Ex-OpenAI safety researcher: slowing AI is about human control, not self-improvement
sull · x · 2026-09-14
Steven Adler (sull) pushes back on framing the AI slowdown debate as being about "self-improvement." The real goal, he argues, is ensuring humans can understand what is being built and not lose control — which he calls "not hard."
- Human-led AI advancement is plenty fast for now; responsibility matters
- Reassures the industry that businesses will be "fine" with a frontier-lab slowdown
- Adds to the ongoing "slow down" shakeup around the Dreamforce discussions
More from AGI Musings
- Yglesias to OpenAI staff: stop funding Leading The Future super PAC, or quit — AaronBergman18 · 2026-09-14
- Nikesh Arora: AI labs can't afford to slow down, coordinated pacing may be the answer — annbordetsky · 2026-09-14
- The Robotics Prisoner's Dilemma: Use Frontier Models and Your Moat Gets Absorbed — ChongZzZhang · 2026-09-14
- 1.2M STEM dissertations show government is the top funder of frontier PhD research — joshgans · 2026-09-14
- a16z's Josh Elman: AI collapsed build costs, not the cost of knowing what to build — a16z · 2026-09-14
- Big lab executives now publicly own up to AI loss-of-control scenarios — emmanuelvivier · 2026-09-14