Ex-OpenAI VP: human AI researchers have 'two years' left in the driver's seat
The man who built OpenAI's o1 and o3 reasoning models now runs a lab whose stated goal is to make researchers like him obsolete — and he has put a number on how fast it will happen.
Jerry Tworek, the former OpenAI research vice president who led the o-series reasoning models, said in a recent interview that human researchers have roughly two years left to play a meaningful role in AI research before the entire loop — from idea generation to validation — is automated. Tworek left OpenAI in April after nearly seven years to found Core Automation, which bills itself as "the world's most automated AI lab." He told the Core Memory podcast that a dark joke now circulates among researchers: "Our jobs only have a few days left, so let's get as much done as we can while we still can." In his telling, the punchline is turning into a timeline.
The claim is specific rather than apocalyptic. Tworek splits AI research into two parts — idea generation and execution — and argues agents have already absorbed most of the execution work. The bottleneck left for humans is deciding what to try next. His lab has compressed a full experimental cycle from a month down to a day, a 30-fold speed-up, and he expects the remaining human role to narrow to the people setting direction. He paired that with a sharper, more provocative number: an OpenAI researcher told him there may be only 30 to 50 people in the world who truly understand, end to end, how to train and deploy a frontier model. Everyone else — including most staff at the biggest labs — is supporting those few dozen brains. The two-year countdown, he said, applies to them, not to humanity generally.
Tworek's most pointed argument is architectural. He calls training Transformer models to compete with OpenAI and Anthropic "a losing game" for new labs, and blames the Transformer itself for the field's stagnation — it can't keep learning after deployment, and fine-tuning brings catastrophic forgetting. In seven years at OpenAI, he said, there were only three or four serious attempts to replace the architecture, each crushed by the accumulated momentum of Transformer research or absorbed into it. Cheap AI coding agents, he argues, finally make those abandoned paths cheap enough to explore again — which is exactly what Core Automation is built to do. We profiled the wave of lab alumni chasing this thesis back in August — Why 37 AI-lab alumni founded startups in 2026.
What to watch: whether Core Automation ships something the incumbent labs can't ignore before the year is out — that would turn Tworek's prediction from provocation into benchmark.
Do you buy the two-year timeline, or is this an AI-lab founder marketing his own thesis? Tell us in the comments.
