Nate B Jones
OpenAI’s 700 Math Papers Look Like Superintelligence. That Doesn’t Mean Your Job Is Next.
Ask the tutor about this video
Sign in with a verified email and ask the tutor anything about this video.
TL;DR
Research summary
▶ The math results and the central question
The speaker opens by saying OpenAI released more than 700 math research papers and that almost every result came from a single prompt using an unreleased model. They question whether this demonstrates general superintelligence, arguing that AI’s strong mathematical performance has not meant that coding agents can be trusted near production. They describe engineers using evaluations and zero trust environments, and frame the gap between verifiable coding and real engineering work as relevant to understanding AI’s effects on jobs. They preview a discussion of the mathematical results, narrow superintelligence, work, and skills.
▶ What the mathematical research contributes
The speaker distinguishes a math exam, where answers are known, from mathematical research intended to establish results others can build on. They describe one result as improving a theoretical bound for matrix multiplication, which AI systems use extensively, and another as proving that, for a particular class of elastic materials, interior information can be learned from surface measurements. That result could help inspection without cutting an object open and motivate further work with incomplete or noisy measurements. The speaker says mathematical discoveries may eventually contribute to cheaper computing, better inspection methods, and improved software, and that AI could speed the path from theory into engineering. They say the model remains unreleased and inaccessible to them and the audience; according to OpenAI, each described result took about three hours of thinking. The speaker calls that superintelligence.
▶ Why mathematics may suit narrow superintelligence
The speaker defines narrow superintelligence as exceeding human abilities in particular intellectual tasks while other abilities develop unevenly. They suggest mathematics is especially suitable because success can be stated clearly and correct answers can be recognized, even when reaching them requires difficult reasoning across multiple branches. The speaker says OpenAI reported expanding to open research problems because existing mathematical evaluations were saturated, and describes the resulting tension with mathematicians whose careers may have centered on unsolved problems. They say there is no evidence OpenAI set out to hurt mathematicians, while noting that a lab’s need for hard tests can collide with researchers’ long term work. They caution that mathematical breakthroughs do not establish how well the same capability transfers to engineering judgment or business intuition.
▶ Coding reveals limits in transferring ability
The speaker argues that coding is a useful test of transfer because labs have invested heavily in it, commercial demand is substantial, and code can be run and tested. Despite AI being very good at parts of programming, the speaker says engineers still need zero trust environments, evaluations, restricted access, bug finding, and attention to maintainability. In a document search example, tests and useful answers could appear successful while a revoked permission might leave an old index or saved answer exposing private content. The engineer must identify and trace that risk; knowing what question to ask is part of the job. Experienced engineers also handle interactions with older code, changing assumptions, and maintainability over time. Faster feature generation can increase work for the small number of engineers able to check and maintain the result. The speaker says AI is useful for code generation but does not currently provide the same value in designing maintainable systems at architecture scale, and suggests gains in mathematical reasoning do not transfer one to one to systems engineering.
▶ Business value depends on discovering the real problem
The speaker extends the search example to business use. Even after the feature’s privacy issue is fixed, customers may find conflicting documents and be unsure which to trust. Improving search further could leave the underlying problem intact; someone may need to determine who owns the information, which version is current, how it stays accurate, and whether workflows across teams must change. The speaker says the person who discovers this blocker contributes business value because the discovery can change what the company decides to do. They argue the missing information could not have been included in the original instructions before the problem was discovered, and say they have seen similar situations repeatedly. In their view, discovering the question is often more important to business intuition than anything else, and they say there are not many examples of LLMs doing that. Faster development helps when it lets a customer try something sooner and the team learns; producing features quickly without discovering why the customer is stuck can waste effort.
▶ Uneven capabilities and changing work
The speaker asks how people in technical and business work can use AI that is especially good at generation but may be weaker at tasks people expect from a generally superintelligent system. They say it remains an open question whether this will change over the next few years, and argue that superhuman ability in one area does not guarantee immediate transfer elsewhere. They point to context dependent business intuition and system design judgment, and say AI has not improved as expected at challenging assumptions or reframing questions, which the speaker attributes in part to training that emphasizes helpfulness. They say AI is changing engineers’ roles and that more engineers are not touching code directly, but conclude this does not mean they are out of work; it means they are changing how they work. The speaker expects the math results to change mathematical practice and the movement of theory into products, but says mathematical ability does not automatically provide business judgment or system design.
