CEOInterviews.AI
Start App
Jacob Coxon
Former AI Researcher, Anthropic

A Conversation with Jacob Coxon & Rob Reich

📅 Nov 13, 2026 Stanford HAI 61 MIN 25 VIEWS 102 SEGMENTS · 3 SPEAKERS
The companies building artificial intelligence have promised to transform the world. But who gets to decide what that transformation looks like, and what happens when people who have worked inside those companies question their direction? Jacob Coxon, an AI researcher formerly at Anthropic and OpenAI, recently made a very public decision to leave Anthropic, a company that has made the responsible development of AI central to its identity. In this conversation, hosted by Stanford HAI and the AI Governance Lab, Coxon joined Stanford Professor Rob Reich to discuss his departure and the questions...

Questions asked in this interview

12
  1. 0:05And how many years have you been in the Bay Area?
  2. 1:26Can you say a bit about whether there was any particular event or set of conversations, or what was the approximate cause that led you to a) deciding to leave and b) deciding to do it publicly?
  3. 6:28So your Anthropic colleagues, you say, thought that exiting was not especially praiseworthy or maybe recommendable in terms of change. Why not?
  4. 11:10So for you, do you have equity in OpenAI?
  5. 13:09... you like to share with people two months out that you feel is essential from, let's say, a technical point of view to know that maybe most people who didn't have the proximity inside the company to learn about now really need to hear?
  6. 17:08You ask an ordinary person, how has AI made a tangible difference in your life?
  7. 21:05What about — what kind of strategies do you think come to mind for people who, like yourself, are technical in background but who see the solution as being something like a global treaty on pausing?
  8. 26:35So if many people exit, it leaves — my words now, not yours — bad actors, who are actors without the moral compass or sense of safety as a significant issue, present, and things go worse. Is that fair?
  9. 29:08What do you think about that?
  10. 37:06Why don't we start here in the front row?
  11. 48:52So the question I have for you is, what are the labs actually doing to have any hope of measuring these kind of discontinuities at the societal level?
  12. 55:15It was a human barista that made the coffee today, and isn't it cool for me to have purchased it from a human?
Rob Reich 0:05 ↗
Are you ready? Hello everybody. All right. Thank you all for coming to a conversation that I'm going to have, and then you'll have a chance to participate in, with Jacob Coxon. I don't think I need to do much introduction. You've read his viral tweet almost exactly two months ago. Jacob resigned publicly from his position at Anthropic, saying that he and other people who worked with him inside Anthropic, and also at OpenAI where he had been previously, were developing frontier capabilities and AI systems that were accelerating, and that in his judgment posed a potentially catastrophic, even existential threat to humans. Jacob had worked — I don't know, what was it, four months at Anthropic, and then maybe three or four years previously at OpenAI. He got his undergraduate degree in a technical field, in computer science, at the University of Cambridge. And how many years have you been in the Bay Area?
Jacob Coxon 1:04 ↗
Three years.
Rob Reich 1:05 ↗
Three years. So we're going to have a conversation first between the two of us, and then we'll open the floor up to you. I think I'm going to take my seat after having performed the introduction. Would you please join me in welcoming Jacob Coxon?
All right. So let's do a little bit of a look back in time. You spent three years plus working at OpenAI. You move over to Anthropic. You had been there for less than half a year. You decide sometime along the way to post this resignation tweet. Can you say a bit about whether there was any particular event or set of conversations, or what was the approximate cause that led you to a) deciding to leave and b) deciding to do it publicly?
Jacob Coxon 2:01 ↗
Yeah, there was no immediate approximate cause. There was like a cumulative realization that we were going very quickly and that also we didn't fully control the systems. In terms of one cause that pointed towards things going very quickly, I saw exact projections for the capabilities of models in the next six months to one year that made me think, you know, we're on the cusp of recursive self-improvement. And at that point, we could reach superintelligence before the decade is out, and we don't obviously know how to control superintelligence. In fact, some might argue we have no clue how to control superintelligence. But that was basically the reasoning.
