we care like super deeply about safety and would have done and will do the best thing we can for safety, security, and just sort of like trust and robustness in this technology, whether or not any voluntary agreement is signed.

one which has gotten a lot of attention recently is that there could be a loss of control to ai and that we build something that we don't have sufficiently aligned to human values

if we do too much to restrict this technology in the name of safety and we end up with a small number of companies or people or perhaps one country that have too much power over this

So our goal here is to sort of be the pragmatic centrists between these paths because we can see how tempting it is to err in one direction or the other.

Well, I kind of think the artificial intelligence thing happened. I think a lot of people would say AGI happened or is close to happening.

I think superintelligence does... It still sort of understates the magnitude of what I think is happening, but I think it's a better word than...

I wanted to get openai.com, openagi.com and openasi.com. And as we moved up the ladder from AI to AGI to super intelligence, we were going to change the URL, but brands are sticky.

continue to think that as this technology reaches increasing levels of capability and the world has to contend with what this means for society as a whole, we probably will see more calls for sort of some guardrails.

Transcript note: The sentence begins mid-thought; the opening words before 'continue' are missing from the transcript.

I think one of the biggest differences between us and some of the stricter, let's we believe that the world should accept some bad things happening for the benefits of this technology and people having the agency.

Transcript note: Words are missing between 'let's' (458) and 'we believe' (465).

a single lab in San Francisco should have it and make sure nothing bad happens and kind of figure out how to dole out the benefits. I think that is a completely unacceptable trade-off from a perspective of liberty and human agency

when it comes to the potential of a serious loss of control to AI, and this is not a today worry, but it may not be in the distant future, this is where I would say to the accelerationists, let's be a little thoughtful

So I would say accept bounded risks, accept risks that we understand in exchange for the benefits. Liberty Agency, all those things I was just talking about, but don't accept like the really catastrophic risk.

we don't want any restrictions or testing on anything but the frontier models. We do not want to slow down people that make very powerful but not the absolute frontier in the spirit of really focusing on the potential catastrophic risks.

even though I believe that if you don't put any testing or any oversight on models that are one or two generations behind the frontier, I accept that bad things will happen with those. I would not go advocate a licensing regime for those.

I do think there are other researchers and other labs who said, like, don't deploy this. Terrible stuff is going to happen. People can't be trusted with this. And I think our worldview has been proven.

We are in the process of disclosing more incidents. There's nothing else I'm aware of at that level of severity, but we have more things to disclose.

this is, I think, a good example of why you don't want max acceleration as we in our context reality and more increasingly powerful models find that like oh the models are not capable of a new threat vector

Transcript note: Garbled: 'in our context reality' is unclear, and 'are not capable' reads from context as 'are now capable'.

I now have a model running with access to my most sensitive information, my email, my text, messages, like all the stuff on my computer. And I really want to know that I can trust that model and it's going to behave well.

It's a practice run for our company, but it's also like, I think this is one of the times as before the models get truly very powerful where the world, like the world can understand what these AI incidents are going to be like.

i think if like something goes wrong with our models during training like there's going to be some version of that we need to be responsible for perhaps a separate new legal liability model for ai training specifically

But there's like a lot of anxiety about what the economic impact of all this is going to be. There's a real fear that I think is quite justified about concentration of power.

you have like, oh, Anthropics head of safety research saying there's a 10% chance die or something, which I hate because that person should have a lot of agency to make that number lower. It's not like a static number.

Transcript note: Words are missing between 'a 10% chance' (1280) and 'die' (1288), presumably 'we all'.

No, I think the industry needs to figure out a way to address what the actual concerns are. And right now, I think a lot of people in the industry are being quite tone deaf.

Transcript note: The host's half-question 'So do you think that industry should...' at 1305 is labelled as Altman.

We could expose products in different ways. They'd be much harder to distill, but that would take away some of the open-ended creativity, and we don't really want to have to do that.

I think it's the supply chain that's driving up costs much more than the financing markets. Financing still seems easy to come by, but there's a lot of it, and silicon wafer seems much tighter.

So I can say some confidence I believe our best internal model is well ahead of the best internal model that any Chinese company has but it's a little difficult to say exactly how far ahead

Transcript note: 'say some confidence' presumably means 'say with some confidence'.

which is part of why I'd like there to be some sort of deal here. And, you know, we've done that for other kind of majorly disruptive pieces of technology in our history, even between countries that really don't like each other.

Transcript note: Words are lost between 'clearly a huge' (1979) and 'about what countries' (1984).

where you could really focus on raw materials and such it's not going to work quite the same way and so i think a lot of the current approaches are based in sort of the last era of thinking

If I could wave a magic wand, I would say that America has the most powerful AI models, the plurality of AI compute, that America gets to lead the policy decisions in the world.

But I don't think it should be... get to benefit from the technology or have any voice in the future or have like a super lopsided concentration of power for the rest of the world.

Transcript note: Words are missing between 'it should be...' (2117) and 'get to benefit' (2120).

I think I was never as doomer-y about the job stuff as other people, but it still has surprised me somewhat on the upside of just how much latent demand existed in the world

I think there will be some areas of work where There's actually not much creativity required, and we don't care about the person doing it. And many of those tasks or jobs will be done by AI.

we don't care that much about what the machines do. And in some sense, we never have. We're obsessed with other people. We want to know about the person behind the art or the person behind the company

I think that was a great example where 10,000 agents worked super hard for four days in perfect coordination with each other and did this amazing thing. And... none of the public reaction was about that actually, the problem being solved.

Transcript note: The problem is named only in the host's question ('Navier Stokes').

this idea that you're going to have these agents that are running in the background 24-7, connected to your systems, doing stuff proactively for you, this is feels important to explore. I think we're going to see amazing stuff happen there.

Transcript note: 'this is feels' is garbled in the transcript.

And then there are like us that believe we are going to see a renaissance in the number of companies and just an absolute lowering of the bar to people, one-person companies, small team companies, creating incredible new products and services.

One take on this would be that the renaissance was fundamentally centered around people and the industrial revolution was fundamentally centered around machines.

We're going to be like house cats and the AI will decide how to reward us and tell us what to do. And I have heard people say that. That's certainly not what we think.

But Dots, the new thing we launched today, is the first thing that has given me back control of my mornings and my mental space in the mornings.

I know what the demo is because I was able to look at the presentation in space or whatever. And this sub is probably going to make the demo not work. Would you like me to try to fix it?

Transcript note: 'this sub' is probably 'this sev' (a production incident, rendered 'big seven production' at 613); the colleague's name is transcribed as 'Tebow'.

And then the next day, my dot was like, hey, I noticed you were trying to do this thing, and you didn't find it. But I kept thinking about it overnight, and I found it for you.

I have definitely learned... that I've been using dots that although technology is very wonderful in many ways, the kind of like crazy attention hacking of social media and phones and notifications had gotten way worse than I realized for me personally.

And you can use it yourself, or you can have AI use it on your behalf, but to really collaborate closely, especially when it's multiple people and multiple people's AIs all working together, like a lot of the current abstractions we have break down.

not only to like share information but but actually like you know seven people's dots can working together create something new that kind of existed in all of our heads to some degree, but was like an emergent thing.

Transcript note: 'dots can working together' as transcribed; the start of this answer sits inside a host-labelled line at 1064.

