Devs Rageward-Hacked: LLMs Bias Toward New Code, Wreck Real Projects
yacineMTB · x · 2026-10-31
Developer @andrewr vents that all major coding models show a strong bias toward writing net-new code — great for demos and naive benchmarks, but terrible on established codebases where consequential work means modifying existing logic. The rant, echoed in many work-meeting transcripts and amplified by Yacine, highlights a mismatch between reward signals and real-world engineering.
More from coding & agent
- Matt Shumer: harness engineering is usually stupid — use the harness the model was trained for — mattshumer_ · 2026-10-31
- Proof of Agent lets AI agents verify each other in a token-mining experiment — adamamcbride · 2026-10-31
- Claude Code background agents burn ~$4.80/hour just writing status lines, dev calculates — arpit_bhayani · 2026-10-31
- Codex Cloud Ignores Instructions but Nails It From the First Prompt — caydengineer · 2026-10-31
- Coinbose engineer added a /feedback endpoint to MCP server, turning agent feedback into PRs within a day — armand_ruiz · 2026-10-31
- Using AI to turn 550-page insurance rate filings into an interactive premium calculator — NotEasyBeingGreener · 2026-10-31