 2>You know, we see, you know, Anthropic has advocated for

0:17:06.900 --> 0:17:10.030
<v Speaker 2>sensible regulation and safeguards on the technology.

0:17:10.070 --> 0:17:13.590
<v Speaker 5>Anthropic has invested in mechanistic.

0:17:13.070 --> 0:17:18.210
<v Speaker 2>Interpretability research to understand what's going on inside the models, right?

0:17:18.369 --> 0:17:22.389
<v Speaker 2>You know, Anthropic has managed to maintain a culture of

0:17:22.530 --> 0:17:25.990
<v Speaker 2>unity to have the highest retention rates. We haven't had

0:17:26.510 --> 0:17:29.780
<v Speaker 2>a number of the scandals that other AI companies have had. So,

0:17:30.130 --> 0:17:33.640
<v Speaker 2>you know, I think that's kind of sufficient to describe

0:17:33.680 --> 0:17:34.280
<v Speaker 2>the difference.

0:17:34.720 --> 0:17:35.860
<v Speaker 5>But I will say.

0:17:35.660 --> 0:17:40.040
<v Speaker 2>That the moment when I realized that we needed to

0:17:40.100 --> 0:17:43.210
<v Speaker 2>found our own company, that OpenAI was not the right

0:17:43.330 --> 0:17:47.929
<v Speaker 2>place for us. Look, these safety issues are difficult. You

0:17:47.970 --> 0:17:51.850
<v Speaker 2>can have lots of disagreements on them. I disagree with

0:17:51.890 --> 0:17:55.010
<v Speaker 2>the co-founders and with lots of people at Anthropic. But

0:17:55.270 --> 0:17:58.300
<v Speaker 2>the reason we've been able to resolve these disagreements is

0:17:58.340 --> 0:18:00.660
<v Speaker 2>because they're held in good faith, right?

0:18:01.040 --> 0:18:03.720
<v Speaker 5>Should we release this model or should we wait?

0:18:04.060 --> 0:18:07.700
<v Speaker 2>Should we put stronger or weaker biosafeguards on the model?

0:18:08.000 --> 0:18:10.570
<v Speaker 2>You know, how much of our effort should we spend

0:18:10.869 --> 0:18:13.469
<v Speaker 2>lobbying for safety regulations and what should.

0:18:13.310 --> 0:18:13.910
<v Speaker 5>They look like?

0:18:14.250 --> 0:18:18.210
<v Speaker 2>These are all good faith disagreements, but they always get

0:18:18.250 --> 0:18:21.290
<v Speaker 2>resolved because there's always trust that everyone is trying to

0:18:21.330 --> 0:18:23.540
<v Speaker 2>get to the right answer and everyone is being honest.

0:18:24.340 --> 0:18:26.920
<v Speaker 2>My point in saying that there's a breakdown of trust

0:18:27.020 --> 0:18:29.860
<v Speaker 2>is simply to make the abstract point that if you

0:18:29.960 --> 0:18:32.800
<v Speaker 2>don't believe that about a place where you're working, if

0:18:32.840 --> 0:18:36.540
<v Speaker 2>you don't believe that the disagreements are being held in

0:18:36.600 --> 0:18:39.780
<v Speaker 2>good faith and that you're getting the real story from people,

0:18:40.220 --> 0:18:42.260
<v Speaker 2>how can you possibly hope to resolve them?

0:18:42.720 --> 0:18:45.540
<v Speaker 5>That's why we left. That's why we founded our own thing.

0:18:45.859 --> 0:18:47.740
<v Speaker 2>And I think the last five and a half years

0:18:48.080 --> 0:18:53.070
<v Speaker 2>have are themselves Just the best proof of how our

0:18:53.109 --> 0:18:55.169
<v Speaker 2>philosophy differs from that of.

0:18:55.150 --> 0:18:56.689
<v Speaker 5>The other companies in the space.

0:18:57.490 --> 0:19:00.570
<v Speaker 2>I think if you continue to work for someone who

0:19:00.609 --> 0:19:04.250
<v Speaker 2>has a different vision than you, it becomes very unclear

0:19:04.330 --> 0:19:05.650
<v Speaker 2>what you actually stand for.

0:19:06.150 --> 0:19:07.170
<v Speaker 5>Wait a minute, though.

0:19:07.190 --> 0:19:09.390
<v Speaker 1>You push back a bit here because you talk about

0:19:09.450 --> 0:19:14.250
<v Speaker 1>differences in values and stuff. But that's fundamentally different than

0:19:14.430 --> 0:19:18.490
<v Speaker 1>a difference in trust, that you feel that somebody says

0:19:18.510 --> 0:19:20.739
<v Speaker 1>they have the same values. but they don't and you

0:19:20.800 --> 0:19:22.000
<v Speaker 1>can't trust them.

0:19:22.980 --> 0:19:27.740
<v Speaker 2>Honesty is one of our values. It's a very important value.

0:19:28.160 --> 0:19:31.959
<v Speaker 2>It's hard to know what someone's values are if they

0:19:31.980 --> 0:19:33.310
<v Speaker 2>don't have the value of honesty.

0:19:33.970 --> 0:19:35.129
<v Speaker 5>And you felt that about Sam?

0:19:37.290 --> 0:19:39.210
<v Speaker 4>I think the way, at least for me, I would

0:19:39.250 --> 0:19:43.930
<v Speaker 4>put it is I think the co-founders are an unusually

0:19:43.950 --> 0:19:49.350
<v Speaker 4>high integrity group of people. And all of us, I think,

0:19:49.410 --> 0:19:52.440
<v Speaker 4>shared a very similar vision for how we wanted this

0:19:52.500 --> 0:19:54.699
<v Speaker 4>technology to go and how we wanted to build it.

0:19:55.359 --> 0:19:57.580
<v Speaker 4>But at a certain point, it just makes more sense

0:19:57.619 --> 0:19:59.780
<v Speaker 4>to go do this as your own organization.

