Ex-OpenAI Leader: Human AI Researchers Have Maybe Two Years Left

新智元 · wechat · 2026-08-27

Jerry Tworek, former lead of OpenAI's reasoning models, suggests in an interview that human involvement in AI research may become obsolete within two years. He divides research into "ideation" and "execution," with the latter now largely handled by agents, boosting efficiency by 30x. He estimates only 30-50 people globally truly understand end-to-end frontier model training and deployment.

Regarding technology, he argues Transformer is unlikely optimal due to its inability to learn post-deployment and limited context windows. Yet, in seven years at OpenAI, serious attempts to replace the architecture happened only 3-4 times, hindered by resource allocation and inertia. His breakthrough came with support from Jakub Pachocki, enabling the scaling of reinforcement learning and the development of o1.

Now running CoreAutomation, Tworek used coding agents to improve a GPU kernel by 60x in four weeks at a cost of $100k, replacing expensive human experts. He predicts a post-AGI world where humans work not to keep the world running, but to pursue "excellence" through continuous learning and creation.

Original post →

Ex-OpenAI Leader: Human AI Researchers Have Maybe Two Years Left

More from AGI Musings