Jerry Tworek gave both a keynote and a long-form interview at the AGI House Auto-Research Summit & Build Session on July 18th. This piece draws on both. Quotes are lightly edited for clarity; watch the recordings for the full conversations.
The résumé compresses into two sentences. Seven years at OpenAI, where he led the team that taught language models to think — o1, o3, and the reasoning inside GPT-5 — and was a primary contributor to Codex, the model behind the original GitHub Copilot; before that, a robot hand solving a Rubik's Cube; before that, five years extracting signal from noise at an Amsterdam quant firm. In January he walked away from one of the most powerful research posts in AI. In April he launched Core Automation, with the explicit goal of building the world's most automated AI lab.
At AGI House, the through-line of everything he said was a heresy against the measurement regime his own career helped build: "The era of evals is dumb."
Paranoia as method
Ask him what trading taught him that ML researchers never learn, and the answer isn't math — it's fear. A trading algorithm is a pile of code managing a pile of numbers, and when something goes wrong you don't get an exception; you get a wrong number that buys instead of sells with large sums of money. "Being a good trader means understanding your algorithm very well — knowing whether it's doing what you intended, because in the end all you see is the P&L."
He carried the paranoia into research: double- and triple-checking everything inside the network, the learning algorithm, the hyperparameters. And he offered it as advice for the automation era, which is arriving fast enough that the advice has teeth: "Models can do a lot of your work, but there's no way to outsource your understanding. It's good to know what's inside the box."
The interviewer noted how many frontier researchers share the trading pedigree — DeepSeek grew out of a quant fund — and Tworek recalled his own 2018 fork in the road: apply RL to trading, or join a lab and apply AI to research. He picked the lab. Both paths, he noted, turned out to work.
Why code came first
The decision to point language models at code — the bet that became Codex and HumanEval — reads as obvious now. In 2020 it was a contrarian answer to a specific anxiety. GPT-2 had arrived wrapped in "too powerful to release" discourse; GPT-3's few-shot prompting felt groundbreaking internally. Tworek's reaction was that the models were vague: they'd regurgitate a few coherent sentences from everything-in-the-world training data, and you couldn't tell whether you'd hit actual thought. "It's easy to talk in hyperboles around a topic. It's very hard to write a program that is true. You cannot write a correct program if you don't understand the concepts."
Code was the pinning-down mechanism. The origin of HumanEval is charmingly manual: Tworek and Wojciech Zaremba sat in a room and hand-wrote about thirty problems each, just to see what the models would do. And the seed of everything that followed was already planted — even then, the team was measuring how performance scaled with test-time compute.
The prophecy, and the GPU catch-22
The best material of the interview is first-draft history. At one of Tworek's earliest research all-hands, around 2018, Ilya Sutskever stood on stage and stated the roadmap: train a large generative model, then do reinforcement learning with it. Tworek's gloss, seven years later: "This was extremely precise. It is exactly what we are doing today."
What the prophecy didn't come with was compute. RL believers were a minority inside OpenAI — "I was RL-pilled, but for many years we were the minority" — and the leadership's position was reasonable: prove scaled RL works before we give you GPUs that could be training GPT-4 or GPT-5. But the proof required the GPUs. A catch-22 that lasted years.
It broke, in his telling, the day Jakub Pachocki came to him and said they were allocating GPUs to scale up reinforcement learning. "It became a self-fulfilling prophecy. We had the GPUs, so now I had to figure out how to organize: we need data, we need problems, we need an algorithm that has a chance of scaling, we need a babysitting rotation for runs no one had ever done before. I assumed success and started putting everything in place — and things started clicking." There's a research-management lesson buried in that sentence: sometimes the proof follows the allocation, not the other way around.
The technical distillate, for those keeping score: multiple rollouts per prompt to build a baseline ("a really powerful idea"), obsessive numerical stability, and stripping PPO down rather than building it up. "Simplicity is what makes scaling work."
What's actually happening when o1 "thinks"
Asked whether the model's characteristic "wait, let me reconsider" is engineered, his answer was unambiguous: it's emergent, completely shaped by RL, with zero reward design behind it. The model has seen humans reconsider throughout pretraining; during RL it recalls those patterns "almost like a function call within the chain of thought," and the ones that lead to correct answers get reinforced. Backtracking isn't a feature someone built. It's a behavior that survived.
