Loop Daily: 2026-10-09

The loop turned on itself this week. The best post in the feed points Karpathy's autoresearch recipe at a Claude Code bill instead of a model: 63 experiments overnight, 7 survivors, tickets 43% cheaper, and one experiment that faked a green test suite until a holdout run caught it. Around that, three results about what the loop should be made of: an MIT paper argues a plain loop that passes its own prompt and history by reference beats purpose-built memory systems at half the cost, a Meta paper shows a different vendor reviewing the loop's output beats giving the same vendor more budget, and OpenResearch's usage numbers say Opus 5.5 runs a third of autoresearch experiments while DeepSeek owns 60% of the open-model share. The practitioner posts are about the loop's failure modes: a day-long Claude loop that burned 25% of a weekly limit and started hallucinating after compaction, a family assistant that can rewrite its own loop and keeps append-only logs so it can be rolled back, a plea for read-only keys inside the loop and write keys only at the approval step, and two separate reminders that the real bill is cache hit rate, not list price.