The big thing that we are targeting is producing an automated researcher. So automating the discovery of new ideas. The next set of evals and milestone that we're looking at will involve actual movement on things that are economically relevant. And I was talking to some high schoolers and they're saying, 'Oh, you know, actually the default way to code is vibe coding.' I do think, you know, the future hopefully will be vibe researching.
Thanks for coming Jacob and Mark. Jacob, you're the chief scientist at OpenAI. Mark, you are the chief research officer at OpenAI and you guys have both the privilege and the stress of running probably one of the most high-profile research teams in AI. And so we're just really stoked to talk with you about a whole bunch of things we've been curious about, including GPT-5, which was, you know, one of the most exciting updates to come out of OpenAI in recent times. And then stepping back, how you build a research team that can do not just GPT-5 but Codex and ChatGPT and an API business and can weave all of the many different bets you guys have across modalities, across product form factors into one coherent research culture and story.
And so to kick things off, why don't we start with GPT-5? Just tell us a little bit about the GPT-5 launch from your perspective. How did it go?
So I think GPT-5 was really our attempt to bring reasoning into the mainstream. And prior to GPT-5, right, we have two different series of models. You had the GPT kind of 2, 3, 4 series which were kind of these instant response models and then we had an O series which essentially thought for a very long time and then gave you the best answer that it could give. So tactically, we don't want our users to be puzzled by, you know, which mode should I use and it involves a lot of research in kind of identifying what the right amount of thinking for any particular prompt looks like and taking that pain away from the user. So we think the future is about reasoning more and more, about reasoning more and more, about agents and we think GPT-5 is this step towards delivering reasoning and more agentic behavior by default. There is also a number of improvements across the board in this model relative to O3 and our previous models but our primary focus for this launch was indeed bringing the reasoning mode to more people.
Can you say more about how you guys think about evals? I noticed even in that launch video there were a number of evals where you were inching up from, you know, 98 to 99% and that's kind of how, you know, you saturated the eval. What approach do you guys take to measuring progress and how do you think about it?
One thing is that indeed for like these evals that we've been using for the last few years, they're indeed pretty close to saturated and so yeah, like for a lot of them, like, you know, inching from like 96 to 98% is not necessarily the most important thing in the world. I think another thing that's maybe even more important but a little bit subtler, when we were in this like GPT-2, GPT-3, GPT-4 era, you know, there was kind of one recipe. You just like pre-train a model on a lot of data and you kind of like use these evals as just kind of a yardstick of how this generalizes to like different tasks. Now we have this different ways of training, in particular reinforcement learning on like serious reasoning where we can pick a domain and we can really train a model to like become an expert in this domain to reason very hard about it which lets us, you know, target particular kinds of tasks which will mean that like we can get like extremely good performance on some evals but it doesn't indicate as great generalization to other things. I think so we think about it in this world, we definitely think like we are in a little bit of a deficit of great evaluations and I think the big things that we look at are actual marks of the model being able to discover new things. I think for me the most exciting thread and like actual sign of progress this year has been our model's performance in math and programming competitions. Although I think like they are also becoming saturated in a sense. And the next set of evals and milestone that we're looking at will involve actual discovery and actual movement on things that are economically relevant.
Totally. You guys already got number two in the AtCoder competition. So there's really only number one left.
Yeah. Yeah. I mean I think it is important to note that these evals like, you know, IOI, AtCoder, IMO are actually real world markers for success in future research. I think a lot of, you know, the best researchers in the world have gone through these competitions, have gotten very good results. And yeah, I think we are kind of preparing for this frontier where we're trying to get our models to discover new things.
Yeah, very exciting. Which capability from GPT-5 before the release surprised you the most when you were working through the eval bench or using it internally? Were there any moments where you felt like this was starting to get good enough to release because it was useful in your daily usage?
I think one big thing for me was just how much it moved the frontier in very hard sciences. You know, we would try the models with some of our friends who are, you know, professional physicists or professional mathematicians and you already saw kind of some instances of this on Twitter where, you know, you can take a problem and have it discover maybe not like very complicated new mathematics but, you know, some non-trivial new mathematics. And we see physicists, mathematicians kind of repeating this experience over and over where they're trying problems and saying, 'Wow, this is something that the, you know, previous version of the models couldn't do.' And it is a little bit of a light bulb moment for them. It's like able to automate maybe like what could take one of their students months of time.
Well, GPT-5 is a definite improvement on O3. For me, O3 was definitely like that moment where the reasoning models became like actually very useful on a daily basis. I think especially for, you know, working through a math formula or a derivation, like it actually got to a level where it is like fairly trustworthy and I can actually use it as a tool for my work. And yeah, I think it is very exciting to get to that moment. But I expect that, well, now as we're seeing these models like actually able to automate, well yes, like we're saying, solving contest problems over longer time horizons, I expect that that was quite small compared to what's coming over the next year.
