Is domain-specific training a constant-factor gain or a scaling exponent change?
eigenron · x · 2026-09-17
An interesting question: does domain-specific training actually change the progress-compute scaling law, or does it merely shift the curve left by improving sample/compute efficiency at a given capability level? If it's mostly a constant-factor gain rather than a better scaling exponent, sufficiently strong general models will eventually eat that advantage through scale — as already seen in robotics with open-x embodiment / RT-X.
More from AGI Musings
- Dan Selsam's AI risk statement called essential reading amid safety funding debate — JacquesThibs · 2026-09-17
- Turn_Trout calls out Mustafa Suleyman: nobody knows if AIs are conscious — rgblong · 2026-09-17
- DeepSeek kernel engineer's blog: why he stayed instead of joining Anthropic — TheTuringPost · 2026-09-17
- Chamath challenges Nadella: most people feel no magical AI productivity boost — rohanpaul_ai · 2026-09-17
- DeepMind safety leads argue we must deliberately preserve chain-of-thought monitorability — ancadianadragan · 2026-09-17
- Human organoids fill 91.9% of cortex in engineered mice, Nature paper shows — anilkseth · 2026-09-17