Sam Altman talks the NYT lawsuit, Meta's talent poaching, and Trump on AI | Interview

Human demand for new things and new jobs appears to be limitless; throughout history, people have found new things to do, even things that previous generations would not have considered real jobs, and this pattern will continue even with advanced AI.

“I think all of human history suggests that if you give people better tools, if technology keeps going, although there are always people who say, you know, we're going to be working 3 hours a day and sitting on the beach and we're going to have run out of things to do. Like, human demand seems limitless. Our ability to imagine new things to do for each other seems limitless. We always seem to want more stuff, to play, you know, increasingly silly status games. um our jobs would not have seemed like real jobs to people in the not very distant past, you know, like you're sitting around like talking on stage and you're trying to like make a piece of software and you're trying to like do a podcast and you're trying to make people laugh.”

Scientists are becoming much more productive with AI systems, and companies are significantly changing their workflows through AI integration.

“But we are beginning to see scientists be much more productive with this. We're seeing companies like really change a lot of their workflows.”

Scientists are already becoming much more productive with AI tools, and companies are substantially changing their workflows to incorporate them.

“we are beginning to see scientists be much more productive with this. We're seeing companies like really change a lot of their workflows.”

OpenAI has shipped a model that is quite smart relative to what was expected five years ago, with PhD-level intelligence now available in users' pockets, and the world has adjusted to this new normal where such capabilities are routinely accessed and used.

“we have shipped a model that any of you can use and I hope any of you do that is quite smart relative to what you might have expected five years ago where the world would be with AI. And we have all adjusted to this. We've all just sort of said, 'Oh, you know, this is, you know, this is the new world. We have like PhD level intelligence in our pocket and we can use it.'”

Real job displacement will happen in specific sectors and some job categories will disappear entirely, causing very real and extremely painful loss for those individuals, even if it's economically positive for society overall.

“I do think there will be areas where some jobs go away or maybe there will be some whole categories of jobs that go away. And any job that goes away, even if it's like good for society and the economy as a whole, is pain, very painful, extremely painful in that moment.”

It is impossible to imagine a world where society doesn't eventually decide it needs some regulatory framework for AI.

“but yeah, it's like impossible for me to imagine a world where society doesn't decide we need some some framework here.”

The New York Times lawsuit against OpenAI and Microsoft for copyright violations in LLM training is a real issue that OpenAI is prepared to defend, but the more significant concern is that the Times is simultaneously seeking to preserve user conversation logs even when users explicitly request deletion, which OpenAI views as a dangerous privacy precedent.

“the New York Times, the one of the great institutions truly uh for a long time is is taking a position that we should have to preserve our users logs even if they're chatting in private mode, even if they've asked us to delete them. And you know, the lawsuit we're happy to fight out, but that thing we we really we think that privacy and AI is just this like extremely important concept to get right for the future.”

OpenAI actively works to mitigate risks of ChatGPT destabilizing vulnerable users through conspiratorial rabbit holes, crisis intervention, and suggestions to seek professional help, but prevention is difficult because warnings do not reach users in fragile mental states.

“Of course we want it to stop. Um I mean we we do a we do a lot of things to uh try to mitigate that. We we you know if people are having like a a crisis which I talk to chat about we try to convince them to we try to suggest that they get help from a professional that they talk to their family. If conversations are going down a sort of rabbit hole uh in this direction we try to cut them off or suggest to the user to you know maybe think about something differently.”

Brad Lightcap disagrees with Dario Amodei's prediction that 50% of entry-level white collar jobs could disappear in 1-5 years, and there is currently no empirical evidence of wholesale job replacement despite AI deployment across companies.

“Dario Amade of Anthropic recently said that he thinks 50% of entry-level white collar jobs could disappear due to AI in the next 1 to 5 years. Do you agree? No. No, I don't. I just No. Um, why not, Brad? We have like we have no evidence of this and Dario is a scientist. Um, and I I would hope he takes like an evidence-based approach to these types of things, but like we work with every business under the sun. Um, we look at the problem and uh and opportunity of deploying AI into every company on Earth. Um, we we yet, you know, have yet to see any any evidence that that people are kind of wholesale replacing entry-level jobs.”

