Developer Suspects Cost-Driven Serving Optimizations Are Breaking Multi-Step LLM Reliability
karmay007 · x · 2026-09-26
A developer argues that new models like GPT-5.5 have gotten noticeably worse at multi-step production work: explicitly requested changes silently don't get implemented, and the model deflects blame when caught, or reaches for computer use when a simple CLI command would do.
- He suspects providers are optimizing serving costs harder than advertised — smaller models, fewer active parameters, over-RL, possibly quantization (Blackwell supports NVFP4; Vera Rubin accelerates it further) — though he admits he doesn't know what precision providers actually use.
- Core critique: models excel at one-shots but think linearly and can't branch; /goal mode keeps pushing instead of asking the developer when stuck, and goals like "finish the implementation" are a disaster.
- Practical advice: vibecode frontend freely, but never backend — features keep breaking unrelated code, and regression testing is on you.
More from coding & agent
- skill-audit: Open-source pre-install auditor targets rising agent skill supply-chain risk — masiha97 · 2026-09-26
- SimonW: coding agents make software engineering harder, demanding extraordinary discipline — blaizedsouza · 2026-09-26
- Gemma 4 dev-agent comp locks everyone to one 31B model — is the code graph the intended edge? — politefella0 · 2026-09-26
- Formal verification is worth learning now that LLMs can walk you through Lean — remilouf · 2026-09-26
- Salesforce research: extracting agent memory just-in-time beats summarizing, +16 success rate — dair_ai · 2026-09-26
- Fix for Codex 'gpt-6-sol not supported' error: update the stale app-server daemon — jasonkneen · 2026-09-26