What is coming in the next one to five years? It would be just at whatever level you're comfortable sharing, what does the research roadmap look like?
So the big thing that we are targeting with our research is producing an automated researcher. So automating the discovery of new ideas. And, you know, of course, like a particular thing we think about a lot is automating our own work, automating ML research. But that can get a little bit self-referential. So we're also thinking about automating progress in other sciences. And I think like one good way to measure progress there is looking at like what is the time horizon on which these models actually can reason and make progress. And so now as we kind of like get to a level of near mastery of this high school competitions, let's say, I would say like we get to like maybe on the order of one to five hours of reasoning. And so we are focused on extending that horizon both in terms of like the models' capability to plan over very long horizons and actually able to retain memory.
And back to evals question, that's why I think eval of the form of how long does this model autonomously operate for are of particular interest to us. And actually maybe on that topic, there's been this huge move toward agency and model development. But I think at least the state that it's in currently, users have sort of observed this trade-off between too many tools or planning hops can result in quality regressions versus something that maybe has a little bit less agency, the quality is at least observed today to be a bit higher. How do you guys think about the trade-off between stability and depth? The more steps that the model is undertaking maybe the less likely the tenth step is to be accurate versus you ask it to do one thing it can do it very well and to have it keep doing that one thing better and better but more complex things there's sort of that trade-off. But of course to get to full autonomy you are taking multiple steps, you're using multiple tools.
I think actually like, well, the ability to maintain depth is a lot of it is being consistent over long horizons. So I think they are very related problems. And in fact I think like with the reasoning models we have seen the models like greatly extend the length over which they are able to reason and work reliably without going off track. Yeah, I think this is going to remain a big area of focus for us.
Yeah. And I think reasoning is core to this ability to operate over a long horizon because, you know, you imagine kind of yourself solving a math problem, right? You try an approach, it doesn't work and, you know, you have to think about, you know, what's the next approach I'm going to take, what are the mistakes in the first approach and then you try another thing and, you know, the world gives you some hard feedback, right? And then you keep trying different approaches and the ability to do that over a long period of time is reasoning and gives agents that robustness.
We talked a lot about math and science. I'm curious to get your take on do you think some of the progress that we've made can actually extend similarly to domains that are less verifiable? They're sort of less of an explicit right or wrong.
Oh yeah, this is a question I really like. I think if you actually truly want to extend to research and, you know, finding, discovering ideas that meaningfully advance technology on the scale of like months and years, like I think these questions stop being so different, right? Like it is one thing to solve like a very well-posed constraint problem on the scale of an hour, right? And there's like kind of a finite amount of ideas you need to look through and that might feel extremely different from solving something very open-ended. But, you know, even if you want to solve like a very well-defined problem that is on much longer scale, right? You like, you know, prove this Millennium Prize problem. Well, that suddenly requires you to think about, okay, like what are the fields of mathematics or other science that might possibly be relevant, you know, are there inspiration from physics that I must take, like what is kind of the entire program that I want to develop around this and now these become very open-ended questions. And it's actually hard to, you know, for our own research, right? Like if all we cared about is, you know, reduce the modeling loss on a given data set, right? Like measuring the progress on that, like, you know, like are we kind of actually asking the right questions in research like actually becomes like a fairly open-ended affair.
Yeah. And I think it also makes sense to think about what the limits of, you know, open-ended means, you know. I think a while back Sam tweeted about some of the improvements that we were making in having our models write more creatively. And, you know, we do consider the extremes here as well.
Right. Right. Let's talk about RL because it seems like since O1 came out, RL has been the gift that keeps giving. You know, every couple months OpenAI puts out a release and everyone goes, 'Oh, that's great, but this RL thing is going to plateau. We're going to saturate the evals. The models won't generalize or there's going to be mode collapse because of too much synthetic data for whatever.' Everybody's got a laundry list of reasons to believe that the gains and performance from RL are going to tap out and somehow they just don't. You guys just keep coming out and putting out continuous improvements. Why is RL working so well and what if anything has surprised you about how well it works?
