Hello, I'm Sam. This is our chief scientist, Yakob. And we have a bunch of updates to share today about OpenAI. Obviously the news of today is our new structure. We're going to get to that near the end, but there's a lot of other important context we would like to share first. Given the importance of a lot of this, we're going to go into an unusual level of transparency about some of our specific research goals and infrastructure plans and product, but we think it's very much in the public interest at this point to cover all of this. So our mission at OpenAI, in both the nonprofit and our new PBC, is to ensure that artificial general intelligence benefits all of humanity. As we get closer to building this, we have new insights into what that is going to mean. There was a time earlier on in OpenAI where we thought that AI or AGI would be sort of this oracular thing in the sky and it would make all these wonderful things for us, and we now have a sharper view of that, which is we want to create tools and then we want people to use them to create the future. We want to empower people with AI as much as possible and then trust that the process that has been working for human history of people building better and better things with newer and better tools will continue to go on. We can now see a vision where we help build a personal AGI that people can use anywhere with all of these different tools, access to all these different services and systems to help with work and personal life. As AI gets better and better, as AI can even do things like discover or help discover new science, what people will be able to create with that to make all of society better and their own lives more fulfilled, we think should be quite incredible. There are three core pillars we think about for OpenAI: research, product, and infrastructure. We have to succeed at the research required to build AGI. We have to build a platform that makes it easy and powerful to use. And we have to build enough infrastructure such that people can use at a low cost all of this amazing AI that they'd like. Here's a little cartoon of how we think about our world. At the bottom layer here, we have chips, racks, and the systems around them, the data centers that these go into, and the energy. We'll talk more about the first three today and energy another time. Then we train models on top of these. Then we have an OpenAI account on top of that. We have a browser now called Atlas and we have devices coming in the next few years that you'll be able to take AI with you everywhere. And we have a few first-party apps like ChatGPT and Sora, and we'll have more over time. But mostly what we're excited about is this big puzzle piece in the upper right. We're finally getting to a world where we can see that people are going to be able to build incredible services with AI, starting with our API with apps and a new enterprise platform that we'll have over time, and way more, and people will be able to fit all of the current things in the world and many more into this new AI world, and we want to enable that, and we believe the world will build just a huge amount of value for all of us. That's kind of what we see the economic picture looking like. But one of the things that we've thought about for a long time and we really see happening now, or starting to happen now, glimmers of it, green shoots, whatever you want to call it, is the impact that AI will have on science. Although the economic impact from that previous slide will be huge, for the long-term quality of life and improvement and change in society, AI that can autonomously discover new science or help people discover new science faster will be one of the most important things and something that we're really trying to wrap our heads around. So I'm going to hand this over to Yakob to talk about research, and as I mentioned, we're going to share a lot about our internal goals and our picture of where things are.
Thanks Sam. At the core we are a research laboratory focused on understanding a technology called deep learning. A particular focus of ours is understanding what happens as you scale up training deep learning systems. One consequence we discuss a lot is AGI, artificial general intelligence. But we find that in some way even this maybe understates the magnitude of the possible progress and change here. In particular, we believe that it is possible that deep learning systems are less than a decade away from superintelligence, systems that are smarter than all of us on a large number of critical axes. This is of course a serious thing, with a lot of implications to grapple with. One particularly focusing impact of this technology and the technologies leading up to it, and something that we organize our entire research program around, is the potential to accelerate scientific discovery and the development of new technology. We believe that this will be perhaps the most significant long-term impact of AI development, and it will fundamentally change the pace of progress on developing new technologies. Thinking about how far along we are towards these goals, one good way to think about progress is to look at the time horizon that it would take people to accomplish the task that the models can perform. This is something that has been extending rapidly over the past few years. Where the current generation of models is at right now is about five hours. You can see this by looking at the models matching the best people in competitions such as the International Olympiad in Informatics. We believe that this horizon will continue to extend rapidly, in part as a result of algorithmic innovation and in part scaling deep learning further, particularly scaling along this new axis of in-context compute, also called test-time compute, where we really see orders and orders of magnitude to go. This is roughly how much time the model spends thinking. If you look at how much