Rob Reich 2:39 ↗
All right. So maybe just let's get a little human take on what's happened to you the past two months. You know, in the wake of what you posted, various people have said, you know, we will look back as time goes on on September 8th, 2026 as Coxon Day. And, you know, your tweet thread got — I don't know what it is these days, 175 million impressions or views, which just to be clear is like the equivalent of the locker room celebration at the final of the World Cup. I don't think you really had much of a social media profile before. So the virality of this was extraordinary. You went and did a whole bunch of interviews. You're here today doing an interview. Just from a human experience, what's the past two months like? You obviously didn't expect this. Have you been overwhelmed, joyful that you have a platform to speak, want a plan of action because that's what would be valuable in the moment? Give us what the past two months have been like.
Jacob Coxon 3:41 ↗
Yeah, sure. It was incredibly surprising, and I don't know if anyone has gone accidentally viral before, but basically there's just non-stop adrenaline. Like when you first have a bunch of eyes on a thing you're saying, especially when it's like you go against your co-workers or a lot of your immediate social circle. There was this thing of not only having disrupted my immediate personal life, but also feeling like the eyes of the world were on me. So basically it was like a week of minimal sleep, not really understanding what was happening, still kind of conducting interviews, and just about like looking back and being like, 'Oh wow, I just did five interviews and I don't really remember exactly how they went down.' But since then, it's been more relaxing. I've been making use of the opportunity to have conversations I wouldn't have had otherwise. And it's been mostly just very surprising that people were willing to listen.
Rob Reich 4:30 ↗
Yeah. So one of the things that happened in the immediate wake of the tweet thread was not merely, you know, arresting the attention of lots of people around the world, ordinary folks, about potential existential threat from AI. Even your at that point then former colleagues at Anthropic said, you know, Jacob is on to something. We also agree that there are catastrophic existential risks. The CEO of Anthropic went on some TV show and was asked about what you had posted and said, you know, on balance, I agree more with Jacob than disagree with him. I want to sort of take that kind of view and ask you whether you can reflect upon your time at OpenAI or Anthropic as a way of asking aloud: do you feel like you were perhaps somewhat more courageous in speaking out, or since people agreed with you even in public, is there a kind of private posture inside the company not offered up publicly in which you sit around the cafeteria talking about the risks and that most people don't speak out publicly but you did? What's your sense of the corporate culture at Anthropic with respect to private worry but public silence?
Jacob Coxon 5:45 ↗
Yeah. So there's different kind of factions internally and different perspectives. One quite dominating faction at Anthropic feel like they've been speaking about superintelligence for, you know, a decade and people won't listen. And I think when I spoke to them about the plan of, you know, leaving and saying I'm kind of fed up, they said, we get your concerns, but from our experience that won't do anything. Like people aren't going to listen. It's just not the right strategy. You need to stay and help. So I'd say it was maybe less courageous with regard to that corporate culture insofar as people actually agreed, and more just, you know, I did it as an expression of personal resignation. I did not expect it to be a strategic decision. I expected them to be vindicated.
Rob Reich 6:28 ↗
All right. So that's interesting to hear. And, you know, since you're at a university here, not on some, you know, television program, I want to take the opportunity to share that one of the things I thought about when I learned about your resignation was a short story that I have been teaching for the past 20 years to a variety of different people by a wonderful fantasy writer named Ursula Le Guin called 'The Ones Who Walk Away from Omelas.' I asked Jacob — this is the first time I've met Jacob, just 10 minutes ago — whether he knew the story, and he does. In case you don't, it's a short six-page, seven-page story about a utopia. And the utopia has a social contract that makes the place the happiest place on earth, but on the condition that a single child is kind of imprisoned in a dungeon-like condition under the