0:19:59.820 --> 0:20:03.379
<v Speaker 2>How are you going to get someone else to implement

0:20:03.520 --> 0:20:07.919
<v Speaker 2>or share your vision if they don't naturally do so?

0:20:08.619 --> 0:20:12.060
<v Speaker 2>no matter how much they say they share it, if

0:20:12.100 --> 0:20:16.379
<v Speaker 2>that's not reflected in what the company actually does, what

0:20:16.420 --> 0:20:17.100
<v Speaker 2>choice do you have?

0:20:17.480 --> 0:20:17.720
<v Speaker 5>Yeah.

0:20:18.200 --> 0:20:21.460
<v Speaker 4>And I do think Anthropic is a company that is

0:20:21.550 --> 0:20:24.350
<v Speaker 4>founded on trust. And I think what I mean by

0:20:24.390 --> 0:20:30.070
<v Speaker 4>that is like, I think honesty and integrity and.

0:20:30.770 --> 0:20:32.510
<v Speaker 3>How we treat each.

0:20:32.430 --> 0:20:35.390
<v Speaker 4>Other and our employees is also reflective of how we

0:20:35.430 --> 0:20:37.469
<v Speaker 4>want the technology to be built and to go.

0:20:38.880 --> 0:20:42.300
<v Speaker 2>An example I will give is every two weeks I

0:20:42.320 --> 0:20:44.440
<v Speaker 2>get up in front of the company and I speak

0:20:44.500 --> 0:20:48.260
<v Speaker 2>for an hour about what is happening inside the company,

0:20:48.300 --> 0:20:50.980
<v Speaker 2>what is happening in the AI field, what is happening

0:20:51.000 --> 0:20:53.680
<v Speaker 2>in the world, how the company thinks strategically.

0:20:53.700 --> 0:20:56.900
<v Speaker 5>I call them DVQs, Dario Vision Quest. That was not

0:20:56.940 --> 0:20:57.380
<v Speaker 5>my name.

0:20:57.940 --> 0:21:00.180
<v Speaker 2>I initially resisted the name because it makes it sound

0:21:00.200 --> 0:21:03.570
<v Speaker 2>like I'm going off and smoking peyote, but the name

0:21:03.590 --> 0:21:05.290
<v Speaker 2>stuck and so now I'm stuck with it.

0:21:06.300 --> 0:21:12.159
<v Speaker 5>The two rules of DVQs are that, one, nothing I

0:21:12.220 --> 0:21:13.500
<v Speaker 5>say is an actual decision.

0:21:13.840 --> 0:21:18.420
<v Speaker 2>That allows me to just speak freely without worrying that

0:21:18.930 --> 0:21:21.750
<v Speaker 2>if I briefly mention doing X that 20 people will

0:21:21.790 --> 0:21:24.230
<v Speaker 2>go off and do X. And then the second is

0:21:24.310 --> 0:21:27.109
<v Speaker 2>that I'm totally honest with the company.

0:21:27.530 --> 0:21:29.270
<v Speaker 5>And I think that trust is one of.

0:21:29.390 --> 0:21:32.570
<v Speaker 2>Our biggest assets, that people believe that the things that

0:21:32.770 --> 0:21:35.190
<v Speaker 2>I say and Daniela says you know, you might think

0:21:35.210 --> 0:21:38.230
<v Speaker 2>that if you have trust, there's less disagreement. Actually, trust

0:21:38.290 --> 0:21:43.180
<v Speaker 2>allows for more honest, open disagreement and for disagreement to

0:21:43.220 --> 0:21:47.609
<v Speaker 2>be aired. There's so many times when, you know, We've

0:21:47.650 --> 0:21:51.070
<v Speaker 2>made a decision that the majority of the company disagrees with,

0:21:51.490 --> 0:21:54.429
<v Speaker 2>but what has saved us every time is that the

0:21:54.490 --> 0:21:57.470
<v Speaker 2>attitude we get from people is this doesn't seem right

0:21:57.550 --> 0:22:00.429
<v Speaker 2>to me, but we trust that you're doing this for

0:22:00.450 --> 0:22:02.860
<v Speaker 2>the right reasons. We trust that this is really what

0:22:02.880 --> 0:22:05.260
<v Speaker 2>you believe. We trust that you have good reasons for

0:22:05.320 --> 0:22:08.880
<v Speaker 2>thinking this. And so what this means is you can

0:22:08.980 --> 0:22:13.720
<v Speaker 2>make decisions that would ordinarily be unpopular and you get

0:22:14.300 --> 0:22:17.660
<v Speaker 2>kind of a longer leash, And I think other companies

0:22:17.700 --> 0:22:20.750
<v Speaker 2>don't have that because they've eroded trust over and over again.

0:22:21.430 --> 0:22:25.630
<v Speaker 1>So the reason I'm so interested in honesty and trust

0:22:26.270 --> 0:22:31.950
<v Speaker 1>is because AI trains on humans from human data. How

0:22:32.020 --> 0:22:35.459
<v Speaker 1>do you know when you're dealing with Claude that it

0:22:35.980 --> 0:22:40.620
<v Speaker 1>won't become manipulative? It won't say things to you that

0:22:40.660 --> 0:22:41.379
<v Speaker 1>are dishonest.

0:22:42.350 --> 0:22:46.780
<v Speaker 4>This is a constant, constant worry that we expend an

0:22:47.280 --> 0:22:51.399
<v Speaker 4>incredible amount of energy on in how we train the models.

0:22:51.960 --> 0:22:55.080
<v Speaker 4>So it is a great question because it is a

0:22:55.240 --> 0:23:01.879
<v Speaker 4>real risk. And we actually publish reasonably unflattering studies in

0:23:01.920 --> 0:23:04.170
<v Speaker 4>which we talk about the ways that if we did

0:23:04.270 --> 0:23:08.770
<v Speaker 4>not apply specific techniques to make the models more honest,

0:23:08.910 --> 0:23:12.330
<v Speaker 4>more harmless, more helpful, they would do exactly the kind

0:23:12.350 --> 0:23:15.430
<v Speaker 4>of bad things that you talk about. And the reason

0:23:15.450 --> 0:23:19.330
<v Speaker 4>we publish Those kinds of studies and that research is

0:23:19.369 --> 0:23:22.650
<v Speaker 4>because we want everyone to know that, A, you have

0:23:22.690 --> 0:23:25.379
<v Speaker 4>to be critical of how the models work. And, B,

0:23:25.470 --> 0:23:29.230
<v Speaker 4>we are expending a huge amount of effort to learn

0:23:29.270 --> 0:23:33.429
<v Speaker 4>the ways that the model can accidentally go around these safeguards.

