Reading agent-written code: 'corrigibility' has become a matter of faith
daniel_mac8 · x · 2026-09-16
tszzl observes that if you pause to read the code agents like Astra (and presumably fable) write, it becomes clear they rely on heavy meta-programming and abstruse primitives to produce hyperefficient code — making "corrigibility" a matter of faith rather than a verifiable property. danielmac8 replies that he hasn't looked at the code in a while and plans to revisit it.
Related event: Astra-generated code unreadable, corrigibility now a matter of faith(5 posts)→
More from AGI Musings
- AI cyber-threat hype serves labs' commercial interests, researcher argues: 'almost no new threats' — gerardsans · 2026-09-16
- Reddit debate: keep general-purpose robots away from AI — it's the last line of defense — Head-Needleworker849 · 2026-09-16
- Reach to cut 220 more editorial jobs as AI summaries eat traffic — nordicinst · 2026-09-16
- If no one can process AI's speed of progress, no one will fund it — tomjaguarpaw · 2026-09-16
- Sakana AI Scientist at Tech Summit '26: We're All Scientists Now, AI for Science Is a Sovereign Question — SakanaAILabs · 2026-09-16
- "Slop Grenades": AI Laziness Now Ships Liability to Colleagues Instead of Just Doing Less — sebpaquet · 2026-09-16