AI safety researcher lays out 11 positions: reject 'accelerate and die' and blanket safetyism alike
jachiam0 · x · 2026-09-16
An AI safety researcher published a rapid-fire synthesis of his current positions on AI, safety, and progress:
- Rejects both extremes: neither "accelerate and die" nor safetyism that halts technological progress entirely.
- Supports concrete safety regulation: codifying best organizational practices around incident information management, training records, and data provenance.
- Takes existential risk seriously: national security strategy should prioritize detecting, defeating, and deterring rogue AIs, including kill switches — though he considers kill switches alone insufficient.
- AI for social good: use current and near-term AI to slash the cost of basic goods and services, raising everyone's floor above today's 1% ceiling.
- US-China competition: wants the US to win, but not via a race to the bottom — favoring a stable, mutually beneficial relationship.
- Questions policy concepts: "superintelligence ban" may not be well-formed since today's AI is already superhuman in many respects; "frontier pacing" is incomplete because frontier models unlock disruptive scientific discoveries in other fields. He finds the vulnerable world hypothesis's problem formulation plausible (without Bostrom's panopticon solution).
He also signed the "Pacing the Frontier" letter, arguing governments need levers to pace the frontier. A representative voice for the "responsible acceleration plus institutionalized safety" position.
Related event: Researcher Answers Dario's Slowdown Call: Responsibility, Not Rationing(3 posts)→
More from AGI Musings
- vikhyat predicts alternating compute oversupply and undersupply on the road to AGI — TheZachMueller · 2026-11-02
- Sam Altman: AI has surpassed top scientists, and 2026 will be a turning point year — kimmonismus · 2026-11-02
- DeepSeek design doc: a person can only follow 3-5 agent sessions at once, attention is the real bottleneck — heyshrutimishra · 2026-11-02
- AI researcher: we have no reliable way to tell whether an AI system is conscious — dioscuri · 2026-11-02
- Toby Ord: track both what models can and can't do to see real capability gaps — tobyordoxford · 2026-11-02
- Forward Deployed Verifier will soon be as valuable as FDE, argues poster — vaibhavbetter · 2026-11-02