time the model's currently spent thinking about problems, and if you think about how much compute, how much time you would like to spend on problems that really matter such as scientific breakthroughs, you should be okay using entire data centers. So there is really quite a way to go there. Anticipating this progress, we of course make plans around it internally, and we want to provide some transparency around our thinking there. So we want to take this perhaps somewhat unusual step of sharing our internal goals and goal timelines towards these very powerful systems. These particular dates we absolutely may be quite wrong about them, but this is currently how we think, plan, and organize. As a research organization that is working on automating research, we are thinking about how this impacts our own work, how AI systems that accelerate development of future AI systems will look, and how they can empower research like alignment. We are making plans around getting to quite capable AI research interns that can meaningfully accelerate our researchers by expanding a significant amount of compute by September of next year. We believe that is actually quite close. Then we look towards getting a system capable of autonomously delivering on larger research projects, a meaningful fully automated AI researcher, by March of 2028. As we look towards these very capable systems, we think a lot about safety and alignment. In fact, a lot of our work, both on deployments and safety and on understanding deep learning and development capability, we can think of as preparation for these very capable models. Safety is a multifaceted problem. The way we generally structure our thinking are these five layers, ranging from factors that are most internal to the model to ones that are most external. At the core, what we believe is the most important long-term safety question for superintelligence is value alignment. You can think of value alignment as what is the thing that the AI fundamentally cares about. Can it adhere to some high-level principles? What will it do if it's given unclear and conflicting objectives? Does it love humanity? The reason we believe that this high-level objective or principles driving the AI are so important is that as we get to systems that are thinking for very long, as they become very smart, as they tackle problems that are at the edge or perhaps beyond human ability, getting to complete specifications becomes quite difficult. So we have to rely on this deeper alignment. Then there's goal alignment. Does the agent interact with people? How does it interact with people? How does it follow instructions? Then reliability. Can the AI correctly calibrate its predictions? Can it be reliable on easy tasks, express uncertainty on hard ones, can it deal with environments that are a little bit unfamiliar? Then we have adversarial robustness, which is very related to reliability but about adversarial settings. Can the AI withstand targeted attacks from human or AI adversaries? The outer layer is systemic safety, which are guarantees about the behavior of the overall system that do not rely on the AI's intelligence or alignment. For example, this can be security, what data does the AI have access to, or what devices it can use. We invest in multiple research directions across these domains, and we have seen quite a lot of progress also come from just the general development and improving understanding of deep learning as a whole. I want to take a slightly deeper technical dive here and talk about a particular direction. Value alignment is a hard problem. It's definitely not solved yet. However, there is a new promising tool that aids our study of it, and that is chain-of-thought faithfulness. It's something we invest in very heavily. Starting from our first reasoning models, we've been pursuing this new direction in interpretability. The idea is to keep parts of the model's internal reasoning free from supervision, so don't look at it during training, and thus let it remain representative of the model's internal process. We refrain from guiding the model to think good thoughts and let it remain a bit more faithful to what it actually thinks. This is not guaranteed to work, of course. We cannot make mathematical proofs about deep learning. This is something we study. But there are two reasons to be optimistic. One reason is that we have seen very promising empirical results. This is a technology we employed a lot internally. We use this to understand how our models' propensities evolve over training. Also, we have had successful external collaborations on investigating the models' propensity schemes, for example. Secondly, it is scalable in the sense that we make the scalable objective not adversarial to our ability to monitor the model. Of course, an objective not being adversarial to the ability to monitor the model is only half the battle. Ideally, you want it to help with monitoring the model. This is something we're researching quite heavily. One important thing to underscore about chain-of-thought faithfulness is it's somewhat fragile. It really requires drawing this clean boundary and having this clear abstraction and having restraint in what ways you can access the chain of thought. This is something that is present at OpenAI from algorithm design to the way we design our products. If you look at the chain-of-thought summaries in ChatGPT, if we didn't have the chain summarizer, if we just made the chain of thought fully visible at all times, that would make it part of the overall experience, and over time it would be very difficult to not subject it to any supervision. Long-term, we believe that by preserving some amount of this controlled privacy for the models, we can retain the ability to understand their inner process, and we believe this can be a very impactful technique as we move towards these very capable, long-running systems. I'll hand back to Sam.
