Google DeepMind has once again used large language models to find new solutions to long-standing problems in mathematics and computer science. This time, the company has shown that its approach can not only solve unsolved theoretical puzzles, but also improve a number of important processes in the real world.

The new tool from Google DeepMind, called AlphaEvolve, uses the Gemini 2.0 family of LLMs to generate code for a wide range of different tasks. LLMs are known for being more or less successful at coding. The key feature is that AlphaEvolve evaluates every suggestion made by Gemini, discards the bad ones, and improves the good ones in an iterative process until it has created the best possible algorithm. In many cases, the results are more efficient or more accurate than the best existing (human-written) solutions.

AlphaEvolve offers the following advantages:

– A self-improving agent that combines the generative power of Gemini Pro and Flash with automated evaluators and an evolutionary algorithmic loop.
– Generates complete algorithms in code, executes them, evaluates them, and refines them iteratively based on performance
– It is capable of making data centers, chip design, and AI training more efficient.

You can find the complete article here: Google DeepMind’s new AI agent uses large language models to crack real-world problems | MIT Technology Review

Learn more from a customer project where SCX Demand AI was implemented at a customer's site based on multi-agents and LLMs: https://scxpartner.com/works/den-markt-hoeren-bevor-er-spricht-wie-ein-supermarktnetzwerk-mit-ki-die-nachfrage-neu-definierte/