AI hashtagagents, evolution is coming for you
Ok, by evolution I mean evolutionary computing (EC), meaning algorithms inspired by natural evolution such as genetic algorithms, evolution strategies, genetic programming and more.
Why? Everyone, their cousins and their kitchen sink (why not? kitchen sinks powered by an openAI operator do not sound that far-fetched) is talking about AI agents taking over the world, or to paraphrase someone famous, agents are eating AI for breakfast. Everything has to become agentic to have even a chance of funding. Don't get me wrong, I am happy to reinvent myself as agentic if need be. I am actually quite agentic but that's for another post. And while I can see the promise of the concept, it remains just that: a concept. Lots of smart people, and also lots of people, are racing to define flavors of agents, agentic workflows and multi-agent orchestration. What I think is the most exciting agentic opportunity is... evolutionary algorithms.
Let's unpack: at the most basic level, it seems that a 2025 agent is an AI that can take actions in the real world. That in itself is a big deal, I guess, with lots of ramifications. But where it gets yummy is when actions have consequences and agents can improve themselves by learning from the feedback. And what better way to evolve than by... evolving? I think EC is a very natural and effective framework for adapting agents and multi-agent systems. My passion being scientific discovery, let me just point at what David Ha the team at Sakana AI have built in the last few months, including their AI Scientist. I will come back to that and to their artificial life generator in another post, but their most recent piece will resonate with AI practitioners : "The AI CUDA Engineer is an agentic framework that leverages frontier LLMs with the goal of automating the conversion of standard PyTorch code into highly optimized CUDA kernels. Through the use of evolutionary optimization, and leveraging concepts in evolutionary computation, such as ‘crossover’ operations and ‘innovation archive’ to discover promising ‘stepping stone’ kernels, our proposed framework is able to not only automate the process of converting PyTorch modules to CUDA kernels, but our highly optimized CUDA kernels often achieve speedups that have significantly faster runtime."
Or, how to evolve DeepSeek efficiency in your sleep.