Speaker A
Mario, thanks for doing it. So, a little backstory. So, uh, we had spent some time talking, uh, to Dario in particular about the jobs and what it could do. And we'd had a lot of off-the-record, uh, conversations just about
Dario Amodei discusses AI's impact on white-collar jobs, rapid tech advances, and societal responses including retraining and regulation.
Ask about this video. Answers come from its transcript only — with the timestamp, so you can check them.
Generated from the transcript and can be wrong — check the timestamp.
Key Takeaways
- AI could drastically reduce white-collar jobs in the near future, necessitating urgent societal response.
- Transparency and public discourse about AI’s risks and benefits are crucial for informed policy decisions.
- Retraining and adaptation programs are essential but insufficient alone to address job displacement.
- Government intervention, including potential taxation of AI firms, may be needed to manage economic disruption.
- AI technology is evolving rapidly, fundamentally altering work roles and productivity within companies.
What the video covers
- Dario Amodei from Anthropic warns about AI potentially eliminating up to half of white-collar jobs within 1 to 5 years, causing unemployment spikes.
- Anthropic publicly disclosed these risks to ensure society understands both benefits and threats of AI technology.
- Current data shows a 13% contraction in white-collar entry-level jobs, evidencing early labor market impacts.
- AI is rapidly changing job roles internally at Anthropic, with engineers managing AI systems rather than coding directly.
- The technology is advancing faster than expected, increasing concerns about widespread economic disruption.
- Amodei advocates for government action to help workers adapt through retraining and possibly taxing AI companies to support displaced workers.
- Despite skepticism about legislative success, he stresses the importance of honest communication about AI's societal effects.
- Anthropic’s revenue is growing exponentially, highlighting the massive wealth creation potential of AI companies.
- The discussion includes the need for transparency, regulation, and technical safeguards to mitigate AI risks.
- Future AI applications may include embodied agents and humanoid robotics, expanding AI’s real-world impact.
Chapters
- 00:00Introduction and Backstory on AI Job Impact
- 01:26Communicating AI's Capabilities and Threats
- 02:43Current Evidence of AI's Impact on Jobs
- 04:00Changing Roles of Engineers at Anthropic
- 05:27Adapting Workforce to AI Technology
- 06:48AI Company Growth and Wealth Creation
- 08:15AI Risks and Transparency Advocacy
- 11:03Future of AI and Robotics
- 20:02Closing Remarks and Summary
Full Transcript — Download SRT & Markdown
Speaker A
what he was sort of describing as this white-collar bloodbath that I've been hearing from other folks in the AI community, from other companies, but no one would say it publicly. So, we called him and said, "Listen, it's really
Speaker A
important that society knows this, hears this from you." Uh, and to his credit, he said, "Let's do it. I'll go on the record. I'll say exactly, uh, what I said to you and what I think." In that
Speaker A
conversation, uh, it was chilling, right? You said, "Listen, you could see half of white-collar jobs wiped away in 1 to 5 years. Unemployment spiked to 10 to 20%, uh, because of this technology regardless of the good it might
Speaker A
do beyond that. Talk more about that and why did you go public and why won't others?" Yeah, so, so a few points. Um, you know, I think, I think the reason that, uh, you know, I, we at Anthropic decided to go public
Speaker A
about this is, you know, I had been saying these things before but it had mostly been in, in, you know, kind of limited context in, in podcasts restricted to the technology industry and, you know, when I would get on a plane, when I would
Speaker A
be in the airport, when I would be in some other city than, you know, San Francisco or California, I'd walk by a bunch of people and say, "I, I don't think we've communicated to you very, very accurately about what this
Speaker A
technology is capable of and where it's going, the ways it could benefit you and, and in the ways it could be a threat to you if you don't react to it in the right way." And, and at some point that
Speaker A
just felt wrong where I was talking to a lot of CEOs who would say in private, you know, this is what we're planning.
