Optimise what you can't differentiate
User preference, pass/fail test suites, latency, cost. If you can score it, a population can climb it. No gradient required.
Evolutionary training for language models
Don't train a model.
Evolve one.
Quicksort AI breeds a population of LLM variants, ranks them on the thing you actually care about, and lets the fittest write the next generation.
01 / The loop
One generation,
in five moves.
Each bar is a model variant; its height is how well it scores on your evaluation. Watch the population get ranked, culled and rebred. The average climbs because selection keeps what works.
- SpawnSeed a population of variants from your base model.
- EvaluateScore every variant on your own tasks, judges or metrics.
- SortRank the population by fitness. Pivot, partition, done.
- SelectThe bottom half goes extinct. No hard feelings.
- MutateSurvivors breed. Children inherit, recombine and vary.
fittestweakest
02 / Why evolve
Gradient descent follows the slope.
Evolution explores the map.
Parallel by nature
Every variant evaluates independently. Add more hardware and the search gets wider, not just faster.
Search the whole recipe
Hyperparameters, adapter shapes, data mixes, merge weights and prompts evolve together instead of being tuned one at a time.
Every model has a family tree
Each checkpoint knows its parents and the mutations that made it, so you can trace exactly why a model behaves the way it does.
03 / The genome
A training run,
written as DNA.
Everything that defines a variant lives in its genome. Mutation nudges a few genes; selection decides whether the nudge was a good idea. Try it.
variant g.0001
illustrative
no mutations yet
Be in the
first generation.
We're onboarding a small founding cohort of teams. Leave your email and we'll reach out.