▶ Skills, tool use, and human agency
The speaker advises against stopping domain learning out of fear that AI will soon become equally capable in every area. They recommend combining experience, passion, and personal aims with a chosen area of impact, treating AI as a powerful tool and learning how to use it. The speaker says that focus on business impact helps determine when to use AI and when not to trust it, and cites engineers, junior engineers, and sales teams as examples of people working through how to apply AI effectively. They emphasize that people have agency in choosing where AI is used in business and personal life and in judging whether it succeeds. In their view, AI has to work within the messy conditions of human workplaces, databases, and CRMs.
▶ A milestone with limits and an uncertain next step
The speaker calls the mathematical results a milestone in AI and mathematics, expresses excitement, and empathizes with researchers who spent years on problems that are now solved. They urge a clear eyed view of narrow, uneven superintelligence and say mathematical and formal logic strengths do not imply that AI can immediately design engineering systems without guardrails, produce highly persuasive prose, or understand business judgment. They argue practical applications depend on making and following through on good decisions, not simply on higher IQ, and refer to Alexander Wang’s comments about execution and focus. The speaker says a math result cannot remove the human desire to make an impact. They describe the present as a world of narrow superintelligence, say it may offer some comfort because the areas where it appears reveal what labs are focusing on, and speculate that if labs have to choose problems such as Navier–Stokes, they may be running out of problems elsewhere. They close by saying they do not know what comes next and invite comments on the math results.
Full transcript
[0:31] coding agent near production. In fact, most of them don't. They put a bunch of evals in place and they have a zero-trust environment. That difference should be something you pay attention to because isn't code checkable? Every lab is pouring money into models that can code. There is easy verifiability here. Does it run or does it not? If the model is superhuman at math, you would think it would be crushing your codebase by now. And it's not. That gap is the most useful thing that you can understand
[1:01] about AI right now because it tells you where your job stands and why. So, today we're going to look at why you actually ought to care about math, even if you didn't take calculus, what we learn from all of the mathematical insight this week, and what this doesn't tell us about work. Because I think we're looking at a world with narrow superintelligence and some of the consequences of narrow superintelligence are very surprising. And we're going to follow that into code, we're going to follow that into business work, and then we're going to
[1:31] talk about your skills. What's worth learning in this world? How do we use the abilities these systems give us? And how do we respond and get better at things that we actually care about? So, let's start with what these papers actually contain. We're going to get a little nerdy, but not too nerdy. On a math exam, someone already knows the answer, right? In mathematical research, you're trying to establish something that people haven't actually figured out yet. So, a useful result becomes something that other researchers can
[2:01] build on. Eventually, it can change the methods available to engineers and the products that we all use every day. Let's take something super nerdy like matrix multiplication. Maybe you learned it in high school, maybe you didn't. That's the arithmetic involved in multiplying grids of numbers and AI systems do piles of it. And one of these results improves the theoretical bound on how much math you need as these grids grow. Another nerdy result concerns learning what's inside a material from measurement taken at the surface, like
[2:32] measuring the surface of a beach ball and learning what's inside. For particular class of elastic materials, the proof basically says you can do that. You can measure the surface of the beach ball and learn what's inside it. If you're trying to inspect an object without cutting it open, that's super useful knowledge. It tells you the information you need is present in the surface measurements. And now you can work on recovering it under real-world conditions where measurements are incomplete and noisy. And there is reason to care about this, even if math is like your least favorite subject. The
[3:02] discoveries are going to reach you eventually through cheaper computing, through better ways of inspecting things, through improved software. And so, a model might improve the tools available in your industry with discoveries like this math moment, but that doesn't mean it's an expert at running a business in your industry, right? In other words, the discoveries from the math papers today may get all over the world, but that doesn't mean that magically speaking a model that can solve these math papers is going to transfer its judgment everywhere we work
[3:33] day-to-day. But one of the things that I want to say really clearly is that even though this feels a long way away from day-to-day, it may get there surprisingly fast precisely because we have AI. In other words, AI may make it possible for us to pull this theory into engineering, into the kinds of technology that we can use day-to-day much faster than we would have 10 years ago. And so, that's part of why I'm excited.