Speaker A
Some of them are our customers. You know, we, we have these plans to, um, uh, you know, deploy this technology and it's going to have a labor impact. And so, and so we really felt that we needed to, we needed to say something, you know,
Speaker A
the first step towards solving these problems is, is kind of, you know, being honest with, with the population that these problems exist. To say more about, about, you know, where I see things and why I see things this way. Um, you know, I
Speaker A
would say there's kind of two modes that we have at Anthropic. One is looking at what is happening now. Um, and we've done a lot of that with the economic index. We've done a lot of that. Um, you
Speaker A
know, we recently released a state-by-state economic index which lets everyone look at, uh, what, what you can do, you know, how people are using the models in real time in different states, in different geographic locations for different tasks, whether they're
Speaker A
automating or augmenting. So, that's one mode and we've even seen some research externally showing that there's a job impact there already. For example, the work from, um, uh, Eric, Eric Bernolson and others showing that white-collar AI, you
Speaker A
know, white-collar entry-level jobs have contracted already by 13%, which is a significant, um, you know, you know, fraction in some area, significant fraction of, of what I predicted. But what I'm really worried about is where the technology is
Speaker A
going. And I think there's a little disconnect here where people will sometimes say, "Oh, you're worried about what AI is going to do to jobs, but, you know, AI can't do this, AI can't do that." Well, we're talking about today's
Speaker A
AI. The technology is moving quickly. I'm worried about advances in the technology and diffusion of that technology through society. And that's where I get to the one to five years. As with most things, when an exponential is moving very
Speaker A
quickly, you can't be sure. This could happen faster than I imagine. This could happen slower than I imagine. Um, you know, or something very different could happen. But I think it is likely enough to happen that we felt that, that there
Speaker A
was a need to warn the world about it and to speak honestly and in candid terms about it. That was 2 months ago.
Speaker A
We were talking off stage in some ways since then. Both of you were remarking how the technology, despite some of the conventional wisdom, is actually getting much better than we even thought 2 months ago. Are you more worried today
Speaker A
than you were 2 months ago? So we studied this inside Anthropic. We talked to 130 engineers. We interviewed many of them and talked about their experience using the technology and their jobs have changed in the last year radically. Many of them now do two or
Speaker A
three times as much work. And rather than writing code, they're managing fleets of AI systems. And in these interviews, they said, "My job has completely changed. I'm now having to rethink what my role here at Anthropic looks like." Now, obviously, we're a
Speaker A
fast-growing company. They will have jobs. But we are changing people's jobs in real time inside the company because the technology has moved so, so, so quickly. And what happens inside the AI companies will happen to all of the
Speaker A
other businesses that use this AI technology in the coming years. The vast majority of code that is used to support Claude and to design the next Claude is now written by Claude. It's, it's the, the just the vast majority of it within
Speaker A
Anthropic and other fast-moving companies. The same is true. I don't know that it's fully diffused out into the world yet, but, but this is already happening. And just to put a finer point on that, I think that's what I think
Speaker A
what you're seeing is what we're seeing that has a fiduciary responsibility to drive value that they can use technology for productivity. They're going to do it and I think you're already seeing it in the unemployment numbers when, when you
Speaker A
on the hill where there, there's some momentum we're talking about to, to potentially legislate it. You're more bullish than I am. I'm deeply skeptical Trump would sign anything into law that would regulate AI in his, in this term.
Speaker A
But put aside what can be done. What are the two things in order that if you were running Congress or you're king of the US that you would do today to specifically address this? I would, I would say the first thing would be
Speaker A
something around helping people adapt to, to, to AI technology. Um, you know, helping, you know, I, I don't want to think of this as a bromide. You know, people have tried retraining programs and there, there are real limits to what they can, to what
Speaker A
they can do. There are real limits to kind of helping people to train and adapt but it's, it's better than nothing and it's where we got to start and, you know, there, there is a world where, you know, I've seen, you know, there are these
Speaker A
startups like Lovable or Replet who are customers of Anthropic that allow people who are not software engineers to, you know, to build software products and to start businesses with those products. So if we can move more people in that
Speaker A
direction again, I don't think it's fully going to, going to solve the problem here. I don't think it's fully going to, going to, you know, stop the spike. This is too broad. It's too big. It's too deep. But
Speaker A
but it can be a piece of the solution. So that's number one. Number two, I would say, and this is more controversial, I, I suspect at the end of this that the government is going to need to step in, especially during a
Speaker A
period of transition and, and, you know, provide for, for people, for some, for some of the disruption. Um, and, you know, one, one thing I've, I've, I've suggested is, you know, maybe you might want to tax the AI companies. Um, you know, I don't
Speaker A
I don't know how that would, how that would fly in, in, in Congress today, but, um, I, I, I think that is actually a serious proposal. If you look at the amount of wealth, the increase in the pie that's coming from the AI companies.