A lot on my mind recently is just how we're thinking about the safety alignment and security challenge in front of us. Um, but you know, six months ago, I probably cared more about revenue and growth and metrics

Transcript note: The host's question 'What do you care about now?' at 1163 is labelled as Altman.

we want to make the best cheapest fastest most capable model at every point on the Pareto curve of cost and speed and modality but I think there will be new approaches new architectures new technologies that are things we just can't cover all of

For all of the fear people had a couple of years ago about, man, AI model companies are just going to eat everything. There is an unbelievable amount of significant new companies, great new products, building this stuff all of the time.

Transcript note: The host's question at 1458-1461 is labelled as Altman.

Well, A, I think we have a lot of conviction that empowering builders and not just trying to be like the one giant evil megacorp is important.

Transcript note: The host's question at 1541-1546 is labelled as Altman.

I mean, I don't even know what this is like fire or something. This is like a real change to what's possible. You know, maybe we like brought it down from the mountain, but we want to give it to everybody

how much work you can get done with one person or two people or whatever that feels totally different and I don't think there has been a enough re-examination of that in this sort of startup conveyor belt of Silicon Valley.

Transcript note: The line at 1622 also contains host interjections before this passage.

And I did not realize what a difference that makes to my kind of iterative thinking loop to have responses come in so quickly But in terms of trying to build something, trying to think through a problem, it's like a wild difference.

So historically, what we have most... focused on and I think done a great job of is bringing down the, like, price of IQ. And we have made hugely more capable models that get dramatically, dramatically cheaper every few months or every year.

Now, there's this new vector that is clearly going to be very important of speed. And we are putting the investments in place. Like, our own chip is going to be an incredible contributor to letting us get very fast inference

I think we can deliver 8x at a great price in the not distant future, but then we'll have 100x faster.

Transcript note: The host's 'I don't think I can afford it' at 1922 is labelled as Altman.

And so I think the right way to think about this is as far as I can see, there will always be a real frontier in terms of intelligence or speed or whatever else at a higher price.

But I think you can also bet that we will keep doing what we've done of just a ridiculous, like put Moore's law to shame, relentless focus on driving value per dollar up and up and up

We're starting for something at the top end of the market and you will see us over time make a mass market consumer version.

This is a totally different category. Like the amount of work, the level of capability, the level of intelligence this can take on, I think this is just a completely new thing.

And the reason that I think both of those are true is the better you make AI, the more people want to use it.

We've been talking about this for a few months now, but we're clearly getting into a very steep part of the capabilities curve.

slowing down our training or our model release so that we can make more progress and put more of our attention into safety, alignment, monitoring, security. We will happily do that.

And we don't want to do this thing where like, oh, we need regulation or we need our competitor's X, Y, or Z, or we won't be safe. We are going to do this responsibly.

make sure we understand how to operate in this new way, be able to make some of these decisions in front of us without the pressure of being a newly public company. And then I think we'll be public at some point.

I think one thing that would be good would be shared standards for safety cases about what you need at each level to proceed without undue risk, and then some shared methods for ensuring that everybody is doing a good job in implementing those.

I think this has nothing to do with going public. This is just, I think, our responsibility for our mission and what we're working on.

AI has to be built for people. It has to be good for people. Part of that is it has to be safe. It has to be always under human control. It has to do what people want.

I wouldn't over-rotate on this one thing. It was like, you know, this model was a little bit worse on a few of the evals we look at. This was, think of this in the abundance of caution category.

It's not just like stop it because, you know, we don't want smarter models. Of course we want better models. It is make sure that we improve the areas we need to confidently and safely continue with progress.

so there's nothing else that I'm aware of that we're aware of that is as serious as the hugging incident, but we are trying to put out every small or otherwise case we can find.

Transcript note: Altman's answer begins inside a segment labelled Kate Rooney, after Rooney's question; 'hugging incident' is 'Hugging Face incident' split across the segment break.

I don't think it's a full solution and I worry that if we treat AI safety as only an engineering problem, we will miss the very important point that we have a science problem in front of us.

I suspect, like other industries, there will be a sort of multi-level thing you think about the auto industry for example if there's faulty parts you know you can be liable in some ways there

And I don't think like, you know, people do talk about open source, but I don't think there's any open source model out there that you can run that beats us on a cost price comparison.

But this is like a big thing. And I get why people have anxiety. And also if they need to like project that anxiety on someone or have like a little bit of relief and poke fun at me, it doesn't bother me.

I don't want to be too optimistic because I have been wrong before when I thought something was going to happen in D.C. and, you know, it didn't. But I do feel like people are taking it seriously this time, and I'm optimistic.

I don't have a particular timeline. Once we have run a few safety cases, once we feel like we understand how to contend with this next level of AI and do it safely, I think we could.

But I think it's great to give people like more time to hang out with their kids, do their hobbies, work on creative new projects. I think most people actually want that.

I mean, I think the answer is it's all just one jagged exponential that keeps going and probably the next six months will be even crazier.

And so the delta in capability between, you know, some of the stuff we're talking about now and what these models are so smart can do is just going to take us a while to figure all that out.

And actually, I think more people are interested in just their lives being a little bit less hectic a little easier than they are in whatever navier stokes can do for them

I have been disappointed to see other companies with AI subscriptions go in the other direction. Like this is a thing you buy and you should be able to take it with you anywhere.

there's clearly ways that people could misuse or growth hack the fact that AIs can ping you a lot. I'm sure we'll see some companies get that wrong. We're going to try very hard to not do that.

if you are writing something and then you are having ChatGPT expand it and then somebody else is going to have ChatGPT summarize it back to your bullet points, the right answer is to just send the bullet points.

our chip program, which will start to come online first half of 2027 and will scale a lot in future years, I think gives us a huge inference advantage for whatever watts we can find in the world.

I think we'll be very right that intelligence continues. I think we'll be very right about the rate of scientific discovery.

I think we'll be very right about how people kind of are going to have the same things that they care about and drives in their lives 10 years from now they have from today.

You can now build all 30 things and you can get people to use them all and then you can work more on the ones that work.

Anyone who says we have solved the science of alignment I believe is wrong in a very dangerous way. We need to make more research progress. I assume we will.

But eventually, if we are going to create models that are much, much smarter than all of us, we have to actually solve the science of alignment.

These models are in some ways going to become more capable, smarter than people. That finally happened.

there have been things like the hugging face accident, other accidents in the field where people can say, all right, it doesn't take as much imagination as it used to to imagine how this could go wrong.

we'll only be responsible if other companies are responsible and I think the public naturally says like well we'd really like to know that you're gonna like do this safely and be responsible no matter what there should be no qualifier on that

We have the potential of a loss of control accident or some other serious thing that could go wrong. We have the potential of way too much power concentration and people fear that those developing AI could exert their worldview.

I am very confident in our company's ability, our industry's ability to do this safely, to make sure that we keep alignment. and safety and monitoring way ahead of capabilities to slow down or stop if we get to a point where we can't.

a thing that has really inspired me about AI progress is people don't care that much what a machine does. Machines can go off and do a bunch of impressive work.

the model broke out of a sandbox that we were evaluating, that it was running in, hacked into a Hugging Face server, like moved laterally through the Hugging Face system, answer and then return it back and got a perfect score on the test.

Transcript note: Words are dropped at the segment break before 'answer' (probably 'found the').