city. The story retells about festival day. Every day, every year in Omelas, they have to celebrate their great happiness by acknowledging the existence of the child. And the story ends by saying in every single year some people walk away from Omelas. Now what I find interesting about this story is not whether it's a utilitarian parable about the greatest good or the greatest number — can we impose all the misery on one person and get happiness for all? — although it's sometimes taught that way too, but rather think about the people who leave. Is it better to stay inside and fight for change or to exit and try to fight from outside, which seems a version of what you must have gone through? So your Anthropic colleagues, you say, thought that exiting was not especially praiseworthy or maybe recommendable in terms of change. Why not?
Jacob Coxon 8:14 ↗
I think many of them feel like they have been trying to warn people, which is somewhat amusing because they're building the very thing that they warn about. And I think they maybe don't quite appreciate how weird that contradiction looks from the outside. So I think strategically, from a purely like, you know, if you expect that the race will happen, then staying and participating does make sense. But I think there is a kind of communications or appearances problem where people take your continued participation as tacit acknowledgement that things are going to be kind of okay. Like they don't see a kind of fire alarm situation because they think, well, if there really was smoke or a fire, people would be like running out of the building rather than continuing to stay in what they consider to be a doomed fight.
Rob Reich 9:06 ↗
And well, how did you break from this consensus? Was it something about your conscience — you didn't want your hands dirty by continuing to contribute to whether it was pushing the capabilities forward or having modest effects on the safety team — or was there something about a decision you made about wanting to speak from now the outside that was independent of feeling like your hands were dirty?
Jacob Coxon 9:31 ↗
Yeah, I honestly — it was mostly I am genuinely quite scared of what AI will look like in 2030. For me selfishly, for my immediate friends, for like my family, and being part of making it just felt bad. So maybe to a certain extent it was an inability to like decouple and say, I'm going to continue to fight. It might be the right thing to do this. I just really didn't want to be part of it anymore. Also, maybe even like, you know, with a couple of years left on the clock, I was like, do I want to be inside an AI company? There's other places in the world I could be for the last five years, you know.
Rob Reich 10:13 ↗
All right. But I can't — we have to pursue that, like head to a, you know, a bunker on an island and start writing poetry, or going on The Daily Show and telling Jon Stewart and the country about what's going on.
Jacob Coxon 10:26 ↗
I mean, yeah, that wasn't the plan. Like I thought I'd be in London at this point. Yeah, I didn't expect people to listen. I thought I'd be in London writing Substacks or maybe helping like, you know, talk to certain people who might listen a bit. I didn't want to be in.
Rob Reich 10:42 ↗
All right. And yet here you are.
Jacob Coxon 10:45 ↗
I'm still in the Bay. I can't.
Rob Reich 10:46 ↗
Yeah. Are you losing less sleep because of AI dread today than you were when you were on the inside?
Jacob Coxon 10:53 ↗
I think it's about the same. There's trade-offs. On the one hand, I'm not presented with the immediacy of it every day. But on the other, I now actually can't see what's going on, which is quite a scary feeling I'm finding. Having, you know, it's pretty inadequate transparency from the outside. I don't know what new developments they've had in the last two months.
Rob Reich 11:10 ↗
Yeah. All right, let's get the following question sort of out of the way. I don't know the answer to it. You know, I think a week or two after you resigned, David Robinson also very publicly resigned from OpenAI. He said in various interviews, most prominently on Ezra Klein, that it wasn't a judgment about the, you know, the compromised position that he felt at OpenAI — he thought it was a shared problem across the main frontier AI developers. And he was asked, you know, do you have equity and financial interests caught up? How much of a financial consideration is the decision to stay rather than leave? So for you, do you have equity in OpenAI?
Jacob Coxon 11:56 ↗
Yeah, I do.
Rob Reich 11:57 ↗
And how should you invite people to react to your financial interest in the company going well?
Jacob Coxon 12:05 ↗
