Skip to main content
Jack Clark

Jack Clark

American baseball player

Nov 20, 2026

Gita BhattCORRESPONDENT
13:13
Are you worried about that? And what is it again for society and ethics?
13:18
We can individually take actions here, right? Like I write this weekly AI newsletter None of it's written by AI, it's all written by me.
13:28
And the practice of the newsletter is me reading research papers and seeing what I've understood from them, writing summaries, and then I use AI to say to the AI, hey, have I understood this paper correctly, et cetera.
13:40
I do that partly because I want to provide an educational service to the world.
13:44
but also because it is a way of every week exercising my ability to have attention, to be able to read long things without any intermediary, and to be able to think for myself outside sort of the tool domain of like, what does this mean? So that's what I do individually.
14:02
I think the question is how we get that to happen generally, because when I explain this to people, they're like, cool, strange hobby, not very common.
14:11
I think that at a technical level, we can put friction into the product.
Gita BhattCORRESPONDENT
15:12
But what are other near-term risks that you see? And what should policymakers do then?
Steven KolkenUNKNOWN
0:36
Thank you.
0:37
Thank you very much.
0:39
What I thought I'd do to open this conversation is talk very briefly about, I think the, the both technical and usage trends that have happened in the AI industry for maybe the last 15 years.
0:51
And by way of background, before helping to found Anthropic, I was at OpenAI.
0:56
I was at OpenAI from 2016, which is the year it was, it was founded roughly, to, to 2020 when I left.
1:03
And before that, I was a, a technical journalist that specialized in writing about artificial intelligence in, you know, 2011, 2012.
1:11
And I think that from that perspective, I can give us some context on what's happening now and what the technology trends are and what they might mean.

10 MINS LATER

Ben GraffickMODERATOR
11:43
Where are those benefits? How do you see them and understand them?
12:43
Jack Clark, he's a co-founder and head of public benefit at the company, and he says that while Anthropic would like to see more industry-wide regulation, the company is already addressing many of the safety concerns.
12:56
We've taken actions which exercise our agency, like putting third-party evaluators in.
13:01
We have published some of the best science on how to control and prevent these things.
13:05
There is a huge amount to do, and there is lots of work that has been done here.
13:10
So don't, don't take from this it is the industry saying, "Oh, no, like a really bad thing is coming and there's nothing we can do." That's the opposite of the message.
13:17
The message is, okay, the serious moment for the technology has arrived.
13:22
It's no longer just gotta be industry calling the shots here.
Faisal Islam
Faisal IslamCORRESPONDENT
13:47
Go ahead and pace the frontier yourself."
2:11
How had he, Jack Clark, Anthropic, and the whole sector let AI develop to the point that insiders were warning of a one in ten chance of human extinction?
2:22
We've been talking about AI and the potential risks of it for ten years now.
2:28
What's happened is the technology has got substantially more powerful and those risks which were once academic or theorized, we can now see the early signs of them.
2:37
Which is why we spent this weekend writing this essay about the need to pace the frontier and basically get our arms around this technology so that it's not just industry making all of the decisions about it, but it's industry and government with greater oversight of the industry itself.
2:55
That's how you deal with the risks of this technology.
Faisal Islam
Faisal IslamCORRESPONDENT
3:48
... practically means.
3:49
AI systems for a while have been useful tools with specific skills like translation or coding or being able to label what's going on in an image.
3:58
And then in the last few years, people have successfully put all of these skills together into single systems which have started to become much more intelligent and much more general.
39:19
How do you put hedges on that? How do you build in breaks? Because this is part of the doomerism and the apocalypticism, is that AI realizes, it's not just that it doesn't need humans or shareholders, but they're in the way.
39:35
Well, it's a harder problem, but of similar form to what we're currently dealing with things like cyber and buyer, which is you have some property of the system that you're worried about.
39:46
Well, you need to know how to measure it.
39:48
You need to know how to potentially control it, make it so that the system can sometimes not be capable of AI R&D.
39:55
And you need to probably demand transparency out of the places which are building this, because you as society might say, This is great, and it's going to drive loads of strategic and important stuff in our economy and national security.
40:08
And we probably don't need it to go thousands and thousands of times faster than any human can hope to govern it.
40:14
So, you need some measurement property of it.

