The Era of Pre-Training is Evolving
deanwball · x · 2026-07-12
The core argument is that pre-training scale was previously aimed at better fitting human language, but its primary role now is likely to provide the neural capacity for post-training capabilities like sample-efficient reasoning and long-horizon planning.
The author believes this makes "buying cutting-edge capabilities purely by scaling compute" much harder, as low-hanging web-scraped data is largely depleted. The necessary post-training data is more akin to continuously generated "learning by doing" micro-experiments from numerous diverse reward environments and user-agent OODA loops. The text notes this data is incredibly expensive, comparing its value to a "$6 billion SpaceX / Cursor acquisition." Finally, it emphasizes that beyond model distillation, it's nearly impossible to jump-start a frontier agent from scratch directly to the finish line.
More from AGI Musings
- CS professor: don't hide your AI use from future employers, lead with it — lemire · 2026-10-26
- Engineer: AI-cheap engineering will change which manufacturing jobs get a yes — burhop · 2026-10-26
- FixItBench: the joke benchmark where robot-agent teams race to repair electron microscopes — teortaxesTex · 2026-10-26
- Instructor advice: record and save all your teaching materials for future AI use — prof_g · 2026-10-26
- Jason Furman: Building Econvitals pre-AI would need ~50 people, 3 years, $20M — HarrySurden · 2026-10-26
- Dune's 1960s Warning on Handing Thinking to Machines Resurfaces in AI Debate — umeshonai · 2026-10-26