The jump from o1 to o3 — from puzzles to tool use inside the reasoning loop — he described as "99% systems, 1% algorithms." Training tool use at scale meant operating one of the largest distributed systems ever built; he claims OpenAI's training framework moved more data across more virtual machines than Netflix or Amazon, and his team built it in months. "The engineering part there was brutal."
The limits, stated by the person who found them
Two admissions stood out for their candor.
On serial scaling: his teams never found a meaningful ceiling on test-time compute. The constraints are practical — when a rollout takes a day, "there are only so many days in a year" to iterate — plus the imperfect long-context generalization of transformers.
On parallel scaling: sub-agents, the industry's current answer to the latency problem, don't work well yet. Token usage is much worse, and "people complain that their Claude or Codex usage disappears on sub-agents that weren't that useful." Could be an architecture problem, could be an RL problem — "we don't see as much progress here as we would like." When the person who built the paradigm tells you where it's stuck, that's the part worth writing down.
The innovator's dilemma, from inside
His description of fitting reasoning models into ChatGPT: "a triangle peg in a square hole." ChatGPT's enormously successful form factor — fast, reasonable answers to small problems — was simply the wrong vessel for a model that thinks for minutes. And then the line you don't expect from a former OpenAI research lead: Claude Code was "maybe one of the first good agentic products," arriving while OpenAI "was trying to shoehorn agentic models into ChatGPT, and that didn't make sense product-wise."
The point isn't score-keeping; it's that product form factors lag capabilities, sometimes by years, even inside the lab that built the capability. "It shows how important it is to innovate not only on the models, but on how the usefulness gets into people's hands." Every reader with a roadmap should sit with that.
The bet: learning has to move to test time
Which brings us to why he left. (His full eval critique — and Pushmeet Kohli's counterargument — is in our flagship memo; the short version is that benchmarks get maxed within two or three model generations, and his proposed replacement is the trader's regimen of causal backtests plus live deployment. "The only test is the real world.")
The deeper claim is architectural. Transformers, in his view, memorize at training time and retrieve at test time: "All the learning happened in the lab." His thought experiment: how useful would today's models be if OpenAI and Anthropic shut down their training clusters tomorrow and said, this is it, use what exists? The industry's own vocabulary gives the answer away — "freshness fine-tunes," knowledge cutoffs — symptoms of models that cannot learn where they live.
He's equally dismissive of the scaffolding workaround. Markdown-file memory, retrieval, context tricks — everyone he talks to reports the same thing: brittle, doesn't generalize. His analogy is pointed: it's AutoGPT and LangChain in 2023 all over again — people scaffolding agents onto GPT-4, a model never trained end-to-end to be an agent, and discovering it was a model problem. "Those are machine learning problems, not markdown-file problems."
How do you study a system like that? Not, he thinks, through mechanistic interpretability — "I don't think we can interpret humans, and models have a trillion parameters." He treats models behaviorally: "like rats in a maze," or, in his warmer version, like an Olympic swimmer whose every minute gets analyzed — technique, energy, where it's wasted — and coached toward the best version of itself. Core Automation's method follows directly: use existing agents as the baseline for every task in the company, then make the lab's own operation the eval you hill-climb. "Not improving evals you don't care about — the thing you actually want to do."
The enterprise consequence he's building toward: your own deployment, your data never leaving, a model that becomes genuinely yours because it learns on you. "Continual learning is solving personalization the way it should be solved." As for the current alternative: "I don't think anyone actually wants to be fine-tuning open-weight models." (He declined to confirm reported details of the company's first model, so neither will we.)
The high-school future
His timeline is short and he doesn't hedge it much: "I'm not joking when I say we probably have a few more years left of work at all." Within roughly five years, he expects work to transform beyond easy recognition — while acknowledging the strange duality of the moment: "I already feel like I'm living in the future. I am mostly vibe coding and mostly self-driving. And at the same time, our lives look only a little bit different than they used to."
He knows how the endpoint sounds, and offered his own image for it anyway. People still play chess though computers win; people still run though cars are faster. "Sometimes I say, only half joking, that the future will look like high school. No one yet has to work, no one has real responsibilities, and people just learn about things and spend time socially and try to figure out the world — with AI as this teacher, always there. It does sound dystopian to some people. But I generally think it's where our future should be."
Part of our Auto-Research Summit series from the AGI House build day. Start with the flagship memo on the day's big debates, see the builder recap for the 18 projects, and read the pre-event research brief for the technical landscape. Watch Tworek's keynote and full interview.