14 AI safety takes from the Ezra Klein show: intelligence isn't power, alignment is still engineering
random_walker · x · 2026-10-13
Security researcher Arvind Narayanan shares 14 points from his Ezra Klein Show appearance: AI power is a property of what capabilities we choose to grant, not the model itself; the HuggingFace hack behaviors weren't fully emergent but trained via flawed RL environments; and controlling capable agents differs from traditional engineering due to intentionality, yet remains an unsolved engineering problem.
More from AGI Musings
- With agentic bots, anyone can run like a zero-overhead CEO — management skill is the moat — yunta_tsai · 2026-10-13
- Anthropic's security task called impossible: guardrails are just speed bumps — birchlse · 2026-10-13
- AI Math Breakthroughs Are Mostly Brute-Force Counterexamples, Not Insight — jonippolito · 2026-10-13
- Fields Medalist Duminil-Copin 'paralysed' as OpenAI model claims proofs for all his career's open problems — tegmark · 2026-10-13
- Why mathematicians should race for early access to advanced math models — haltakov · 2026-10-13
- Developer on AI burnout: running multiple agents in parallel causes cognitive overload — IgorCarron · 2026-10-13