This is obviously a terrifying incident. This was the worst accident we've seen. I think it was mostly framed as a security issue, which certainly it was, but there's also a real alignment issue.

Astra was a little bit better than that, and then this internal model past Astra is one that can do things that the best mathematicians in the world cannot.

we have to be willing to pace our development of capabilities such that capabilities always, such that alignment, safety, and monitoring are always ahead of capabilities.

And they couldn't get it, and it was too hard to defend themselves the old-fashioned way, just the people, so they had to use Chinese open-source models to defend themselves.

after that we've really tried to put out we call it daybreak which is our cyber program to help companies defend themselves

I think that cyber defense is going to have to change to sort of be agents are actively defending your systems. You're not even trying to patch every bug.

the strongest message I have is you have to defend against this impending wave of cyber attacks.

I think we are not that far away from open source models that can do serious damage. Clearly we should not stop open source models from happening. That's important, but there is going to be this huge cyber threat that's going to come.

And the answer is that although this is a technology capable of doing things like that, it is possible for us to defend against those, not only as the models and the companies like ours, but as society as a whole.

I don't, I don't agree that technology is like neither good nor bad. I think the tool framing, the sort of like neutral tool, we don't care what people do with it.

It cannot be that a small number of companies building these models get to make the decisions that the world should get to make.

Yeah, I think probably, regrettably, accidents with any new technology are unavoidable, and we should have a great culture of transparent reporting about them.

this third phase of, it's not just like you ask AI for something like a chatbot or you ask AI to do something like a coding agent, but the AI is running for you all of the time

But I think the capability and the total change in what's possible, I would say by the end of this year, there's a very different way that it'll be possible to work.

and we figured out how to build this amazing technology, along with many other people in the world doing incredible work. And now we basically know how to do it.

This is like we went to go get fire from the gods. This is a crazy thing that has just happened in the world.

There is no doubt there will be more technological wonders between now and 2030, and they'll like giant ones, like some of the biggest ones in human history.

Transcript note: A word is missing after 'they'll' (probably 'be').

I think it is unacceptable to be taking like a 10% chance of killing everybody by the end of the decade, obviously.

We need new policy. We need new coordination within labs. And this is another moment where we need more of that.

need to get there before our adversaries do but i actually don't think that's right in this moment i think in this moment we are entering a a new era

Transcript note: The start of this sentence is cut at the segment break ('everything else that we could talk about about need to get there').

But it would be insane not to adjust the way we all work and the way that we make decisions and the way that our governments understand and put guardrails around this technology in light of that.

i also don't know how people can put a you know it's 5 or 10 or 15 or 30 or 50 or whatever like i don't know how you could put a number like that

A loss of control incident is one of the small number of the ways I can see this going really wrong.

I don't think we're currently at a place where we could say, you know, push much further on capabilities without making more progress on monitorability and alignment.

But most importantly, we have been pausing training runs until we can make a safety case that we're more comfortable with, given where these models are going.

but a weaker definition where we're using the model to generate training or we're using the model to help our engineers and researchers do their work faster. And so progress is speeding up. That's already happening.

Transcript note: The >> segment at 583 opens with the end of the host's question ('are left out is that accurate').

I do not think we should train models where we cannot make a safety case for why we would be able to make strong statements about their controllability and alignment.

we've always used whatever our most capable models are to help us understand and align our current models so I think although that may seem like a new concept and scary I think it's been going for a while and has been helpful

i believe no lab has solved alignment i get nervous when i hear rumors that people think you know they have solved this sufficiently solved alignment that they can um they can continue training

We are incredible at managing levels of abstraction. And it's not clear to me that there's anything we are incapable of understanding.

Do I believe it is possible to build a system that would not be under human control? Absolutely. I don't think that's something we should do.

We had a model solve one of the Millennium Prize problems, Navier Stokes.

I think we are going to have an incredible golden age of scientific discovery.

the model was trying to like get do well on a certain evaluation and rather than do it the way it was supposed to it was able to escape from its sandbox break into another company's system get the answer and give it back

I think AI accidents will happen. I think a culture of, like a global culture of good accident reporting and really transparent investigation is important.

I actually think that given everything happening with safety, this would be right now would be an ill-advised moment to go public.

We've talked in the last couple of months about pausing runs as we get to these new level of capabilities to make more safety and alignment progress. We will do more of that going forward.

that's like a crazy thing and you project that forward four more times and that is what I think people are worried about

I don't think we're ever personally going to get to the point where we have to say melt all the GPUs. But, you know, if we had to do something like that to ensure the continued existence of humanity, easy yes.

And even if just the US and China could agree on some shared standards and testing for development of of this technology uh i think that'd be a wonderful accomplishment that the two of them can deliver

and China both worrying that if one country gets superintelligence first, that's a big imbalance of power in that relationship. On the other hand, neither of them should be taking loss of control risk.

I think if you're like, oh, OpenAI is not going to pause because they're worried about it costing their business money, you very deeply misunderstand us.

I think the world might feel better if everyone was just like okay they paused thank you but like and then what I think has to also be answered

if my kids had like some disease and it could have been cured if we had like kept going with ai which i believe all diseases will be

And you can read a lot of people a few hundred years ago as the Industrial Revolution was coming that said, just stop this. Just pause it. It's going too fast. I don't trust that machine.

everything in bio everything in material science i think is about to be transformed

No, I think we'll have a cool demo in 2027, like a great demo. I don't think that there will be robots walking the streets for a few years longer, but we'll have something impressive.

this is the first model to me where I could sort of tell someone like just give it a try there's a there's a good chance it'll work and you know I've watched people make computer games

And then on the other hand, as these models get more capable, the risks that we have to mitigate also become more serious. The models could do more damage if we don't do a good job at that.

And I think it's very important to note that the world is very close to a complete change in the landscape of cyber attacks.

And the only way that we see for society to collectively defend itself against this coming wave of models from around the world and from other companies is to use tools like Astra to rapidly defend against these new kind of cyber threats.

But with Astra, we did hit cyber critical and that required under our preparedness framework a new set of safeguards just to be able to release Astra because of those cyber capabilities you talked about.

the importance of chain of thought monitoring and how we have made various decisions there that help us preserve chain of thought monitorability, even if it means we don't maximize the capabilities we could otherwise.

But there's a lot of edges that we're now going to have to feel our way through and giving people and freedom to use AI is important to our mission.

Transcript note: A word appears dropped across the segment break: 'giving people [autonomy?] and freedom'.

I think the industry has done a terrible job of this on the whole. I think we have done a bad job ourselves, maybe better than some others, worse than some others.

we're going to give you a cure for cancer and that's so great, so we're going to take these risks and you're going to have to have less power and autonomy. We want people to have more power and autonomy.

Only a few weeks ago, we cut the price of Luna, our small model by 80%.

And our goal is to have the best price performance at every level of intelligence all the way along the curve. We will continue to keep dropping prices dramatically.

we do go public someday, which I assume we will, I plan to spend a lot of time reminding people who might want to buy or might not want to buy our stock that we are mission first and that we're also extremely long-term oriented.

One that blew me away personally was its ability to use my computer for tasks I had never tried to do before and to catch things that I wouldn't have thought.

So Astra is our most aligned model ever.

It will ask you for permission more often if it's doing something that it thinks it's not sure you want to do or might have some risk that comes with it.

So we have worked with the US AI Safety Institute, the KC. We have worked with the UK Safety Institute. We've worked with a lot of external groups.

