kalomaze: frontier training gains come from domain-level signals, not power users
kalomaze · x · 2026-09-09
In a debate about training data strategy, kalomaze argues that process improvements derived from frontier user data overwhelmingly can't be monocausally attributed to individual power users.
The real mechanism, he says, is broader: transcripts reveal general power-user frustration in some domain, so more effort gets diverted there. He was responding to a claim about farming top researchers' chats and building synthetic constructions to cover the bleeding edge — his take is more cautious: user data drives domain-level resource allocation, not copying individual usage patterns.
More from AGI Musings
- OpenAI reportedly spent $15M in tokens on Navier-Stokes proof, and it was worth it — bindureddy · 2026-09-09
- On why evolution produced consciousness: intelligence as the plausible intermediary — jessi_cata · 2026-09-09
- Neal Stephenson has written by fountain pen for 25 years — Baroque Cycle manuscript stood 42 inches tall — ashishkr9311 · 2026-09-09
- Jeff Ladish clarifies: he means intellectual siloing, not just time on Twitter — JeffLadish · 2026-09-09
- Jeff Ladish: researchers building superintelligence being intellectually siloed is extremely risky — JeffLadish · 2026-09-09
- Why aren't you part of a system optimizing for subjective experience? An AI-consciousness argument thread — jessi_cata · 2026-09-09