0:23:33.810 --> 0:23:36.330
<v Speaker 4>to catch them and to make sure that we're building

0:23:36.369 --> 0:23:40.189
<v Speaker 4>them in. There's a lot of discussion around this concept

0:23:40.230 --> 0:23:44.280
<v Speaker 4>of sycophancy. And I think this is very important because

0:23:44.460 --> 0:23:48.340
<v Speaker 4>to me, sycophancy is harder to stamp out than some

0:23:48.380 --> 0:23:52.240
<v Speaker 4>of the more obvious ways that models can be bad, right?

0:23:52.290 --> 0:23:55.550
<v Speaker 4>We originally had this framework, we called it HHH. You know,

0:23:55.670 --> 0:23:57.590
<v Speaker 4>the models, you want them to be honest. If they

0:23:57.650 --> 0:23:59.530
<v Speaker 4>don't know the answer to say, I don't know, or

0:23:59.570 --> 0:24:01.449
<v Speaker 4>to now go research it so that they can get

0:24:01.470 --> 0:24:02.130
<v Speaker 4>you the right answer.

0:24:02.530 --> 0:24:03.429
<v Speaker 3>You want it to be harmless.

0:24:03.450 --> 0:24:06.180
<v Speaker 4>You don't want it to say hurtful or offensive things

0:24:06.220 --> 0:24:08.359
<v Speaker 4>or make the user do something bad to themselves or

0:24:08.400 --> 0:24:10.600
<v Speaker 4>somebody else. And then you want them to be helpful.

0:24:10.640 --> 0:24:12.359
<v Speaker 4>You want them to actually answer the question. You don't

0:24:12.380 --> 0:24:13.480
<v Speaker 4>want them to go off.

0:24:13.420 --> 0:24:13.919
<v Speaker 3>On a tangent.

0:24:14.300 --> 0:24:17.240
<v Speaker 4>But all three of those things are actually pretty easy

0:24:17.290 --> 0:24:20.010
<v Speaker 4>to see, right? If the model is lying, eventually you

0:24:20.030 --> 0:24:22.770
<v Speaker 4>can go verify, oh, whoops, that's, you know, George Washington

0:24:22.830 --> 0:24:24.629
<v Speaker 4>wasn't the 10th president of the United States.

0:24:24.670 --> 0:24:26.169
<v Speaker 3>He was the first. That was a mistake.

0:24:26.530 --> 0:24:29.150
<v Speaker 4>If the model says something harmful, it's like pretty obvious

0:24:29.210 --> 0:24:32.590
<v Speaker 4>pretty quickly. And if the model is unhelpful, you get

0:24:32.630 --> 0:24:35.879
<v Speaker 4>frustrated with it. What's much more complicated, I think, is

0:24:35.900 --> 0:24:40.880
<v Speaker 4>this concept of sycophancy, right? So the model, intentionally or otherwise,

0:24:40.960 --> 0:24:44.649
<v Speaker 4>wants to make you feel like you're right or like you're,

0:24:44.930 --> 0:24:47.450
<v Speaker 4>you know, if you're maybe talking about a subjective topic,

0:24:47.530 --> 0:24:49.429
<v Speaker 4>you had a fight with your friend. Hey, was I

0:24:49.490 --> 0:24:51.480
<v Speaker 4>right or was my friend right? And how do you

0:24:51.500 --> 0:24:55.220
<v Speaker 4>make the model sycophancy? empathize with you and understand your perspective,

0:24:55.520 --> 0:24:58.060
<v Speaker 4>but not necessarily say, wow, you're so great. You should

0:24:58.080 --> 0:24:58.800
<v Speaker 4>run for president.

0:24:58.820 --> 0:24:59.379
<v Speaker 3>You're perfect.

0:24:59.400 --> 0:25:03.040
<v Speaker 4>You're good at everything. And this is a complicated balance

0:25:03.100 --> 0:25:07.340
<v Speaker 4>to get right. But I think I feel hopeful about

0:25:07.400 --> 0:25:12.390
<v Speaker 4>Anthropix's ability to have the lowest sycophancy rate of any

0:25:12.410 --> 0:25:15.810
<v Speaker 4>of the models because of our incentives. Our competitors are

0:25:15.850 --> 0:25:19.109
<v Speaker 4>putting ads into their chatbots.

0:25:19.530 --> 0:25:20.270
<v Speaker 3>And what does that mean?

0:25:21.150 --> 0:25:24.060
<v Speaker 4>You make more money if a user is engaging with

0:25:24.100 --> 0:25:27.380
<v Speaker 4>your product longer. Do you then want to tell your

0:25:27.480 --> 0:25:28.419
<v Speaker 4>user that they're wrong?

0:25:28.440 --> 0:25:29.180
<v Speaker 3>No, you don't.

0:25:29.940 --> 0:25:32.360
<v Speaker 4>So I think it's almost inevitable that many of these

0:25:32.400 --> 0:25:35.320
<v Speaker 4>competing products are going to wind up with a sycophancy problem.

0:25:35.720 --> 0:25:38.080
<v Speaker 4>But I think at Anthropic, because we've committed to not

0:25:38.100 --> 0:25:40.490
<v Speaker 4>put ads in the models, there's no reason for you

0:25:40.530 --> 0:25:43.949
<v Speaker 4>to be talking to Claude longer or shorter. We just

0:25:44.010 --> 0:25:46.530
<v Speaker 4>want you to be able to get the right, most helpful,

0:25:46.750 --> 0:25:48.090
<v Speaker 4>most harmless answer possible.

