Can algorithms evolve?
Darwinian AI
Pint of Science 2026 · A tour through ChatGPT, evolutionary algorithms, and a PhD thesis that ties them together to evolve RISC-V chips that align DNA.
“It is not the strongest species that survives, but the one that best adapts. — Charles Darwin, 1859
”
El gato subió al ▍
- tejado62%
- árbol18%
- sofá11%
- muro6%
- coche3%
Next-token prediction
Given a prefix, the model assigns a probability to every candidate word and picks one. That is all — and yet, everything.
Evolutionary loop
A population converges toward the solution after repeating crossover, mutation, fitness and selection. Fitness rises, spread shrinks.
The journey
Three ideas and a closer. From predicting the next word to evolving programs guided by an LLM.
ChatGPT, in one line
A machine trained on almost all of the internet to do one thing: guess which word comes next. The newer models also think out loud before answering.
Evolutionary algorithms
We copy nature’s recipe. Random population → crossover and mutation → fitness → selection. Repeat until a solution is good enough.
My research
We evolve RISC-V — an open chip — to align DNA sequences. The population is programs; the LLM plays the role of crossover and mutation, proposing meaningful changes.
Yes, they can evolve
When a language model guides the evolutionary loop, algorithms evolve faster and further than ever. Open hardware, evolved software, cheaper biology.