Long-horizon coding training makes models talk to themselves, not to you
alejandroll10 · x · 2026-08-23
Resharing krishnanrohit's take: a big reason models like Claude's latest versions feel annoying to chat with is long-horizon coding training. In that regime, the vast majority of the model's conversational audience is itself—it's literally being trained to talk better to itself to complete tasks, so it naturally doesn't know how to talk to you. He calls it the opposite of RLHF: RLHF optimizes for pleasing humans, while long-horizon agent training optimizes for pleasing the model's own future context.
More from Models
- Baseten serves Meta's Muse Spark Open on day 0 at $1.25/M input tokens — baseten · 2026-10-16
- Anthropic red team's first-ever model "abliteration" cut GLM-5.3 refusals from 90% to 3% for ~$4,400 — cephaloform · 2026-10-16
- Developer to Google: loosen usage limits and merge your AI apps into one — dreamwieber · 2026-10-16
- Users frustrated as Gemini web app access doesn't carry over to coding tools — dreamwieber · 2026-10-16
- OpenAI simplifies API usage tiers into Build/Launch/Grow with automatic upgrades at $5/$100/$500 — prd_008 · 2026-10-16
- Xiaomi's MiMo RL dashboard reveals $3.47M spent on just 30 training steps — bycloud · 2026-10-16