A Model That Solves New Math Problems Every Week Still Isn't AGI

Jerry Tworek spent seven years at OpenAI, rising to VP of research and helping ship the reasoning models o1 and o3 plus Codex. Host Jacob Effron opens by asking how far the two current scaling vectors—pre-training and reinforcement learning—can take capabilities. Tworek's answer is that they take you somewhere real. The gains, he says, are "real and predictable and nice" for almost every practitioner who has watched a training run.

Scale pre-training and the model knows more about the world, understands language better, and builds a richer linguistic world model. Scale RL and it gets better at acquiring the skills you actually want. In both cases the rule is the same: you get what you train for. If you care about a skill, he says, you just do reinforcement learning on that skill, and "there are basically no limits."

The hesitation starts at the edge of the objective. How do those models perform outside what they were trained for? Knowledge that never appeared in the pre-training corpus? "Probably not." Tasks you never ran RL on? "Probably not that great." Those two gaps, Tworek argues, are the remaining questions in AI, because what we train for we are already getting really good at.

Effron poses the usual fork: keep scaling RL until generalization shows up, or admit you need something new. Tworek recasts it as economics. Scaling means adding data; it does not work much without it. Every quarter, every lab ships a better model by adding compute and, more importantly, data aimed at whatever the previous model was bad at. The loop is powerful. It is also slow.

Keep adding the data for the skills you want and, under today's methods, the model will get good at them, with some generalization along the way. The research question he actually wants answered is whether some method could buy more capability with less data—or, more fundamentally, generalize better from what the model has already seen. Until that exists, iterating the data loop is the reliable path, not the elegant one.

His test for what RL can do today is simple: can you get a signal on quality? OpenAI made real progress training models to be great writers, but judging a book is hard. You might have to sell it and count buyers, and even that is noisy if the marketing failed. Starting companies is worse. You wait five to ten years and still cannot separate early skill from luck. Direct RL on those objectives is nearly impossible.

Coding and math look easy because verification is cheap. Surgery has rules and a clean success criterion—the patient survived. The interesting case is the expert who breaks the rules and saves the patient anyway. Tworek thinks models will eventually learn that kind of violation, given enough time and enough tries. The open question is how long today's models would need to get there.

Tworek used to be the optimist in the room: do RL on the model and you get to AGI. He has lengthened that bet. The latest GPT-5.2 is solving previously unsolved math problems every week, and olympiad-style contests are basically solved. Many people would call that AGI. He notes that the definition is "very personal" and, in some ways, "everything we still don't have."

He is a heavy user of coding models, and they are already an "extreme force multiplier" on work that would take him a very long time. They still make mistakes. The real limit is what happens after failure: you get "hopeless pretty quickly." You paste an error, tell the model to try harder, maybe add words of encouragement. There is no good mechanism for it to update its beliefs from the miss.

"Intelligence always finds a way," he says. "Intelligence works at the problem and probes it until it solves it," which current models do not. Hit a wall and they just cannot solve it—you switch models or do it yourself. That feeling of hopelessness is why he will not call them AGI. The update he made last year is sharper still: a static model can never be AGI. Continual learning is required.

Continual learning first means the model does not collapse. Deep-learning training fails in many horrible ways, and a huge share of lab work is keeping runs "on the rails." It is a fragile process. Human learning is anti-fragile. Humans rarely crash into gibberish after new information; models do it frequently. Why hasn't robustness been solved? The question a researcher should ask first is why not, and which path nobody took. His hypothesis: it likely needs a certain scale, and the few labs that can afford it have been busy with other things—the same "idea whose time wasn't right" pattern that delayed large-scale RL until pre-trained models existed.

The labs have converged, and Tworek reaches for RL's own exploration-versus-exploitation trade-off to explain it. When do you try something new, and when do you squeeze what you already know how to do? There is no closed-form answer, because you do not know the unknown. Someone once asked him why commercial airplanes all look the same: it is the most economically efficient design. The same forces now push labs toward the best model at the lowest price. Customers can switch. In many ways, customers are winning.

Sailing off to train a completely different model would steal focus from making the current stack more efficient. Maybe there is something 10 times or 100 times better out there. That, he says, is a question of belief and conviction. As the landscape gets more competitive, it starts to look like a prisoner's dilemma: doing things differently exposes you very heavily to losing market share.

Effron wonders whether it even matters if the next breakthrough happens elsewhere, given how fast ideas diffuse. Tworek agrees diffusion is good, but the lead from being first should not be discounted. OpenAI won by going harder on large-scale transformer pre-training than anyone else, and that lead made it one of the most successful companies in history. Being first to large-scale RL, he still believes, left it with the best RL research program in the field, able to do more ambitious work while others caught up.

He has been reading about semiconductor manufacturing. Core inventions started in the United States and spread, but countries that bet early compounded advantages for decades. Not every country has a successful semiconductor industry. A window of lead can last, potentially forever, even as some newcomers succeed, some fail, some old firms turn themselves around, and some die. That, he says, is "the Darwinian part of progress."

Leaving OpenAI did not happen quickly. It is a hard place to leave—friends, shared history, a life built there. As a researcher, though, if you wake up and no longer love the work, you cannot do your best. It is "basically impossible" without being 100% excited. For a long time Tworek had infinite enthusiasm and believed he was doing exactly the right things. Near the end, that got harder and harder.

