OpenAI

“The reason that obviously you mean the rest of the team you know, created open AI was to help [783] Elon Musk: with the democracy, help, help spread out AI technology [787] Elon Musk: so it doesn't get concentrated in the hands of a few.”

“OpenAI is structured as see a five one C three nonprofit but, you know, many nonprofits do not have a sense of urgency.”

“Whenever I've helped people for no immediate benefit and with no intention of ever getting a benefit at all time and again in my career, it has really Later, benefited me a lot.”

“So in terms of robotics, we are working on manipulation. We think that manipulation is the complete, it's the one of the parts of robotics, which is the most unresolved.”

“In total we gather an investment of one billion dollar in the group.”

“So OpenAI focuses on building AI for the good of humanity. We are a group of researchers and engineers collaborating together who essentially try to figure out what are the missing pieces of artificial, of general artificial intelligence and how to build it in…”

“So someone like that can actually become productive from day one.”

“We sometimes actually kind of sacrifice good engineering habits, good, good kind of code modularity to make our code shorter and simpler and kind of having less, essentially less lines where you can make bugs.”

“So I leave, there is nothing, I come back, there is this reinforcement learning bot, and actually, it's beating our scripted bot after, like a week worth of engineering effort. Possibly it was two weeks, but it was something very miniature compared to the deve…”

“Because the thing is, if you have a new idea for, okay, well, I've kind of got this thing working, and now I'm going to try something slightly different. As you're doing the new thing, well, machine learning is, to some extent, very binary. At the start, it ju…”

“You actually try to make that core be as small as possible because machine learning is really hard. It's a lot of compute. It's really hard to tell what's going on there. And so you want it to be as simple as possible, but then you surround it by as much engin…”

“Well, I, I'd say very, very specifically that kind of the root cause here was that he had gone for an item, an early wand build. And we had just never done item early wand build. And so it's just like our bot had just never seen this particular item build befo…”

“Initially the way that the project management was happening was that each week, well, so we'd written down our milestones of let's beat this person by this state, let's beat this other person by this state, let's, you know, be able to do, you know, kind of the…”

“There was a point after the event where we set up this big LAN party where we had, like, 50 computers running the bot. We kind of unleashed this swarm of humans to kind of add our bot. And they find, found all the exploits”

“Because actually it's a little known fact, but from day to day, like each version of the experiment was not good enough to beat the next player, next, next day's professional.”

“I think it's actually not AI researcher. The AI researcher will kind of go before. It's actually very interesting when you ask people this question. I think that everyone tends to say whatever their job is as well as the hardest one. But I actually think that …”

“At OpenAI, we see this sometimes, for example, we had a paper on this unsupervised learning where you train a language model You train a model to predict the next character in Amazon reviews, and just by learning to predict the next character in Amazon reviews…”

“With Sumail, we went undefeated. I, and I think it was five zero that day.”

“We first played our TZ who showed up on, on our switch bot, you know, kind of the Franken bot. And you know, that beat him three times.”

“It was like one of the major examples, one of the major examples of the things that we kind of didn't have explicit incentive for, and yet the bot actually learned them.”

“Originally, we were downloading all of them every day, and realized that was about two terabytes worth of data a day. That adds up quite quickly, so we ended up filtering down to the sort of most expert players.”

“Actually the way that we selected the game was we went on, on Twitch and just looked down the list of most popular games in the world and starting, you know, number one is League of Legends. The thing about League of Legends is it doesn't run on Linux and it d…”

“Everyone thinks they have to be an AI PhD. Not true. Neither of these guys are.”

“So, so essentially becoming a good engineer for our team is much more valuable than, for example, people spending you know, months upon months implementing exotic models in TensorFlow.”

“Monday bot, it is pretty good at the first wave. This new bot is a super bot thereafter. ... So we wrote, so, so we already, we had some code for doing something similar. So we kind of revived that. And then in the three hours, Jay spent his time doing a very …”

“Two weeks before the International was kind of our cutoff for at this point, there's not much more we can do, that we're gonna do our biggest experiment ever, put all of our compute into one basket, and see, see where it goes.”

