Risto Miikkulainen: Neuroevolution and Evolutionary Computation | Lex Fridman Podcast #177

Lex Fridman talks with Risto Miikkulainen about evolutionary computation and neuroevolution, the idea of evolving rather than designing intelligent systems. They explore how evolution discovers surprising, creative solutions by removing human biases, from optimizing plant-growing recipes to evolving neural network architectures. The conversation ranges across social structures as the root of language, theory of mind in robots, predator-prey co-evolution, and whether evolved agents could ever develop languages we could translate. It closes on diversity, exploration, mortality, and finding meaning as one agent in a larger evolutionary engine.

• Evolutionary optimization discovered that basil doesn't need to sleep, 24-hour light made it bigger and tastier, surprising even the biologists who assumed darkness was a necessary constraint.
  00:28:34
• Students' evolved tic-tac-toe AI won a tournament by making moves millions of squares away, causing rival programs to expand memory until they crashed.
  00:39:25
• Evolution learns to walk by tolerating agents that fall flat repeatedly, eventually discovering controlled falling that becomes running, something reinforcement learning is too conservative to find.
  00:53:02
• In simulation, a robot spontaneously developed something like a theory of mind, faking a move to trick its opponent into wasting energy and losing.
  00:58:46
• Miikkulainen argues a society can tolerate a minority of cheaters and liars and still function, and that diversity (even destructive variation) is fundamental to evolution.
  01:30:51
• Lex notes AI systems that eventually pass the Turing test will be ones exceptionally good at lying, calling it a terrifying concept.
  01:31:24
• Growing neural network topologies via evolution, rather than using fixed human-designed architectures, is largely unexplored and potentially powerful.
  01:08:28

• Sharks have been around 400 million years, longer than trees have existed.
  00:05:08
• Researchers modeled hyenas that band together (around 30) to push lions off a kill, balancing fear of the lion against affiliation with each other.
  00:17:34
• Hyena emotional states can be quantified after the fact from neurotransmitter chemicals left in their droppings.
  00:23:51
• There's a real paper, 'The Surprising Creativity of Digital Evolution,' collecting anecdotes of evolution exploiting bugs and loopholes.
  00:38:22
• Today's autonomous robots are afraid of everything, treating humans as ballistic objects to avoid rather than partners in a dance.
  01:01:17
• Multi-task learning works even across unrelated domains, you can learn better vision by learning language, or better language by learning DNA structure.
  01:17:52
• Star Trek envisioned many devices we now have, and motivated many people to enter computing and AI.
  01:25:41
• Novelty search rewards solutions just for being different from prior ones, with no fitness goal, yet still discovers useful things like efficient robot walking.
  01:44:46
• It takes roughly 10,000 hours to become an expert, but exploration across diverse fields should precede that commitment.
  01:50:30