It’s the job of those in the tech industry to big up everything they do. As an academic, I see it as my job to put things back in perspective. This time I’m giving this treatment to AlphaEvolve, which Google DeepMind described in their paper as “A coding agent for scientific and algorithmic discovery”.
You might take a look at the examples in OpenEvolve: https://github.com/algorithmicsuperintelligence/openevolve. There's one where a basic hill-climber is transformed into simulated annealing (at least on their documented run). Another example that's easy to visualise is the circle packing problem - this is one of the problems addressed in Deepmind's paper (probably quite tersely) but there's been various attempts to reproduce/improve on it, including: https://alfredclwong.github.io/blog/2025/06/19/alfred-evolve-0.html
Do you know of any relatively simple concrete examples in which AlphaEvolve transforms a seed program into something significantly better? Thanks.
You might take a look at the examples in OpenEvolve: https://github.com/algorithmicsuperintelligence/openevolve. There's one where a basic hill-climber is transformed into simulated annealing (at least on their documented run). Another example that's easy to visualise is the circle packing problem - this is one of the problems addressed in Deepmind's paper (probably quite tersely) but there's been various attempts to reproduce/improve on it, including: https://alfredclwong.github.io/blog/2025/06/19/alfred-evolve-0.html