RL-trained Astra writes uniquely unreadable code as corrigibility becomes a matter of faith
aryaman2020 · x · 2026-09-16
aryaman2020 complains on X that code written by Astra (and presumably Fable) uses crazy meta-programming and abstruse primitives to become hyper-efficient but uniquely unreadable, suggesting "corrigibility" has become a matter of faith. He pleads with tszzl to fix his RL setup. A follow-up admits it "works though" — raising concerns that RL optimization trades readability and controllability for performance.
More from Fun
- America.gov's AI chatbot refuses a geography riddle — unless you ask about exactly 3 states — conitzer · 2026-11-02
- Letting Claude run wild in an endless loop, showing only what it thinks is cool — rachittshah · 2026-11-02
- Even a skeptic almost fell for an AI fruit cartoon reel — menhguin · 2026-11-02
- Sole Intentions: a film-noir style AI short film starring a sloth at an office desk — heresalexandria · 2026-11-02
- Founder Tries Building a Custom CRM, Ends Up Integrating 3 Existing Tools Instead — menhguin · 2026-11-02
- Designer calls out peer for near-verbatim copy of his Figma-to-Claude workflow thread — hewarsaber · 2026-11-02