kalomaze: transfer extrapolation should be a learned function, not a theory
kalomaze · x · 2026-09-12
Open-source researcher kalomaze threads on X about why extrapolating transfer is probably something that must be learned: his intuition stems from CoT/reasoning being a decomposition of problems into subroutines that compose generically, so good strategies transfer across very distinct domains.
In the quoted post he argues you don't need a strong mechanistic theory of transfer (which may not exist in closed form) before training — you need a strong learned function of transfer extrapolation for a given policy, which is constructible.
More from AGI Musings
- Skimming Anthropic's S-1: Winner-Take-All Not of a Category, but the Whole Software Ecosystem — mobileraj · 2026-10-31
- CS312 course post explains sharpness and the edge of stability in LLM pretraining — tatsu_hashimoto · 2026-10-31
- NBER paper: training your AI replacement can actually raise workers' pay — joshgans · 2026-10-31
- Trata's Hedge-Bench: no model tops 50% on real-world financial forecasting — ycombinator · 2026-10-31
- Anthropic's internal Claude model claims 2^o(n) 3-SAT algorithm, purportedly refuting the ETH — ctjlewis · 2026-10-31
- Foundry investor: durable companies are built on data loops, not copyable features — alvelda · 2026-10-31