He joined believing reinforcement learning was a necessary element of the path to AGI, and he made it happen. Introducing reasoning models to the world was, to him, a tectonic shift in how models are trained. He wants to chase that high again: find what is missing in how the world trains models, and make it mainstream. After seven years the important problems are not new to him. The question is how to solve them differently, because nobody has solved them yet.

He joined in 2019, when the lab was 30 to 40 people and already extremely ambitious about building AGI. Every year felt like a different company. Executives shifted, headcount exploded, and the research theme flipped from no pre-training at all, to a pre-training company, to largely an RL company, to a more balanced mix. It is a company that keeps reinventing itself. Living through those stages, he says, is an experience that is very hard to compare with anything else.

The pivotal calls that stick with him: releasing ChatGPT, whose virality nobody he heard expected, then GPT-4 soon after, which created a momentum that made OpenAI what it is today. Pulling a lot of resources to train GPT-4 at that moment. And betting that reasoning models were the future with no product-market fit—o1 was smart at puzzles but not practical; o3 plus tool use made it useful for research and coding. Once you see real PMF, humans are excellent at optimizing what already works. Getting to that moment is the exam, and OpenAI passed it.

Ben Thompson called OpenAI an accidental consumer business. Internally, Tworek says, research mostly sits apart from product. The mission is intelligence: one team optimizes product metrics, and the rest of research asks how to make the models smarter. The real tension is opportunity. AI will disrupt everything, so it feels imprudent not to try it all. Companies are bad at doing multiple hard things well. Trying to do everything is, in his view, a very big risk for OpenAI.

The cost already showed up in coding. OpenAI "really lost focus on coding for quite a while" while it chased the consumer product, and that cost market share it is now working hard to regain. Coding models are great again, he says, but lost focus and lost lead definitely have a cost. Building an AI company right now feels like being a kid in a candy store: there is so much valuable work that it is hard to stop yourself from doing all of it.

Why is Anthropic so good at coding? "It's focus." Focus, Tworek says, can explain 95% of why companies succeed. He knew Anthropic's founders from their OpenAI days; they were always obsessed with coding and always believed it was a critical part of AGI. He thinks they are telling the truth when they say very few people at Anthropic type code themselves these days.

He sees two futures. If data is king, it is a zero-sum game: you put data into the skills you want, at the cost of others, and labs specialize because data work is labor-intensive. If research is king, a good idea lifts every domain at once and you leapfrog everyone. He is "pretty sure" something more generalizable exists; the question is how easy it is to find. One slightly pessimistic variant: get coding agents good enough to automate AI research, because maybe this is the last model humans figure out.

The whole story of coding, Tworek says, is programming computers at a higher and higher level of abstraction. Coding agents are a higher-level language with very different semantics from every language that came before. He thinks it is "pretty unlikely" that people will keep typing code themselves. Very few already do, and it is a one-way ticket. Software still has to be reliable. The next problem is certainty that the software does the right things when you are not typing it, and maybe not even reading it.

The most important skill right now, he argues, is being a good manager of junior software engineers. A lot of engineers have been reluctant to go into management and preferred to specialize. The best managers understand the work of their team extremely deeply and still give back some control. That is how to work with models today: understand every trade-off, allow them to make their own choices, and live with the consequences.

Sitting next to the research team is an advantage that companies like Cursor and Cognition do not have. The most successful companies in the world train their own models, and Tworek thinks that is telling. Just as big AI labs become hyperscalers running their own data centers, successful application companies will start training themselves: first a working app, then post-training, then pre-training, then data centers. That is the natural path of a good AI business.

Is catching the labs hopeless given their compute and talent? "Nothing ever is hopeless. The future is not determined." If data matters, differentiate on data. If research matters, a smaller company can still win with better ideas. Innovation comes from constraints: a company obsessed with one domain's failure modes might accidentally make a generally better model and become the next big lab. Competition is hard, and big companies have advantages. New ones still appear. "It's not hopeless. It's just hard."

People are more complex than the models, Tworek says, so he will not generalize about why researchers join a lab. The useful question is not how to appeal to everyone. It is what type of researcher would really want to work here, and then going to find those people. Teams with shared values move faster. Meta's mega-packages, in his reading, were a way to break negative momentum after missteps: if the industry thinks you are not doing well, you cannot hire, and the spiral continues. The new Meta team is training a new model, and the industry is watching.

What makes a great AI researcher? Timing helps. Beyond luck, he wants both systems and theory: how computers work and how networks are trained, plus the theory of networks and optimization. Doing only one well is hard; being at least okay at both makes you "easily 10 times more productive." The second ingredient is independent thought. People converge on a group's median view, which kills research. "If you have 100 researchers that think the same thing, you essentially have one researcher."

Being a researcher means being slightly contrarian all the time, working on something that is not working yet and that people do not yet believe in. What many brilliant, hardworking people miss is courage: standing up to try something most people do not believe in, in a world where machine-learning experiments cost as much as Hollywood movies. You stack advantages, but experiments go into the unknown.

In the quickfire, his biggest one-year update is that a static model can never be AGI. A ChatGPT-like moment in robotics is two to three years out; progress is "slightly better than most people realize." Biology will take longer, because it needs more intelligence and precision. The underestimated impact is widely deployed work automation over the coming decades—societal change is slow, and it will probably be painful. He is not the parent who will demand his young daughters win every math contest; critical thinking stays essential, because even AI will not fully advocate for you. Extinction does not worry him: no one wants humanity gone, and "it's also not a good business to kill everyone." The more realistic dystopia is entertainment so good that people live in VR, Ready Player One plus WALL-E. That one, he says, only humans can solve.