Solutions That Evolve: Building Self-Improving Systems with Genetic Algorithms
summary
Genetic algorithm encodes a Shoots-and-Ladders strategy as a 299-gene chromosome (679 total alleles) producing about 10^103 possible strategies, runs six mutated players per generation with ~1,000 simulations each via a rules engine, preserves winners and occasionally mutates — it outperforms a shortest-path baseline but only modestly and risks local minima. A second system targets semantic distance using text-embedding-ada-002 (1536‑dim vectors), evolving prompts/constraints and measuring cosine distance; exploratory services push improved strategies into production (Kafka) for continuous improvement.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 4:23Genetic algorithms for problem solving
- 5:26Snakes and Ladders with GA
- 11:51DNA for game strategy
- 22:46Implementing evolving algorithm
- 29:37Semantic distance with embeddings
- 33:31Measuring semantic distance
- 40:48Challenges in semantic space
- 43:17DNA for semantic distance
- 46:31Evolving word selection
- 51:35Self-improving systems in production
- 53:45Traits of GA problems
- 55:31Summary
related talks
48mTight Genes: Intro to Genetic Algorithms
46mTight Genes: Intro to Genetic Algorithms
46mTight Genes: Intro to Genetic Algorithms
24mGenetic Algorithms in Go
30mEvolutionary Algorithms Perfecting the Art of "Good Enough"
46mBeing Creative with Genetic Algorithms and Typeclasses
54mAmazing Algorithms for Solving Problems in Software
29mSurvival of the Safest: Automating LLM Defense with Genetic Programming