[4:03] And I think that it gives us a chance to actually improve upon our world in probably hundreds of ways all in one day. And the model that produced all this, it's still unreleased. I don't have access to it, you don't have access to it. What researchers have in their hands is just the work this model produced. And what OpenAI reports is that this model is so powerful that any given result that I just described only took about 3 hours of thinking. So, the model did something in 3 hours with some of
[4:34] these problems that mathematicians can't do their whole lives. That's superintelligence. There's no other way to describe it. And so, what can we learn here? Why do we care? Let's set aside the math piece, let's pretend we don't care for a second. What can we learn about AI and ourselves? I think narrow superintelligence is the right way to understand what we're seeing. It's a system that can exceed human abilities in very particular kinds of intellectual work, while the rest of the abilities develop very unevenly.
[5:06] And that makes the math piece especially interesting because you can state exactly what success would mean and then spend your entire life trying to achieve it. And the question can be perfectly clear along the way, while the reasoning required to get there is extraordinarily difficult, convoluted, takes mastering multiple branches of mathematics at once. When you train a system by rewarding correct answers, being able to recognize a correct answer helps a lot.
[5:36] And math gives you that possibility alongside problems that get extremely hard. And my suspicion is that it is unusually well formulated for the kinds of superintelligent reasoning gains that we are seeing right now. OpenAI says it expanded into open research problems because existing mathematical evaluations were saturated. The tests just weren't hard enough anymore. The lab needed harder questions, and those questions were things that people had built their careers around and had not yet solved. And that helps explain the
[6:07] collision with the mathematics community. I don't think you need to imagine a world where OpenAI is setting out to hurt mathematicians. We have no evidence of that. The researchers at the lab needed to find out how capable their model had become. They had to reach for difficult problems, and to somebody else that problem represents years of work in the contribution they hope to make. And to the lab, they're just looking to find something hard enough to test the model on. But first, there's a big assumption here that I think we need to examine about our own work. A breakthrough in
[6:38] mathematical reasoning doesn't tell us very much about how the same capability transfers to other areas of work, like engineering judgment or business intuition. And I'm going to pause on the engineering piece cuz I think coding is one place that this gets really interesting that we haven't talked about very much at all. Coding has obviously had a tremendous amount of investment from the labs. There's enormous commercial demand for a machine that can code better. You can run code, write tests, check whether it produces the right answer. So, if verifiable rewards
[7:09] and enough investment were enough to get to superintelligence, coding is exactly where you would expect overwhelming machine superiority to show up first. But we don't see that. AI is already extraordinarily good at parts of programming, yet I've been in conversations all week about systems engineering required to make AI-generated software work. We're talking about zero-trust environments, restricting what the AI can access or change instead of assuming it will behave correctly. Evals, evals, evals.
[7:39] We're talking about finding bugs and keeping the code sustainable as the system grows. And the engineers putting guardrails in place are putting them in place because these are senior engineers who know better. The human contribution is the thing making the value here. And we don't get to attribute all of that value back to the model just cuz the model ends up producing a bunch of code. Here's a simple example of the kind of problem I'm talking about. Let's imagine a company asked an AI to build a feature that lets employees search documents. So it builds the feature. The
[8:10] answers are useful. The tests pass. It's verifiable. Then an engineer asks, what happens when someone's permission to see a doc gets revoked? Could an old search index or a saved answer still reveal the contents on accident? If the answer is yes, well the company has a privacy problem, right? The feature looked successful because nobody had tested the condition that really mattered when it came to risk and security profiling. The engineer ended up recognizing that condition. The engineer followed it through the system. The engineer knew
[8:40] what needed to change. You can write a test once you know how to ask the question. Knowing to ask the question in this example is part of the intelligence required to do the job. And so is understanding how that answer might depend on other parts of the system beyond the code the model had just written. And this is why I find coding such an important challenge to the idea that we're all on the same path to superintelligence. Ask experienced engineers whether they've run out of difficult problems. They haven't.