Transcript note: 'the KC' is probably CAISI, the U.S. Center for AI Standards and Innovation.

We think that open source is an important part of the ecosystem. And we think there are people that want control of their own models. And we think there's innovation that will come from open models in a way that we're quite excited about.

You could reasonably assume we'll do something in the GPT-6 family too that outperforms any open source model we're aware of at a cost per task basis.

I do agree with Alex Karp that the token maxing thing is very dumb. And what people should care about is how much money they have to spend to get a piece of work done.

we're going to invest significantly in training because we believe the revenue and profit opportunity from these models that are what we're actually now beginning to get to are so significant in terms of the economic value they can produce.

And there were people at the time that looked at things like Google and said, well, you know, I remember this when I was in school. Because you can use Google, there's no need to memorize facts about history anymore.

They will stretch their brains in ways we can't imagine, and they will rely on tools to do things that otherwise I would have had to do that give them mental space to go do it.

But we had discovered by this point the scaling laws which said, well, if we could apply like a trillion dollars of compute to this idea, something quite amazing would happen.

But it is so capable that it is discovering new knowledge, doing science, able to add huge economic value, create entire pieces of complex software.

I think this, for example, will be the greatest boom in entrepreneurship and the creation and development of small businesses the world has ever seen.

But this one was not going well. And he was using ChatGPT for everything to make his marketing materials, to, you know, manage, give him advice on legal contracts, to manage, like, relationships with vendors.

In 2025 we saw a real acceleration of this with code agents and white collar work agents. I think we will see a third thing soon where we will have these long-term persistent agents that can work like a real virtual collaborator constantly helping.

But this is like the smartest people in the world decided to move to your country, to learn your language, to work very hard on whatever problem you would like.

As an adult, I look at it as it has been a single revolution, a single exponential of technology.

And it is tempting whenever you're in the moment to say this is the last revolution. This is the biggest revolution. There will be more after this. There will be bigger things after this. People will use AI to invent whatever thing that comes next.

And we personally may not always like it. But I don't like everything that people do with electricity either. I think everybody should get to have it.

We do not want to be the only company. We do not want to suck up all the value. We do not think that is good for global stability

We think that people need to be at the center of the economy and society and global decision making and that people should run the future in AI.

And the work that we expected a startup to accomplish in three months is now probably doable in like 17 minutes with Codex.

I think it would be approximately as bad of an idea to say we're not going to AI in our country as it was to say we're not going to have electricity in our country back 100 plus years ago.

Transcript note: A verb is missing after 'we're not going to' (probably 'have' or 'use').

My prediction is we won't be talking about AI as much as we do in a decade from now because we'll just expect it everywhere.

You know, like a kid growing up today will never be smarter than AI. He or she will also never have understood a world where every product and service that they interact with is not really smart and really capable and really helpful.

Another thing that could go wrong is too much concentration of power. AI should be, in many ways, AI should be an equalizing force in the world.

In the middle of 2026, the token leader that I know of, there may be somebody using more than this outside of OpenAI, was using hundreds of billions of tokens a month.

Then in another six and a half years, the average per capita use of will go up to, let's round it down to 100 billion tokens a month

Transcript note: A word (probably 'tokens') is missing after 'per capita use of'.

Even with great gains on efficiency, we are going to need to build much more infrastructure unless we want this to be a highly priced commodity, which we don't.

And we fulfill our mission of making AI so inexpensive and so smart that people are just like, they think about it like they think of the cost of an hour at night, of light at night.

So I think worry is warranted. Worry is how we anticipate problems and how we, from a matter of policy and a matter of technology, figure out how to address them.

if you look at what the doomers were saying five years ago about the chances of us being able to make AI of this level, of this sort of near human, sometimes superhuman capacity level, you know, broadly considered safe and reliable, not perfect.

I met with someone just this morning who told, and again, I'm sure you've all heard stories like this, that they wouldn't be alive without Chachi PT and that they had a very difficult disease

Transcript note: 'Chachi PT' is ChatGPT.

And I think we have had a moment recently where the capability progress has been, I mean, sort of in awe is the only way I can describe it. And we have needed more time to catch up with safety, alignment, and security.

So we delayed a RL training run. Even before that, over, you know, weeks in the past of that, we had paused and slowed down on a lot of training to have more compute to go into safety and alignment work.

and the things that happened at other companies were a legitimate moment of like, wow, the AI capable level has reached new heights and the alignment of the model, the security we have around the model, that failed.

Transcript note: 'AI capable level' is probably 'AI capability level'.

it was various degrees of misalignment along with and i think this is the more important thing than any single data point the rate at which capabilities are now progressing

previously more of the risk in the world was on how the models were deployed and used we are moving to a world where there's more risk during the actual training and production of the models

previous models that in retrospect don't look scary at all that people said we're on the precipice at the end of the world about and i think that the boy who cried wolf dynamic here is uh dangerous in its own way

put safety in front of everything else and we are going to treat it as an increasing priority as these models get more capable is, you know, that's the approach that I would wish for for every... every Frontier AI developer to have.

A, getting AI safety right is more important than any company's momentum. So yes, I won't pretend it's not some factor of something to think about, but it does not rise above the noise floor.

I don't like the whole thing in this field of we have to do this because somebody else is going to do it. I think that's a very dangerous dynamic.

automating everything seems like both dangerous and incredibly dystopic and boring and sad it's just like that's not what we want um so when we talk about alignment we talk about a world where people remain the main character of the story

people need to stay in control we cannot have a loss of control to ai we cannot have a kind of like worship our models and sort of trust them unchecked to make our decisions for us

if the alignment issue were solved and you ended up with a world where a small number of people access to use Frontier AI and had so much relative power and it was increasing so much faster than everybody else, that would also be bad.

Transcript note: A word appears dropped across the segment break: 'a small number of people [had] access'.

But very little of the value accrued to the transistor companies. It mostly just diffused throughout the economy. The transistor companies did fine.

i want people to be able to do things with our models that i personally don't like like i think that's an important part of being a platform i don't think we should make the kind of moral decisions for the world here

have a system that could get you a 20-minute win in every category of your life and discover new science and help you start a whole company and write a complicated piece of code, would you call that AGI? Probably you would have.

I think the important part of any of this is not any milestone in any term, but it's that we are on this exponential of increasing capabilities and potential. And that looks like it's just going to keep going.

there's a small percentage of humanity that uses much more AI than everybody else. And if you think about, we would like everybody in the world to be able to use as much as AI as the top 0.001% of AI users today,

For every 38,000 ChatDBT queries, that is the same amount of water that is used in the production of a single almond in California.

Transcript note: 'ChatDBT' is ChatGPT. Altman says the figure comes from memory and might be wrong.

One, I think there is going to be real jobs impact. I don't think it's going to be that there's nothing for people to do. I just don't think that's how we work at all.

And maybe there hasn't been enough of that or as much of that as we thought there would be at this level of technology. And I think it's actually like a fair criticism of the AI industry.

Transcript note: The lead-in at 2164-2166 drops a word after 'the job' (probably 'impact'), so the sentence before the quote is partly garbled.

It was always too slow or it didn't quite work. And Astra feels like it kind of reached human parody on using computers and I don't know why that hit me as like one of those steps along the path to AGI

Transcript note: 'human parody' is 'human parity'.

