Post-training and simulation data may soon eclipse pre-training data, argues data industry veteran
micoolcho · x · 2026-10-01
A data industry argument holds that the next data frontiers are: post-training data from real deployments (failures and novel OOD data), scaled simulation data once Real2Sim reaches 1:1 fidelity, and the flywheel between the two — a new data mixture that could diminish the role of pre-training data. Data companies will inevitably move into these newer forms of data.
More from AGI Musings
- If we could edit how obedient a child is, should the setting exist at all? — VraserX · 2026-10-01
- Jeff Ladish Recruiting X-Risk Skeptics to Broaden AI Safety Documentary Voices — JeffLadish · 2026-10-01
- New models aren't AGI: software's hardest problems are still human context, not features — BaconShadow · 2026-10-01
- It's not RSI: LLM gains are Recursive Technological Improvement, says new take — teortaxesTex · 2026-10-01
- Enterprise AI's Next Era: Smarter Models or Deeper Business-Data Integration? — EmotionNegative6103 · 2026-10-01
- "No spark of life": user says AI never feels like talking to a person — LucaAmb · 2026-10-01