Why the Future of AI Isn't Just Bigger Models. It's Models That Evolve | Risto Miikkulainen of Cognizant

What It Covers

Risto Miikkulainen, VP of AI Research at Cognizant AI Lab and UT Austin professor, explains how evolutionary computation — specifically population-based search, neuroevolution, and evolution strategies — solves problems that gradient descent cannot, enabling creative AI solutions across finance, medicine, scientific discovery, and multi-agent decision-making systems.

Key Questions Answered

• •Evolution Strategy for LLM Fine-Tuning: Cognizant AI Lab demonstrated that evolution strategies can optimize billions of parameters in pretrained models like LLaMA and Qwen without gradient descent. Rather than backpropagation, a cloud of candidate solutions explores the parameter space, rewarding configurations that perform better on target tasks. Oxford and NVIDIA have since replicated this approach, validating it as a viable alternative to RLHF for fine-tuning specialized model behavior.
• •Population-Based Search vs. Gradient Descent: Gradient descent optimizes a single solution incrementally, making it vulnerable to local minima in jagged loss landscapes. Evolutionary methods deploy 30 to 1,000 parallel agents spread across the solution space, using recombination of high-performing candidates to make large jumps. This produces solutions human designers would not anticipate — a result so reliable it has a dedicated "human competitive results" competition at the Genetic and Evolutionary Computation Conference.
• •Quality Diversity as a Discovery Engine: Explicitly rewarding novelty — independent of performance — produces "stepping stone" solutions that unlock further evolutionary progress. Combining novelty rewards with performance rewards, called quality diversity, is now standard in advanced neuroevolution pipelines. Practitioners building creative AI systems should implement dual fitness functions: one scoring task performance, one scoring behavioral distance from all previously seen solutions in the population archive.
• •LLMs as Evolutionary Operators: LLMs can replace traditional crossover and mutation operators by accepting two parent solutions as input and generating a recombined offspring in natural language or code. This means any domain representable in language — molecules, ML architectures, scientific hypotheses, trading strategies — becomes evolvable. Sakana AI applied this to automated research, producing a paper accepted at a major ML conference, demonstrating end-to-end AI-driven scientific discovery is operationally feasible today.
• •Neuroevolution for Metacognition and Continual Learning: Current transformer architectures lack mechanisms for self-knowledge and continual adaptation. Miikkulainen proposes using neuroevolution to discover novel neural architectures inspired by hippocampal circuitry — starting with navigation and spatial memory tasks — where evolved networks must answer whether they actually know a fact versus confabulating. This neuroscience-grounded approach targets two unsolved AI problems simultaneously: catastrophic forgetting and calibrated uncertainty about internal knowledge states.
• •Pandemic Decision-Making as a Deployable Template: Cognizant built a working system that ingested global case, death, and hospitalization data alongside government intervention records to generate next-day policy recommendations — school closures, masking, contact tracing — for any country with available data. Iceland used the system in fall 2021 to guide school reopening decisions, with recommendations reaching the health ministry. This surrogate-model-plus-evolutionary-search pipeline is replicable for any domain where outcome data and intervention diversity exist.