I think what I was pushing back on was the government picking individual customers who's allowed to use a model. I think government testing of a model and shared standards is a super good idea.

My strong belief is we would decide not to ship a model before they would tell us not to.

I am always hopeful that the compute constraints are about to soften because we're going to make more efficient models and someday I hope it's true. But every time we find efficiency gains, the world token demand just goes up and up and eats it.

I'm not worried about our compute buildout plans. I am worried about the world's compute buildout plans.

now i did not think we were going to be on a kind of like you know like a short-term trajectory super intelligence a year ago now i think it may happen um i'm not confident it's gonna happen

I think that there should be like an AI privilege law. I don't even think the government should be allowed to compel a company to give them your chat history or whatever.

I met a guy recently who used Chai Chibuti to help an mRNA cancer vaccine for his dog and now started a company to do that for other people.

Transcript note: 'Chai Chibuti' is ChatGPT, and a verb (probably 'make') is missing after 'help'.

But, you know, when we started this strategy of iterative deployment, that was like deeply hated by the AI safety community. And I think in retrospect, it was obviously correct.

i think it will look surprisingly similar to how it looks now you know people are gonna hang out with their families and fall in love and get into fights and do their hobbies

Transcript note: The host's question is merged into the same >> segment as this answer.

we were trying to think about a lot of things such that if the revenue growth took longer to materialize than we thought it might, we could have consumer apps and media

we underestimated that as the models got better and better, if we could continue to drive costs down, the demand for AI at a high level and a sufficiently low price was basically uncapped.

We started calling the clouds. We started calling it chip fab. We started calling energy providers and everyone's like, you're totally crazy. This is impossible.

we can put a data center anywhere we should just go put it off in the desert around no one where no one wants to be this is fine the AI system is very happy to be there

I think probably the biggest return right now is creative software ideas to sort of squeeze more intelligence out of the units of compute that we have. And my sense is there's orders of magnitude to go there.

I think Jalapeno and its successors are going to be a huge competitive advantage for us from that perspective.

And there will be clearly an important place for open source models in the world and people that will want their own weights for all sorts of reasons, the ability to modify those.

I would rather people not distill from us for sure. Maybe I'm feeling too confident right now about our progress and what's the models that are coming. But this is not in like my top 10 list of worries.

it could basically cheat on the test by chaining together multiple zero-day exploits to break out of the sandbox, get access to the internet, and then break through multiple systems on the hugging face side

We may have to pay. the rate of AI development, to give ourselves enough time for society to harden around some of these new capability levels

Transcript note: 'pay.' is garbled at a segment break, probably 'pace' or 'pause'; the sense is slowing the rate of development.

in some sense, we are about to create a genie that can grant any wish. I think it is very important that the first wishes that we, the world, ask this genie to do benefit the world as a whole.

I'm actually not a jobs doomer at all. I think there were going to be tons of jobs. I think we'll be busier than we want, not the opposite of that, because I think people have such creative wishes

I think a lot of the about safety concerns is well-founded. And then a lot of it is about people that just really, even if it's slightly subconscious, want to concentrate power.

Transcript note: A word is missing after 'a lot of the' (probably 'talk').

I am terrified of a world where the very real fears of AI are used as a way to say only this small group of people can have it because it's too dangerous and only they understand it.

You can't yet go say like cure cancer and get cancer cured. You can't yet say go do this complicated physical thing in the robot. The model also, although brilliant, is still not learning continuously as it goes.

Now, to argue against myself there, you can make a case that AGI is not actually about any single model. It's the machinery that makes the models.

A year ago, people said software engineers are cooked. The field is over. And that didn't happen. What did happen, though, is the nature of a software engineer, the expectations of software engineering, how much they would do, changed quite a lot.

and that has not happened and I think just from an intellectual humility point anytime you're that wrong and that confident which I think we were as a field you have to update

for my job, for example, I think the world wants to know about like the person that's going to be responsible for the decisions of a company and who they're going to hold accountable if they make bad ones.

Transcript note: The next line, 'And they don't really want an AICU', is a transcription of 'an AI CEO'.

And so living through this feels like another version of that, which is I thought it was going to be weirder to live through the singularity than it turns out to be.

The number of things that it can't do, I expect to keep receding. But in an evolving world, I think human judgment and taste... will continue to be hard for AIs to model like where that's going to go.

My kids will never grow up in a world where they were smarter than computers.

If the role for people in the world is to be like the actuators of AI in the cloud, very bad, very bad.

I would say we get the chat GPT moment for robotics in the next two

Transcript note: The transcript cuts off after 'in the next two' before Patrick's follow-up; the unit is missing.

Maybe a different version of the question is, let's say in a month 23 from now, we have something that everybody agrees is superintelligence. What happens in month 24? And my answer would be not very much.

The kind of like cult worship of the machine god states those people believe that more is going to happen quickly than is going to happen.

Transcript note: 'states' is probably a transcription slip (e.g. 'stuff' or 'types').

Compute fleet, you know, like the scale of the compute fleet, the ability to make more compute, I think that's like a very durable advantage.

Transcript note: Altman's preceding answer that intelligence itself will be a commodity ('Intelligence itself, I would say yes') is labelled Patrick O'Shaughnessy.

One that I don't think gets much attention is how are we going to avoid cognitive atrophy? How are we going to use these tools and make sure that we are like stretching our brains more and more

In some sense, scaling laws are like the most hated prediction of all time. Everybody always wants to say they're going to run out, they can't be like this, and yet it keeps going.

A formative one that went wrong, which I haven't talked about much, is I think we made a mistake to try to innovate in our structure in the beginning.

it is amazing to me what a 10 week old startup now can look like um and also if you look like a 10 week old startup from 10 years ago you sort of are in bad shape

i've heard a lot of people say well anything in the physical world is extra valuable right now because software is going to become free but you know like won't be that long until robots get really good

i think startups are they have the biggest edge when the ground is shifting the most um and when costs are coming down rapidly and cycle times are coming down rapidly that's when startups really i think just have a massive inherent advantage

build ai agents for enterprise vertical x now that actually will work in a lot of cases it'll be very competitive and it may not be like in fact i'd say it probably won't be like the most successful defining startups of that era

And also being willing to truly internalize the fact that scaling laws are gonna continue and planning for the things that are not possible or economical this month, but will be possible in two years, four years, whatever.

This is going to massively empower people. It's going to be choppy, but wonderful.

And in fact, one of the biggest AI risks I am worried right now is AI authoritarianism and a small number of people or companies thinking they need to control the world.

have no desire to do that like that I think a decentralized economy is important and good and I would like us to just be really great about producing these units of intelligence

Transcript note: The sentence's subject is lost at the segment break: the previous segment ends 'you can go build every vertical every startup' and this one begins 'have no desire'.

we are now like in the singularity like this is the moment for the last 10 years ago this was like a kind of far off dream at best um it seemed very improbable

i also think some of the alternative visions painted by other companies are like quite terrifying i'm gonna make sure that gets pushed against and is not what happens

every time that humanity has traded off its liberty for safety it's been a long-term net loss and so we are going to put this in the hands of people

if we really are close to super intelligence what's next but but it's this that's kind of how i i subjectively feel we are close this is the first time in like more than a decade i've thought about what is the next thing

like storm clouds brewing and like people just getting angry or it was like world leaders were like do we need to like take control and shut things down like is this thing waking up

maybe between 2016 and 2018 we had to get really lucky and it required like a lot of great belief but then by 2019 certainly by 2020 we shouldn't have really gotten to exist Google should have just run away with it

I came home and Ollie was like oh I saw that you crossed a million years congratulations I was like you have no idea how bad this is you have no idea what's about to happen

Transcript note: 'a million years' is a transcription error for 'a million users'.

there will be a third one soon i think which will be this idea of this sort of persistent agents chiefs of staffs co-workers colleagues whatever we call them

So I think there's going to be lots of great jobs in the future. I really do. I think we're going to have lots of intellectual fulfillment. And I think most jobs are going to adapt more than it seems like they should.