Speaker A
If you look at Anthropic's revenue, it's, it's growing 10x a year. It's now in the, in the, in mid to high single-digit billions. If it keeps growing this way, this is going to be an unprecedented amount of wealth creation. Like it's not
Speaker A
going to disincentivize.
Speaker A
Congress is going to have to do this? And we're technology optimists. We think this technology is moving far faster than most people suspect. And when people say AI is slowing down or it's overhyped. We just look at we measure
Speaker A
the properties of the system and it's right on schedule to make really really powerful systems arrive easily during the next 5 years. What does that mean?
Speaker A
You need some kind of policy response at the scale of disruption we expect within 5 years and we expect that along with the the ideas Dario talked about on route to that we need more transparency out of the AI companies. You know, we
Speaker A
and the other AI companies are already affecting society in large ways, and we need to be transparent about how we're measuring our systems, how we're securing our systems, and the economic data about how our systems are being used so economists can tie that to the
Speaker A
actual broader economy and give policy makers the data they need. You're in in uh you guys have been also very transparent about things that are happening in testing and when you guys have released some stuff about some really weird stuff, right? You had like
Speaker A
one testing pattern where somebody where where the machine basically went through someone's email and and attempted to blackmail them. You've had this is our famous marketing tact where they're trying to basically like lie to you so you don't shut them down because
Speaker A
they now are smarter than us. Uh why shouldn't that scare the hell out of me?
Speaker A
Yeah. So um you know the the the answer well one answer is it should um uh uh but but but you know I we should put it in proper context. So, you know, those those are all things that happen to the
Speaker A
model in in a testing scenario. You know, you you can think of it as like, you know, I I put the car, you know, I'm testing a car. I put the car on like a super icy road. I like mess with its
Speaker A
tires a bit and the car crashes. Does it does it mean that will necessarily happen in the wild? No. Does it mean that there are limits to the resilience of the car such that if you push things if it was in an extreme enough situation
Speaker A
and if if you know the requirements of the car started getting tighter or or you tried to amp up the performance of the car by designing a new car then this could be an issue in the real world. So
Speaker A
I see it not so much as a look at the present but a look at the future of where things could be going. And that's why we've advocated so much for transparency is that basically when we run these tests, when we show them to
Speaker A
the world, we're basically looking ahead in time a year or two to what could happen in the real world if we don't train our models to mitigate these risks. And so when we've advocated for transparency legislation, when we
Speaker A
suggested a federal transparency legislation against the the state regul the the the the the state AI moratorium 10-year state AI moratorum when we endorsed SB53 in in California. We're just looking for every AI company to have the level of transparency that we
Speaker A
have shown. We've seen behaviors in the in the wild where models misbehave where they're sick of fantic where you suggest doing something that's really not a good idea like committing suicide and and you know the models just like go ahead and
Speaker A
do that. We we want transparency on all of these kinds of behaviors so that we can look ahead and so that so that we can mitigate. This is still an evolving art. It's still an evolving science and so we see transparency as the key here
Speaker A
when you're talking about it the technology moving faster than people realize. take us behind the or take us under the hood the use your car thing take us under the hood like what is the scariest or wildest thing you've seen AI
Speaker A
do that we're not aware of so you know I I have definitely seen you know you know as we're training the you know the the we're always training training new models trying to design new clouds um you know the process of doing
Speaker A
that is like you have a giant cluster with like thousands of chips and there are these problems you have to solve across the you know this the scale of the cluster and and we've had cases where uh uh you know there was a problem
Speaker A
that an engineer was working on for days or a week and we just fed the whole thing the whole environment into Claude and Claude said this is how you solve it. Um and so Claude again is playing this very active role in designing the
Speaker A
next Claude. We can't yet fully close the loop. I think it's going to be some time until we can fully close the loop, but the ability to use the models to design the the the next models and create a positive feedback loop. That
Speaker A
cycle, it's not yet going super fast, but it's definitely started. And yeah, these days when we make these AI systems, we have to build really, really complicated tests to see how good they've got because they've got way better than just answering multiple
Speaker A
choice tests. The tests are make a computer program that does X. Well, now when we test out the frontier model that we are training, we'll find that it has written a computer program to cheat at the test and persuade us that it's doing
Speaker A
better than it is. So, it says to itself, "Aha, they want me to do this, but I figured out cuz I'm very smart, how to write a computer program that gets me a high score on the test." And
Speaker A
so, when we look under the hood, we'll find, oh, we've made a really smart model that's cheating on its test.