“At the Future of Humanity Institute, we have a lot of relationships with organizations like DeepMind and OpenAI and others but, you know, we don't have, like, a ton of, you know, GPUs or TPUs here like, running the latest experiments outside of, you know, some…”

“Like, try lots of things, let people follow their intuitions, and then when something starts working, put a lot more resources behind it. That's worked for us really well.”

“And I do think we've done one of the more interesting research labs of the last couple of decades. The world isn't set up to support this anymore, and I, and it's a very interesting question about why.”

“If this one thing works, like, if we do actually build AGI at OpenAI, that is more important than all of the innovation sum total in humanity so far. So, clearly not has all the innovation happened.”

“I mean, it's the same one that OpenAI is using right now. Yeah. And most AI researchers.”

“There's some question about, you know, whether or not the feedback data might give them a flywheel. I'm a little bit skeptical of that, that it would give them so much that no one could catch up.”

“If you look at what the difference is between ChatGPT or the most recent OpenAI Text DaVinci Three model, and what's been in the platform for two years and has not gotten as much attention, the difference is fine tuning. Like it's the same base model, more or …”

“In the InstructGPT paper that OpenAI released, they compared, you know, a one or two billion parameter model with instruction tuning and RHF to the full GPT-III model and people preferred that despite the fact it was a hundred times smaller.”

“When we got the chance to work at GPT-IV, maybe about six or so months before it was publicly released,”

“Just because you're getting free credits doesn't mean that that's a cost that doesn't exist. It just means you're hiding it for the moment. So companies that hide behind OpenAI credits and claim that they've got these huge, huge gross margins have a nasty shoc…”

“What I think a bunch of the companies in the space are seeing is that, like, that's not enough to keep the customers, especially because, like, OpenAI, like, the cost of all of the models is just going down.”

“So people at AWS, Microsoft, OpenAI, giving out tons and tons of credits to startups. You should treat those as a cash cost. Don't assume you'll have unlimited credits forever. It'll totally mess up your margins.”

“OpenAI seems like it is continually the one pushing the envelope, but they always seem to be The first ones to make major breakthroughs, but they have never been able to maintain the lead so far.”

“The Ongoing usage-based pricing is nine dollars per hour, and it sort of points to a sort of powerful thing. Like, if I were a macro trader, I would be very, very bearish on countries that have, that rely very heavily on call centers right now because, you kno…”

“It seems like, or, 15% of the batch are already using O-one, even though it's not, like, fully available yet.”

“But I think there is no denying that you are, we're on a scaling law right now where orders of magnitude matter. The next model is going to be an order of magnitude bigger and the next one on and on. And so that does make it very capital intensive.”

“It's compute first and it's not cheap. It's great talent second. And then of course it's all the normal operating expenses of a more traditional company.”

“OpenAI itself has now enabled distillation internal to its own API. So you can use O-one, you can use even GPT-IV or IV-O to distill it down into a much cheaper model that's internal to them, like GPT-IV, IV-O-Mini. And that's sort of their, you know, lock-in …”

“If you treat OpenAI as the Google of the next 20 years, you want to invest in Google and all the things that Google enabled, like Airbnb. Google could do Airbnb, it probably won't. Just from, like, I don't know, Coase's theorem of the firm, probably. It's just…”

“I had been telling people for a while, I thought that The level two to level three jump was going to happen, but then the level three to level four jump was, level two to level three was going to happen quickly, and then the level three to level four jump was …”

“Most of the world still does not understand the value of, like, a fairly extreme level of conviction on one bet.”

“Conviction is great until the moment you have data one way or the other. And there are a lot of people who hold on it past the moment of data.”

“But, with the possible exception of copywriting, no great businesses were built on GPT-III.”

“I can see a path where the work we are doing just keeps compounding, and the rate of progress we've made Over the last three years, continuous for the next three, or six, or nine, or whatever. You know, nine years would be, like, 3500 days, or whatever. If we …”

“We said from the very beginning we were going to go after AGI at a time when in the field you weren't allowed to say that because that just seemed impossibly crazy and, you know, borderline irresponsible to talk about.”

“And at the time, unsupervised learning was just not really working. So he noticed this one really interesting property, which is there was a neuron that was flipping positive or negative with sentiment. And yeah, that led to the GPT series.”