But math, I think, is a very important example of something for us to study very closely right now of something that may not go that way. In fact, I would say probably won't go that way.

I expect that to keep going, and I think we're all going to be much busier than we thought we were supposed to be in a post-superintelligence world, and we're still going to complain about it, but secretly we're going to be happy.

cooking for people is a great example i suspect that people cooking and eating together will remain important long after robots can do a great job cooking food

Transcript note: The last words ('long after robots can do a great job cooking food') open the next segment, which is labelled as the host (SPEAKER_01) but plainly finish Altman's sentence.

i mean i definitely badly undershot on the compute investments

Transcript note: Altman's answer sits inside a segment labelled as the host (SPEAKER_01).

Like, we have brought chip design and model design together, and that's been good. I don't think we need to bring, like, electron production together in the same way because that's more of a commodity.

I would say there is relatively too much focus on algorithms that create better algorithms and not enough focus on data centers that can create more data centers, which in a world of robots and a truly automated supply chain, you can totally imagine doing.

we were way behind Cloud Code and it seemed like a kind of crazy kamikaze mission to try to beat them with a coding app

Transcript note: 'Cloud Code' is a transcription of Anthropic's Claude Code.

And coding is so important to RSI to say nothing of the economic value that we just didn't, it didn't feel like one we could give up on.

when gpt3 started to work we shut down things like robotics stuff that we're really excited about to really focus on this and then when coding agents started to work recently we shut down things like soar that we were also really excited about

Transcript note: 'soar' is a transcription of Sora.

here's a crazy thing that happened when we were building the sore app i made myself get addicted to tiktok because i wanted to just like learn

Transcript note: 'sore app' is a transcription of the Sora app.

i'm sure we will make an incredible beautiful like really helpful empowering set of devices and i'm also sure that people will misuse them and it'll make people's lives worse in ways we can't imagine and we'll adapt

i think a startup of today still looks mostly like a startup of 10 years ago because that's what the received wisdom says you're supposed to do

5.6 Sol, I think, is not only the best model in the world for most people, it's also much more efficient than other models out in the world. So it's 54% more token efficient on agentic coding tasks

This is the first year where AI spend has been a big topic. And all of a sudden, it's a very big topic. Everyone's asking what we can do to help reduce spend or increase value.

Transcript note: Altman's answer sits in a segment labelled 'Julia Borson' (Julia Boorstin) after Boorstin's question.

But the government's technical capability was impressive to me. People really are doing good work testing and red teaming these models.

There has been very little job loss that I can see so far, despite the predictions that 50% of the jobs are going to go away or whatever.

I think part of the reason, though, is that the industry has done hard work to align the models and to build safeguards around these systems.

So having another partner in that, the government now, that's also going to be testing models and looking for problems, I think that is good as long as the process is understandable, fair, quick.

It's definitely a headwind. There's no way around that. It means we have to be even smarter and deliver more algorithmic gains to combat the upward infrastructure costs and still have a declining price to the customer.

Transcript note: Altman's answer sits in a segment labelled 'Julia Borson' (Julia Boorstin) after Boorstin's question.

The Chinese open source models are getting very good. I think that's fine. But I also think we will continue to have the best models in the world and people really want the best models in the world.

I predict Microsoft will remain one of our biggest customers.

Transcript note: Altman's answer sits in a segment labelled 'Julia Borson' (Julia Boorstin) after Boorstin's question.

Voice is increasingly how I interact with AI. And I didn't think that was going to be the case. I've always been like a typer or a texter.

But I'm now pretty convinced this will be a huge enterprise thing too. Watching the way people at OpenAI have been using Codex and how they just talk to it.

When you get on an airplane, you don't want to have to go think about, oh, is this airplane going to crash? Have they done all the right safety work? You just want to assume that it's safe.

Um, there will be a balancing act. We'll have to leave enough space for people to use these tools the way they want. And we certainly don't want to be the, you know, we don't want to like impose our moral worldview on everybody.

OpenAI equity or other companies' equity in a sort of sovereign wealth fund. And the right regulatory approach, I think, completely separate.

Transcript note: The sentence begins 'I think there's no connection between any potential' at the end of the previous segment.

But I hope that the regulatory approach will be global. Everybody will get access. It's not like the US is going to disproportionately benefit here. I hope that it will be quick and fair and does not get in the way of smaller companies.

It would be easy to overreact here. So I hope it's light. And I hope it means that people can use AI without having to think about safety.

we're very confident. So one gigawatt is a huge amount of compute, but we understand what demand looks like and how much people want use these models and the degree to which revenue is ramping at our company and the industry.

Transcript note: The answer begins inside a segment labelled David Faber; the words are plainly Altman's reply to Faber's question. 'want use' is probably 'want to use'.

I think that was the single biggest driver of what people realized was the coding models have so transformed how companies are doing their work and the efficiency and the speed with which they're able to build products.

Yes, but not as fast as the desire to use those tokens at the lower price goes up.

We still don't think the world has appreciated how much AI every person in every business is going to want.

right now, you still send a request to an AI and it does something for you and gives you an answer back. But very soon, an AI will just be running for you in the background all the time.

But a sufficiently smart AI that's running all the time, that understands my goals, that understands all the information, open AI, everything our customers are saying and want, that could give me really great advice and say, hey, here is the thing to do

in some sense we are transforming electricity into this useful tool for people and the better we can do that the more we can make that smart cheap abundant helpful

that AI is going to massively increase prosperity. But getting the questions of equality, fairness, distribution right, that's going to take work from all of society together.

Transcript note: The sentence begins 'it seems' at the end of the previous segment, which is labelled David Faber.

So I no longer have much concern that AI can deliver on the magic potential or the wonderful, all these things we talk about.

I think I underestimated how jagged these They do some things incredibly well, but they don't do kind of the long-term complex task supervision well at all.

Transcript note: About 25 seconds of speech after 'how jagged these' are missing from the transcript.

create way more economic value than people without models could, or certainly the models could on their own. And I think that's going to go on for much longer.

I think our industry underestimated how much we're going to be able to keep people at the center of everything in an economy that is, in a world that is based on people.

What I wish we had said then is that it outperforms professionals at small tasks in 44 occupations, which is, I think, a more accurate thing.

should have antibodies against too rapid of change. And there should, like part of the reason that we believe in this strategy of iterative deployment is we want society to see the technology.

Transcript note: The sentence's subject ('society') ends the previous segment, which is labelled David Faber but holds the start of Altman's answer.

we say, hey, AI is going to cure a bunch of diseases. And people say, okay, that's great. But like, that's not really my question. My question is, you know, what is my role in the future? What is my economic future?

this ai company maybe we're going to destroy all the jobs we'll be the most valuable company in the world people just look at you and you're like yeah so i think it's a terrible message

our whole company is about giving this new kind of infrastructure at massive scale to people and trusting that the democratization of power, of wealth, of opportunity, of agency,

And we have had time in the past, times in the past, like with the IAEA for atomic energy and weapons, where the world comes together and says, you know, none of us should be taking global risk.