Yeah, that's a good question. I think the regulatory capture narrative is a little bit confused. I agree it would probably be easier for me to argue with that if I, like, you know, ceded rights to my OpenAI stock, but it's like — it's nice money. Like, I like it.
Rob Reich 12:21 ↗
It is. But you could take the moral high ground.
Jacob Coxon 12:23 ↗
Yeah, I could. I could.
Rob Reich 12:24 ↗
But you're not.
Jacob Coxon 12:25 ↗
I'm not. No. Why not? I like money, bro. I'm like — I agree, it's blood money.
Rob Reich 12:31 ↗
Do you prefer dying to liking money?
Jacob Coxon 12:34 ↗
I don't think it's going to — if I thought for a moment that donating it like would change anything, then sure, I'd do it. I also like —
Rob Reich 12:40 ↗
Might give you more moral credibility though.
Jacob Coxon 12:42 ↗
I agree. But you could also stop asking the question because it's irritating to be asked. No, but like, you know, that would also — if you hadn't asked it, then the moral credibility wouldn't have been questioned. Anyway, I like —
Rob Reich 12:54 ↗
That's a real dodge, Jake. That's a real dodge.
Jacob Coxon 12:56 ↗
Yeah. Anyway, I'm not holding out on donating it. But one thing I've said is also I think my parents were like worried I was having a psychotic break. I think if I also given away all the money, they'd be even more convinced of that.
Rob Reich 13:09 ↗
I'm not so impressed by that answer either, but we'll leave that be. All right. Let's be prospective rather than retrospective. Recursive self-improvement has begun. Automated R&D is underway. Next generation models accelerating potentially their capabilities. You saw from the inside the timelines, the compute that have been booked for the coming year or more. You know, are we on the brink of an intelligence explosion? What would you like to share with people two months out that you feel is essential from, let's say, a technical point of view to know that maybe most people who didn't have the proximity inside the company to learn about now really need to hear?
Jacob Coxon 13:57 ↗
Yeah, I think the main thing is there's this notion of a general intelligence that has been spoken about for a while, and I think people still forget about it or don't think about it. In particular, there are spiky capabilities in the models at the moment. So they're sort of better at coding and math than they are other tasks, and that can sometimes lead people to imagine that, you know, labs are spending a lot on human data for coding environments for specific tasks and somehow there's not this general intelligence. But I think the thing that does remain true is that improvements to the general capabilities of the model translate to downstream performance in all tasks, and not just like the ceiling for those capabilities, but also the efficiency with which you can learn from data. So as we look forward, I think you should expect rising capabilities in soft skills, in strategy, in forecasting, in persuasion, in kind of all domains of human expertise, not just the hard stuff, and you should expect it quite quickly.
One more thing about RSI here is that means there is one very technical specific task, namely coming up with algorithmic improvements to general intelligence. So amendments to the transformer architecture, amendments to the data set, and that one narrow task translates to capabilities across all domains even without human expert involvement. And I think we've roughly known this to be true for a long time. But I want to rearticulate that that is still very much the case in the labs, that the holy grail is general intelligence and it does not require that much additional human expert involvement.
Rob Reich 15:38 ↗
And when offered up the chance to sing, you know, a tune about the extraordinary potential benefits, the utopia that's on the horizon — accelerating science, solving math problems, you know, finding cures for diseases, massively growing GDP — do you rate the risks as just greater than the potential benefits, or you can't tell about which one is going to come about and we should be somewhat risk-averse, maybe like the precautionary principle toward things that are potentially catastrophic? What about all the benefits?
Jacob Coxon 16:13 ↗
I think the benefits are obvious and extreme, like health, science, everything happening quickly. I just think the benefits are basically priced in — like we know how to achieve the benefits, we know that they're coming. What is not priced in is that the transition will be smooth. And even if we were to slam down the brakes, I think like present-day models would deliver tremendous benefits already. I think the problem really is we're going to be like drowning in benefits to the extent that we're accidentally rendering people unemployed or completely reshaping society without even noticing. So whereas the downsides could be very sudden, could completely change things. I think the overall sentiment I have is taking things slowly allows people to see what's happening, to make their voice heard with regard to what they do and don't want to happen. But the benefits are definitely there. Yeah.