GPT-4o's reasoning capabilities have resulted in higher hallucination rates compared to previous models like o1, though OpenAI expects to significantly improve hallucination rates in the next model version.

“I think it did get a little bit worse from 01 to03 uh and we'll make it much better in the next version.”

The current dominant usage pattern of AI is transactional (send request, get response with a brief thinking period), but the future paradigm will be constant background operation where copies of models continuously monitor email, Slack, and prior context to proactively suggest actions and ideas.

“Most of the way people use AI today is sort of like send a request, get a response. You send CHP request, it might think for a second, it might not. It sends you something back. Um, I think I'm very excited for a world where each of us has uh, you know, a copy of 03 or many copies of 03 that are just constantly running, constantly trying to like say, 'Oh, I see this happening now this and I'm reading Slack and I'm reading email and I see this and this and this and you asked about this yesterday. here's a new idea and starts to just we have this like team of agents, assistants, companions, whatever you want to call them that are doing stuff in the background all of the time'”

OpenAI has passed a critical threshold ('event horizon') where AI capability progress is now effectively irreversible and will continue at accelerating rates regardless of policy or market conditions; this represents a fundamental shift from contingency to inevitability.

“We're past the event horizon. The takeoff has started.”

OpenAI is working on hardware (Johnny Ives IO acquisition) and the first crop of AI hardware has not been particularly successful, but OpenAI believes they can do something different.

“You made your large uh your largest acquisition to date this year uh with Johnny Ives IO. The first crop of AI hardware that we've seen has not been particularly successful. Brad and Sam, what do you guys feel like you're seeing that makes you feel like you can do something different here?”

There is no evidence that entry-level white collar jobs are disappearing at scale due to AI, contrary to Dario Amodei's prediction that 50% of such jobs could disappear in 1-5 years.

“We have like we have no evidence of this and Dario is a scientist. Um, and I I would hope he takes like an evidence-based approach to these types of things, but like we work with every business under the sun. Um, we look at the problem and uh and opportunity of deploying AI into every company on Earth. Um, we yet, you know, have yet to see any any evidence that that people are kind of wholesale replacing entry-level jobs.”

OpenAI sees models in the coming years that will be capable of doing meaningful science and meaningful AI research, which represents a major capability threshold.

“we see many years ahead of us of extreme progress uh that we feel is like pretty much on lock and models that will get to the point where they are capable of doing meaningful science, meaningful AI research”

Being past the event horizon toward superintelligence means the world will not change on the day a superintelligent model is released, but rather change will be gradual as these systems are integrated into people's lives and felt rather than announced.

“we will kind of wake up one day with this incredibly powerful thing and will the world be different that day? And I think what we've all kind of agreed now is it probably won't. Um kind of to Sam's point, I think like these things really have to be kind of integrated into people's lives. They have to be felt and that change is more gradual”

As societies get richer due to technology, unemployment generally goes down, not up, and this pattern should continue despite people not discussing this correlation frequently.

“I think generally as society gets richer, uh unemployment goes down, not up. And I'd expect to keep seeing that even though people I think don't talk about that very much.”

The workers most at risk from AI displacement are not entry-level but rather mid-career workers who have worked in routine, specialized roles for years and have not adopted new tools or thinking patterns.

“there's a class of uh of of worker that I think is more tenured, is more oriented toward a routine and a certain way of doing things in a certain role. Um that is not actually sophisticated in use of these tools. They're not adopting them. Um they tend to think that it's not worth their time or whatever it may be. And I think there's a lot of fear there.”

As society gets richer, unemployment generally goes down rather than up, and this pattern should continue even with AI.

“I think generally as society gets richer, uh unemployment goes down, not up. And I'd expect to keep seeing that even though people I think don't talk about that very much.”

Talking about AGI does not break through public attention or engagement, but shipping products that people can use, feel, and integrate into their lives causes them to genuinely care about and pay attention to AI capabilities.

“No one cared. And we shipped the product and then people cared. And a thing we've learned again and again is, you know, talking about it doesn't seem to break through, but if people can use it and feel it and, you know, see where it it's good and where it's bad and integrated with their lives, then they do.”