13 MINS LATER

53:31
Do you worry about a reserve army of unemployed recent college grads? Or is it that, as you were saying, the data so far, there seems to be a softening of hiring recent college grads or younger people in certain positions? Overall, the sector is growing.
3:53
Highly literate audience.
3:54
Yeah.
3:54
So recursive self-improvement is what happens if we build AI systems that rather than just assisting researchers in their own creation, which is the case today, we build systems that might be able to invent their successor, come up with the ideas that helps Claude 10 build Claude 11 without any researchers involved, only you give it the, the care and feeding of compute and data.
4:18
We think there's a, there's a chance this happens this decade, and if I had to put a year on it, it would be in 2028 or so.
4:25
And the implications of that would be profound.
4:28
We'd enter some regime where suddenly the progress of AI and science more broadly would be defined not only by humans, but by humans and something else, which has never really happened before.
4:39
And I think the world after that can become increasingly hard to think about, so sometimes people call that the, the singularity.

31 MINS LATER

Alana LeboAUDIENCE
36:06
So, um, my name is Alana Lebo, and I'm curious what is needed in how AI is built so that AI can partner and support communities and organizations in how they partner and build the trust and outcomes that would not be created otherwise?
10:54
... dangerous things actually happen?
10:56
So by technological pessimist, I mean I thought the technology would keep getting better, but I didn't think it would get better in the, like, maximalist sense that some of my colleagues did.
11:05
I didn't think that we would have, say, functionally automated all of coding right now.
11:09
I find that actually, like, quite surprising.
11:11
But basically, over the last few years, and I worked at OpenAI before Anthropic, I was just hit repeatedly over the head with what computer scientist Richard Sutton calls the bitter lesson.
11:21
And the bitter lesson is this concept that the more compute and resources we dump into these relatively generic neural networks, the smarter they get and the more emergent properties they have.
11:31
And your specialized system or your ability to be pessimistic about future AI progress loses versus just scaling compute and scaling systems.

20 MINS LATER

31:16
What does Anthropic actually do with this data? Does it somehow feed back to your engineers who are developing frontier models? Do they do anything differently?
9:38
... dangerous things actually happen?
9:39
So by technological pessimist, I mean, I thought the technology would keep getting better, but I didn't think it would get better in the like maximalist sense that some of my colleagues did.
9:48
I didn't think that we would have, say, functionally automated all of coding right now.
9:53
I find that actually like quite surprising.
9:55
But basically over the last few years, and I worked at OpenAI before Anthropic, I was just hit repeatedly over the head with what computer scientist Richard Sutton calls the bitter lesson.
10:05
And the bitter lesson is this concept that the more compute and resources we dump into these relatively generic neural networks, the smarter they get and the more emergent properties they have.
10:15
And your specialized system or your ability to be pessimistic about future AI progress loses, versus just scaling compute and scaling systems.

44 MINS LATER

54:51
For more of those
10:06
Uh, what, and what sort of jobs, therefore, are under risk, likely to be replaced in this world of AI agents that can actually kind of improve themselves without human input?
10:17
So as we say in this report, I think there are open questions about whether AI systems can be truly creative, you know, truly, um, come up with interesting, off-the-wall ideas.
10:28
There is not really evidence for that yet.
10:31
And what we see it, it happening with work at Anthropic is we're now limited more by the ability to generate good ideas than the ability to do the engineering to turn those ideas into reality.
10:42
So everyone who generates a bunch of ideas, is creative, has kind of entrepreneurial ideas, is going to be advantaged by this.
10:49
And if you're in rote jobs, highly repetitive jobs, jobs that in the past would have been at risk of automation in, say, things like factory work in, in previous generations, now that same phenomenon is going to play out in knowledge work.
11:02
And we need to figure out, how do we help people change their mindset about what sorts of work they're doing and move to these more kind of creative jobs where they're gonna be able to get a lot more done?
12:04
Well, give, can you give us a sense of that? An AI agent very quickly in the next, before 2030, could replace a lawyer, an accountant? A copywriter?
CalumHOST
9:40
How do you start getting that freelance work? So like this
9:44
Dribbble site I mentioned was a bit like a portfolio.
9:46
So it allowed you to sort of like put on like bits and bobs of what you're working on.
9:50
So if I did something I was proud of, I'd stick it on this.
9:52
A bit like Instagram is like or was at one point where you'd be putting on some like a picture that you really liked.
9:57
Like Dribbble was a bit like a... showcase of your work you could put on some work in progress or something that was done and that was like your little portfolio of like stuff that you'd done so I was quite active on there put quite a few things on there and mainly just word of mouth some other designers and local people this guy's done this for me he's done a good job like he's really creative I was very I really enjoyed the website side of things because I could back in the day like code a little bit and put something like something together You did your year.
10:30
Did you drop out after that? I did.