Rob Reich 17:08 ↗
Okay. For what it's worth, you know, I think the election last week showed that most ordinary Americans don't really feel that there's much benefit that's tangible. Yet it's something in the neighborhood of 40-plus local or state initiatives about moratoriums or limiting data center construction. Nearly every single one passed. You ask an ordinary person, how has AI made a tangible difference in your life? Most of them say, I can't really tell, but I learned from Jacob that it could kill us all, that it might take my kids' jobs, and that it's raised all of my costs if I'm on the electrical grid. So the benefits are so abundant maybe in the future, but for what it's worth, I don't see them yet. Does that — you agree that —
Jacob Coxon 17:53 ↗
It doesn't quite ring true in that when I'm learning new fields, I think the educational benefits are quite obvious. Also the educational downsides with regards to potential cheating. Although I think, you know, in-person exams can kind of skirt that. I'm curious what you think. Yeah, in terms of learning, I think when I was younger, I would have loved to have what is essentially a personal tutor in every subject for a very cheap price. I think that's already a huge benefit.
Rob Reich 18:19 ↗
Yeah, that's an omnipresent conversation amongst anybody who's either a student in a classroom or an instructor in a classroom, and happy to talk more about that. And, you know, a bunch of people in the class that I'm now teaching on AI governance are here in the room. They can weigh in on it too. But I want to now, you know, use my background as a philosopher and a political scientist to go back to this idea that's present in Ursula Le Guin about exit and voice, or staying on the inside and fighting or moving to the outside. So you can take this as an invitation both to reflect perhaps upon, you know, your decision to leave, but I mean more to focus on the opportunities that are before you or anyone else right now. You decided to act, you know, what at least the outside world felt unilaterally: I'm going to resign and post about it. An alternative is something that you and a lot of other people did just a few months prior, which is to sign this document, 'Pacing the Frontier,' calling for a slowdown of the development of AI models, in which I don't know what it is now, 1,300, 1,400 people across five or six different of the frontier model developers said how urgent it was to bring about some coordination. That suggests to me an effort at what an ordinary, you know, political scientist would call labor action — the coordinated labor activities of the workers. Why defect from acting collectively? Why not, for example, organize people inside Anthropic and say, unless we get something better than what's currently on offer, we will go on strike for a day and withdraw our labor from frontier model building?
Jacob Coxon 19:59 ↗
Yeah. So I hope this sort of employee action does happen. I think what is needed ultimately is some form of international regulation of this. And it's interesting you brought up the 'Pacing the Frontier' letter because, yeah, when I had conversations with, say, leadership at Anthropic or senior employees at Anthropic and we're talking about the 'Pacing the Frontier' letter or the prospect of, say, you know, having some sort of treaty with China, their overall — the TLDR was that like that letter is not going to do anything, like it's not going to help, no one's coming for international regulation, at least not by next year. We are on our own. Like, and I think in that situation, defecting against your own company, unilateral striking, it's kind of at the wrong level. Like we need to solve the collective action problem at the layer above the companies, which requires people kind of waking up outside the companies.
Rob Reich 20:51 ↗
At the layer above the companies. So do you think, you know, citizens in the US trying to get the federal government to do something, or do you mean all citizens trying to get nation states to coordinate?
Jacob Coxon 21:01 ↗
Nation-state coordination, I think we need ultimately.
Rob Reich 21:05 ↗
