Data flywheel is fake: only deployment data where models fail is valuable, argues researcher
chris_j_paxton · x · 2026-11-12
Kyle Vedder sparked discussion by arguing the naive "more deployment = more data = better model" data flywheel narrative is fake: only deployment data where the model fails is actually valuable for training. Chris Paxton amplified the take, saying it hits on a point he has been making. The argument directly challenges the common industry strategy of collecting data at scale from product deployments.
More from AGI Musings
- Adam Scholl: strong evidence AI is useful, weak evidence it generalizes — gleech · 2026-11-12
- "SF is right long-term, NYC right near-term, nobody has a clue mid-term" on AI — austinc3301 · 2026-11-12
- Andrew Critch: stop calling extinction risks "improvement"—RSM vs RSI matters — AndrewCritchPhD · 2026-11-12
- Canada tech vacancies down 46.9% from peak; software engineer roles fell 72% — RichardsonDx · 2026-11-12
- Any US State Could Legalize AI Safety Coordination via Parker Immunity — TinfoilTricorn · 2026-11-12
- The new Turing test: can a human with AI pass as intelligent? — cocktailpeanut · 2026-11-12