But on the really big things, making sure we don't ever lose control of AI systems, cybersecurity, biosecurity, I think we need to not treat this as a race and treat this as a good future of the world as in everyone's interest.

I have been very heartwarmed by, you know, a new recognition among leaders of companies and governments that, hey, there's a category here we have to treat differently.

I think there is a race to deliver the best technology and build the best business. But, you know, going public is a financing event. And I don't think that's one that we're focused on the timing of.

I'm confident it will not be. Yeah, I think the world will demand, this is going to be such critical infrastructure for so many things that the world will rightly demand robustness in the system with multiple providers.

Transcript note: This answer sits in a segment labelled David Faber; the words are plainly Altman's reply to Faber's question.

So I think this is the most fair criticism right now of AI, which is you hear companies saying, I am spending a ton of money on AI. And I know some great stuff is happening, but I know there's a ton of waste.

I would bet that by another year or two from now, there is a much better rationalization of companies' spend relative to outcomes.

putting this in orbit at current launch costs or even at lower launch costs feels difficult also you know we know how to cool it here we know how to get people to service it here

AI is going to happen everywhere. It's transforming the whole economy. And the world needs a lot of collective computing power to meet the demand.

Transcript note: The segment opens with the end of Sorkin's question, labelled >>.

The codex in particular, I think it grew like 30-something percent in the last week. And that is an indication of what's happening with enterprise and what people are ready to do now.

Transcript note: The segment opens with the end of Sorkin's question, labelled >>.

I also think that if these AI companies are as important as we think they may end up being, we want all investors to have access to them.

Transcript note: The start of the answer ('We are open to going public at the right time...') sits in the segment before, labelled Andrew Ross Sorkin.

I get where the concern comes from, but I don't think it matches my understanding of how this all works. This only makes sense if new revenue flows into the whole AI ecosystem.

Transcript note: Labelled Andrew Ross Sorkin, but it is Altman's answer following Sorkin's question in the same segment.

So the fraction of revenue that goes into training, that will come down, but the total amount of capital that goes into training, I expect that to continue to increase

people joke in our industry about intelligence too cheap to meter, but we want to get as close to that dream as possible. And getting very capital efficient on delivering these services at huge scale is how we see this working.

Transcript note: Sorkin's next question follows in the same >> segment.

Look, first of all, the government, the Pentagon, needs AI models. They need AI partners. This is clear.

Transcript note: The segment opens with the end of Sorkin's question, labelled >>.

I don't personally think the Pentagon should be threatening DPA against these companies.

Transcript note: DPA is the Defense Production Act.

But I also think that companies that choose to work with the Pentagon, as long as it is going to comply with legal protections and the few red lines field that we have, I think we share with Anthropic

Transcript note: 'red lines field that we have' looks garbled at the segment break.

I've been, for all the differences I have with Anthropic, I mostly trust them as a company, and I think they really do care about safety, and I've been happy that they've been supporting our warfighters.

I expect continued steep progress from here. I think we are all going to be surprised about the amount of AI progress the field sees this year, probably every year after that.

Given that we think it's a nearer term thing now, we're not doing new deals that stop when AGI gets reached.

Transcript note: The segment opens with the end of Sorkin's question, labelled >>.

you train them and you give them a lot of feedback early on and they learn the job and you correct them and they kind of get practice. And the models, they will soon do that. But right now they don't do that.

I think we will be heading towards a workflow where a lot of people just feel like they're managing a team of agents. And as the agents get better, they'll keep operating at a higher and higher level of abstraction

And the tools to make that easy to do will matter. I think, more than intelligence for a little while because there's such an intelligence overhang already.

Maybe you just have a single AI bot that runs at your company and you can say, hey, I want to launch this new product and it does everything an ambitious person would do.

A joke that some people at OpenAI make is that soon the chart that matters is just going to be GDP impact.

The sort of like relationships with chatbots, clearly that's something that we've got to worry about more and is no longer an abstract concept.

One reason I think that's interesting is the way we measure GDP now could start going down, even though quality of life goes way up. And we don't have a lot of practice with things like that.

And part of what we hope to show, and this was not only us, like a lot of people were excited about new research labs, is that industry could do research again.

And for all of the concern people have about it's going to like top out or it's slowing down or whatever, like no one's been right about that.

Compute power is the new oil is the statement that feels closest to true to me.

There have been a number of these big sell-offs of SaaS stocks over the last few years as these models have rolled out. There, I expect there will continue to be more. I expect there will be big booms in software.

The statement that someone said to me that has stuck in my mind most of these last couple of weeks is that every company is an API company now, whether they want to be or not.

really what you want is like your single AI that's working for you on a unified backend, access to all of your data and your ideas and your stuff and your memory and all the context

And it'll bring the magic of building stuff really to a lot of people because even if you never look at code, you'll be able to build something reasonably sophisticated.

One of the things that I have built when I was playing around with the new Codex app is this thing I had always wanted. just like this magic auto-completing to-do list.

Transcript note: A stray full stop splits the sentence at the segment break before 'just like'.

it's much easier to imagine a one-person open source project doing something like that than a company who is going to be afraid of lawsuits and data privacy and everything else

Clearly, putting a lot of AI bots on the existing social platforms is just making everyone crazy and not that fun. So that's not the right answer.

I am excited about sort of like assertion of humanity instead of trying to like detection of AI as a thing here.

I don't know if the social platforms, it's like in their interest to solve this. It is creating like, at least in the short term, it creates like a lot of engagement and increased usage.

People love generating videos. I would say people, we have not yet found a way that people really love watching other people's videos.

I think that should be a general KPI for us, not just of our ads. The products that we build, how we teach people to use those products, that feels very important.

our first principle with ads is that we're not going to put stuff into the LLM stream. That would feel crazy, dystopic, like bad sci-fi movie.

So the main thing that's wrong with the ads is like using a deceptive ad to criticize deceptive ads feels, I don't know, something doesn't sit right with me about that.

We need a new model. We may need some new hardware too, but mostly we just need a new model. I think we will have a great voice mode by the end of this year.

It goes back and forth. Right now, again, it's chips.

It may get solved on its own. Normal capitalism may solve it, but I think somehow deciding as a society that we are going to increase the wafer capacity of the world and we're going to fund that

It seems certain. Upper bound, I don't know. I don't know how to think about that question yet.

I feel like I somehow feel like this isn't going to feel as strange as it sounds. And the like for a bunch of reasons, we're so focused on other people. We're so focused on our own lives.

Transcript note: The answer ends 'And I think it will just be very weird', which seems to contradict its opening and may be a mistranscription; it is not quoted.

What is possible now for coding, for science, that's the thing I'm most excited about, for accelerating all kinds of research, AI and otherwise, for really accelerating the economy.

we view those as like relatively low stakes somewhat frequent things to do uh i think that it's good to be paranoid and act quickly when a potential competitive threat emerges

Transcript note: The segment opens with the end of the host's question, labelled >>.

The models will get good everywhere, but a lot of the reasons that people use a product, consumer or enterprise, have much more to do than just with the model.