Okay, that's very hard to follow with the rest of this, and obviously that's the most important part of what we have to say. To reiterate, we may be totally wrong. We have set goals and missed them miserably before. But with the picture we see, we think it is plausible that by September of next year we have sort of an intern-level AI research assistant, and that by March of 2028, which I believe is almost 5 years to the month after the launch of GPT-4, we have a legitimate AI researcher. This is the core thrust of our research program. There are two other areas we want to talk about: product and then infrastructure. On the product side, as we make this incredible progress with deep learning, we want to make it useful to people to sort of invent the future as we mentioned. What that's looked like traditionally for us is an AI super assistant inside of ChatGPT, but we're now really going to evolve to a platform that other people will build on top of, and all of the pieces that need to fit in of the world will be built by others. Before we go talk about that, wanted to just show a quick video of how people are using GPT-5, some of the ways people are using GPT-5 in ChatGPT today.
Yakob will join me back for Q&A in a little bit, but we're going to have one special guest on before the end. We love that. We want much more of that. We want that everywhere. So we want OpenAI to be a platform that people and companies can build on top of. We can sort of see our way now to an AI cloud where this is not just in ChatGPT, not just services that we create, but we want to expose our technology for as many people to build the things that people will depend on and use and create with as possible. I think this idea is originally from Bill Gates, at least that's where I first heard it, that you've built a platform when there's more value created by people building on the platform than by the platform builder. That's our goal. Next year, we really think we can now take this technology and this user base and this sort of framework we've built and get the whole world to build amazing new companies and services and applications on top of it. To do that, there will be many things that we have to evolve towards, but there are two foundational principles as we move towards being a platform that I wanted to touch on. One is about user freedom. If this is going to be a platform that all sorts of people are building on and using and creating with, people around the world have very different needs and desires. There will of course be some very broad bounds, but we want users to have a lot of control and customization of how they use it. I made one of my many stupid mistakes when I tried to talk about this recently. I wish I had used an example other than erotica. I thought there was an understandable difference between erotica and pornbots. But in any case, the point we're trying to get across is that people need a lot of flexibility and people want to use these things in different ways, and we want to treat our adult users like adults. In our own first-party services, we may have tighter guidelines, but AI is going to become such an important part of people's lives. The freedom of human expression is going to need to be there. Along with that, we think that world will need to think about privacy in a different way than they have for previous kinds of technology. Privacy is important for all sorts of technology, of course, but privacy for AI will be especially important. People are using this technology in a different way than they've used the technologies of the past. They're talking to it like they would to their doctor, their lawyer, their spouse. They're sharing the most intimate details of their lives. We need strong technical protections on that privacy, but we also think we need strong policy protections of that privacy. We've talked about concepts like AI privilege, but really strong protections if AI is going to be this fundamental platform in people's lives seem super important to us. Okay, and then I want to go on to infrastructure. I know there's been a lot of confusion about where we are in our infrastructure buildout, and we figured we would just be super transparent about that. Where we are today, all of our commitments total a little bit over 30 gigawatts of infrastructure buildout. That's about a $1.4 trillion total financial obligation for us over the next many years. This is what we've committed to so far. We of course hope to do much more, but given the picture we see today, given what we think we can see for revenue growth and our ability to raise capital, this is what we're currently comfortable with. This requires a ton of partnerships. We've talked about many of our great chip partners. There are people building the data centers for us, land, energy. There will be chip fab facilities. This is already getting to require quite a lot of supply chain innovation. We're thrilled to get to work with AMD, Broadcom, Google, Microsoft, Nvidia, Oracle, SoftBank, and many others to really make this happen. But this is still early. If the work that Yakob talks about comes to fruition, which we think it will, and if the economic value of these things happen and people want to use all these services, we're going to need much more than this. So I want to be clear, we're not committing to this yet, but we are having conversations about it. Our aspiration is that we can build an infrastructure factory where we can create 1 gigawatt a week of compute, and we aspirationally would like to get that cost down significantly, to like $20 billion per gigawatt over the 5-year life cycle of that equipment. To do this will require a ton of innovation, a ton of partnerships, obviously a lot of revenue growth. We'll have to repurpose our thoughts about robotics to help us build data centers instead of doing all the other things. But this is where we'd like to go, and over the coming months we are going to do a lot of work to see if we can get here. It will be some time before we're in a financial position where we could actually pull the trigger and get going on this. 1 gigawatt is a big number. I figured we would show a little video to put this into perspective. This is a data center that we're building in Abilene, Texas. This is the first Stargate