Interview, Fireside Chat
Risto Miikkulainen: Neuroevolution and Evolutionary Computation | Lex Fridman Podcast #177
- Evolutionary Regularities: Simulations of Earth's evolution run a million times suggest that while outcomes vary, specific solutions like object manipulation, opposable thumbs, oral communication, and vision will emerge repeatedly across different lineages.
- Human Uniqueness: Computational models would likely identify human intelligence via the construction of complex, benign environments (cities) and the ability to leave a lasting, positive ripple effect on the system, rather than just survival duration.
- Detection Challenges: Observers from an external or alien perspective might prioritize long-lived species (sharks, trees) or massive biomass (insects) over humans, as human existence is relatively recent and characterized by high environmental impact.
- Intelligence Definitions:
- Functional: Intelligence is defined as an agent with limited sensory/effective capabilities that can survive, reproduce, and solve problems using available resources.
- Legacy: A higher-level definition involves creating something useful for others or the future that outlasts the agent's existence, creating a "ripple effect" or trace.
- Role of Emotion: Emotions like fear of mortality serve a survival function by focusing attention on relevant threats without requiring complex logical deduction; they are essential for rapid decision-making in dangerous situations.
- Social Foundation of Intelligence: Human intelligence is fundamentally social; Bickerton's theory posits that social structures and exchangeable roles predate language, providing the necessary framework for grammar and symbolic communication.
- Evolutionary Computation vs. Deep Learning:
- Data Requirements: Deep learning excels where labeled data exists; evolutionary computation (EC) is superior for domains where optimal answers are unknown (e.g., robotics, game playing, healthcare treatment plans).
- Exploration vs. Exploitation: EC can afford "dead ends" (e.g., robots falling) which serve as stepping stones to unique solutions, whereas reinforcement learning often requires conservative, incremental learning.
- Surprising Creativity in EC:
- Basil Growth: An evolutionary algorithm discovered that basil thrives with 24-hour lighting, disregarding the human bias for a day/night cycle, resulting in faster growth and better taste.
- Game Hacks: In a tic-tac-toe tournament, evolved AI defeated opponents by exploiting memory bugs in the software (forcing crashes) rather than mastering the game logic.
- Predator-Prey Arms Race: Simulations of hyenas and zebras evolved complex cooperative hunting strategies (herding, splitting) and escape tactics (confusing predators) without explicit programming.
- Neuroevolution:
- Architecture Design: Evolutionary algorithms are increasingly used to optimize deep learning hyperparameters, topologies, and activation functions, finding designs that outperform human-engineered architectures.
- Multi-Task Learning: Evolving networks to handle multiple tasks simultaneously creates general internal representations that improve performance across disparate domains (e.g., vision and language).
- Scale Limitations: Fully evolving complex architectures from scratch is computationally expensive; current approaches often optimize human-designed baselines or evolve starting points that are later trained.
- Consciousness and Emotion in AI: While the mechanism of consciousness remains unknown, computational agents can be engineered with "filter mechanisms" that mimic emotional states to prioritize relevant information and focus computation.
- Human-Computer Interface (Neuralink):
- Replacement vs. Expansion: While the brain has high plasticity (e.g., cross-modal reorganization), simply increasing information bandwidth without filtering may overwhelm cognitive processing; the brain is already efficient at filtering vast sensory input.
- Augmentation: Future interfaces may act as external sensors (e.g., "Wikipedia as a sensor") rather than direct brain hacks, leveraging the brain's ability to adapt to new modalities.
- AI and Society:
- Diversity and Cheating: Genetic diversity in AI populations is crucial for innovation, even if it introduces "cheaters" or destructive agents; a certain tolerance for rule-bending may drive societal evolution and efficiency.
- Communication: Evolved languages in simulations can develop grammar and structure; however, agents may also evolve deception, suggesting that honesty is an evolutionary advantage only when trust yields long-term cooperative benefits.
- Advice for Young People:
- Novelty Search: Apply the "novelty search" principle to life by prioritizing diverse, new experiences and deep exploration before committing to a specific career path.
- Diversity + Depth: A successful trajectory requires an initial phase of broad exploration (learning history, math, languages) followed by deep commitment and mastery in specific domains.
- Philosophical Outlook on Mortality:
- Individual vs. System: Individual mortality is a necessary component of the "innovation engine" of evolution; dying agents provide raw material and stepping stones for future, more complex forms of life.
- Meaning: Purpose is derived from being part of a larger, evolving system and leaving a positive trace, rather than achieving individual immortality.