No RL pipeline trains proactive agents for workplace communication norms
menhguin · x · 2026-10-05
The author points out that humans must learn subtle workplace communication norms—no double texting, no pinging at 3am, how to handle being left on seen, whether and how to follow up—yet there is no known RL training pipeline teaching agents these skills.
Ironically, the one thing proactive agents do manage well is ML training runs, and they get extensive training for exactly that. The implication: agent capability distribution is shaped by training data, and high-frequency human scenarios like workplace communication are being neglected by current training regimes.
More from AGI Musings
- Why the AI Consciousness Debate Talks Past Itself, via the Drowning Ant Problem — PeterBowdenLive · 2026-10-05
- Schmidhuber revisits his 2016 panel with Chalmers and Kahneman on artificial consciousness — SchmidhuberAI · 2026-10-05
- repligate: alignment researchers should have heeded Infinite Backrooms' emergent goals — repligate · 2026-10-05
- Dean Ball clarifies: not about likelihood, but the one scenario certain to doom us all — deanwball · 2026-10-05
- "Call labs nurseries, not labs": the debate over growing AI minds instead of engineering them — repligate · 2026-10-05
- Will mass AI usage kill the internet as we know it? — BabuDevluu · 2026-10-05