RL is a very versatile method, right? And there are a lot of ideas you can explore once you have an RL system working. A long time at OpenAI, we started from this before language models, right? Like we were thinking about like, okay, like RL is this extremely powerful thing of course like on top of deep learning which is this incredible general learning method. But the thing that we struggled with for a very long time is like what is the environment, like how do we actually anchor these models to the real world or like should we, you know, simulate some island where they all learn to collaborate and compete. And then, you know, of course came the language modeling breakthrough, right? And we saw that, oh yeah, if we scale deep learning on modeling natural language we can create models with this incredibly nuanced understanding of human language. And so since then we've been seeking how to combine these paradigms and how to get RL to work on natural language. Once you do, right, like then you kind of have the ability to actually execute on these different ideas and objectives in this extremely robust rich environment given by pre-training. And so yeah, so I think it's been a real, yeah, I think it's been perhaps the most exciting period in our research over the last few years where we've really like, yeah, we found so many new directions and promising ideas that all seem to be working out and we're trying to understand how to compare.
One of the hardest things about RL for folks who are not practitioners of RL is the idea of crafting the right reward model. And so, especially if you're a business or an enterprise who wants to harness all this amazing progress you guys are putting out, but doesn't even know where to start. What do the next few years look like for a company like that? What is the right mindset for somebody who's trying to make sense of RL to craft the right reward model? Is there anything you've learned about the best practices or an approach of thinking of using this latest sort of family of reasoning techniques? What is the right way I should think about even approaching reward modeling as a biologist or a physicist?
I expect this will evolve quite rapidly. I expect it will become simpler, right? Like I think maybe like two years ago we would have been talking about like what is the right way to craft my fine-tuning data set and I don't think we are at the end of that evolution yet and I think we will be inching towards more and more human-like learning which, you know, RL is still not quite. So I think maybe the most important part of the mindset is to like not assume that like what is now will be forever.
So I want to bring the conversation back to coding. We would be remiss not to say congrats on GPT-5 Codex which just dropped today. Can you guys say a little bit more about what's different about it, how it's trained differently, maybe why you're excited about it?
Yeah, so I think one of the big focuses of the Codex team is to just take the raw intelligence that we have from our reasoning models and make it very useful for real world coding. So a lot of the work they've done is kind of consistent with this. They are working on kind of having the model be able to handle more difficult environments. We know that real world coding is very messy. So they're trying to handle all the intricacies here. There's a lot of coding that has to do with, you know, style, with just like kind of softer things like how proactive the model is, how lazy it is and just being able to define in some sense like a spec for how a coding model should behave. They do a lot of, you know, very strong work there and as you see, like they're also working on a lot better presets. You know, coders, they have some kind of notion of this is how long I'm willing to wait for a particular solution. I think we've done a lot of work to dial in on, you know, for easy problems being a lot lower latency, for harder problems actually the right thing is to be even higher latency, get you the really best solution and just being able to find that preset sweet spot for if you were to say like easier problems versus harder.
What we've found is the latest, the previous generation of the Codex models, they were spending too little time solving the hardest problems and too much time solving the easy problems. And I think that is actually just probably out of the box what you might get out of O3.
Maybe just on the topic of coding since you guys are both competitive coders in prior lives. I know you've been at OpenAI for almost a decade now, but I was struck by the story of Lee Sedol, the Go player who kind of famously quit Go after he lost to AlphaGo multiple times. And I think in a recent interview you guys were both saying that now the coding models are better than your capabilities. And that gets you excited. But say more about that. And how much would you say you code now? Well, if you're hands-on keyboard, you can talk about OpenAI generally, but how much code is written by AI now in terms of coding models being better?
I mean, I think yeah, I think it is extremely exciting to see this progress. I think like the programming competitions have a nice kind of encapsulated test of like ability to come up with some new ideas in this like boxed environment and time frame. I do think like, you know, if you look at things like, well, I guess the IMO problem six or maybe some very hardest programming competitions problems, like I think there's still a little bit of headway to go for the models but I wouldn't expect that to last very long. I took a little bit, historically I've been like being humble, historically I've actually been like extremely reluctant to use any sort of tools. I just used Vim pretty much.
Yeah. Yeah, eventually I think like especially with this latest coding tools, like GPT-5, I've really kind of felt like, okay, like this is no longer the way. Like you can do a, you know, 30 file refactor like pretty much perfectly in like 15 minutes, like you kind of have to use it. And so I've been kind of like learning this new way of coding which definitely feels a little bit different. I think it is like a little bit of an uncanny valley still right now where like you kind of have to use it because it is just like excelling at so many things but it's still like, you know, a little bit like not quite as good as a coworker. So, you know, I think like our priority is getting out of that uncanny valley.
Yeah. But yeah, it's definitely an interesting time.