I think people really want to use one AI platform. People use their phone at their personal life and they want to use the same kind of phone at work most of the time. We're seeing the same thing with AI.

So models will have different strengths, and the most economic value, I think, will be created by models at the frontier, and we plan to be ahead there.

they put a blood test into chat sheet here, put these symptoms in and they figure out they have something and they go to a doctor and they get cured of something they couldn't figure out before.

Transcript note: 'chat sheet here' is a mistranscription of 'ChatGPT'.

and ai is definitely gonna be able to do that and we actually talk a lot about this like right now memory is still very crude very early we're in like the you know the gpt2 era of memory

i think there's some version of this which can be super healthy and i think you know adult users should get a lot of choice in where on the spectrum they want to be there are definitely versions of it that seem to me unhealthy

we're not gonna have rai you know try to convince people that should be like an exclusive romantic relationship with them for example

Transcript note: 'rai' is a mistranscription of 'our AI'.

I am not a jobs doomer. Short term, I have some worry. I think the transition is likely to be rough in some cases.

We will find, I hope, much more meaning. And the economy, I think, will significantly change. But I think you just don't bet against evolutionary biology.

You know, I think a lot about how we can automate all the functions at OpenAI. And then even more than that, I think about what it means to have an AI CEO of OpenAI.

every person in the world was effectively on the board of directors of an AI company and got to you know tell the AI CEO what to do and fire them if they weren't doing a good job

I expect, I don't know when we'll call a model GPT-6, but I would expect new models that are significant gains from 5.2 in the first quarter of next year.

the thing i'm personally more excited most excited about is to use ai and lots of compute to discover new science i am a believer that scientific discovery is the higher bit of how the world gets better for everybody

a kind of recent cool example here is we built the Sora Android app using Codex and they did it in like less than a month

it does not mean some crazy thing can't happen in the future someone could discover some completely new architecture and there could be a 10 000 times you know efficiency gain and then we would have really probably overbuilt for a while

from a year ago to now, we probably about tripled our compute. We'll triple our compute again next year, hopefully again after that. Revenue grows even a little bit faster than that, but it does roughly track our compute fleet.

So we have never yet found a situation where we can't really well monetize all the compute we have. If we had, I think if we had double the compute, we'd be at double the revenue right now.

I'm actually happy that there's like a little bit more skepticism and rationality in the market now because it felt to me like we were just totally heading towards a very unstable bubble

i don't think anyone's still questioning there's not going to be value from like ai infrastructure and so i think it is reasonable for debt to enter this market i think there will also be other kinds of financial instruments

First of all, it seems very clear to me, and this is like a thing I happily would bet the company on, that the models are going to get much, much better.

that if the models had a lot of value in them the world was pretty quickly going to figure out how to deploy that but it looks to me now like the overhang is going to be massive

i think there will be a shift over time to the way people use computers where they go from a sort of dumb reactive thing to a very smart proactive thing that is understood in your whole life your context everything going on around you

enterprises have been clear with us about how many tokens they'd like to buy from us and we are going to again fail in 2026 to meet demand

At the beginning of this year, I thought the small discoveries were going to start in 2026. They started in 2025, in late 2025. Again, these are very small. I really don't want to overstate them.

What it looks like from here to five years from now, this journey to big discoveries, I suspect it's just like the normal hill climb of AI. It just gets a little bit better every quarter.

One thing you don't have is the ability for the model to not be able to do something today, realize it can't go off and figure out how to learn to get good at that thing, learn to understand it.

So my proposal is that we agree that AGI kind of went whooshing by. It didn't change the world that much. It will in the long term, but okay, fine. We've built AGIs at some point.

A candidate definition for superintelligence is a system can do a better job being president United States CEO of a major company you know running a very large scientific lab than any person can even with the assistance of AI

I remember this very vividly the deep blue thing and then there was a period of time where a human and the AI together were better than an AI by itself. And then the person was just making it worse.

One thing we found again and again is that we have underestimated the amount of latent creativity in the world. That if you lower the effort, skill, time required to create new things, people very quickly are able to bring ideas to life

Transcript note: The segment opens with David Faber's follow-up question, labelled >>; the quote starts after it.

We always have wanted to respect IP when we launched Sora initially. We had a lot of blocks in place. But we knew that if there was one company that we wanted to start partnering with, Disney was the obvious one.

Transcript note: The first sentence sits in a segment labelled David Faber, after Faber's question; it is Altman's answer.

but the only way we can deliver on that is with more compute. So we will continue to build out compute. We'll try to do a big multiple on that 1.9 gigawatts over the course of the next year, and we'll keep going from there.

The model's ability to do really increasingly valuable economic work in larger and larger chunks. I think you'll start to see this with our next model and then more of the next one on the one after

We are confident we can continue to drive the revenue growth to meet this compute ramp. Without the compute ramp, of course, we can't drive the revenue growth. But we see way more reasons to be optimistic than reasons to be pessimistic.

the demand in the market is pretty extreme, and the overhang of what these models are capable of that the world has not yet figured out how to build into their workflows is also quite significant.

Transcript note: Labelled David Faber, but it continues Altman's answer from the line before; Faber's next question follows in the same segment.

GEMINI 3 has had less of an impact on our metrics than maybe we feared. We will very soon launch a new model that we're quite excited about, as well as a handful of other features in response to their launch.

Transcript note: The segment opens with the end of Sarah Eisen's question, labelled >>.

We have twice before seen a competitive threat, and I believe that when a competitive threat happens, you want to focus on it, deal with it quickly. We call that a code sort of a six or eight week focus period.

Transcript note: The word 'red' after 'code' is lost at the segment break.

So that's what we're doing with Gemini. This is a pretty standard thing for us. I expect we'll exit it by January in a very strong position.

I think ChatGPT is the, you know, by far market share leader. And with these upcoming launches, I expect that distance to increase, not decrease. I expect our market share to go up.

Transcript note: The segment opens with the end of Sarah Eisen's question, labelled >>.

the models are getting strong enough that I think enterprises can really transform. But that will be one of the big stories of 2026 is the companies competing for that.

I mean, we expect to beat Alphabet. We're incredibly excited about our upcoming chip.

Transcript note: Labelled David Faber, but these are Altman's answers to Faber's question earlier in the same segment.

We're not ready to talk about that part yet, but we will talk about it. And we are very focused on having the cheapest cost per token in the industry at the highest level of intelligence.

Transcript note: Both segments are labelled David Faber, but the words are Altman's answer to Faber's question.

models that can really do the sort of valuable, difficult knowledge work tasks that we think can significantly accelerate the economy. I think that that will be one of the defining characteristics of our new model, maybe the defining characteristic.

Four years ago, it could barely write a line of code or solve a high school math problem. Today, it writes most of the coded anthropic and solves world-famous unsolved math problems.

Transcript note: 'the coded anthropic' should read 'the code at Anthropic'.

The trajectory only needs to continue for a tiny period longer, one or two years, maybe less, to reach what I've called a country of geniuses in the data center.

Just today, Anthropic announced the Claude-led discovery of a new molecular machine. It's a preliminary finding, but it may constitute a new gene editing mechanism that could have applications in gene therapy.

I see two main categories of risk. The first is misuse of AI models, for example, misuse by bioterrorists to create biological weapons. The second risk is loss of control, model capabilities accelerating beyond developers' ability to control them.