the goal is to have an agent that can:
- Understand a complex problem description.
- Generate initial algorithmic solutions.
- Rigorously test its own code.
- Learn from failures and successes.
- Evolve increasingly sophisticated and efficient algorithms over time.
So it only functions with Gemini? Couldn’t this methodology be adapted to any model?
Yeah, it’s just an example. In their Visions for the Future section they mention supporting other LLMs.
the endpoint design of deepseeks api, i think, is cross compatible with OpenAI (or am I thinking of something else?), what about google? I haven’t looked at the codebase yet but I wonder how quickly you could hack in Deepseek support.
A lot of providers (including deepseek) offer an openai compatible api, actually very surprised this project doesn’t use that.
I’m trying to figure out exactly what this does, seems like it’s just some ABCs and a general framework for writing prompt loops with some logging glue?
All the code is placeholder in the actual agent modules.
If you skim the paper, Alpha Evolve is basically just a genetic algorithm which iterates on the output of the agents.
