OpenAI Is Paying $445,000 to Prepare for the Moment AI Trains Itself
OpenAI is advertising a job that should not exist.
The role — Researcher, Recursive Self-Improvement Preparedness — pays between $295,000 and $445,000 a year to prepare for a future where AI systems can train themselves without human involvement. The listing appeared on OpenAI's careers site this week, according to Business Insider, and it offers compensation more typical of a senior quant at a hedge fund than a mid-level research hire at a software company.
Jack Clark, cofounder of rival lab Anthropic, put the odds at roughly 60% that AI systems will be able to conduct research without human involvement by the end of 2028. Clark, who cited evidence from benchmarks including CORE-Bench, PostTrainBench, and MLE-Bench (tests that measure how well AI performs tasks that previously required human scientists), framed recursive self-improvement as requiring global coordination infrastructure, including a hotline between rival labs, before it arrives.
The concern is not hypothetical. METR, the research organization that tracks how long tasks frontier AI models can autonomously complete, has found that task length has been doubling approximately every seven months since 2019 — six years of consistent exponential improvement. Under its updated methodology, TH1.1, the post-2023 doubling time is 131 days. Since 2024 alone, the figure is 89 days — a compression that suggests the pace is accelerating, not plateauing.
If the trend holds, generalist autonomous agents capable of week-long tasks arrive within two to four years. Month-long projects by the end of the decade. Sam Altman has stated OpenAI has set internal goals of an automated AI research intern running on hundreds of thousands of GPUs by September 2026, and a true automated AI researcher by March 2028. Sam Altman has stated OpenAI has committed roughly $1.4 trillion in compute infrastructure and 30 gigawatts of power capacity to get there.
The job listing offers no answer to the structural problem it creates. The last humans who know how to train AI are being hired at peak market rates — right when training becomes automatable. There is a historical precedent that should give anyone watching this pattern pause: programmers built the tools that later automated programming tasks; engineers designed factories that eventually replaced engineering work. The expertise required to build the automation gets priced at its scarcity peak, not after. Critics will note that OpenAI posts hundreds of roles and that $295K-$445K reflects the extreme tightness of the current AI safety talent market, not necessarily a confession about timelines. That reading is plausible. The market rate for the humans who know how to train AI may simply be high because few people have the skill — not because those skills are about to disappear. If the researcher succeeds in preparing for recursive self-improvement, they may have automated their own role out of existence before the ink dries on the offer letter.
OpenAI declined to comment beyond the job posting. Anthropic did not respond to a request for comment.
What to watch: whether other frontier labs post similar roles, and whether the 89-day doubling figure survives contact with the real-world compute and data constraints that have slowed previous capability jumps. Clark has said the Anthropic Institute is committing to monthly reports on how AI tools are reshaping its own research — one sign the labs are treating the coordination problem as urgent and operational, not theoretical.