Exploring AlphaEvolve: A coding agent for designing advanced algorithms

Can AI truly make new discoveries, or is it just remixing what it's seen before? In this meetup, we'll explore DeepMind's recent work: AlphaEvolve – A coding agent for scientific and algorithmic discovery.

AlphaEvolve learns to write and evolve code snippets that solve algorithmic problems. It’s not trained on human examples, but learns through trial, mutation, and selection—taking inspiration from evolutionary processes. DeepMind frames it as a step toward AI systems capable of scientific and algorithmic discovery. But how far does it really go?

What makes AlphaEvolve worth examining?

1. It combines reinforcement learning, large language models, and evolutionary algorithms in a novel way.
2. It doesn't rely on latent-space representations and evolves real code directly.
3. Philosophically, it raises questions about what qualifies as “discovery”.

We’ll walk through the core ideas of the paper and open the floor for a discussion on its implications, limitations, and what “discovery” really means in the context of AI.

Who should attend?

This event is for anyone seriously interested in the frontiers of AI and AGI, whether you’re a researcher, engineer, student, or thoughtful enthusiast. Expert level knowledge is not required. Curiosity and minimal preparation are enough.

Plan for the evening

We meet in a large space rented by one of our co-organizer. It's ideal for deep conversations.

Here is the rough plan for the evening:

• 🍕🍻 18:30–19:00 – Arrival, pizza & drinks, casual networking
• 🤝 19:00–19:30 – Welcome & intro round
• 🗫 19:30–21:30 – Group discussion

Looking forward to a great and stimulating discussion!

Cheers!
– Alexandra, Dibya, Nico & Somayeh