site. We're doing several of these now around the country, but this one is the furthest along. There are many thousands of people that work here every day just doing the construction at the site. There's probably hundreds of thousands or millions of people that work in the supply chain to make all this happen, to design these chips, to fab these chips, to put them together. There's all of the work that goes into this for energy. There's an enormous amount of stuff that has to happen for each gigawatt. We want to figure out how we can make this way more efficient and way cheaper and way more scalable so that we can deliver on the infrastructure that the research roadmap requires and that all of the ways that people will want to use this need. To enable that, we have a new structure. Maybe you saw before this crazy convoluted diagram of all of the OpenAI entities. Now it's much simpler. We have a nonprofit called the OpenAI Foundation that is in control, where the board sits, and owns a slice of our PBC, public benefit corporation, called OpenAI Group. So a nonprofit in control, public benefit corporation sits under it. We hope for the OpenAI Foundation to be the biggest nonprofit ever. As I mentioned a few times now, science is one of the ways that we think the world most improves, along with the institutions that broadly distribute the benefits of that. Science will not be the only thing that the nonprofit funds, but it will be an important first major area of the things that we do. The nonprofit will govern the PBC. It will initially own about 26% of the PBC equity, but that can increase over time with warrants if we perform really well. It will use these resources to pursue what we think are the best benefits of AI, given where the technology is and what society needs. The PBC will operate more like a normal company. It will have the same mission. It will be bound to that mission. In matters of safety, it will only be bound to that mission, but it will be able to attract the resources that we need for that gigantic infrastructure buildout to serve the research and product goals that we have. The initial focus of the foundation, we'll do more things over time, but we want to knock something out of the park first, hopefully, is a $25 billion commitment to use AI to help cure disease. There are a lot of ways this can happen: generating data, using a lot of compute grants to scientists, and also for AI resilience. AI resilience is a new and very important area, and I'd like to invite our co-founder Wojciech up to talk about what this will look like.
Hello, glad to be here. Thanks for being here. The term AI resilience is a little bit broader than what we historically thought about AI safety. With resilience, we think that advanced AI comes with risks and disruptions, and we would like to have an ecosystem of organizations that can help to solve a number of these problems. Let me give you an example to better illustrate it. We all believe that AI will advance in biology, and as it advances in biology, there is a risk that some bad actor could use AI to create manmade pandemics. On the safety level, the mitigation would be to make sure that the models block the queries that have to do with virology. However, if you consider the entire AI industry, it's very likely that even if OpenAI blocks it, someone could use different models out there and still produce pathogens. With resilience, we don't want just to block it but also have a rapid response if the problem would occur. When I think about the risks and disruptions, there are many. Mental health is one of them, bio is another, job displacement might be another. We think that we need an ecosystem. A good analogy that I like is cybersecurity. At the beginning of the internet, it was a place where people didn't feel comfortable putting their credit card numbers because it was so easy to get hacked. When there was a virus, people were giving each other a call to disconnect the computer from the internet. We got a long way. Now there's an entire infrastructure of cybersecurity companies protecting critical infrastructure, governments, corporations, and individual users to such an extent that people are willing to put the most personal data online, to have life savings online. Cybersecurity got really far, and we think that something analogous will be present for AI. There will be an AI resilience layer, and I'm really excited that the nonprofit will help out to stimulate it, to create such an ecosystem.
So am I. I think this is an important time to be doing this, and I'm very excited that you're going to figure out how we go off and make it happen. Again, these are not the only things that the nonprofit will fund, but we're excited about these as the first two: using AI to develop cures and treatments for diseases, and this new AI resilience effort as we figure out what the deployment of AGI into society is going to look like. So we mentioned that those are our three pillars, but what if this all works? We think it is plausible that in 2026 we start to see the models of that year begin to make small discoveries. By 2028, medium or even larger discoveries, and who knows what 2030 and 2032 are going to look like. If AI can keep advancing science as has happened in the past, we think the future can be very bright. Of course, we think it's very important that humans can self-determine our way through this future. But the open space that new scientific advances give us is quite impressive. We asked Sora to help us imagine a radically better future by looking at the past. We are particularly interested in how the history of science builds on itself, discovery after discovery. This is what we hope will happen with AI. This is going to be 200 years of science. But if you can do these 200 years of compounding discoveries, the scaffolding building up on each other, not in 200 years but in 20 years or in 2 years, and if you look at how much this is accelerated, think about what could be possible. You can imagine a world where a radically better future becomes quite possible. You have a data center here that is discovering a cure for cancer. A data center there that's making the best entertainment ever. A data center here that's helping you find your future husband or wife. This one is