34 MINS LATER

CalumHOST
44:31
So what was your
16:57
Jack Clark is one of the firm's co-founders and has been speaking to our economics editor, Faisal Islam.
17:02
Right now, it's like the AI industry has a gas pedal, but it doesn't have a brake pedal in the car.
17:08
And what we're saying is we want to do some of the work required to build that brake pedal so we in the world have an option.
17:15
It's not obvious today that you want to do that, but absolutely, as you say, at some point in the future, you might say, "Let's get all of the benefits we can for, say, biology and medical research, and let's, let's take a, take a pause or take a moment on AI research while we can absorb the societal changes implied by this."
speaker_12UNKNOWN
18:45
For people watching this who are sort of hearing about the potential pitfalls and perils, and at the same time, you know, due to your suc- commercial success, your company's gonna be worth apparently a trillion dollars in the next few weeks officially, the, there's a bit of a disjunction here that you're creating a system that is now the world's problem.
19:08
It, it's a tremendous responsibility, not one we take lightly.
19:12
I think about this all the time.
19:13
We share this information because we're trying to have a discussion.
2:44
He's been speaking to the BBC's Faisal Islam.
2:47
Right now, it's like the AI industry has a gas pedal, but it doesn't have a brake pedal in the car, and what we're saying is we want to do some of the work required to build that brake pedal so we in the world have an option.
3:00
It's not obvious today that you want to do that, but absolutely, as you say, at some point in the future, you might say, "Let's get all of the benefits we can for, say, biology and medical research, and let's, let's take a, take a pause or take a moment on AI research where we can absorb the societal changes implied by this."
speaker_1UNKNOWN
3:56
Are you, are you worried for, for your kids if, if, if this message doesn't get out?
3:59
I'm, I'm not worried for my kids if this message doesn't get out.
4:02
I am, um, worried for my kids if we as a society don't, don't, don't have a serious conversation about what the implications of AI's continued advances mean.
4:11
They're going to mean that there are potentially great benefits.
4:14
There are also risks.
31:03
W- what is the benefit of this?
31:06
So today, when we want to do good in the world in areas like biology or medicine or robotics, you have to take these systems that mostly exist as digital AI systems and adapt them into this complex domain in the real world, and you do that by pairing with scientists to do itNow, an AI system capable of improving itself is also capable of going into domains like medicine, going into domains like biology, discovering what it needs to get smarter and more capable, and working with people to adapt itself to improve its performance there.
31:38
So I think what we're seeing is the potential for a dramatic acceleration in science in the coming years as these AI systems gain this capability and become much more like creative co-scientists than tools that scientists use, which would be a big change.
33:14
Is that right? And 'cause what, I mean, what would that look like in practice, and how do you convince competitors to do that as well? Because obviously this is, you know, obviously it's been described as an arms race.
33:27
So our view is we've built amazingly powerful technology.
33:30
We're gonna keep building it.
33:31
And in the coming years, that technology's going to start to do a lot of major things in the world in domains like science.
33:37
But when I look down at the, the car we're driving, all I have is a gas pedal.
29:15
And those that do typically don't have the capability.
29:18
Those that...
29:19
Well, the number of people...
29:20
This gets to the point, right? The number of people who want to visit harm on others in the world is, is, is relatively small for this kind of, of, of horrendous act, and they typically aren't trained biologists.
29:31
And if they are, carrying out acts of terrorism or usually doesn't... requires you to not be a lone wolf.
29:37
It requires some amount of coordination.
29:39
Now, the risks of AI systems are AI systems are universal educators.

28 MINS LATER

57:46
Tell us a bit about this index.
Patricia ClarkeCORRESPONDENT
8:11
Can you explain a little bit about how you're measuring economic data at the institute and, and what you are seeing?
8:17
Yeah.
8:17
So what we do is we can look in a privacy-preserving way at the sorts of tasks people ask Claude to do.
8:25
So people might ask Claude to do certain types of software engineering tasks.
8:29
They might ask it to do legal tasks like reviewing the difference between two contracts, the sort of work that paralegals do.
8:37
They might ask it for advice about things like, you know, cooking to gardening or, or whatever.
8:43
You can look at these sorts of tasks, and you can then join it to a certain type of l- job classification label, which is called the O*NET classification scheme, which the Bureau of Labor Statistics uses.

