Blog Argues Labs Should Train RL Over a Distribution of Human Effort Levels
voooooogel · x · 2026-11-19
A vgel.me post argues labs should sample human effort levels alongside model reasoning effort during RL training, just as they do for low/medium/high thinking budgets. Users oscillate between deep pair-programming, phone-level steering, and full autonomy — and models handle those mode switches poorly, especially inside forked subagent contexts (with a real conversation with an Opus 5.5 subagent as evidence). Training over the distribution of user engagement is the next step for collaborative models.
More from coding & agent
- Aristotle formally proves a property over all runs of the 6k-LOC Rust e-graph library egg — satnam6502 · 2026-11-19
- Replit's agent crawled 1,000+ DocuSign flows to build free e-sign rival Ternsign — amasad · 2026-11-19
- Inworld open-sources 48kHz real-time voice isolation models under Apache 2.0 — Scobleizer · 2026-11-19
- GPT5.5 ports DSPy-style axllm to Zig in 45 minutes for $16 — dosco · 2026-11-19
- Pichai Spotlights Photorealistic 3D Lab Modeled in Antigravity via Blender MCP — sundarpichai · 2026-11-19
- Google's Gemini 4 Neon goes GA in GitHub Copilot, tuned for agentic coding — intellectronica · 2026-11-19