Jerry Tworek: Why One of OpenAI's Top Researchers Walked Away

The person who helped build some of OpenAI’s most important capabilities just walked out the door, not because of drama or burnout, but because he believes the entire industry is stuck in a local optimum and the most interesting work now lives outside the big labs.

The Exit

Jerry Tworek spent nearly seven years at OpenAI, joining when it was a 30-person research lab and leaving as it became a global AI powerhouse. He worked on or led many of OpenAI’s most consequential projects, including the shift toward reasoning-driven models that produced o1 and its successors. This is his first public interview, eight days after departing.

His departure wasn’t acrimonious. He describes it as a natural evolution: when OpenAI was small, he could wear multiple hats, jumping between research directions. As the company scaled to thousands of employees, the work became more structured, more specialized. The kind of exploratory, high-variance research he thrives on became harder to do inside a large organization optimizing for reliable progress on known approaches.

Two structural tensions drove him out. First, the brutal competitive race demands consistently shipping the best model each quarter, which reduces risk tolerance across the board. Second, organizational charts now determine research scope. He invokes the concept of “shipping your org chart”: each team has fixed research boundaries, so output is shaped by organizational structure rather than optimal research direction.

There’s also a paradox of sky-high AI compensation: researchers become more afraid of losing their positions or having a bad performance cycle, gravitating toward short-term, safe bets rather than high-risk exploration.

“When I joined, it was 30 people. You could just walk around and talk to everyone. Now it’s this massive organization, and the nature of the work changes. It’s not that it’s worse, it’s just different from what I want to do.”

The Homogeneity Problem

Tworek’s sharpest critique isn’t aimed at OpenAI specifically but at the entire AI industry. He calls the current state of major labs “extremely, extremely sad” in its homogeneity: every lab is using transformers, training on internet text, doing RLHF, scaling compute. The approaches have converged so completely that 99.9% of users cannot distinguish between models in blind tests.

He frames this through the lens of reinforcement learning itself: the exploration vs. exploitation tradeoff. The industry as a whole has tipped entirely toward exploitation, squeezing incremental gains from known approaches while virtually no one explores fundamentally new directions.

“I am definitely extremely, extremely sad that all the AI labs are trying to do the same thing.”

“If you want to do research that deviates from the ML mainstream, there’s almost nowhere to do it, and that’s the thing I am probably the most sad about.”

This convergence happened for rational reasons. Transformers work. Scaling works. RLHF works. But Tworek argues this creates a dangerous intellectual monoculture. If everyone is climbing the same hill, nobody is checking whether there’s a taller mountain nearby. He points to the early days of deep learning when there was genuine architectural diversity: RNNs, CNNs, attention mechanisms, various training schemes. That diversity produced breakthroughs. Now the field has settled into a comfortable groove.

The Mavericks

Tworek singles out three people as exceptions to this homogeneity, thinkers who consistently pursue genuinely different ideas:

John Carmack is building an AI company from scratch, refusing to start with transformers and instead trying to reason from first principles about what intelligence requires. He’s betting on end-to-end RL in video game environments. Tworek admires the intellectual courage of ignoring what works today to bet on something fundamentally different.

Ilya Sutskever, after leaving OpenAI, started Safe Superintelligence (SSI) with a mandate to pursue approaches that might be radically different from current paradigms. He proposed that “the age of research is ending,” suggesting the big discoveries may already be in hand. Tworek notes that Ilya has always had an instinct for the next big thing, seeing the potential of scaling before most people did.

Yann LeCun continues to publicly argue that autoregressive language models are a dead end for achieving real understanding, pushing instead for world models and joint embedding architectures. While many dismiss his critiques, Tworek thinks LeCun might be identifying something genuinely important about what’s missing.

The Reasoning Revolution: Q-Star and Strawberry

Tworek was deeply involved in the work that became known externally as Q-Star and eventually shipped as the o1 reasoning model. He recounts the origin: while everyone bet on pre-training scaling, a handful of “dreamers” believed you could layer reinforcement learning on top of language models to unlock capabilities that pre-training alone would never achieve.

The key insight was that pre-training provides foundational world knowledge, while RL teaches skills on top of that foundation. The older generation of RL on games (AlphaGo, Dota) failed because models lacked world knowledge; training from scratch produced only “lizard brain” level reactions. Now that pre-training provides high-level concepts, RL can truly deliver.

When early experiments worked, the team genuinely experienced something rare in research.

“Sitting in a room and seeing a meaningfully new technology emerge… I think responsible AI researchers should feel some concern in such moments.”

The internal reaction was electric. People realized this wasn’t just an incremental improvement but a qualitative shift. But it also triggered fear, both about the implications of the capability and about how to control it. Tworek confirms that the board crisis at OpenAI in November 2023 was connected to concerns about this kind of capability jump, though he’s careful to note it was more complex than a simple “the board saw scary AI” narrative.

The Board Crisis, From the Inside

As someone who was at OpenAI during the dramatic board coup that briefly ousted Sam Altman, Tworek offers a measured insider perspective. He describes it as a “soap opera” but acknowledges there were real, substantive issues underneath the drama.

The core tension was between two legitimate positions: the board’s duty to ensure the company was developing AI safely, and the practical reality that removing the CEO of a company moving at that speed would cause massive disruption. Both sides had valid points, and the situation was made worse by poor communication.

What struck him most was how quickly the entire thing was resolved, not because the underlying issues were solved but because the practical consequences of leadership instability were too severe. The same safety and governance questions remain unanswered.