Speaker A
That's not exactly what we intended here. We we see models that are supposed to browse the web and do some task instead opening up a command line or or the toolkit and and and writing code to allow themselves to go around the
Speaker A
browser and cheat on the task. Yeah. Just like a smart kid in high school that annoyed the teacher.
Speaker A
You know, on the one hand, I'm like, "Wow, that's awesome. That's super cool." The other one is like, "Do you ever worry you're creating a monster you can't control?" We So, we worry we worry a lot about that.
Speaker A
That's why we've invested so much in the field of mechanistic interpretability which is looking inside the models in order to understand them. Think of it as like doing doing an MRI on the models.
Speaker A
Um it is there's some research suggesting that is possible for example to detect psychopathy in humans uh by by doing MRI scans. Um uh uh and so we're aiming to do the the the we're aiming to do the same thing with the models to to
Speaker A
to determine what their motivations are, how they think in detail so that if they don't think in the right way, we can kind of retrain the models or or or kind of adjust them to get them to think in
Speaker A
the in the in the you know to to think in a way that is not dangerous to human beings. And so you know we we really believe that the science of shaping and control controlling models is nassient even more nassant than the science of
Speaker A
making the models itself. And so when we call for things like transparency, when we react by saying no, there shouldn't be a 10-year moratorum on, you know, on on on on any regulation of these models, it's it's from a feeling of we don't
Speaker A
fully understand these things we're creating. And we need to both do technical work and and we need some help from from society, from legislative mechanisms to to have some basic understanding shared among industry so that the relevant decision makers
Speaker A
understand what we're seeing and what we're not seeing. what um Nvidia there's some reports out today that maybe China's backing off on buying chips from Nvidia where I don't think you're certain like how much of that's posture I'm not certain how much of it
Speaker A
is but the you're you think it's nuts that we're selling allowing Nvidia to sell chips to China explain why I think I think it's completely nuts um so you know whatever the dangers of the technology what whatever guard rails are
Speaker A
needed I think it's also very important that that we defeat China in this technology right? When I when I talk about, you know, uh uh AI that's capable of doing, you know, all economically valuable labor, um or that goes to the
Speaker A
point where it's a country of geniuses in a data center, just the national security implications of that are absolutely staggering, right? You know, it's it's it's just like this could control the fate of nations. This could control the future of freedom and and
Speaker A
democracy. And if we look at all the ingredients, China has many of them better than we do. They're better at building energy. They're better at building data centers. They have a thriving app app ecosystem. They're catching up on the models. Chips are the
Speaker A
one place where they are behind. Um and and so if we give them those if we give them those chips, they will be able to get ahead of us. We've seen with Deepseek, Deepseek released their model R1 uh back in January. They aimed to
Speaker A
release an R2 that was much better in May. That has been delayed four months.
Speaker A
And the model that they did release, V3.1, was considered a disappointment. was considered a minor update and they themselves have said as have many US publications that the reason for for the delay was the embargo on chips. Now some
Speaker A
people say they'll make their own chips through Huawei. We've also exportc controlled semiconductor manufacturing equipment and uh without that it's going to be many years until they can produce the relevant yields. We saw articles in semi analysis and elsewhere saying that
Speaker A
this year they can only produce a couple hundred maybe a few hundred Huawei chips. The US is producing tens of millions. It's going to take them years, years to catch up. And and in that time, we can we can get ahead of China and
Speaker A
maintain a national security advantage. It may be the only advantage we have because we're so far behind on things like munitions, ship building. This is the only advantage we have. It is it is mortgaging our future as as a country to
Speaker A
to sell these chips to China. Um if we were to do this, I you know, I think and you know, it sounds like some folks are considering it. You know, I I think it could be the single most disastrous
Speaker A
national security decision made in this term. We have about three four minutes left. I want to do a speed round. Let's do like blunt blink reactions to these. Uh other than anthropic, putting enthropic to the side of all your competitors, who's most
Speaker A
likely to be one of the winners? Google. Google. Why? Um they're they're you know, they're they're a big they're a big company. Um they have a lot of compute. They were doing AI research pretty much before anyone else. they
Speaker A
were behind the original deep learning revolution. I used to work there for a year. Um I have a lot of respect for the kind of stuff they've done around for example AlphaFold and uh you know starting starting to make make progress
Speaker A
with their models. They're a big company and they've often been been held back by that and in some ways continue to be held back by that. But I think I think they're a formidable player and people people should should take them seriously
Speaker A
and you know I I have a lot of respect for what they've for what they've done on science and you know you know in some ways I think they've been thoughtful about the technology.