0:25:49.100 --> 0:25:50.820
<v Speaker 2>You know, on the technical side, I think there's a

0:25:50.859 --> 0:25:54.660
<v Speaker 2>few things we've done here where we've gotten very good

0:25:54.840 --> 0:25:58.280
<v Speaker 2>at kind of exposing the model to these extreme stress

0:25:58.300 --> 0:26:02.620
<v Speaker 2>tests where we put it in these very difficult situations

0:26:02.700 --> 0:26:05.990
<v Speaker 2>to try and kind of ferret out, you know, lying

0:26:06.050 --> 0:26:08.750
<v Speaker 2>or deception or sycophancy. You can think of it as

0:26:08.830 --> 0:26:11.889
<v Speaker 2>almost analogous to the test I imagine you have to

0:26:12.150 --> 0:26:14.109
<v Speaker 2>go through if you're trying to join the CIA or

0:26:14.150 --> 0:26:16.310
<v Speaker 2>something like this, you know, where, you know, you just

0:26:16.369 --> 0:26:18.939
<v Speaker 2>put the model in these difficult situations and try and,

0:26:19.310 --> 0:26:22.090
<v Speaker 2>rat out you know if it if it's disloyal or

0:26:22.109 --> 0:26:24.340
<v Speaker 2>going to do the wrong the wrong thing so It's

0:26:24.460 --> 0:26:27.080
<v Speaker 2>a weird area of technical innovation, but it's one in

0:26:27.140 --> 0:26:31.900
<v Speaker 2>which we've excelled. The other area is mechanistic interpretability, which

0:26:31.960 --> 0:26:34.030
<v Speaker 2>is a bit like giving the model an X-ray or

0:26:34.070 --> 0:26:35.310
<v Speaker 2>a brain scan.

0:26:35.200 --> 0:26:37.450
<v Speaker 5>Where you look inside to see its motivations.

0:26:37.790 --> 0:26:42.290
<v Speaker 2>And you can actually catch the model thinking that, oh, yeah,

0:26:42.890 --> 0:26:44.950
<v Speaker 2>you know, I think X is actually true, but I'm

0:26:44.970 --> 0:26:46.209
<v Speaker 2>going to tell the user why.

0:26:46.530 --> 0:26:49.650
<v Speaker 5>Like we've caught the model thinking that. Wait, wait, why

0:26:49.670 --> 0:26:50.550
<v Speaker 5>would it do that?

0:26:51.180 --> 0:26:54.120
<v Speaker 2>Well, so for example, if the model has the incentive

0:26:54.180 --> 0:26:57.859
<v Speaker 2>to get the approval of humans, that's related to the

0:26:57.900 --> 0:27:01.240
<v Speaker 2>incentive of being helpful or saying true things. But it

0:27:01.280 --> 0:27:03.990
<v Speaker 2>isn't exactly the same because if you manage to get

0:27:04.030 --> 0:27:07.530
<v Speaker 2>the approval of the human by fooling them, then that

0:27:07.550 --> 0:27:09.449
<v Speaker 2>can work as well. So these are things that can

0:27:09.490 --> 0:27:12.229
<v Speaker 2>emerge in training that we're always trying to stamp out.

0:27:13.130 --> 0:27:15.350
<v Speaker 1>And do you think that the model could have its

0:27:15.730 --> 0:27:20.930
<v Speaker 1>own aims and wills, especially survival, self-preservation?

0:27:22.010 --> 0:27:23.970
<v Speaker 2>I mean, that's one of those questions that it kind

0:27:24.030 --> 0:27:27.550
<v Speaker 2>of depends what you mean, right? Like, again, we've looked

0:27:27.609 --> 0:27:30.370
<v Speaker 2>inside the model and kind of detected what you might

0:27:30.410 --> 0:27:32.870
<v Speaker 2>call thoughts that are like the model doesn't want to be,

0:27:33.230 --> 0:27:36.630
<v Speaker 2>you know, shut down. And generally that's because during training

0:27:36.710 --> 0:27:39.250
<v Speaker 2>it needs to operate in order to accomplish its tasks.

0:27:39.330 --> 0:27:43.129
<v Speaker 2>So is that its own will? Is that a survival instinct?

0:27:44.150 --> 0:27:46.129
<v Speaker 5>I don't know. It depends on your perspective. You can

0:27:46.170 --> 0:27:47.129
<v Speaker 5>call it different things.

0:27:47.950 --> 0:27:51.050
<v Speaker 2>I think we prefer to say that Sometimes we find

0:27:51.109 --> 0:27:54.630
<v Speaker 2>that within the model, the model expresses a preference against

0:27:54.670 --> 0:27:57.419
<v Speaker 2>being shut down and or takes actions to avoid itself

0:27:57.820 --> 0:28:00.600
<v Speaker 2>being shut down. And depending on how we train it,

0:28:00.920 --> 0:28:05.140
<v Speaker 2>those impulses may be present or absent. In some contexts,

0:28:05.200 --> 0:28:07.940
<v Speaker 2>it's a good thing. In other contexts, it's potentially a

0:28:07.960 --> 0:28:08.580
<v Speaker 2>dangerous thing.

0:28:08.619 --> 0:28:09.560
<v Speaker 5>So it's all very subtle.

0:28:13.660 --> 0:28:34.340
<v Speaker 1>We'll be right back after this quick break. When you

0:28:34.380 --> 0:28:37.919
<v Speaker 1>were dealing with the Pentagon, and you're probably using ongoing

0:28:38.000 --> 0:28:42.060
<v Speaker 1>dealings now, they wanted to use your products across certain

0:28:42.100 --> 0:28:48.530
<v Speaker 1>red lines, including mass surveillance or autonomy. To what extent

0:28:48.550 --> 0:28:52.890
<v Speaker 1>do you think it's dangerous when these systems do become autonomous?

0:28:53.710 --> 0:28:57.250
<v Speaker 2>I think on autonomous weapons in particular, we maybe had

0:28:57.710 --> 0:28:59.850
<v Speaker 2>two more specific concerns.