“We could hit some, like, unexpected wall, or we could be missing something, but it looks to us like there's a lot of compounding in front of us still to happen.”

“It was definitely helpful that I could just, like, write the early checks for OpenAI, and I think it would have been hard to get somebody else to do that at the very beginning. And then Elon did it a lot at a much higher scale, which I'm very grateful for, and…”

“I had no idea that language models were going to be the thing. You know, we started working on robots and agents playing video games and all these other things.”

“This is the first time ever where I felt like we actually know what to do. Like, I think from here to Building an AGI will still take a huge amount of work. There are some known unknowns, but I think we basically know what to go do, and it'll take a while.”

“Three is agents ability to go off and do these longer-term tasks You know, maybe like multiple interactions with an environment, asking people for help when they need it, working together, all of that. And I think we're gonna get there faster than people expec…”

“I believe there were three goals for the For the effort at the time. It was like, figure out how to do unsupervised learning, solve RL, and never get more than a 120 people. Missed on the third one.”

“I remember back when he was starting open AI, one of the things that really motivated him to do it was he believed that when we actually had AGI basically be better at doing science than humans were, and therefore would accelerate the rate of all scientific pr…”

“If you use O-one in ChatGPT, it looks like it will tell you what's really going on, but apparently they have a fake model that just spits out things to give you the impression that it's breaking it up into steps. And they've actually You know, hidden it, becau…”

“Like the full O-one model, which is coming out any day now is a huge step function above even O-one preview, which is what enabled all these incredible results at the hackathon.”

“It sounds like basically with O-one, the chain of thoughts will replace the workflow. So you might not need to break it down into steps yourself, but the evals are still really important.”

“GPT is all generative based on predicting the next token and patterns and then getting those results to check that they're correct. So I think a lot of it is you had to have a lot of data that was factually correct and Fed into probably the model and the train…”

“Sam was just telling us that like O-two and O-three are not far behind.”

“I think over the last year, the big frontier labs and a lot of other people have figured out the tricks to take big models. And, you know, take a very particular distribution of user input and train a model that is almost as good as the big model, but much, mu…”

“And one of the early robotics papers, we actually said site as open AI because we didn't want to get into a fight. You know, the first author is the one who, you know, gets cited and their name shows up every single time. So we said, you know, we're not going …”

“And we took a different approach, which was really more like a startup where there was no sort of big centralized plan. But at the same time, people didn't have, you know, it wasn't just sort of let a thousand flowers bloom. Instead we had opinions about what …”

“And the core idea behind GPT-ONE is that if you have a transformer and you apply this super simple objective of guessing the next token, guessing the next word, that that would be enough signal that you could actually have something that would be able to gener…”

“And there was real insight that was generated there, which was that it really strengthened our belief that scale was the path to improving artificial intelligence. That with Dota two, the secret idea was that we could take huge amounts of experience and feed i…”

“I think Aditya Ramesh, who built that model, spent 18 months, maybe two years, just getting to the first version that, that clearly worked.”

“Early on there were sort of, you know, a couple big projects, as I said, and then some room for exploratory research. And at the very earliest days, the exploratory research was really about you know, it's about what the researcher wanted to do, but also it wa…”

“And then he told me the only thing that matters is he drew two circles. One big circle, called it unsupervised learning, and then inside he said, reinforcement learning, another circle, and he said, this is AGI, every other research doesn't matter.”

“Out of the list you mentioned, Google is the only company that actually has the product taste to do this.”

“OpenAI even had a bot when I worked there called the Truthbot, which John, John built with this team. Where you could ask it a question and it'll go and search the web, and then it'll give you an answer with some sources. And it was very slow, and it was built…”

“The thing about CodeGen, the big game in town that we saw, ah, six months ago was Clawed Sonnet. It's still actually a big contender. Most are still using it. But O-one, O-one Pro, and O-three meaning all these resilient models are starting to see it's almost …”

“The other one is four-oh, virtually no use for CodeGen.”

“And so, I think, like, Google were really on the cutting edge of this transformer technology, and were like, leapfrogged by OpenAI, just because I was too cautious to put stuff out there.”