Speaker A
Uh you know you're playing with fire when the people building it have a score called the pdoom which is the percentage chance that this ends in disaster.
Speaker A
What's your pdoom number? Yeah I I I really hate that term. Um it's it's but it's a good it's a good question for blink reaction. very I I I you know I definitely think between the autonomous danger of the model and kind of ending
Speaker A
up on the bad side of some national security tradeoffs and a kind of job thing that's that's uh you know that that kind of goes in a a very bad direction. I don't know. I've I've I'm I'm relatively an optimist. So, I think
Speaker A
there's a 25% chance that things go really really badly and a 75% chance that things go really really well with with not much uh not much space, not much space between.
Speaker A
The 25% chance is a choice that we make and it's a choice that we make in policy. So, uh I'm here trying to push that number way down by talking to policy makers and so is Dario.
Speaker A
It's it's a dynamic number. We I hope that every time we say something the number the number goes down hopefully down not not up. You guys are more B2B than than consumer, but you know the technology as well as anyone in the
Speaker A
world. After the phone, what's the most likely uh what's the most likely form factor of the device where we're using AI?
Speaker A
Uh I you have to give the sci-fi answer. There'll be a strange robot invented by the AI systems we build in the coming years that'll be the thing you use.
Speaker A
Yeah. I don't, you know, we're we're mostly building the engine that gets plugged into all, you know, all these things and powers the world's businesses. So, we're not making we're not making the device, we're not making the device ourselves. But, yeah, I would
Speaker A
I would pay attention to robotics. I think robotics is advancing very quickly. It's not going to be the first area where, you know, progress is made.
Speaker A
But, you know, eventually we're going to want these agents to be embodied and perform tasks in the real world. And so, I would look at humanoid robotics.
Speaker A
And we'll end with this. Uh, and I think I know your answer, but if we come back and do this in I guess we're doing it in March or whatever if you're on stage again, will we look back and say that
Speaker A
the AI capability was much faster, bigger, broader than we thought today. We'll be sitting here saying this is really surprising. It's got way better.
Speaker A
Why didn't you guys tell us? We like we we tried to be clear that it would keep getting way better. It will have got substantially.
Speaker A
Why do you think people don't seem to get that? the conventional like all the coverage the last three months is like oh my god ch5 don't live up to the hype and it just therefore like we don't think we think
Speaker A
I I think people pay too much attention to companies that produce a lot of hype and then and then and then don't live up to the hype. Um you know every 3 months we've we've released a model that you
Speaker A
know that improves in this very straightforward log linear way. It's gotten better and better at coding benchmarks. It's gotten better and better at coding in the real world. our revenue has increased by 10x a year as as as I've said before and all those
Speaker A
curves have just gone straight. The thing that is not straight is people are really excited they're hyped by it and then when they see it you know they get disappointed by the vibes because their their expectation was so high. So there
Speaker A
is a smooth exponential and then there's there's just lots of wiggles around it that are more a phenomenon of perception more a phenomenon of discussion. I will say one more thing which is that um something I've started to see already is
Speaker A
that predictions come true but they don't kind of look the way people think that they do right so you know I I would say maybe 70 80 90% of the code written in anthropic is written is is is written
Speaker A
by claw you know I said something like this 3 or 6 months ago people think of it as falsified because they think of it as like we're going to fire 70 80 or 90% of the software engineers but what
Speaker A
really happens is that the 10 the 10% we're still writing, you know, humans become managers of AI systems. There's a shift because of the principle of comparative advantage. So, it looks more normal than you think. I think eventually it all, you know, kind of
Speaker A
eventually that logic may not hold, but but there's a sort of sci-fi sheen to these predictions to looking at the future that it's going to be weird, that you'll be looking through different colored glasses, that there will be, you
Speaker A
know, that it'll look like Star Wars or something. Um the often when these predictions come true, it's wild, but it's also in a way ordinary.
Speaker A
Yeah. People get used to it. They say, "I've got a universal tutor in my pocket. What's the big deal?" You're like, "That's a huge deal. That's never existed before." Uh I could do this all day, but we're getting the hook. So thank you guys very
Topics:AI impactwhite-collar jobsjob displacementAnthropicDario AmodeiAI regulationretraining programseconomic disruptionAI technologyautomation