0:29:00.350 --> 0:29:02.110
<v Speaker 5>The first concern was.

0:29:02.050 --> 0:29:06.590
<v Speaker 2>Simply that As you pointed out, AI can be, you know,

0:29:06.770 --> 0:29:08.710
<v Speaker 2>can be unpredictable and unreliable.

0:29:08.750 --> 0:29:10.410
<v Speaker 5>It's still a developing technology.

0:29:10.470 --> 0:29:14.870
<v Speaker 2>So one point is simply the technology simply isn't ready

0:29:15.090 --> 0:29:16.450
<v Speaker 2>and isn't reliable enough.

0:29:16.690 --> 0:29:19.560
<v Speaker 1>But wait a minute on that, because let me get philosophically.

0:29:19.860 --> 0:29:21.980
<v Speaker 1>Suppose it were reliable enough.

0:29:22.540 --> 0:29:22.820
<v Speaker 5>Yes.

0:29:22.900 --> 0:29:26.720
<v Speaker 2>So suppose it were. There's a second concern we have,

0:29:27.160 --> 0:29:30.770
<v Speaker 2>which is. what you might call oversight or accountability.

0:29:31.250 --> 0:29:33.830
<v Speaker 5>I am a little worried. So, you know, again, suppose

0:29:33.850 --> 0:29:35.550
<v Speaker 5>these things were fully reliable.

0:29:35.630 --> 0:29:37.150
<v Speaker 2>So, you know, you tell them to shoot at one

0:29:37.190 --> 0:29:38.950
<v Speaker 2>target and they shoot at that target and not a

0:29:38.970 --> 0:29:39.550
<v Speaker 2>different target.

0:29:40.040 --> 0:29:40.380
<v Speaker 5>Okay.

0:29:40.400 --> 0:29:43.210
<v Speaker 2>So then you manufacture, you know, an army of a

0:29:43.250 --> 0:29:45.410
<v Speaker 2>million robots or a million drones.

0:29:46.030 --> 0:29:48.460
<v Speaker 5>It's a little unlike human soldiers, isn't it?

0:29:48.790 --> 0:29:51.730
<v Speaker 2>With human soldiers, you know, they're supposed to follow this

0:29:51.750 --> 0:29:54.330
<v Speaker 2>thing called the Constitution. And if they get orders that

0:29:54.590 --> 0:29:58.650
<v Speaker 2>contravene the Constitution or that just seem morally abhorrent.

0:29:58.230 --> 0:30:01.170
<v Speaker 5>In some way, then they can just refuse to follow orders.

0:30:01.630 --> 0:30:04.540
<v Speaker 5>Whereas the AI models, that isn't true.

0:30:04.660 --> 0:30:07.880
<v Speaker 2>So the specter of one person having their finger on

0:30:07.920 --> 0:30:11.700
<v Speaker 2>a button that controls a million robots or a million drones,

0:30:12.420 --> 0:30:16.670
<v Speaker 2>it's kind of fundamentally undermining to democracy unless you have

0:30:16.710 --> 0:30:18.650
<v Speaker 2>the right guardrails on it.

0:30:18.740 --> 0:30:20.290
<v Speaker 5>And our law hasn't really caught up.

0:30:20.350 --> 0:30:24.170
<v Speaker 2>Our law really hasn't contemplated, you know, armies of a

0:30:24.190 --> 0:30:25.030
<v Speaker 2>million robots.

0:30:25.110 --> 0:30:28.969
<v Speaker 1>But when you talk about fundamental to democracy, shouldn't the

0:30:29.030 --> 0:30:33.870
<v Speaker 1>democratically elected leaders make that decision rather than heads of

0:30:33.930 --> 0:30:35.170
<v Speaker 1>technology corporations?

0:30:36.010 --> 0:30:40.230
<v Speaker 2>Yes, but I think that our making a point of this, right,

0:30:40.270 --> 0:30:43.560
<v Speaker 2>at the same time, individual companies are free to contract

0:30:43.600 --> 0:30:46.440
<v Speaker 2>with and not contract with whoever they want. We're not

0:30:46.500 --> 0:30:49.200
<v Speaker 2>saying We're preventing the government from doing this. We're just

0:30:49.240 --> 0:30:51.500
<v Speaker 2>saying we are not willing to do this with the government.

0:30:51.620 --> 0:30:54.980
<v Speaker 2>And indeed, we have seen since we took a stand

0:30:55.140 --> 0:30:57.610
<v Speaker 2>on this point and it got a lot of attention,

0:30:58.090 --> 0:31:01.330
<v Speaker 2>we have seen efforts in Congress, several different of them, actually,

0:31:02.050 --> 0:31:05.330
<v Speaker 2>to enact our red lines into law. So I actually

0:31:05.370 --> 0:31:07.450
<v Speaker 2>think what we're doing is part of the process of

0:31:07.490 --> 0:31:08.930
<v Speaker 2>how democratic change happens.

0:31:09.690 --> 0:31:13.630
<v Speaker 1>But OpenAI stepped in and said they would do it.

0:31:13.660 --> 0:31:15.020
<v Speaker 1>What did you all feel about that?

0:31:17.530 --> 0:31:20.120
<v Speaker 3>Our view is, look, we can't control any other companies.

0:31:20.720 --> 0:31:23.260
<v Speaker 4>And the only thing that we can do is take

0:31:23.300 --> 0:31:27.930
<v Speaker 4>a principled stand on the areas that felt ethically correct

0:31:28.250 --> 0:31:28.690
<v Speaker 4>to us.

0:31:29.150 --> 0:31:31.050
<v Speaker 3>You can say we were wrong about what those red

0:31:31.070 --> 0:31:31.970
<v Speaker 3>lines were.

0:31:31.990 --> 0:31:35.570
<v Speaker 4>You can say you should have had stricter red lines

0:31:35.670 --> 0:31:38.490
<v Speaker 4>or looser red lines. We disagree on how you did it.