[9:10] They're dealing with changes that interact with older code, assumptions that stop being true, and decisions that determine whether a system will still be maintainable in 6 months. A coding benchmark can stop being a useful test, and you can saturate that out, and we have. Even though there is still plenty of very, very difficult engineering that only humans are able to do at this point. And that distinction ends up mattering a lot because generating a feature faster is valuable, but generating more features that a small
[9:41] number of engineers understands how to check and maintain, that just creates more work for those people, right? That happens to become the whole job of engineering, and that's exactly what we see all over the industry right now. In other words, engineering is a place where we would have expected super intelligence to show up because super intelligence should show up in a verifiable domain where we've put a lot of investment, and you should be able to make something that feels like a superhuman engineer. But we don't see that. In fact, we see
[10:12] the opposite. We see engineers having to guardrail around the weaknesses of AI even as AI gets faster and faster and faster and faster at producing code. It's not that AI isn't tremendously useful in engineering, it is. It's just that the utility is around the generation of code, and it is not around designing maintainable code that you can scale at the systems architecture level. Now, I know there are folks at the labs that will say that's just a matter of time, and we're going to get it done.
[10:42] But so why didn't it show up? Why didn't it show up? I think it is because it is difficult to take and transfer gains in IQ and logic that you see in things like mathematics and immediately put them into something like systems engineering. It It does not seem to translate one-to-one. It does not seem to transfer in the same way, and we see it even more when you move farther outside a technical domain, farther outside a formally logical domain into business. And if you follow the same
[11:12] thread into the business story, this gets even more relevant to us. Let's suppose the search feature I just described got fixed by that engineer. It now works beautifully. You could sit with a customer. You can watch them use it. You can watch them find documents that disagree, and the customer doesn't know which one to trust, and the customer or to make a decision, and now you've got a different problem, haven't you? You have now made it easier to find information that is conflicting, that was always in your repository, and you've exposed a bunch of discrepancies
[11:43] your customer has to sort out. Now, in our example, you can choose to spend another month improving search and leaving the problem intact, right? Someone needs to work out who owns the information, which version is current, how it will stay accurate, and you can go through all the business rules, you can get a PM to do that. That might mean changing the way people work across a bunch of teams, and it becomes a different project by the time you're done. The person who discovers that problem has contributed something immensely valuable to the business cuz they found a blocker for the kind of realized value
[12:14] that the customer will pay money for. They found out what was happening in the customer's work and changed the company's decision about what it took to fix it. You couldn't have put that understanding into the original instructions because you hadn't discovered it yet. Getting the missing information is a part of the work, and I know you're telling me, "Oh, this was incredibly obvious." But it's a simplified example of a story I've seen personally play out time and again. Discovering the question is often more important in business intuition than anything else.
[12:44] And I think it's interesting that we actually don't have a lot of examples of LLMs doing that. And this has consequences for how a business uses AI. In our example, the faster development is only useful because it gets you to something a customer can try sooner. If you use that opportunity to learn, you can absolutely go faster overall. You can correct the project sooner. If you keep producing features without discovering why the customer is stuck, you can go real, real fast and waste a
[13:15] whole lot. And so, for all of us who are not mathematicians, who are stuck in this business world, who are stuck in this technical world, what we are trying to figure out is how to work with an AI that is extremely good at generation. I described the coding example cuz we're all familiar with it at this point. But maybe not as good at some of the things that we would expect a super intelligent AI to be good at. And I think one of the great questions of our time is whether or not that is about to
[13:45] change. And the good news is we're all going to live through the next couple years and we'll find out. But one of the things we can take away from this moment in super intelligence is that it is not a foregone conclusion just because something is super intelligent in one area that it will immediately transfer to other areas. And I think the reason is very simple. Not every problem is a function of intelligence. There are things like business intuition, things like good judgment around system design.
