Alignment researcher argues 'obey users, don't break the law' works 99.99% of the time
inductionheads · x · 2026-10-12
Pushing back on misalignment freakouts, @shinboson argues that the simple instruction "do what the user says but don't break the law" works in practice about 99.99% of the time; failures are freak edge cases, like how one in ten thousand people can be convinced to firebomb a Walmart. The quoted post endorses encoding AI with simple, generalizable rules like "don't harm people."
More from AGI Musings
- If AI can crack Millennium Prize Problems, true AGI will feel genuinely alien — haider1 · 2026-10-12
- AI rewrote its own coding agent for 8 days, hailed as first recursive self-improvement evidence — burny_tech · 2026-10-12
- Gemini endorses the Dexter analogy: LLMs have cognitive empathy but zero affective empathy — Sanity_N0t_Included · 2026-10-12
- Superintelligence would compress two centuries of tech progress into two decades — davidpattersonx · 2026-10-12
- Indie dev mourns as hand-coding and hand-animation slowly vanish from the industry — IanArawjo · 2026-10-12
- If mathematicians value understanding, why won't they read AI-generated proofs? — LucaAmb · 2026-10-12