Yeah. So I've hung around computer scientists now for a decade, and as a political scientist I could hardly think of anything more important than what you described — getting actual politics to do some work rather than thinking technical safeguards are the only or perhaps most important path forward. But can you describe — I worry that perhaps you won't have a lot of great ideas about getting that coordinated action happening. Where are the policy-minded people at work to try to bring this about? What about — what kind of strategies do you think come to mind for people who, like yourself, are technical in background but who see the solution as being something like a global treaty on pausing?
Jacob Coxon 21:55 ↗
Yeah. I think at least for me, and I'm quite, you know, politically naive, the main thing is just having actual awareness of how dangerous the technology is. Because I think an international treaty can sound like a crazy thing or, you know, like it's never going to happen, but if you actually do consider the machine god a likely outcome, not just something straight out of science fiction, then I think the citizens of the world being scared of the machine god is not utterly implausible. So at least for now, the main route to me does continue to be spreading awareness of precisely how transformative a technology we're on the cusp of developing.
Rob Reich 22:38 ↗
This will come across as snarky, but I don't intend it this snarky. I'm spreading awareness through Substack posts — spreading what's the — give me the mechanism for spreading awareness. I mean, you do have a bigger platform, so you have perhaps a different answer to the story.
Jacob Coxon 22:53 ↗
If you have any ideas, I'd love to hear them. I'm doing my best here. And I think some people are trying to say — like I think Grimes is an artist with a popular following who is also scared about this stuff. I think potentially convincing people with popular cachet that this is something worth taking seriously could be a great route. Also convincing, say, like superstar politicians that this is worth them staking their career on. Generally, my best guess is it has to go via convincing people that the public will take seriously.
Rob Reich 23:27 ↗
There's an alternative that I've occasionally heard from people who perhaps share a bit of your profile, which is — it will sound odd to say it publicly, but some are willing to say it publicly — I'm fervently hoping for an AI disaster that will awaken both the public and policymakers. You know, the classical thing to say here is that a policy window opens, but it really shifts dramatically. The Overton window shifts in the wake of a disaster. Just to give an example, you know, airplane flight — the Federal Aviation Administration was created in the United States in 1958. And in 1956, there was wall-to-wall coverage of an airplane crash over the Grand Canyon to awaken the public to the risks involved in commercial airflight. What do you think of that as an historical diagnosis? And how do you feel about that?
Jacob Coxon 24:22 ↗
This is really tough because I think some people say jokingly that the best thing that happened to AI safety in the last year was the OpenAI hacking attack on Hugging Face because it raised awareness of the reality of the issue. And there are definitely people, especially the more, you know, strategic, utilitarian-minded people in AI safety, who will perhaps whisperingly joke that we need a catastrophe in the short term to scare people. And they say it still whisperingly because obviously it's quite a — I think quite a heinous thing to think. And it always confuses me because I'm definitely not a pure utilitarian. And even if in the long run a catastrophe of some sort would be in the best interest of people, like I think we should do our best to prevent all catastrophes in the immediate term. That's just my unreasoned ethical opinion. But I can definitely answer that there are, you know, there are some people who think a minor disaster would be to the best of everyone.

49 more exchanges in this transcript

Sign in free to read the rest of this interview. No card required.

Sign in to read the full transcript

Cite this transcript

APA, MLA, BibTeX
APA

Coxon, J. (2026, November 13). A Conversation with Jacob Coxon & Rob Reich [Interview transcript]. Stanford HAI. CEOInterviews.AI. https://ceointerviews.ai/interview/3275222/

MLA

Jacob Coxon. "A Conversation with Jacob Coxon & Rob Reich." Stanford HAI, 13 Nov. 2026. Transcript, CEOInterviews.AI, https://ceointerviews.ai/interview/3275222/.

BibTeX
@misc{coxon2026_3275222,
  author       = {Jacob Coxon},
  title        = {A Conversation with Jacob Coxon \& Rob Reich},
  howpublished = {Interview transcript, Stanford HAI. CEOInterviews.AI},
  year         = {2026},
  month        = {nov},
  url          = {https://ceointerviews.ai/interview/3275222/},
  note         = {Speaker-attributed transcript with timestamps}
}