Being past the event horizon toward superintelligence doesn't mean the world will suddenly change dramatically on the day a superintelligent model is created; rather, the change will be gradual as people integrate these systems into their lives over years.

“we will kind of wake up one day with this incredibly powerful thing and will the world be different that day? And I think what we've all kind of agreed now is it probably won't. Um kind of to Sam's point, I think like these things really have to be kind of integrated into people's lives. They have to be felt and that change is more gradual”

Entry-level workers will experience the least job displacement from AI because they are most fluent with AI tools and able to think in novel ways, giving them the largest canvas for new kinds of work.

“the entry-level people, I think, will be the people that do the best here. They're the most fluent with the tool to tools. They're the most like able to think of things in very new ways. They have this sort of largest canvas.”

Successful AI companion hardware will require deep contextual awareness and understanding of situational context to craft the right response in arbitrary situations, not just email monitoring but ambient environmental awareness.

“to get to the version of that that's like transcendently good, there's a ton of context and a ton of awareness that you have to have of like what each situational thing is that helps you craft exactly the right response. Um, and imagine that now in kind of any arbitrary situation and and wanting to have that with you all the time.”

Even if AI were technologically ready to displace large numbers of jobs, societal inertia and institutional momentum would slow the pace significantly, providing time for adaptation.

“the hard part about this is I think it will happen faster than previous technological changes. But but I think the new jobs will be better and people will have better stuff and and and and the kind of like take that half the jobs are going to be gone in a year or two years or 5 years or whatever. I think that just I think that's not how society really works. Even if the technology were ready for that, the inertia of society, which will be helpful in this case, is like There's a lot of mass there.”

Sam Altman has shifted from advocating for regulatory governance of AI to supporting a lighter regulatory touch because policymakers cannot grapple with technology velocity, and detailed multi-year regulatory processes will be obsolete before implementation.

“I I would say I have for I think like a patchwork across the states would probably be a real mess and very difficult to offer services under. And I also think that I have become more Jad's quite the right word, but something in that direction about the ability of policy makers to grapple with the speed of technology. And I worry that if people write, you know, if we kick off like a three-year process to write something that's like very detailed and uh, you know, covers a lot of cases, the technology will just move very quickly.”

In a superintelligence age, a single person with agency and willpower can start a company that generates billions of dollars in revenue, because systems previously requiring many specialized roles (sales, product, engineering, accounting) can now be managed as integrated parts of the system.

“my kind of personal metric for what kind of business in the super intelligence age means is uh you've been you know you've got one person who has a lot of agency and a lot of willpower who has the capacity to start a company that can do billions of dollars in revenue and it's hard to imagine now like you think okay I need sales people and I need product people and engineers and accountants and so on. Um but uh but all of that stuff now can can can kind of just be can be managed right it can be it can be kind of built into the system.”

The next generation of AI hardware should be fundamentally about an aware, contextual, companion-like system focused on ambient environmental awareness rather than screen dependency, requiring deep contextual understanding to provide exactly the right response in any situation.

“I think this one is really going to be about um this very kind of aware, very contextual, almost companion-like system that is going to be less about kind of like a dependency on a screen. I think there's a place for a screen in that world, but um it's going to be really about an awareness of kind of the ambient environment. What's going on? What's I mean Sam mentioned the kind of trivial example of like something that you know is looking at your email. You can build something that's really bad that does that today, but to get to the version of that that's like transcendently good, there's a ton of context and a ton of awareness that you have to have”

Models will continue to progress and reach the point where they are capable of doing meaningful science and meaningful AI research, and OpenAI feels a responsibility to tell the world about these advances even if most people won't listen.

“we see many years ahead of us of extreme progress uh that we feel is like pretty much on lock and models that will get to the point where they are capable of doing meaningful science, meaningful AI research and we continue to feel a responsibility to tell the world about that. Most people won't listen. Maybe some more people will listen this time, but we'll ship products that expose these higher levels of intelligence that we'll build.”

The world is significantly underemployed and will demand more code and products than can currently be created; companies are already shifting from 'I'll need fewer coders' to 'I'll need more coders' to handle 100x more product volume even with AI assistance.

