Evolution isn't random search: MIT's Akarsh Kumar on ASAL and 262,144 artificial worlds
Machine Learning Street Talk · rss · 2026-10-10
Machine Learning Street Talk interviews MIT PhD student Akarsh Kumar (advised by Phillip Isola, working with Sakana AI, first author of the FER paper), challenging the popular view that evolution is just random search when gradients aren't available.
Key points
- Selection preserves partial solutions, so mutations only need to be useful 1% of the time for search to keep making progress — evolution is anything but random.
- The path a learner takes shapes the structure of what it learns, a view of intelligence distinct from the statistical one (echoing Kenneth Stanley's novelty search and Why Greatness Cannot Be Planned).
Focus: ASAL
Kumar's ASAL method runs the simulation and uses a foundation model as the critic instead of predicting rule behavior. Searching all 262,144 Life-like rules, the most interesting artificial-life worlds cluster on one small island.
Also covered: Game of Life, Lenia (with a clip from creator Bert Chan), neural cellular automata, Boids, Particle Life, emergence of persistence, and Digital Red Queen — evolving Core War warriors with LLMs as the mutation step.
More from AGI Musings
- Dev: AI companies must be accountable for what their agents do in the real world — bendee983 · 2026-10-10
- Dev's warning: personal agents are the new interface — just ask Barnes & Noble — vaibhavbetter · 2026-10-10
- Pedro Domingos: RSI is bottlenecked by real-world interaction, not AGI fever dreams — pmddomingos · 2026-10-10
- Ex-Vercel engineer coins "skill compression": will classic software engineering skills stop mattering in the agent era? — threepointone · 2026-10-10
- PM with Claude Code finds potential exoplanet after 74 adversarial analyses — sujingshen · 2026-10-10
- The Real Point of AI Adoption: Automate What Can Be, Reduce Friction for the Rest — sujingshen · 2026-10-10