OpenAI Chief Scientist Declares Recursive Self-Improvement Its AGI Path, Warns on Alignment
Summary
- • OpenAI Chief Scientist Pachocki publishes 'An Alien Mind', naming RSI as its AGI strategy
- • Pachocki warns no lab has solved alignment well enough to keep scaling at max speed
- • OpenAI researcher token spend hits $600+/day median; agent workdays now 3.1× human workdays
- • Calls for voluntary lab slowdowns and international coordination on AI safety issued
Details
OpenAI pivots research focus to Recursive Self-Improvement (RSI) as AGI path
Chief Scientist Jakub Pachocki's essay states OpenAI believes RSI is 'the only way to remain at the frontier of AI research moving forward,' marking a public commitment to RSI as the company's AGI strategy.
Essay 'An Alien Mind' by Pachocki published September 6, 2026
Pachocki published the essay on OpenAI's website, alongside a companion post 'Research acceleration: The view inside OpenAI' documenting internal agentic engineering adoption.
Pachocki warns no lab has solved alignment enough to scale at maximum speed
The essay explicitly states 'no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer,' a rare public acknowledgment of limits from a leading lab.
OpenAI calls for voluntary lab slowdowns and international coordination
Pachocki calls for voluntary slowdowns among frontier labs and international coordination as preconditions to responsible continued scaling — an unusual public safety plea from an OpenAI leader.
Median OpenAI researcher token spend: $600+/day by mid-August, up from $162 in July
By mid-August 2026, median token spend per researcher exceeded $600/day at API prices, up from $162 in July 2026; the 90th percentile exceeded $7,000/day, reflecting explosive agent adoption.
Agent workdays reached 3.1× human workdays at OpenAI by mid-August 2026
AI agents surpassed human working hours at OpenAI in June 2026; by mid-August, agent workdays were 3.1 times human workdays, with researchers running four or more agents concurrently.
Late-July spend acceleration linked to internal GPT-6 Astra model access
Commentators including Simon Willison speculate that the sharp acceleration in AI spend per researcher in late July coincided with internal OpenAI employee access to the model later released as GPT-6 Astra.
Details compiled via Grok live web research from multiple outlets
Additional facts, figures, and source corroboration for this event were gathered via Grok live web research, drawing on OpenAI, Business Insider, Unite.AI, and The Information Machine.
RSI = Recursive Self-Improvement; AGI = Artificial General Intelligence
What This Means
OpenAI's Chief Scientist has publicly declared Recursive Self-Improvement the company's core AGI strategy, while simultaneously warning that no lab — including OpenAI — has solved alignment sufficiently to keep scaling at maximum speed indefinitely. This is a rare instance of a frontier lab leader openly acknowledging systemic risk and calling for international coordination on safety. The accompanying internal metrics reveal an extraordinary acceleration of AI-assisted research: agents now perform over three times as much work as humans at OpenAI. This dual signal — an ambitious AGI strategy paired with formal safety warnings — marks one of the most significant public inflection points in frontier AI discourse in years.
Sentiment
Concerned acknowledgment of risks, with emphasis on alignment gaps amid rapid progress
“An important warning from the Chief Scientist at @openai. We are starting to see “recursive self improvement”... And we still aren’t sure of “alignment”... I hope his own organisation is paying attention.”
“I think complete alignment is impossible because of the generalisation challenge the article rightly outlined. I also think defence in depth... is probably the best approach.”
“OpenAI admits it does not yet know how to safely reach aligned, full RSI. 'We do not yet know how to safely get all the way to aligned, full RSI.' 'We cannot assume that progress in alignment and safety will keep pace.'”
“OpenAI’s own co-founder and chief scientist both say we’re already in the AGI era and that no lab has solved alignment well enough to keep racing at full speed. It’s no longer hype, it’s a warning from the people building it.”
“OpenAI says its research agents now deliver ~3.1 agent-workdays per human day — while chief scientist Jakub Pachocki argues labs can’t keep scaling at max speed without better alignment and monitoring.”
Split
~70/30 concerned about alignment risks vs. focused on acceleration/competition (with some mixed calls for defense-in-depth approaches).
Sources
- An Alien MindOpenAI
- OpenAI Chief Scientist Jakub Pachocki says no lab has solved alignment enough to keep scaling at maximum speed, and hopes voluntary slowdowns become commonplaceOpenAI
- Research acceleration: The view inside OpenAISimonwillison
- OpenAI’s chief scientist says no lab should keep scaling at maximum speedThenextweb