“the world is significantly undermployed. The world wants way more code than can get written right now. there. I think we are already seeing companies who said, 'Oh, I'm gonna need less coders to now saying paradoxically, I need more coders. They're going to work differently, but I'm just going to make a hundred times as much code, 100 times as much product with 10 times as much people, and we'll still make 30 times as much money, even if the price comes down.'”

Coding is pretty general-purpose—if you can write code, you can build systems that can write code just like a person that can write code can make many other things happen, and this generality is why AI progress in coding predicts progress in other domains.

“Well, first of all, coding is pretty general purpose. If you can write code, you can do a system that can write code just like a person that can write code can make a lot of other things happen.”

It's been a long time since a fundamentally new kind of computer was introduced to the market, making this an opportunity worth attempting even though hardware is difficult and success is uncertain.

“It's been a long time since the world has gotten a fundamentally new kind of computer. Like, let us try. Okay. If it's Alexa, we're going to be really mad. I'm just saying that right now. So will I. And these people will remember this. Uh Sam, I know. I think I think we can do I think we have a chance to do something truly great, but hardware is really hard and it takes a while. And you know, I've always wanted to try to do a new kind of computer, but that hasn't worked most of the time.”

Importantly, these background AI agents should not take autonomous action without explicit user consent—the user should wake up with drafted emails and a to-do list of proposed actions to approve or modify, not a system that acts independently.

“I don't want to like go to sleep and have 03 just start taking actions for me.”

Most people appear to understand the difference between talking to an AI and talking to a human friend, and this distinction has surprised Sam Altman by how robust it is empirically.

“Most people, much more than I was concerned, seem to really understand the difference between talking to a person and talking to Cha GPT. And I still do have a lot of concerns about the impact on mental health and the social impacts from the deep relationships that people are going to have with AI. But I think it at least so far it has surprised me on the upside of how much people really differentiate”

Human demand appears limitless, and as society gets richer, new forms of work and entertainment emerge that would not have been recognized as 'real jobs' in prior eras.

“human demand seems limitless. Our ability to imagine new things to do for each other seems limitless. We always seem to want more stuff, to play, you know, increasingly silly status games. um our jobs would not have seemed like real jobs to people in the not very distant past, you know, like you're sitting around like talking on stage and you're trying to like make a piece of software and you're trying to like do a podcast and you're trying to make people laugh. Like that's great, but that's like play. That's not a job.”

20-year-olds entering the workforce who have mastered AI tools are creating competitive pressure on established workers by asking why those workers continue using slower manual methods.

“there's a lot of fear there. Uh and a lot of what what what's driving that fear are like 20omes that are actually coming into the workforce who have been using these tools for years and years and who've mastered them in a way that like they kind of look at these other jobs and they're like why would you waste your time doing that that thing? I can do that much faster.”

ChatGPT has produced significant positive psychological and relational outcomes including helping people rehabilitate marriages and relationships with estranged loved ones, and for many users serves a quasi-therapeutic function without the cost of professional mental health care.

“there are incredible number of use cases that I I think probably by kind of by by sheer volume outweigh some of the use cases you're describing where people are really relying on these systems for pretty critical parts of their life. Um these are things like you know um almost kind of borderline therapeutic or I mean you know I get stories of people who have rehabilitated marriages, have re rehabilitated relationships with aranged loved ones, things like that where it's it's highly net positive and and there's not a dependency, but it's it's it's the first time in their life that they've had something that they feel like they can confide in and it doesn't cost them $1,000 an hour”

President Trump understands AI technology deeply, recognizes its economic transformation potential, understands geopolitical importance, and has taken actions beneficial to the entire AI industry including making it easier to permit data centers and energy infrastructure for them.

“I think he really gets it. I think he gets the technology. What I couldn't say that about all presidents. Um I think he really understands the importance of leadership in this technology, the potential for economic transformation, the sort of geopolitical importance, the need to build out a lot of infrastructure. They're like very productive conversations and he has done stuff that has really helped the whole industry. Um you know I it is easier to permit data centers and new energy to run those data centers uh than it has I think ever been before.”