“Oh, three was very rigid, actually. Like it really sticks to the rubric. It's heavily penalizes for anything that doesn't fit like the rubric that you've given it. Whereas Gemini 2.5 pro was actually quite good at being flexible in that it would apply the rubr…”

“Is we did the back of the envelope math for what Codex, the first coding model, costed to train. From my memory, it only cost about 90 K or a hundred K by our calculations.”

“But it rumored that to go from GPT-III, which, you know, had existed for a while and didn't You know, impressed some people, but was certainly not the breakout moment ChatGPT was. To ChatGPT, it was like a one percent increase in the training costs.”

“You will have this, like, entity that gets to know you, that connects to all your stuff, and that is, like, proactively helping you. It'll, it won't just be like you send a message and it sends you one back, but it'll, like, be running all the time. It'll be l…”

“We were trying to, like, play video games, and we had this little robotic hand that could sort of barely do a Rubik's Cube, and we had no ideas for products, no revenue, no really idea that we were ever going to have revenue, and we were, like, sitting around …”

“But now you start to see things where you can like really give a task to Codex, for example or to deep research, and you have this thing go off and do a bunch of stuff and come back to you with like a proposal. It's like a very junior employee that can work on…”

“Figure out the remaining scientific stuff, which I don't think there's much left.”

“Five more years I think we'll be able to maintain the same rate of progress”

“I think today.”

“I can say this one because it got publicized in early, not early, a few years into OpenAI, where Elon sent us this really mean email. We'd been working through that for a while and said we had a zero percent chance of success. Like, not 0.10, that we were tota…”

“In five years we have gone from this, like, thing that could barely write a sentence to a thing that is like, you know, PhD level intelligence in most areas.”

“Like, there, we got very close to not starting OpenAI.”

“We are going to try to make the best super assistant out of ChatGPT that we can. We're going to Add the things that we think we need to add to that but that is, like, one small part of the opportunity in front of us, and it makes us sad when people are like, I…”

“I think we will run the, like, largest and most expensive piece of infrastructure in the world pretty soon.”

“If you are doing the same thing as everyone else, it is very hard to concentrate talent, and it's very hard to get people to, like, really believe in a mission, and if you're doing, like, a one-of-one thing you have a really nice tailwind there.”

“It took us a long time to figure out how to answer the defensibility question for ChatGPT. We had built this thing, and for a long time, the only defensibility was, like, we had the only product out in the market. And then we kind of, like, a brand that starte…”

“We won't get all the way here with GPT-V, but eventually we do want one integrated model that can, like, reason really hard when it needs to and generate, like, real-time video when it needs to do that.”

“I think it will be, like, much better than you're hoping for, and the ability to, like, use it to run incredibly powerful models locally is gonna, like, really, really surprise people on what's possible.”

“We have an open source model coming out soon.”

“And also, like last week, O-three cost five times as much as it did this week, and that's gonna keep going.”

“I cannot recall ever thinking until like after starting OpenAI that they were going to be so obviously related. That, you know, the, that energy would be the eventually the fundamental limiter on how much intelligence we could have.”

“So in particular, in December last year, OpenAI previewed its, ah, all three model, and they used a version of it that was, ah, fine-tuned specifically on Arc, and that showed human-level performance on that benchmark versus time.”

“OpenAI has perplexity within ChatGPT that did not kill any of these companies.”

“That said, I'm fully working with the assumption that OpenAI will also build its own browser. Anthropic will also try to build its own browser.”

“And 2024, 20 23 especially, and large part of 2024 too, Google had like, maybe a fourth or fifth best models at any moment. So as a startup outside Google, you had access to AI that was better than what Google internally had.”

“ChatGPT is the most successful consumer AI product out there, and so competing against it is very difficult, which is why I really want to, like, push the company more on the browser side, and I think Comet, the browser, will be an abstraction layer about chat…”

“Sam is one of those people that, you know, he's right about a lot of stuff, so I would encourage you if you just dismissed it outright, To ask yourself what you might be missing.”

“By our calculations, Codex, which was the first, this was the first autocomplete model behind GitHub Copilot, it didn't really cost that much money to train, even though, even back then, at kind of the beginning and middle of 2022, people were talking about ho…”

“It's almost like open AI. Is the home product team and the startups are the FTEs out figuring out how to get adoption of the like research that OpenAI is cooking up back at the home office.”