building rockets and helping you colonize space. This one is helping to solve the climate crisis. We did all this stuff the old-fashioned way. Now with the help of AI, we'll be able to shape what comes next with maybe much more power. We talked a little about AI medicine. We're very excited about robots. We really think energy is very, very important to the world. We want to figure out what personalized education can mean, design novel materials, and probably a ton of other things that we can't even think of yet. As we head into this next phase of OpenAI and, more importantly, this continual progress in deep learning, we thank you for joining us today. We're going to try something new now, which is we're going to just answer questions. If this works, it's something we'll try more in the future. Yakob is going to rejoin for this Q&A. Thank you very much. This is a new format for us, so bear with us as we try it this first time. If this is useful, it's something we'll do again a lot more. We're going to try to just answer questions in the order they are most upvoted. Are we good to go? All right, let's see how this works. You can put questions in the Vimeo link, and we will just start answering them. From Caleb, we're warned that the tech is becoming addictive and eroding trust, yet Sora mimics TikTok and ChatGPT may add ads. Why repeat the same patterns you criticized, and how will you rebuild trust through actions and not just words? We're definitely worried about this. I worry about it not just for things like Sora and TikTok and ads and ChatGPT, which are maybe known problems that we can design carefully for, but we have certainly seen people develop relationships with chatbots that we didn't expect, and there can clearly be addictive behavior there given the dynamics and competition in the world. I suspect some companies will offer very addictive new kinds of products. I think you'll have to just judge us on our actions. We'll make some mistakes. We'll try to roll back models that are problematic. If we ship Sora and it becomes super addictive and not about creation, we'll cancel the product, and you'll have to just judge us on that. My hope and belief is that we will not make the same mistakes that companies before us have made. I don't think they meant to make them either. We're all kind of discovering this together. We probably will make new ones though, and we'll just have to evolve quickly and have a tight feedback loop. We can imagine all sorts of ways this technology does incredible good in the world, also obvious bad ones, and we're guided by a mission where we'll just continuously evolve the product. One thing that we are quite hopeful about in terms of what we optimize for in products like ChatGPT or Sora is optimizing for the very long term, which is naturally aligned with how we think in general about extending the horizon on which the models can work productively. We believe that quite a lot of development is possible there, and we can eventually get the models that really optimize for long-term satisfaction and well-being instead of just short-term signals.
Okay, next question. Will we have an option to keep the for model permanently after adult mode is installed? We don't need safer models; responsible adults. We have no plans to sunset Foro. We are not going to promise to keep it around till the heat death of the universe either, but we understand that it's a product that some of our users really love. We also hope other people understand why it was not a model that we thought was healthy for minors to be using. We hope that we build better models over time that people like more. The people you have a relationship with in your life, they evolve and get smarter and change a little bit over time. We hope that the same thing will happen. But no plans to sunset Foro currently. We have a lot of Foro questions. In the interest of time, we will not go through all of these, but we want people to have models that they want to use. We don't want people to feel like we're routing them around models. We want adults to make choices as adults, as long as we think we're not selling heroin or whatever, which also you shouldn't do. So for people that want to have emotional speech, as we've said, we want to allow more of that and we plan to.
Okay, here's a good anonymous question for Yakob. When will AGI happen?
I think in some number of years we'll look back at these years and we'll say this was kind of the transition period when AGI happened. One way we thought about it, as some said early on, we thought about AGI emotionally as this thing that is the ultimate solution of all the problems and it's this single point for which there is before and after. I think we found that it's a bit more continuous than that. For various benchmarks that seemed like the obvious milestones towards AGI, we now think of them as indicating roughly how far away we are in years. If you look at a succession of milestones such as computers beating humans at chess, then at Go, then computers being able to speak in natural language, and computers being able to solve math problems, they clearly get closer together.
I think the AGI term has become hugely overloaded, and as Yakob said, it'll be this process over a number of years that we're in the middle of. One of the reasons we wanted to present what we did today is that it's much more useful to say our intention, our goal, is by March of 2028 to have a true automated AI researcher and define what that means, than it is to try to satisfy everyone with a definition of AGI. Maybe one other thing to mention: one counterintuitive thing here is that obviously we're working with a pretty complicated technology. We're trying to understand all these algorithms. Maybe initially we imagined that AGI is the moment where you have figured out all the answers and it's the final thing. Now we increasingly realize that there is some curve of intelligence, maybe a multi-dimensional one, and humans are somewhere on it. As you scale deep learning and develop these new algorithms, eventually you inch closer to that point and eventually will surpass it, and we already have surpassed it on multiple axes. That doesn't mean you have solved all the problems around it, which is something we need to seriously think about.