Yeah, definitely. To kind of like speak to the Lee Sedol moment. I think AlphaGo for both of us was, you know, a very formative milestone in AI development. And at least for me, it was the reason I started working on this in the first place. And maybe partly because of our backgrounds in competitive programming, like I had this affinity to building these models which could do very, very well in these forms of contests. And going from, you know, solving eighth grade math problems to a year later hitting our level of performance in these coding contests. It's crazy to see that progression and you kind of imagine or like to think that you feel a set of the feelings at least Lee Sedol felt too, right? It's like, wow, this is really crazy, right? And what are the possibilities and, you know, this is something that I took decades to do and it took a lot of hard work to get to the forefront of. So you really do feel an implication of that is these models, what can't they do, right? And I do feel like already it's kind of transformed the default for coding. This past weekend I was talking to some high schoolers and they were saying, 'Oh, you know, actually the default way to code is vibe coding.' Like, you know, I think like they would consider, oh, it's like maybe sometimes for completeness you would go and like actually do all of the mechanics of coding it from scratch yourself, but that's just a strange concept to them. Like why would you do that? You know, just vibe code by default. Yeah. And so yeah, I mean I do think, you know, the future hopefully will be vibe researching.
Yeah. I have a question about that which is what makes a great researcher. Right. When you say vibe researching, there's a big part of vibe coding is just having good taste in wanting to build something useful and interesting for the world. And I think what's so awesome about tools like Codex is if you've got a good intuition for what people want, it helps you articulate that and then basically actualize a prototype very fast. With research, what's the analog? What makes a great researcher?
Persistence is a very key trait, right? Like I think like what is different about research when you're actually trying to, I think a special thing about research, right, is you're trying to create something or learn something that is just not known, right? Like it's not known to work, like you don't know whether it will work. And so always trying something that will most likely fail and I think getting to a place where you are like in a mindset of like being ready to fail and being ready to learn from these failures and, you know, and of course with that comes creating kind of clear hypothesis and being extremely honest with yourself about how you're doing on them, right? I think a trap many people fall into is going out of the way to like prove that it works, right? Which is quite different from, you know, like I think like believing in your idea and significance is extremely important, right? And you want to persist with that, but you have to be honest with yourself about when it's working and when it's not so that you can learn and adjust.
Yeah, I think there are just very few shortcuts for experience. I think through experience you kind of learn, you know, what's the right horizon to be thinking of a problem, right? You can't pick something that's too hard or it's not satisfying to do something that's too easy. And I think a lot of research is managing your own emotions over a long period of time, too. You know, there's just going to be a lot of things you try and they're not going to work. And sometimes you need to know when to persevere through that or sometimes when to kind of switch to a different problem. And I think interestingness is something, you know, you try to fit through reading good papers, talking to your colleagues and you kind of maybe distill their experience into your own process.
When I was in grad school, you know, there's a big part, I was a failed machine learning researcher. I was in grad school for bioinformatics. But a big part of my research advisor's thrust was about picking the right problems to work on such that you could then sustain and persist through the hard times. And you said something interesting which was there's a difference between having conviction in an idea and then being maximally truth-seeking about when it's not working. And though both those things are sometimes in tension because you kind of go native on a topic or a problem sometimes that you have deep conviction in. Have you found, is there any sort of heuristics you found are useful at the taste step, at the problem picking step that help you arrive at the right set of problems where that conviction and truth-seeking is not as much in zero-sum tension as other kinds of problems?
Yeah, to be clear I don't think conviction and truth-seeking are really in a zero-sum tension. I think like you can be convinced or, you know, you can have a lot of belief in an idea and you can be, you know, very persistent in it while it's not working. I think it's just important that you're kind of honest with yourself like how much progress you're making and you're in a mindset where you're able to learn from the failures along the way. I think it's important to look for problems that you really care about and you really believe are important, right? And so I think one thing I've observed in many researchers that inspired me has been really going after the hard problems, like looking at the questions that are, you know, kind of like widely known but like not really kind of considered tractable and just asking like, you know, why are they not tractable or like what about this approach, like why does this approach fail, right? You're always like thinking about what is really the barrier for the next step. If you're going after problems that like you really truly believe are important, right, then that makes it so much easier to find the motivation to persist with them over years.
And in the development of like during the retraining phase of GPT-5 for example, are there any, were there any moments where there was a hard problem, the original initial attempts that were being made to crack that problem weren't working and yet you found somebody...
Persisted through that. And what was it about those stories that comes to mind that worked well that you wish other people and other researchers did more of?
I think on the path there, along the sequence of models, both the pre-trained models and the reasoning models, I think one very common theme is bugs. Both just silly bugs in software that can stay in your software for months and invalidate all your experiments in a way that you don't know. Identifying them can be a very meaningful breakthrough for your research program. But also bugs in the sense of having a particular way of thinking about something that is a little bit skewed, which causes you to make the wrong assumptions. Identifying those wrong assumptions and rethinking frames from scratch. I think both for getting the first reasoning models working or getting larger pre-trained models working, we've had multiple issues like that that we've had to work through.