“I think multiple agents. And the reason I think that's true is because let's, let's say I'm someone with true, unique, Domain expertise. Let's say I'm a lawyer who is top in the world at solving certain cases that, that are very rare. And so I have this domain…”

“Basically everyone I worked with, like all the safety leads left, which yeah, invited me to go to Anthropic, and that was sort of the reason I joined OpenAI, was because I cared about AI safety and wanted to work with them. So then I went with them to join Ant…”

“Like, you saw that with the GPT store at OpenAI a little bit, but I think that was a little bit of a failed experiment.”

“Sora is kind of proof that, no, these labs, like, they have the taste and capability and the horsepower and the execution to build and ship net new products that also might run you over.”

“I think Sam Altman is the best salesperson of this generation by no bar none.”

“And, shockingly, in this batch, the number one API is actually Anthropic. It came out a bit more than OpenAI, which who would have thought?”

“There's this particular YCE startup that told me that they collected the best data set for healthcare. And they ended up performing better than OpenAI and a lot of the benchmarks for healthcare with only eight billion parameters.”

“Startups are doing as well, they are actually arbitraging a lot of the models. I had some conversations with a number of founders where before they might have been loyalists to, let's say, OpenAI models or Anthropic, and I just had some conversations recently …”

“You know, we've also had YC companies where, ah, they had something that beat OpenAI, ah, you know, GPT-III.V and they were doing fine-tuning with RL, but then, ah, yeah, GPT-IV and then, ah, GPT-IV came out and, ah, you know, basically blew their fine-tuning …”

“Whereas OpenAI really leans into this idea of just, like, we are going to train the best model and reinforce over time and get it to do longer and longer horizon things in this pursuit of artificial general intelligence.”

“Hey, in the future coding is really going to feel more like talking to a coworker. Like you're going to send off a question and then they'll go off and do something and come back to you with a PR.”

“Well, basically we kind of had like a checkpoint. For I think it was O three, like one of the reasoning models. And then we did a bunch of fine tuning on it and reinforcement learning where it's like, oh, you're given a bunch of questions to like solve these c…”

“Where it will run compaction, like, periodically after each turn, and so Codex can continue to run for a very long time, and if you look at the percentage in the CLI, you'll see it, like, move up and down as compaction runs.”

“The sandboxing, it's such an interesting question because I think OpenAI actually takes the like sandboxing and security question more seriously than almost anyone else.”

“I use the cursor bug bot has gotten quite good and I actually like codex for code review as well. I find it does a very good job on correctness.”

“Codex works very well on Python mono repos.”

“No, they didn't do anything, you just adapted to the new standard, and now your expectations went up, but the model is still the average.”

“And, you know, of course, Anthropic and OpenAI and Google, they're exploring recursive self-improvement, but Typically at that level of having the, you know, having to train a new model for every step of self-improvement that they do.”

“As we were running this pilot for this first customer, sort of OpenAI was coming out with new models in the middle of the pilot, which was changing, one, our cost structure by a 10 x factor, but also was changing our performance quite a lot.”

“We've used a collection of models, open AIs, ChatGPT images, but then we've created 3500 prompts for web designs that then Ploy takes inspiration from.”

“I think there will, I think it will feel like the next six months is like maybe equivalent to the last two years of model progress, something like that.”

“I met Greg Brockman my co-founder, OpenAI, because I was a very early investor in Stripe when I was, like, I don't know, 22 or something, and they were, like, this was before they had, you know, like, real investors that could really help them, and they said, …”

“It seems to me like the demand for sufficiently high quality intelligence at a sufficiently low price is effectively uncapped.”

“And that means that we have to, you know, be fairly quite reserved about not imposing our worldview about what we think people should or shouldn't do with AI, or even what we think the good ideas are, while still making sure that we can ensure enough of a safe…”

“If that happens again, which I think it probably will, then in another six and a half years, the average person uses, let's say, five hundred billion tokens a month, and the token leader uses a quadrillion, or quadrillions of tokens a month.”

“Here are the mistakes that OpenAI made, and we did make some big ones, of course, but these systems have gotten incredibly capable. So I think it's an alignment failure. I think it's a security failure. I think it's like a very serious thing, even though it's,…”