Andy Matuschak: don't roll the AI risk die without a real control plan
andy_matuschak · x · 2026-09-29
Andy Matuschak argues the current AI risk debate largely talks past each other. Opposing arguments only support "things will likely go OK," but even a few-percentage-point shift leaves significant probability mass on the most catastrophic outcomes in human history.
His core claim: don't roll the die until there's a real plan to control those risks to an acceptable level—many orders of magnitude from where we are now—and never roll it if mitigation proves impossible. He clusters opposing views into: risks are fake/unworthy of concern, or risks are real but probabilities are low enough to skip mitigation.
Related event: Andy Matuschak on AI Risk: Catastrophic Tail Risk Can't Be Ignored(2 posts)→
More from AGI Musings
- "A kid born this year never gets a drivers license, never gets a job and never dies of old age" — rand_longevity · 2026-09-29
- Caltech Trio on How AI Is Transforming Scientific Discovery, Down to the Quantum Level — AnimaAnandkumar · 2026-09-29
- AI writing hits 25-50% of econ working papers, but submission counts stay flat — paulnovosad · 2026-09-29
- Fei-Fei Li on LLM limits: 'Language is purely generated signal' vs. the physical 3D world — rohanpaul_ai · 2026-09-29
- NBER Study Finds AI Hasn't Taken Recent Grads' Jobs, Defying Predictions — beffjezos · 2026-09-29
- Jevons Paradox Resurfaces in AI Debates: Efficiency Gains Drive Up Total Consumption — burny_tech · 2026-09-29