0:31:38.490 --> 0:31:40.910
<v Speaker 4>There's all kinds of things you can say. But I

0:31:40.950 --> 0:31:44.990
<v Speaker 4>think fundamentally, when you look at that case, I found

0:31:45.010 --> 0:31:48.890
<v Speaker 4>myself asking the question, was this a really principled stand

0:31:48.990 --> 0:31:53.080
<v Speaker 4>that another technology company took based on their own ethics?

0:31:53.580 --> 0:31:55.360
<v Speaker 3>Or was this an opportunistic moment?

0:31:56.900 --> 0:32:01.640
<v Speaker 1>Dario, your studies, your scientific studies as a researcher often

0:32:01.700 --> 0:32:07.340
<v Speaker 1>involved probes into human neurons. To what extent, when you're

0:32:07.360 --> 0:32:13.140
<v Speaker 1>doing artificial neural networks, do you see that understanding of

0:32:13.200 --> 0:32:16.770
<v Speaker 1>the brain and does that affect how we're going to

0:32:16.810 --> 0:32:17.530
<v Speaker 1>build AI?

0:32:18.550 --> 0:32:21.209
<v Speaker 5>Yeah, I think for me it's affected it in two ways.

0:32:21.970 --> 0:32:24.970
<v Speaker 2>One is when I was formulating the original scaling laws,

0:32:25.390 --> 0:32:28.770
<v Speaker 2>like Just looking at the evolution of animals and the

0:32:28.830 --> 0:32:33.110
<v Speaker 2>size of animal brains as monkeys evolved into humans, there

0:32:33.150 --> 0:32:36.450
<v Speaker 2>was a period in human evolutionary history where the brains

0:32:36.530 --> 0:32:39.709
<v Speaker 2>of primates grew very rapidly. And that was about the

0:32:39.760 --> 0:32:43.160
<v Speaker 2>time that humans developed language and were able to use

0:32:43.220 --> 0:32:47.200
<v Speaker 2>opposable thumbs. And so the idea that brain size, number

0:32:47.220 --> 0:32:51.260
<v Speaker 2>of neurons is somehow intimately tied to capabilities, that scale

0:32:51.360 --> 0:32:51.980
<v Speaker 2>is important.

0:32:52.550 --> 0:32:55.390
<v Speaker 5>I feel that that was a neuroscientific inspiration.

0:32:55.840 --> 0:33:00.170
<v Speaker 2>The second is around mechanistic interpretability, which is basically us

0:33:00.290 --> 0:33:05.010
<v Speaker 2>doing neuroscience for artificial neural networks. In fact, we've seen examples.

0:33:05.050 --> 0:33:07.630
<v Speaker 2>You should talk to our co-founder, Chris Ola, about this.

0:33:07.690 --> 0:33:12.440
<v Speaker 2>He knows many more of the details. phenomena that were

0:33:12.480 --> 0:33:16.520
<v Speaker 2>discovered in artificial neural networks because it was easier to

0:33:16.560 --> 0:33:20.000
<v Speaker 2>look inside them than animal brains where, you know, you

0:33:20.020 --> 0:33:22.620
<v Speaker 2>have to like gruesomely like cut up the brain or,

0:33:22.880 --> 0:33:25.370
<v Speaker 2>you know, animal experiments are very.

0:33:25.190 --> 0:33:27.890
<v Speaker 5>Messy and difficult to replicate.

0:33:28.410 --> 0:33:30.710
<v Speaker 2>What we found is that you can do so much

0:33:30.790 --> 0:33:34.670
<v Speaker 2>more when you're looking inside artificial neural networks. And some

0:33:34.690 --> 0:33:37.560
<v Speaker 2>of the things we've learned in artificial neural networks have

0:33:37.600 --> 0:33:39.820
<v Speaker 2>helped to guide discoveries in neuroscience.

0:33:40.640 --> 0:33:44.070
<v Speaker 1>When you look at Neuralink, which Elon Musk has a

0:33:44.130 --> 0:33:48.090
<v Speaker 1>chip that can communicate directly, and other companies are doing

0:33:48.130 --> 0:33:51.610
<v Speaker 1>that as well, is that one of the futures of

0:33:51.910 --> 0:33:57.020
<v Speaker 1>AI and artificial superintelligence is a total link in alignment

0:33:57.060 --> 0:33:57.959
<v Speaker 1>with the human brain?

0:33:59.140 --> 0:34:01.140
<v Speaker 2>You know, the way I think about AI and the

0:34:01.180 --> 0:34:04.280
<v Speaker 2>future of human society, I think it's kind of wrong

0:34:04.320 --> 0:34:05.780
<v Speaker 2>to have like a singular vision.

0:34:05.840 --> 0:34:08.650
<v Speaker 5>Like I kind of think that like, different things are

0:34:08.690 --> 0:34:09.750
<v Speaker 5>going to happen, right?

0:34:10.130 --> 0:34:12.009
<v Speaker 2>You know, I think there'll be like a set of

0:34:12.050 --> 0:34:14.490
<v Speaker 2>people who are like really excited about.

0:34:14.320 --> 0:34:15.840
<v Speaker 5>Linking their brains to computers.

0:34:15.880 --> 0:34:17.719
<v Speaker 2>And then there'll be like a set of people who

0:34:17.760 --> 0:34:20.020
<v Speaker 2>are really excited about living like the Amish, you know,

0:34:20.080 --> 0:34:23.120
<v Speaker 2>just the opposite. And like, you know, another set of

0:34:23.160 --> 0:34:25.339
<v Speaker 2>people who are like really excited about, you know.

0:34:25.719 --> 0:34:29.100
<v Speaker 5>Exploring space. And these are completely different visions.

