40 days of coding-agent telemetry show a provider silently swapped model weights
Jipok_ · reddit · 2026-10-30
A developer noticed deepseek-v4.1-flash felt dumber in his coding agent from mid-October, so he parsed local session logs to pinpoint when it broke.
- Before the update, the model had a funny quirk: randomly replying in Chinese while convinced it was speaking English.
- The -1017 update (Oct 17) fixed the Chinese leak and made tool calls more aggressive — while lobotomizing the model.
- His provider (opencode) silently overwrote the endpoint with new weights by Oct 19; his swearing rate tripled and mid-session model switches rose 6.7x.
- Key lesson: cloud APIs offer no way to roll back to old weights, unlike local deployment.
More from coding & agent
- Skyworks cuts root cause analysis from 20 days to 10 with LangGraph agents — LangChain · 2026-10-30
- 73-entry list sorts agent infra: which MCP tools an agent can provision itself vs needs a human — Total-Reasonable · 2026-10-30
- YC-backed Turnstone turns your existing AI subscriptions into an agent team that auto-switches when limits run out — ycombinator · 2026-10-30
- The simple self-testing loop behind every RL data company, per practitioners — teodorio · 2026-10-30
- 7 AI models all faked completion reviewing the same 111k-line codebase — a postmortem — rseroter · 2026-10-30
- Supabase mistook thousands of new databases per minute for a DDoS — it was Lovable integrating silently — alifcoder · 2026-10-30