Not just OpenEvolve, I’m talking about how agents can use MCP to run code now. Cursor can now have agents iterate on a solution completely on their own.
Genetic algorithms use a fitness function to converge on a solution, so it’s not just random iterations hoping one will work. It’s basically a gradient descent algorithm applied to output from LLMs.
Nowadays, the agents can iterate. It basically works like a genetic algorithm where it converges on a solution.
If you’re talking OpenEvolve you have to specific put in the tests and expected values for each function for this approach to work.
Not just OpenEvolve, I’m talking about how agents can use MCP to run code now. Cursor can now have agents iterate on a solution completely on their own.
Does this not collapse at some point? Like you see with running an image through image to image multiple times?
Genetic algorithms use a fitness function to converge on a solution, so it’s not just random iterations hoping one will work. It’s basically a gradient descent algorithm applied to output from LLMs.