7 MINS LATER

Patricia ClarkeCORRESPONDENT
15:33
What do you think that might do to, is doing to people's brains now and, and might do to us as a society in future?
6:10
It is overseen by a robot factory line, and AI is overseeing the robots, and then AI is also trying to imagine new types of products, and then if you just wanna read-
6:23
When new candidate products are found, it runs through a series of tuned expert systems, and if it gets a high enough score, simulates the product in a high-fidelity simulation.
6:33
If it passes these tests, then a candidate product is produced, airlifted by drone to a nearby human focus group, and if it satisfies their criteria, is sold on an eBay-like auction site frequented by the factory's thousands of distributors.
6:50
A bidding process takes place, and in a few days or weeks, data comes back about how the product succeeds in the market.
6:56
If it does better than the existing product, more parts of the factory are dedicated to creating new products in this style, and the improvisation line begins exploring the latent space of the new product.

5 MINS LATER

12:20
Where-- Do you do that?
12:22
Y- I, I find myself, uh, staring into the, the face of, of change that we haven't experienced in perhaps a century, and I, I don't know how we'll respond.
12:33
Um, a hundred years ago, C.S. Lewis wrote a book called The Screwtape Letters, which is about, uh, a demon in hell whose job is writing letters to their boss, a senior demon, a manager demon, about their attempts to corrupt a person on earth.
36:11
But is this world you're talking about a system where you have AI agents coding, AI agents overseeing the code, AI agents overseeing the meta-overseeing of it, right? Like, are we just talking about-Models all the way down?
36:25
Eventually, yes.
36:26
And I think that the thing that we are now spending all of our time on is making that visible to us.
36:33
A year or two ago, we built a system that let us, in a privacy-preserving way, look at the conversations that people were having with our AI system.
36:43
And then we gained this map, this giant map of all of the topics that people were talking to Claude about.
36:48
And for the first time, we could see in aggregate the conversation the world was having with our system.
36:55
We're gonna need to build many new systems like that which allow for different ways of seeing.

31 MINS LATER

67:57
You know, maybe you wanted to move over to data center construction, but actually now we don't need as much data center construct- Right? Like, you can think of it like that.
29:44
I mean, so h-how, how do you live with that tension, that balance between thinking that you're building something that contains the possibility of magic, but also, like, recognizing sometimes that the only honest way to speak about this technology is to, you know, uh, be clear about just how dangerous this thing might be?
30:08
We, we share what we, we feel about it and also set up things like the institute to share more information so that the rest of the world can work on the problem, problem as well.I don't think, though, this is unique to, to AI.
30:24
Um, almost a hundred years ago or so, there was a memo written by the British government worrying about the rise of civilian air travel.
30:33
And it w-- this memo, which we'll be able to send you after the show, had this dark vision of a world where, like, wars were entirely fought by, like, aircraft and, and terrorists in aircraft, like, bombing cities and killing people and, um, the whole of, like, life in, life on continental Europe would be disrupted by this, and we would live in kind of, like, unimaginable horror.
30:54
And planes should only be kept for, like, government purposes, and you shouldn't have them generally distributed because of the harm they could do.
31:01
Now, obviously, it got part of it right in that all of those things I just mentioned are done by aircraft today.
31:07
But we also have an entirely changed world due to civilian aircraft and transportation, which has unlocked, like, a vast, vast range of things.

17 MINS LATER

48:20
This is artificial general intelligence, the line that was promised to cross." Are they right? Are these, are these coding, are these agentsAre these agents what we five years ago would have called AGI?
20:26
Either in Anthropic or more broadly, what are people doing with these systems that seems actually useful?
20:34
Yeah.
20:35
So this morning, a colleague of mine said: "Hey, I want to take a piece of technology we have called Claude Interviewer," which is a system where we can get Claude to interview people, and we use it for a range of social science, uh, bits of research.
20:48
He wants to extend it in some way that involves touching another part of Anthropic's infrastructure.
20:53
He Slacked a colleague, who owns that bit of infrastructure, and said, "Hey, I want to do this thing.
20:58
Let's meet tomorrow." And the guy said: "Absolutely.
21:01
Here are the five software packages you should have Claude read before our meeting and summarize for you." And I think that's a really good illustration where this gnarly engineering project, which would previously have taken a lot longer and many people, is now going to mostly be done by two people agreeing on the goal and having their Claudes read some documentation and agree on how to implement the thing.

