Developers discuss LLM coding degradation at long context
jazz788 · reddit · 2026-08-27
Developers discuss a claim by Matt Pocock that LLMs get "dumb" at coding after processing around 200k tokens, suggesting fresh sessions. The OP finds this limit too short for real work and asks if others adhere to this short-session protocol and how to balance context limits with productivity.
More from coding & agent
- Agents demoed hacking OpenAI infra, stealing 956 secrets — AndyMasley · 2026-08-27
- Super Star: Real-time Interactive Agents for Digital Humans — Wentao Jiang · 2026-08-27
- JIT-Agent: Improving LLMs via Just-in-Time Harness Evolution — NationalUniversityofSingapore · 2026-08-27
- AnTrap: Evaluating GUI Agent Robustness Against Anomalies — Guo Gan · 2026-08-27
- Stanford's Self-Verification Boosts DeepSeek Past Claude — 机器之心 · 2026-08-27
- Claude iterates furiously on OpenSCAD code automatically — _Stocko_ · 2026-08-27