On AlphaEvolve from Google

Here’s a quick summary of what you should know about AlphaEvolve from Google.

Hey, Cool Supporters!

Here’s a quick summary of what you should know about AlphaEvolve from Google.

As I've always been saying: LMs have a great potential in scientific research because they can suggest solutions to problems in one domain by borrowing techniques used in other domains in their training data.

When combined with automatic solution evaluators (possible in coding and math), good-ish solutions generated by LMs can be further improved and reevaluated.

This doesn't mean that LMs have become capable of original research, of course, but a big chunk of research useful in practice is "standing on the shoulders of giants."

In AlphaEvolve from Google, the principle is to:

1. Pull a specific program (code) whose execution leads to automatically measurable result that you want to improve (e.g. execution time or memory consumption).