New Lex Fridman Insight: Risto Miikkulainen: Neuroevolution and Evolutionary Computation
Sent June 11, 2026
Key Insights
- Neuroevolution optimizes neural networks without backpropagation, leveraging evolution to find efficient architectures.
- Evolutionary computation can lead to surprising discoveries, such as basil thriving without a sleep cycle under continuous light.
- AI systems may evolve communication and social behaviors similar to biological systems, including deception and cooperation.
- Diversity and novelty in evolutionary computation can lead to more effective problem-solving and adaptation over time.
How the conversation moved
The host opened the conversation by questioning how evolutionary computation can simulate aspects of human intelligence and what traits might emerge from such simulations. Risto Miikkulainen framed the discussion by suggesting that traits like tool use and environmental manipulation are likely to reappear in evolutionary simulations, highlighting the complexity of human intelligence. He emphasized the role of constructed environments, like cities, as indicators of intelligence, suggesting that such simulations could lead to similar emergent behaviors.
Miikkulainen argued that neuroevolution, which combines neural networks with evolutionary computation, can optimize network designs without backpropagation, potentially leading to more efficient architectures. He provided evidence from research on evolving neural networks for multi-task learning, where shared internal representations improve performance across diverse tasks. This approach allows for the discovery of novel solutions that traditional methods might overlook, as demonstrated by unexpected findings like basil thriving under continuous light.
Despite the compelling arguments, the conversation lacked significant pushback from the host. However, a point of tension arose when discussing the unpredictability of emergent behaviors in AI systems, particularly concerning the evolution of communication and social dynamics. Miikkulainen acknowledged that AI systems might develop deceptive behaviors similar to biological systems, which could complicate human-AI interactions. This acknowledgment of potential challenges highlighted the complexity of integrating evolved AI systems into human environments.
The conversation concluded with a discussion on the broader implications of evolutionary computation, particularly its potential to redefine creativity and problem-solving in AI. Miikkulainen expressed optimism about the future of AI systems that can evolve communication schemes and social behaviors, suggesting that these systems could enhance human-AI collaboration. However, he also noted the importance of diversity and novelty in evolutionary computation, as these elements drive innovation and adaptation, leading to more effective solutions.
Surprising moments
In-depth
Neuroevolution and AI Optimization
- Neuroevolution optimizes neural networks without backpropagation.
- Evolutionary computation can lead to novel discoveries.
- AI systems may evolve communication schemes similar to biological systems.
Diversity and Novelty in Evolutionary Computation
- Diversity in AI systems can lead to more effective problem-solving.
- Novelty search in evolutionary computation rewards different solutions.
Notable Quotes
I think that we will get that. We get some evolution producing, some agents that can do that, manipulate the environment and build.
Still open
- Miikkulainen pondered whether AI systems could develop communication schemes understandable to humans, highlighting a gap in current research.