A crow bending wire explains what LLM coding agents are missing: feedback loops
RandalSchwartz · reddit · 2026-11-05
The author uses Betty the New Caledonian crow (which bent wire into a hook with zero training in 2002) and Zador's Genomic Bottleneck to argue most AI coding setups breed "10-trillion-parameter moths": more weights but no probing loop or pain reflex. Adding a 20KB Unix Makefile loop and negative pain markers to their agents produced emergent problem solving. The claim: more prompt rules make agents fragile, while small mechanical feedback loops fix hallucinations.
More from coding & agent
- Agent benchmark scores hinge on container configs: 6-point swing from resource limits alone — krishnan · 2026-11-05
- Revolut cut AI coding inference costs ~90% by self-hosting GLM on 56 B200s instead of Claude — sebkrier · 2026-11-05
- VibeCoding a 3D fishing game in 4 days with Kimi K3.1 Pro and DeepSeek: full workflow breakdown — lxfater · 2026-11-05
- Sentry founder says AI agent 'junior' handles hardest coding tasks well — zeeg · 2026-11-05
- How Codex Happened: Inside OpenAI's coding comeback from failed launch to ChatGPT centerpiece — danshipper · 2026-11-05
- Tencent Hunyuan Hy4 lands in OpenClaw with vision, 1M-token context, agent gains — TencentHunyuan · 2026-11-05