Evolving AI Code

Emergent Garden
Aug 12, 2024
4 notes
4 Notes in this Video

Code Evolver as Human–AI Evolutionary Collaboration

CodeEvolver EvolutionaryDesign HumanAICollaboration
00:40

Code Evolver is a web-based tool built for Google’s Gemini contest that lets humans and large language models co-evolve small web programs—HTML/CSS/JS snippets—via generate–select–mutate iterations.

Evolutionary Search in Program Space with Human Selection

Evolution Code Collaboration
01:10

Code Evolver treats large language models as mutation engines while humans provide selection pressure, together exploring the vast space of possible small web programs rather than trying to “prompt engineer” a single perfect solution.

LLM Code Generation Limits and Training Data Biases

CodeGeneration TrainingDataBias EmergentCapabilities
02:10

General-purpose language models like Gemini and GPT serve as Code Evolver’s code engines, but their strengths and weaknesses shape what kind of software emerges from the evolutionary process.

Practical Limits of LLM-Driven Frontend Development

Code Bias Limitations
03:10

The video’s author, a working web developer, is explicit that Code Evolver in its current form prototypes small web toys but does not replace professional frontend engineering, highlighting concrete boundaries on what LLMs plus a thin UI wrapper can realistically do.