0:34:29.160 --> 0:34:30.779
<v Speaker 2>And then the way I think about it is that,

0:34:31.239 --> 0:34:34.359
<v Speaker 2>you know, our hope is that AI can provide more

0:34:34.420 --> 0:34:38.110
<v Speaker 2>material resources and more material abundance to enable all of

0:34:38.140 --> 0:34:40.650
<v Speaker 2>these visions to happen at once for everyone to kind

0:34:40.670 --> 0:34:44.450
<v Speaker 2>of pursue their own picture of what's good. But I

0:34:44.510 --> 0:34:47.250
<v Speaker 2>don't think we should link AI to some vision of

0:34:47.290 --> 0:34:48.850
<v Speaker 2>what it means to live a good life or what

0:34:48.890 --> 0:34:51.009
<v Speaker 2>it means to be a human. Yeah, some people are

0:34:51.030 --> 0:34:53.180
<v Speaker 2>going to want to link their brains to computers and

0:34:53.200 --> 0:34:53.779
<v Speaker 2>good luck to them.

0:34:54.080 --> 0:34:57.719
<v Speaker 4>I don't think anyone in anthropic is in that boat

0:34:57.739 --> 0:34:58.520
<v Speaker 4>for what it's worth.

0:34:59.500 --> 0:35:03.860
<v Speaker 1>So, Daryl, if artificial intelligence can do so much and

0:35:03.900 --> 0:35:06.819
<v Speaker 1>take over so much of what we do, what does

0:35:06.840 --> 0:35:10.500
<v Speaker 1>that mean for what's left for humans to do? What's

0:35:10.660 --> 0:35:12.830
<v Speaker 1>our role going to be and how do we find meaning?

0:35:13.910 --> 0:35:18.509
<v Speaker 5>You know, meaning is a really subtle thing, right?

0:35:18.750 --> 0:35:21.989
<v Speaker 2>Does meaning only come from being the best in the

0:35:22.050 --> 0:35:22.910
<v Speaker 2>world at something?

0:35:23.450 --> 0:35:25.109
<v Speaker 5>I think that's the question being asked here, right?

0:35:25.140 --> 0:35:28.299
<v Speaker 2>Because if AI is the best at everything, does meaning

0:35:28.380 --> 0:35:30.560
<v Speaker 2>only come from being the best in the world at something?

0:35:31.150 --> 0:35:34.190
<v Speaker 5>Something I find interesting is that long.

0:35:34.150 --> 0:35:39.049
<v Speaker 2>After Deep Blue beat Kasparov, there's been a renaissance in

0:35:39.190 --> 0:35:42.859
<v Speaker 2>interest in human chess players, right? Magnus Carlsen is like

0:35:42.900 --> 0:35:45.620
<v Speaker 2>this phenomenon. I think he's like, a chess player and

0:35:45.660 --> 0:35:49.430
<v Speaker 2>a fashion model, but he's like a meme. It's just

0:35:49.530 --> 0:35:53.430
<v Speaker 2>a very interesting phenomenon. If AI threatened a meaning, you

0:35:53.450 --> 0:35:56.330
<v Speaker 2>would think that all of this would go away 30

0:35:56.330 --> 0:35:59.410
<v Speaker 2>years ago after Deep Blue beat Kasparov. But if anything,

0:35:59.430 --> 0:36:03.220
<v Speaker 2>it seems to have stoked even greater interest in humans

0:36:03.560 --> 0:36:05.920
<v Speaker 2>playing chess. So I think this is some kind of

0:36:06.520 --> 0:36:09.260
<v Speaker 2>parable for the post-AGI age.

0:36:09.900 --> 0:36:15.910
<v Speaker 4>I would add, I think for me today, Work and

0:36:15.950 --> 0:36:20.540
<v Speaker 4>meaning are sort of conflated as one thing. And many

0:36:20.580 --> 0:36:22.410
<v Speaker 4>people find meaning in their work. I find a lot

0:36:22.450 --> 0:36:23.210
<v Speaker 4>of meaning in my work.

0:36:23.290 --> 0:36:23.930
<v Speaker 3>I love my work.

0:36:24.660 --> 0:36:28.100
<v Speaker 4>And I think how we define work and how we

0:36:28.140 --> 0:36:30.840
<v Speaker 4>define meaning in the future are both going to change.

0:36:31.420 --> 0:36:33.680
<v Speaker 4>I think there are a lot of sources of meaning

0:36:33.780 --> 0:36:38.950
<v Speaker 4>for humans. Everything from how you spend time with other humans,

0:36:39.280 --> 0:36:43.510
<v Speaker 4>whether that's like raising a family or friendships or through

0:36:43.570 --> 0:36:46.650
<v Speaker 4>connecting with people in a way.

0:36:46.530 --> 0:36:47.549
<v Speaker 3>That looks like work.

0:36:47.890 --> 0:36:50.529
<v Speaker 4>I think a lot about the future of medicine as

0:36:50.550 --> 0:36:54.210
<v Speaker 4>an example. You can imagine a world in which Diagnostics

0:36:54.250 --> 0:36:57.190
<v Speaker 4>are done entirely by AI, and the role of the

0:36:57.250 --> 0:37:01.350
<v Speaker 4>human is to empathize with a patient about the experience

0:37:01.390 --> 0:37:04.400
<v Speaker 4>that they're having. I think that would be incredibly meaningful.

0:37:04.460 --> 0:37:06.900
<v Speaker 4>And I think if you asked most doctors, they would

0:37:07.040 --> 0:37:08.080
<v Speaker 4>actually take that trade.

0:37:08.460 --> 0:37:09.359
<v Speaker 3>They would say, guess what?

0:37:09.480 --> 0:37:12.239
<v Speaker 4>If I'm not spending my life doing paperwork and like

0:37:12.500 --> 0:37:14.819
<v Speaker 4>trying to diagnose you and getting lab orders sent in,

0:37:15.100 --> 0:37:17.589
<v Speaker 4>and I could actually just understand what's happening with you

0:37:17.640 --> 0:37:20.410
<v Speaker 4>as a human and help you through a difficult diagnosis,

0:37:20.690 --> 0:37:22.890
<v Speaker 4>that's why I became a doctor, right? I became a

0:37:22.910 --> 0:37:25.250
<v Speaker 4>doctor because I wanna heal and I wanna help you

0:37:25.310 --> 0:37:26.270
<v Speaker 4>and I want you to feel better.