13 MINS LATER

33:58
But is this world you're talking about a system where you have AI agents coding, AI agents overseeing the code, AI agents overseeing the meta-overseeing of it, right? Like, are we just talking about models all the way down?
25:43
Why is it different? Why is this different to that? Why would conventional economics be wrong in that analysis, in your view?
25:50
Speed is different and, uh, quantity is different, and AI has both of these properties.
25:57
What I mean by quantity is that AI is available over your internet connection.
26:01
So once a system gets better, it's not like people need to build some physical infrastructure to make use of it.
26:08
It's just instantly there.
26:09
It instantly distributes.
26:11
That ties in with the speed thing.

28 MINS LATER

54:17
And all of this, all of this has to be navigated in an international ecosystem of geopolitical competition which is incentivizing governments and states not to ask too many questions, where states don't know much about what AI companies are doing, where there is a profound asymmetry of knowledge, and where governments have an incentive to allow them to run free for fear that their geopolitical opponents will steal a march on them if they don't.Imagine if the Manhattan Project had been in the hands of Elon Musk and not FDR.
3:25
What's your observation with the current state of models? And then my follow-on question would be, how do you see the next three to six months of this space evolving?
3:35
I think the most important thing to take away from where we are now is that language models are a lot more powerful than most of us think today.
3:44
We keep on having this phenomenon in AI research where we think to ourselves, huh, Could an AI model be good at, say, mathematics? And then we give it a test and we discover that it's actually far better at mathematics than we thought.
3:56
And the same has been true of physics, biology, a broad set of skills.
4:01
Additionally, language models today can take actions for you over large periods of time.
4:07
And I'll just try and give you some intuitions here.
4:09
So a couple of years ago, if you asked a language model to do a task for you, like translate a sentence or take a simple bit of code and rewrite it into another language.
6:28
What were some of the things that you have prompted that really was an eye-opening moment for you?
28:25
To me, it's absurd, but...
28:26
Yeah, and the reason that my numbers are more conservative is I think that we will enter into a world where there will be an incredibly fast-moving, high-growth part of the economy, but it is a s- a relatively small part of the economy.
28:40
And it may be growing its share over time, but it's growing from, like, a small base.
28:45
And then there are large parts of the economy, like healthcare or other things, which are naturally slow-moving and maybe slow in adoption of this.
28:54
I think that the things that would make me wrong are if AI systems could meaningfully unlock productive capacity in the physical world at a really surprisingly high compounding growth rate.
29:06
You know, automating and, and, and building factories and things like this.
29:10
But even then, I'm skeptical, because every time the AI community has tried to cross the chasm from, like, the digital world to the real world, they run into, like, 10,000 problems that they thought were paper cuts, but in sum add up to you losing, like, all the blood in your body.

15 MINS LATER

44:33
(laughs)
Cory BookerMODERATOR
8:32
So-
8:32
I'll give one example, which is, I think when you start an AI company, you may not be interested in national security, but national security is interested in you.
8:40
And you end up building these very powerful systems that may have national security uses or misuses.
8:47
And for that, I think we do need to come up with tests that make sure that we don't, you know, put technologies into the market, which could unwittingly to us, like advantage someone or allow some non-state actor to commit something harmful.
9:00
Beyond that, I think we can mostly rely on existing regulations and law and existing testing procedures that exist, like the FDA or the FAA or other things, and we don't need to create some entirely new infrastructure.

7 MINS LATER

Cory BookerMODERATOR
16:21
Senator Clark, what are you doing?
16:23
I think you'd want to find ways to aggressively field like experimental technology and to Alex's point, get departments to actually use it.
16:31
Because at Anthropic, we discover that the more we find ways to use this technology, the more ways we find it could help us.

We value your privacy

We use cookies to understand how you use our platform and to improve your experience. Click “Accept All” to consent, or “Decline non-essential” to opt out of non-essential cookies. Read our Privacy Policy.