OpenAI's former top researcher says Google caught up because OpenAI stumbled

Jerry Tworek joined OpenAI in 2019 when the company had around 30 employees. He worked on many of its most important projects, including the reasoning approach that became known as Q-Star and Strawberry before eventually becoming the o1 reasoning model.

He's now left the company, and in a podcast interview with Core Memory, he explains why: he wants to pursue risky fundamental research, the kind that's no longer possible at a company like OpenAI, where metrics like user growth take priority.

His take on ChatGPT's advertising shows the disconnect between research and commercialization: "It's a business strategy, and I work on training models." The comment supports rumors about a growing rift between AI research and product development at OpenAI.

But Tworek's criticism extends beyond OpenAI. He says all the major AI labs are doing the same thing: optimizing the same Transformer technology with products that barely differ from each other. Most users can't tell the best models apart, and anyone who wants to work outside the machine learning mainstream will struggle to find a place.

Too much money kills the willingness to take risks

Tworek points to several reasons for this lack of innovation. The competition for the best model is brutal, and companies need to show constant strength to hold onto users and justify GPU costs. Rigid organizational structures make things worse when org charts determine what research is possible: Teams work in silos with defined responsibilities, making cross-team research hard to pull off, Tworek explains.

Financial incentives also play a role. High compensation in the AI sector, which Tworek describes as "pretty crazy these days," means researchers don't want to risk their jobs, so they chase short-term wins instead of risky bets.

On that note, OpenAI reportedly recently eliminated a holding period for new employees' shares. The move is likely driven by competition for talent but could also encourage risk-taking. Researchers with less to worry about losing their shares have, at least in theory, more freedom to pursue unpopular or risky research directions.

New architectures and continuous learning as next frontiers

According to Tworek, the industry committed to Transformer architecture six years ago and has been scaling it ever since. This works, but it's not the only path forward. While there are still improvements to be made to the Transformer architecture, he finds entirely new architectures more interesting. This is one of his potential research arcs.

The second is that for a breakthrough toward AGI, continuous learning needs to be cracked first, Tworek says. Humans don't have a separate learning mode, everything happens simultaneously, and as long as models don't learn directly from data, their capabilities will remain limited. Tworek describes continuous learning as one of the last important building blocks before AGI. The topic comes up frequently in the AGI debate, including Sutskever's recent turning point diagnosis.

With his views, Tworek would probably fit well at Google Deepmind or Yann LeCun's new AI startup, even though the latter isn't an AGI believer. The best fit might be Sutskever's SSI, and Tworek even quotes his former boss's views in the podcast. He also suggests he has plenty of opportunities.

Tworek also speaks about his AGI timeline, which has shifted slightly since scaling RL, the training method behind reasoning models, didn't turn out as mighty as he originally believed. Initially, he thought it would lead directly to AGI. Now he sees additional requirements: continuous learning and multimodal perception. But still, Tworek thinks AGI will likely arrive by 2029.

Google caught up because OpenAI stumbled

Tworek sees Google's successful race to catch up with OpenAI as OpenAI's failure. The AI lab made mistakes and moved too slowly, not managing to stay ahead with all the head start it had, Tworek says. And Google at the same time also made a lot of the right calls. When asked about specific problems at OpenAI, Tworek stays vague but suggests that staff departures are sometimes symptoms of deeper issues.

The AI researcher is particularly positive about OpenAI's fiercest startup competitor, Anthropic. The company has impressed him over the past year: with less compute and a smaller team, it has shown focus and strong execution, Tworek says, adding that what Anthropic has achieved with coding models and coding agents is remarkable. The company has built a strong brand and won over a lot of developers as satisfied customers.