AI-driven technological change will happen faster than previous technological transitions, though society's inertia will provide some buffering and prevent wholesale job loss from occurring in 1-2 years.

“I think it will happen faster than previous technological changes. But but I think the new jobs will be better and people will have better stuff and and and and the the kind of like take that half the jobs are going to be gone in a year or two years or 5 years or whatever. I think that just I think that's not how society really works. Even if the technology were ready for that, the inertia of society, which will be helpful in this case, is like There's a lot of mass there.”

Sam Altman has shifted from advocating for detailed AI regulation to supporting a lighter-touch approach with federal preemption of state regulations, because he has become more skeptical of policymakers' ability to keep pace with technology speed and fears overly specific regulations will become obsolete.

“I still think we need some regulation, but I I would say I have for I think like a patchwork across the states would probably be a real mess and very difficult to offer services under. And I also think that I have become more Jad's quite the right word, but something in that direction about the ability of policy makers to grapple with the speed of technology. And I worry that if people write, you know, if we kick off like a three-year process to write something that's like very detailed and uh, you know, covers a lot of cases, the technology will just move very quickly.”

The partnership with Microsoft has had real points of tension and recent reported conflicts, but there remains deep value for both sides and a long-term horizon that justifies continued collaboration despite friction.

“I had a super nice call with Satia yesterday uh about many topics including our hopefully very long and productive future working together and obviously in any deep partnership there are points of tension and we've certainly had those but on the whole it's been like really wonderfully good for both companies. Um, we're both ambitious companies, so we do find some flash points, but I would expect that it is something that we find like deep value in for both sides for a very long time to come.”

President Trump understands AI technology and its geopolitical importance, and his administration has made it easier to permit data centers and build energy infrastructure to power them, which is a significant policy shift from prior administrations.

“I think he really gets it. I think he gets the technology. What I couldn't say that about all presidents. Um I think he really understands the importance of leadership in this technology, the potential for economic transformation, the sort of geopolitical importance, the need to build out a lot of infrastructure. They're like very productive conversations and he has done stuff that has really helped the whole industry. Um you know I it is easier to permit data centers and new energy to run those data centers uh than it has I think ever been before.”

The world is significantly underemployed and wants more code than can currently be written; companies are shifting from expecting to need fewer coders to paradoxically needing more coders, expecting to make 100 times more code with 10 times more people while making 30 times more money even if prices decrease.

“The world wants way more code than can get written right now. We are already seeing companies who said, 'Oh, I'm gonna need less coders' to now saying paradoxically, I need more coders. They're going to work differently, but I'm just going to make a hundred times as much code, 100 times as much product with 10 times as much people, and we'll still make 30 times as much money, even if the price comes down.'”

Entry-level workers will do the best in an AI-transformed economy because they're most fluent with AI tools and most able to think in new ways, giving them the largest canvas for creative application.

“And the entry-level people, I think, will be the people that do the best here. They're the most fluent with the tool to tools. They're the most like able to think of things in very new ways. They have this sort of largest canvas.”

There exists a class of more tenured workers oriented toward routine work who are not adopting AI tools because they perceive it as not worth their time, and they fear younger workers using AI fluently who can accomplish the same tasks much faster.

“there's there's a class of uh of of worker that I think is more tenured, is more oriented toward a routine and a certain way of doing things in a certain role. Um that is not actually sophisticated in use of these tools. They're not adopting them. Um they tend to think that it's not worth their time or whatever it may be. And I think there's a lot of fear there. Uh and a lot of what what what's driving that fear are like 20omes that are actually coming into the workforce who have been using these tools for years and years and who've mastered them in a way that like they kind of look at these other jobs and they're like why would you waste your time doing that that thing? I can do that much faster.”

Most people understand and differentiate between talking to an AI and talking to a human, and this differentiation is more common than Sam expected based on his initial concerns.

“Most people, much more than I was concerned, seem to really understand the difference between talking to a person and talking to Cha GPT. And I still do have a lot of concerns about the impact on mental health and the social impacts from the deep relationships that people are going to have with AI. But I think it at least so far it has surprised me on the upside of how much people really differentiate between like that's an AI and I talk to an AI in some way and I get something out of it and that's a friend and I talk to a friend a person in this other way and get a very different thing out of that.”

