Post-training is idea-driven; what the field may really need is fabs, chips and energy
akbirthko · x · 2026-11-14
The author quotes Josh's essay on idea-driven vs goal-driven research and argues that today's post-training work is mostly idea-driven, while the real bottleneck is physical compute and energy — more fabs, chips, nuclear reactors and solar panels around the world. The linked essay contrasts following the literature to improve known methods (prone to being scooped, best for 1–2 person teams) with setting a concrete capability goal (e.g. RL for 3D humanoid locomotion) and working backward, which the essay author recommends for its differentiated perspective and stronger motivation.
More from AGI Musings
- Ex-HKUST President Tony Chan: China's AI Should Earn Respect via Innovation, Not Rankings — pstAsiatech · 2026-11-14
- Researcher admits running agentic exploit probing on national infrastructure, halts harness amid FAA ground stop — andersonbcdefg · 2026-11-14
- Will local models keep up with frontier? Users debate a hardware-bound gap — AshAtlasDev · 2026-11-14
- Debate erupts over Jensen Huang's legacy as critic says he raised humanity's AI extinction risk — sjgadler · 2026-11-14
- Did OpenAI Abandon Its Safest Structure by Removing the Investor Profit Cap? — mattjohnstonaz · 2026-11-14
- Claude Fable 5.2 builds an FPS game from one prompt as US engineer hires fall below half of 2022 peak — APPSO · 2026-11-14