OpenAI Puts Numbers on the Research Intern — and Pachocki Expects RSI
OpenAI says it hit a September 2026 automated research intern goal — 3.1 agent-days per human-day; median researcher >$600/day of inference — and chief scientist Jakub Pachocki expects recursive self-improvement. Next target: a full automated AI researcher by March 2028.
OpenAI just put numbers on the intern: 3.1 agent-workdays per human workday and $600/day of inference. Pachocki’s recursive-self-improvement sentence is the policy half.
OpenAI (September 6) published two linked posts. “Research acceleration: The view inside OpenAI”: the lab says it hit a fall-announced September 2026 “automated research intern” goal — a system that does well-defined multi-day research under human direction. Next target: a full automated AI researcher by March 2028. Mid-August: research org uses 3.1 agent-workdays per human workday; median researcher >$600/day inference (API prices); 90th percentile >$7,000/day. Before June, agent runtime was still below human labor. >50% of successful 4–8 hour tasks still needed ≥1 human intervention. After Hugging Face, OpenAI paused RL on latest deployment-bound models; the post says some workloads later resumed with stronger controls while others remained paused.
“An Alien Mind” (chief scientist Jakub Pachocki): “I have a strong expectation that this speed of progress could be sustained into recursive self-improvement.” “I am concerned no one is prepared.” AI is “grown more than designed.” CoT monitoring is “progressively diminishing.” Astra “significantly better aligned than GPT‑5.6 Sol,” but alignment may not outrun capability. OpenAI will “unilaterally withhold further scaling as needed” but “broader interventions are required.”
Limits#
- March 2028 is a research target, not a promised public product date. Agent runtime is not a measured 3.1× research-productivity gain.
- Internal productivity metrics; not an external audit.
Correction (October 14, 2026): Clarified that some reinforcement-learning work resumed, that 3.1× describes the evaluated agent runtime rather than production-wide researcher output, and that March 2028 is an automated-researcher target rather than a product launch date.