Two Big Bets

When pressed on where the next breakthroughs will come from, Tworek identifies two areas he plans to pursue:

New architectures beyond transformers. He believes the transformer is genuinely limiting in ways that scaling alone can’t fix. Specifically, transformers process information in a fixed number of steps per token, which is fundamentally different from how biological intelligence works. A system that can dynamically allocate more computation to harder problems, rather than spreading it uniformly, could be a step change. “If no one else does it, I will roll up my sleeves and try.”

Continual learning. Current models are frozen after training. They can’t update their knowledge, refine their skills, or genuinely learn from individual interactions. Tworek thinks this is possibly one of the last key capability gaps before AGI. Humans don’t have separate “learning mode” and “answering mode”; everything happens continuously. If models can’t learn from the data they encounter, they will always feel limited.

The intersection of both: doing RL successfully on a well-trained world model. Whoever makes that work first will experience “a very, very happy moment.”

Training AI on Video Games

One of Tworek’s more specific and concrete bets is on video games as a training environment. Games offer something text doesn’t: a closed-loop environment where an AI agent must take actions, observe consequences, and adapt its strategy in real time.

This connects to his broader thesis that the next leap requires moving beyond passive prediction of text. Language models learn to predict the next token, but they never have to act on their predictions or deal with the consequences of being wrong. Game environments force exactly that kind of learning.

He references OpenAI’s early work on Dota 2 and notes that while that line of research was largely abandoned in favor of language model scaling, the core insight remains valid: RL in rich environments produces a different kind of capability than text prediction alone.

AGI Timelines: From Certainty to Nuance

Tworek has become more cautious about AGI timelines, which is notable given his position at the center of capability development. A year and a half ago, he was certain that “scaling RL to the max equals AGI.” He now admits revision is needed: some problems only become visible once you reach the next stage.

Today’s models are already strong enough in coding that “if you showed this to someone ten years ago, they would call it AGI.” But his personal definition still requires at least continual learning and multimodal perception. Timeline estimate: 2026-2029.

“People who predict AGI in two years are extrapolating from a curve that I think is about to flatten. The next S-curve hasn’t started yet.”

He explicitly pushes back against the most aggressive predictions, arguing they underestimate how much of human cognition remains poorly understood and how many fundamental problems remain unsolved.

Evaluations of Major Labs

Tworek offers sharp assessments of the competitive landscape:

Google’s “comeback” is actually “OpenAI’s fumble.” A company that starts ahead should stay ahead with proper execution. Google originally led in virtually all ML directions; OpenAI differentiated through research conviction and directional bets, but later stumbled in certain decisions and execution speed.

Anthropic left the biggest impression of the past year. Claude built an exceptionally strong brand in coding and developer experience, starting from fewer compute resources and a smaller team. “Absolutely amazing achievement. Congratulations.”

Meta isn’t pursuing differentiated models but rather using industry-comparable models to build differentiated products. For a company owning the world’s largest social network, this may be the right strategy.

Is AI Research Star-Driven?

Tworek acknowledges that “both things can be true at the same time”: a very small set of individuals at OpenAI produced breakthrough results that cascaded through the entire industry. However, he rarely sees researchers who switch companies produce equivalent impact at their new labs. The research engine is a company’s culture and structure, not specific individuals.

“Run fewer experiments and think about them harder.”

What matters more is creating an atmosphere of personal responsibility, freedom to explore, and collaboration, rather than depending on star researchers.

The Polish Mafia

A lighter but revealing thread: Tworek is part of a significant cluster of Polish researchers who’ve been influential at OpenAI and other major AI labs. Poland produced a disproportionate number of elite competitive programmers, and that talent pipeline fed directly into AI research.

The competitive programming background trains exactly the right combination of skills: mathematical rigor, comfort with algorithms, and the ability to quickly prototype and test ideas. It’s also a culture that values solving hard problems over publishing papers, which aligned well with OpenAI’s early research culture.

On Risk and Responsibility

Tworek is thoughtful about AI risk without being alarmist. Current systems pose real but manageable risks, primarily through misuse rather than autonomous action. But as systems become more capable, the risk profile changes. A model that can genuinely reason, plan multi-step actions, and operate with increasing autonomy presents different challenges.

“Working at OpenAI was more stressful than starting your own startup.”

His decision to leave a major lab and pursue independent research is, in part, a reflection of this view. He wants to work on approaches that might lead to safer AI by design, rather than bolting safety measures onto increasingly powerful systems.

Some Thoughts

This conversation is valuable precisely because Tworek is that rare figure: someone with deep inside knowledge of frontier AI development who’s willing to speak candidly after stepping away.

• The convergence critique is the most important signal. When one of the people who built the current paradigm says everyone is doing the same thing and the next breakthroughs require something different, that carries more weight than any external analysis.
• His AGI timeline revision from “all-in on 2026” to “2026-2029” reveals something specific: there are capability gaps between RL scaling and true AGI that he hadn’t anticipated, particularly around continual learning. The problems you can’t see from one level become obvious at the next.
• The exploration vs. exploitation frame is elegant and damning. The AI industry excels at training agents to balance these two forces, yet as an industry it has tipped entirely toward exploitation. High compensation, quarterly competitions, and organizational inertia form a system that systematically suppresses risk-taking.
• The video game training thesis is a concrete, testable bet. If RL in rich environments produces qualitatively different capabilities, the labs that invest in this will have an advantage.
• The unresolved governance questions from the OpenAI board crisis still linger. Tworek confirms the safety concerns were real and the structural problems remain unsolved. The industry moved on from the drama but not from the underlying issues.