Jerry Tworek: humans have "at least two years" left in AI research

Jerry Tworek, who spent close to seven years at OpenAI and left as a VP of research to found Core Automation, sat down with MTS hosts Theo Jaffee and "schisofrenia" for a 46-minute interview titled "Most AI Labs are Simply Cooked". The line that traveled was his estimate of how long human beings remain useful to AI research: "We still have some time. My rough estimate is at least two years for when human researchers are a meaningful part of the AI research discovery process."

Worth reading that sentence carefully. Tworek gave a floor, not a deadline — at least two years — and MTS's own promotion of the clip flattened it into "two years before humans are vestigial in AI research." The mood he describes is real either way: "What you often hear AI researchers repeating right now in the field is, 'Oh, we have a last few days of work, so let's work while we still can, and then we'll all take a break.'" He conceded the pace is costly — "very exhausting, very tiring to be working at this pace for so many years" — and argued it is "probably worth it if we think these are the most important times of our career and our work."

#Why today's agents are not researchers

The more interesting half of the answer is the part the headline drops. Tworek is explicit that current systems are nowhere near independent research, and his diagnosis is about quality rather than volume of ideas.

"Anyone who has used models for doing research has seen that agents generate pretty low quality insights. They are high creativity, but high creativity in a very low quality way in which the ideas are generated. They are diverse. They are generally not very good."

What is missing, he says, is accumulated depth: "There is no such thing yet for a model to represent humans who have spent 10 years researching AI technologies, who have this level of depth of understanding of the objects we are really working with here." Closing that gap requires agents with "a much higher level understanding of a research field and research directions" — which is precisely the capability nobody has demonstrated. In the episode's framing, the agents execute and the humans still generate the insight.

#The bet against the transformer

Tworek left OpenAI in January 2026 — after leading the o1 and o3 reasoning models and building the original Codex — to start Core Automation, pitched from day one as "the world's most automated AI lab," starting with automating its own research. The lab has pulled researchers from Anthropic and Google DeepMind, including former Anthropic and Google Brain researcher Rohan Anil, Gemini research scientist Anmol Gulati, and ex-OpenAI GM Joanne Jang, per Business Insider. Tworek's stated reason for leaving was that this kind of fundamental research was no longer possible inside OpenAI.

The thesis he and Anil laid out on Sequoia's podcast in July is that the architecture, not the compute or the data, is now the constraint: "I think at this moment, what the bottleneck is to better models and to smarter systems is the architecture itself," and "we need to have models that learn at test time... with users on their data, on their real-world tasks, on the real world distribution." He is not dismissive of what transformers achieved — "the majestic thing about Transformer... is that Transformers are economically valuable" — but he thinks scaling RL inside that paradigm did not deliver the general real-world competence he expected it to.

The MTS episode's other chapter markers point at claims the video does not summarize in text: that a lab without a top-tier compute footprint dies, that roughly ten companies had a shot at Anthropic's position, and that only 30 to 50 people in the world understand a frontier model end to end.

#Why it matters

Tworek is an unusually well-placed source for this forecast and an unusually interested one: he is raising money for a lab whose entire premise is that automating research is both imminent and impossible on the current architecture. Both halves of his statement serve that pitch. But the two-year figure lands in the same week OpenAI reported 3.1 agent-workdays of research effort for every human workday and said it is targeting a full automated AI researcher by March 2028 — a timeline that, read as a floor, is not far from his.