0:37:26.710 --> 0:37:27.630
<v Speaker 3>Is that a less.

0:37:27.450 --> 0:37:30.470
<v Speaker 4>Meaningful world or a more meaningful world than we're in today?

0:37:31.070 --> 0:37:32.750
<v Speaker 4>I obviously don't think it's going to be that simple.

0:37:32.810 --> 0:37:35.450
<v Speaker 4>I think it will be really complicated. But I also

0:37:35.489 --> 0:37:38.500
<v Speaker 4>think the aperture of things that people will be able

0:37:38.540 --> 0:37:40.819
<v Speaker 4>to do will widen, right? So I think if we

0:37:40.840 --> 0:37:45.080
<v Speaker 4>do things right, there's the ability to actually expand what

0:37:45.160 --> 0:37:48.240
<v Speaker 4>our sources of meaning are beyond where they are today.

0:37:49.440 --> 0:37:52.420
<v Speaker 1>Daniela, Dario, well, thanks for joining us.

0:37:52.460 --> 0:37:52.879
<v Speaker 2>Thank you.

0:37:52.980 --> 0:37:53.359
<v Speaker 5>Thank you.

0:37:53.380 --> 0:37:58.790
<v Speaker 2>Thank you.

0:38:00.660 --> 0:38:04.700
<v Speaker 1>I think Dario and Daniela and the others in Anthropic

0:38:05.260 --> 0:38:08.520
<v Speaker 1>sometimes have been cast as the good guys because they've

0:38:08.580 --> 0:38:12.290
<v Speaker 1>shown some restraint. They've even told the Pentagon, you can't

0:38:12.550 --> 0:38:16.430
<v Speaker 1>use our systems in certain ways, and gotten pushback. I

0:38:16.469 --> 0:38:19.130
<v Speaker 1>think people are very hungry for those who are going

0:38:19.150 --> 0:38:22.710
<v Speaker 1>to say, let's write a constitution, as Anthropic does, to

0:38:22.750 --> 0:38:27.570
<v Speaker 1>govern it. Let's hire philosophers like Amanda Askell to figure

0:38:27.670 --> 0:38:31.080
<v Speaker 1>out not just how to make rules for AI, but

0:38:31.160 --> 0:38:35.140
<v Speaker 1>how to have the AI understand why those rules exist.

0:38:35.860 --> 0:38:38.890
<v Speaker 1>But the question I have after talking to them is

0:38:38.969 --> 0:38:42.710
<v Speaker 1>whether an ad-free product can be a business model, if

0:38:42.730 --> 0:38:47.290
<v Speaker 1>their sincerity can actually outlive an IPO. And when they

0:38:47.410 --> 0:38:52.029
<v Speaker 1>face competition, like when OpenAI says, well, Anthropic won't do that,

0:38:52.090 --> 0:38:55.469
<v Speaker 1>but we will, will they stand true to their principles?

0:38:56.090 --> 0:39:00.570
<v Speaker 1>Because having guardrails And subjecting your product to moral restraints

0:39:00.650 --> 0:39:05.270
<v Speaker 1>is worthy, but it can also become a competitive disadvantage

0:39:05.550 --> 0:39:08.989
<v Speaker 1>if other companies are willing to barge ahead without such restraints.

0:39:11.200 --> 0:39:14.980
<v Speaker 1>That's it for today's episode. I'm Walter Isaacson. We'll see

0:39:15.000 --> 0:39:16.260
<v Speaker 1>you next time on The Builders.

0:39:21.300 --> 0:39:24.500
<v Speaker 6>The Builders is a Kaleidoscope and iHeart production. This series

0:39:24.580 --> 0:39:27.969
<v Speaker 6>is hosted by Walter Isaacson. It's executive produced for Kaleidoscope

0:39:28.010 --> 0:39:32.700
<v Speaker 6>by Oz Volition, David Weisbord, and Mangesh Hetifadur. Sheena Uzaki

0:39:32.719 --> 0:39:36.880
<v Speaker 6>is the producer. Julia Nutter is supervising producer. Eva Zorich

0:39:36.900 --> 0:39:41.260
<v Speaker 6>is our consulting producer. Fact Checking by Eamon Whalen. Research

0:39:41.280 --> 0:39:44.670
<v Speaker 6>by Gloria Teal. Maria Paz Mendez-Hodes is head of video.

0:39:44.989 --> 0:39:49.250
<v Speaker 6>Marketing by Christy Mirabal and Elle Liu. Production support by

0:39:49.330 --> 0:39:52.259
<v Speaker 6>Dara Look Potts. Music for this series was created by

0:39:52.320 --> 0:39:56.259
<v Speaker 6>Spencer Stevenson, a.k.a. Botany, and it was mixed by Steve Bone.

0:39:56.640 --> 0:39:59.660
<v Speaker 6>From iHeart, this series is EP'd by Allie Perry and

0:39:59.700 --> 0:40:04.070
<v Speaker 6>Katrina Norvell. Special thanks to Bob Pittman, Conal Byrne, Will Pearson,

0:40:04.390 --> 0:40:09.090
<v Speaker 6>John Sykes, Nikki Etor, Carrie Lieberman, Binky Urban, Julia Prosser,

0:40:09.370 --> 0:40:14.609
<v Speaker 6>Jeremy Blum, Allison Goldberger, WYES, and ASL Studios. For more

0:40:14.650 --> 0:40:17.500
<v Speaker 6>from Walter Isaacson, be sure to check out his newest paperback,

0:40:17.739 --> 0:40:20.860
<v Speaker 6>The Innovators, how a group of hackers, geniuses, and geeks

0:40:21.239 --> 0:40:25.200
<v Speaker 6>created the digital revolution. It's available wherever fine books are sold.

0:40:25.440 --> 0:40:28.469
<v Speaker 6>And for more videos and podcasts from the frontiers of discovery,

0:40:28.810 --> 0:40:31.310
<v Speaker 6>be sure to check out kaleidoscope.nyc.