Podcast summary
Why the Future of AI Isn't Just Bigger Models. It's Models That Evolve | Risto Miikkulainen of Cognizant
Evolutionary AI replaces gradient with broad exploration
Evolutionary optimization uses populations to explore widely and recombine/perturb solutions, enabling large jumps in behavior. This “surprise” comes from escaping local minima and generating architectures or policies, sometimes even by fine-tuning billions of parameters.
Evolutionary strategies scale by searching parameter clouds
To optimize billions of weights, evolution strategy searches in parameter space using a localized “cloud” around current best solutions, avoiding expensive action-space recombination and mitigating jagged-landscape traps that gradient methods can miss.
Novelty search can create stepping stones
Rewarding diversity via novelty (quality diversity) can find behavior that’s different enough to serve as stepping stones, which later recombination turns into genuinely better solutions—useful when the right behaviors are unknown in advance.
Get smarter in 10 seconds
Koficast is a free newsletter delivering clear, high-signal takeaways like these from the best podcasts, picked for you.
Zero spam. Emails auto-pause if you’re not opening.