In partnership with
News of the day
1. Google DeepMind's AlphaEvolve uses LLMs for semantic evolution, creating novel VAD-CFR and SHOR-PSRO algorithms that surpass human intuition in MARL. → Read more
2. Deepseek is reportedly training its next AI model on Nvidia's banned Blackwell chips, signaling an imminent, market-shaking release that has major US AI labs bracing for impact. → Read more
3. Elon Musk's xAI will integrate its Grok AI model into classified Pentagon systems, raising ethical and security concerns about AI in defense. → Read more
4. AI tool Claude Code drastically speeds up COBOL modernization, impacting IBM's consulting revenue and stock price. IBM defends its platform value beyond code translation. → Read more
Our take
Hi Dotikers!
Darwin would probably have appreciated the irony: a computer system applying natural evolution principles to create better algorithms. DeepMind's AlphaEvolve takes this idea to its logical conclusion by treating source code as a genome that can be mutated, crossed, and selected.
The difference from traditional optimization tools is fundamental. Classic AutoML adjusts numerical parameters, like turning a radio's volume knob. AlphaEvolve rewrites the entire score. The system uses Gemini 2.5 Pro as a genetic operator to modify the very logic of algorithms, introduce new control flows, and invent symbolic operations nobody had imagined.
The concrete application here concerns counterfactual regret minimization algorithms, the technique that enabled AIs to beat the world's best poker players. AlphaEvolve evolved these algorithms and produced two variants with surprising characteristics. VAD-CFR uses an exponential weighted moving average to detect learning instability and adjust its memory accordingly. Even stranger, it amplifies positive regrets by a factor of 1.1 and waits 500 iterations before starting to average its strategies. No human researcher would have proposed this arbitrary threshold, yet it works. SHOR-PSRO blends two solving approaches with a factor that gradually evolves from 0.3 to 0.05.
Most fascinating is that these evolved algorithms outperform their human-designed ancestors on 10 of the 11 games tested. The student surpasses the master, except the student never had a master. It just had time to try, fail, and start over a few million times.
Alexis
Want to get the most out of ChatGPT?
ChatGPT is a superpower if you know how to use it correctly.
Discover how HubSpot's guide to AI can elevate both your productivity and creativity to get more things done.
Learn to automate tasks, enhance decision-making, and foster innovation with the power of AI.
Meme of the day