Can you give us an idea of how far ahead internal models are compared to deployed ones? I think we have quite strong expectations for our next models. I think we expect quite rapid progress over the next couple of months and a year. But we haven't been withholding something extremely crazy.
One of the ways this often works in practice is there are a lot of pieces that we develop, and they're all kind of hard-won victories, and then we know that when we put them together, we will have something quite impressive, and we're able to predict that fairly well. Part of our goal today is to say that we have a lot of those pieces. It's not like we're currently sitting on this giant thing that we're not showing to the world, but we expect that by a year from now, certainly with this September 2026 goal, we have a realistic shot at a tremendously important step forward in capability. What is OpenAI? Ronin asks, 'What is OpenAI's stance on partnering with labs like Anthropic, Gemini, or xAI for joint research, compute sharing, and safety efforts?' We think this is going to be increasingly important on the safety front. Labs will need to share safety techniques and safety standards. You can imagine a time when the whole world would say, before we hit a recursive self-improvement phase, we really need to all carefully study this together. We welcome that collaboration. I think it'll be quite important. One thing to mention on the chain-of-thought faithfulness that I talked about earlier, we actually have started talking about establishing industry arms and have started some joint investigations with researchers from Google and Anthropic and some other labs. That's something I'm very excited about, and I think that is an example of something where we can really benefit from collaborating across multiple labs.
Anonymous asks, 'Will you ever open source some of your old models like the original GPT-4?' We might do those as museum artifacts someday, but GPT-4 is not a particularly useful open-source model. It's big. It's not that good. We could probably make something that is beyond the power of GPT-4 at a very tiny scale that actually would be useful to people. So for useful things, I expect more things like that. For fun museum artifacts, someday, who knows? I think there could be a lot of cool things like that. Another anonymous, or maybe the same one, asks, 'Will you admit that your new model is inferior to the previous one and that you're ruining your company with your arrogance and greed while ignoring users' needs?' I believe that it is inferior for your use case, and we would like to build models that are better for your use case. On the whole, we think for most users it's a better and more capable model, but we definitely have learned things about the Foro-to-5 upgrade, and we will try to do much better in the future, both about better continuity and about making sure that our model gets better for most users, not just people that are using AI for science or coding or whatever. Y asks, 'Will there ever be a version of ChatGPT meant for personal connection and reflection, not only business or education?' Yeah, for sure. We think this is a wonderful use of AI. We're very touched by how much this has meant to people's lives. All of us get a ton of emails and outreach from users about how ChatGPT has helped people in difficult personal situations or to live a better life. This is what we're here for. This is as important as anything that we do. We love to hear about scientific progress. We love to hear about people that got diagnosed with a disease and got cured. The personal stories are incredibly important to us, and we're thrilled about that, and we absolutely want to offer such a service.
Your safety routing breaks user trust and workflows by overriding our choices. Will you commit to revoking this paternalistic policy for all consenting adult users and stop treating us like children? When do users get control over routing? Where's the transparency on safety and censorship? Why can't adults pick their own models? Yeah, I don't think the way we handled the model routing was our best thing ever. There are some real problems with Foro. We have seen a problem where people in fragile psychiatric situations using a model like Foro can get into a worse one. Most adult users can use those fine, but we have an obligation to protect minor users. We also have an obligation to protect adult users who are not in a frame of mind where they're reasonably likely choosing what they really want and we're not causing them harm. As we build age verification and as we are able to differentiate users that are having a true mental health crisis from users who are not, we of course want to give people more user freedom, as we mentioned that's one of our platform principles. So yes, expect improvement there, and I don't think this was our best work in how we communicated the previous rollout or how we strike the right balance between protecting people and allowing adults to speak about difficult things without feeling policed. You want to say anything there?
Definitely there is a problem where we aim to lay out the high-level policies and guidelines for the model in the spec that we develop for ChatGPT. But the space of situations you can find yourself in is enormous. At some point, establishing the right boundaries really becomes a tough intelligence problem. We are seeing improved results on this metric from reasoning models and from expanding more reasoning on thinking about these software questions and trade-offs. This is a bit more difficult to train for than math problems, for example. This is something that we're researching quite heavily.