[14:17] There are things that are context-dependent in business that it is very very difficult to get AI to be extraordinarily good at. And there's even things where you would sort of expect progress on this, but you don't see it with AI yet. And a good example is you would expect AI to be better at challenging your assumptions. You would expect AI to be better by this point at restating questions and helping you frame them better. But because we have put so much investment into training the AI to be helpful, we've
[14:48] kind of pushed it in the opposite direction. It's not actually getting better at challenging your assumptions over time. And so, that should be encouraging because if you look at the stories like the one I described, those are actual stories I see playing out over and over where people are adding a tremendous amount of value. And is AI changing their role? Yes. Is AI making it so more and more engineers aren't touching code directly? Yes. But does that mean that they're out of work? No. It means that they are changing the way they work because they
[15:19] have a new tool to use. And I think when we look at the maths example, what we see is this is going to change the way mathematicians do maths. This is going to change the way we think about pulling maths from theoretical into engineering problems, into products. But [snorts] it's not necessarily going to immediately mean that it's immediately going to know lots about business judgment and system design. It doesn't transfer that way. The IQ points don't spend that way. And this is why I don't think the lesson
[15:49] is that you can stop learning what you're learning in your domain because you're afraid AI is going to be that good tomorrow. And so if you're trying to figure out, well, where do I spend my skill points, right? Where do I learn that is going to be unregretted in this world? I would encourage you to think about how you can take your experience, take your passion, take what you want to do, and really focus on an area of impact that you can drive, and look at
[16:19] AI as if it were a supercharged tool that you can use in that direction. You absolutely want to learn how to use it. You absolutely want to learn how to go after it, but your focus on the business impact you're driving is what shapes when you use that tool. It's what shapes when you put down that tool and say I don't trust it. And that is a skill that I see happening in multiple areas of the business right now. That's a skill I see when I talk with engineers. It's a skill I see engineers wishing junior engineers had.
[16:49] It's a skill that I see sales guys trying to figure out as they think about how to get a sales engineered pipeline to actually set them up with the right conversations at the right time to close deals. And I think one of the things that I want you to take away from this as you think about this moment is that we have a surprising amount of agency in choosing how to use AI in the business world. We have a surprising amount of agency of choosing how to use it in our personal lives. We get to choose where
[17:20] we want to apply this tool and where we don't, and we get to say where it's actually successful and where it's not. And one of the things that I love about that is that we are making AI conform to the human condition. We are saying if AI is going to exist in our world, it actually has to exist in all the mess of what it means to be human, all the mess of our workplaces, all the mess of our databases, and our messy CRMs. It has to actually work. And so, I get that it's really cool that
[17:51] we can get to superhuman mathematics results. It is a milestone in artificial intelligence and absolutely a significant milestone in the history of mathematics. I'm excited, and I also have a lot of empathy for the folks who worked their whole lives on these problems, and now that they're solved, like that's a tough place to be. But in the midst of all of that, the thing that I keep remembering is that I want to have a clear-eyed view of what it means to be narrowly superintelligent. I don't want to assume that just because we have an AI that
[18:22] sort of has this really spiky edge around mathematics and around formal logic, doesn't mean that it's got an imminent ability to design engineering systems in ways that we don't need guardrails for anymore. Like, no. It's not how it works. It doesn't mean that it's imminently able to produce really, really, really excellent prose that we find incredibly [snorts] persuasive. I know the internet is full of LLMisms, and we've been working on that for a long time. It It doesn't mean that it's
[18:54] imminently able to understand business judgment because these things are not necessarily a function of IQ points, right? And I think that sometimes the labs have been chasing IQ points. And what we're finding is a lot of our practical applications, it's not a function of whether you have 150 IQ versus 140. It's about whether you can actually make the right decision and follow through on it. One of the things that I've been thinking about a lot was actually a remark that Alexander Wang made as he was talking about Muse and kind of what they've been focused on. And he talked about the importance of
[19:24] execution. He talked about the importance of going after something and really focusing on it and getting it done. And I thought about that because anyone who's been in the business world will tell you that kind of focus, you can't trade that off. You have to have the fire in the belly. You have to go after something. It's not a function of IQ, is it? It's not a function of superintelligence, quote unquote. It's a function of the human desire to go and make an impact in the world. And that, you can't take that away. A math result won't take that away. And so, I'm not sure what tomorrow will bring, but it looks like we're living in
[19:55] a world with spiky narrow superintelligence, and we should plan accordingly. And I think in some ways it should give us more comfort than you might think. Because the place where it appeared tells us a lot about what the labs are actually focusing on and what they're able to get out of these models. If they have to pick a problem like Navier-Stokes, it's because they're running out of problems everywhere else. So, what's next? I don't know that they have a perfect answer to that.
[20:25] We'll see. I'll be buckled up along for the ride along with you. Let me know what you think of the math results in the comments.