MIT & Sakana researcher on artificial life: learning paths shape representations more than endpoints
kenneth0stanley · x · 2026-11-08
The latest MLStreet Talk episode features Akarsh Kumar (MIT CSAIL, Sakana AI) on artificial life: searching the space of possible universes with foundation models, why evolved networks find structure that SGD misses, and LLMs evolving Core War programs.
Kenneth Stanley joined the discussion, and the takeaway highlighted in this repost: the path taken during learning may shape the representations we acquire — 'how we get there' can matter more than 'where we end up.' This raises an open question: when does learning through a sequence of stepping stones produce more transferable representations and better generalization than learning the same information in a different order?
More from Research
- ML Professor Fleuret: Transformers Minimize Cross-Entropy Without Understanding a Thing — francoisfleuret · 2026-11-08
- Library of Layers: Transformers That Reorder Their Own Layers at Runtime Match Baselines — LotfiSanae · 2026-11-08
- SHA-256 collision extended to 38 steps, paper promised as researcher posts both messages — jedisct1 · 2026-11-08
- Xiaomi MiMo next-gen moves to causal encoder-decoder: 5x less prefill compute at 1M context — joey00072fp8 · 2026-11-08
- Multi-provider model workflows will deliver the next math breakthroughs, predicts Andrew Critch — sebkrier · 2026-11-08
- Axiom Math formally verifies Stacks Project: 7,654 pages into 16M lines of checked code — danielrock · 2026-11-08