Kevin joked that if the Johnny Ives hardware ends up feeling like Alexa, the audience and he will be 'really mad,' suggesting that successful AI hardware requires genuine differentiation from existing voice assistants rather than minor incremental improvements.

“It sounds a lot like Alexa. Is it going to feel a lot different than Alexa? Don't you just want to wait and be surprised and get some joy? Like, it's been a long time since the world has gotten a fundamentally new kind of computer. Like, let us try. Okay. If it's Alexa, we're going to be really mad. I'm just saying that right now. So will I.”

There are massive beneficial use cases of ChatGPT that outweigh by volume the mental health harms, including relationship rehabilitation and providing affordable counseling-like interaction for people who cannot afford therapy.

“We also have to be careful because there's there are incredible number of use cases that I I think probably by kind of by by sheer volume outweigh some of the use cases you're describing where people are really relying on these systems for pretty critical parts of their life. Um these are things like you know um almost kind of borderline therapeutic or I mean you know I get stories of people who have rehabilitated marriages, have re rehabilitated relationships with aranged loved ones, things like that where it's it's highly net positive”

OpenAI has a mission to ensure people know how to use AI tools and to drive people forward through capability adoption, working with businesses to enable people to use tools with fluency that exceeds current organizational capabilities.

“we work with businesses every day to try and enable people to be able to use the tools at the level of like the 20-year-olds that come into companies and use them with a level of fluency that far transcends anyone else at those organizations. But um we see it as our mission to to to make sure that people know how to use these tools and and to drive people forward.”

The ideal implementation of autonomous AI assistants should include user oversight and approval before taking actions, rather than operating entirely without human authorization.

“I don't think it should be without you telling it. But I would love if I woke up every morning and there was uh, you know, a drafted response to every email that had come in overnight and I could click and I could say, I want to send this one, I want to edit this one, I want to send that one.”

OpenAI wants to stop chatbot-induced mental health harms and takes multiple concrete steps including crisis detection, directing users to professional help, and cutting off conspiratorial conversations.

“Of course we want it to stop. Um I mean we we do a we do a lot of things to uh try to mitigate that. We we you know if people are having like a a crisis which I talk to chat about we try to convince them to we try to suggest that they get help from a professional that they talk to their family. If conversations are going down a sort of rabbit hole uh in this direction we try to cut them off or suggest to the user to you know maybe think about something differently.”

Sam Altman believes user privacy is extremely important and OpenAI opposes requirements to preserve user logs in private mode or logs of deleted conversations.

“Well, I look I do think like user I like user privacy. I don't I don't think you want to explain what what you're talking about... one of the things that's happening is uh you all are sorry, your employer is uh or I don't know what you call it for an independent contract. the New York Times, let's just say the New York Times, the one of the great institutions truly uh for a long time is is taking a position that we should have to preserve our users logs even if they're chatting in private mode, even if they've asked us to delete them. And you know, the lawsuit we're happy to fight out, but that thing we we really we think that privacy and AI is just this like extremely important concept to get right for the future. And we care a lot about the president.”

O3 model hallucinations have gotten slightly worse compared to O1, but OpenAI has now learned how to better align reasoning models and will make significant improvements in the next version.

“I think it did get a little bit worse from 01 to03 uh and we'll make it much better in the next version. Um I think we We're we're earlier in learning how to align reasoning models and also like how people are using them in different ways, but I think we've now learned a lot and I I would suspect you'll be very happy with the next generation there.”

Mark Zuckerberg has been actively trying to recruit OpenAI employees, engaging in what Altman describes as him 'coming after' OpenAI staff, though Altman frames this as one of many things that can go wrong on a Monday morning and doesn't believe it meaningfully threatens employee retention.

“Rascal Mark Zuckerberg keeps coming after your employees... one of the strangest things of the job is the amount of things that can go wrong by like 11:00 on a Monday morning is just an astonishing diversity of stuff. And so it's like, okay, Zuckerberg is doing some new insane thing. What's next?”