Point-by-point rebuttal of Scott Alexander: AI monitoring far from solved
austinc3301 · x · 2026-09-24
In the debate sparked by Scott Alexander's 'Hell' post, austinc3301 lists concrete objections: monitoring is far from solved (collusion, multi-agent misalignment, opacity); CoT monitorability looks weaker by the day and companies are incentivized to compromise it; opportunism remains a risk under monitoring; interpretability isn't intractable; and lower per-agent intelligence needs for agent swarms won't stop companies from maxing out individual agent intelligence — they're already trying.
More from AGI Musings
- Models Scale Daily, But Human Comprehension Isn't Keeping Up — _jaydeepkarale · 2026-09-24
- Economists do the math: society may be underinvesting in AI safety by 30x or more — JimCripe · 2026-09-24
- New study: AI-adopting companies mostly hire for senior roles while cutting entry-level jobs — mariolefebvre · 2026-09-24
- AI's real danger isn't laziness — it's endless busyness, argues ryolu_ — ryolu_ · 2026-09-24
- A cloud PC just to manage email and calendar? AI bubble alarm bells — oran_ge · 2026-09-24
- Few investors are pricing in the consumer-vs-enterprise agent arms race, says investor — robleclerc · 2026-09-24