PERSON DIRECTORY
Matej Balog
Matej Balog appears in 2 indexed conversations across Machine Learning Street Talk, No Priors. This directory brings every appearance, source, TL;DR, digest, and transcript into one searchable feed.
No Priors Ep. 120 | With Google DeepMind’s Pushmeet Kohli and Matej Balog
Sarah GuoPushmeet KohliMatej Balog
AlphaEvolve turns algorithm discovery into iterative code search: Gemini generates executable candidates, evaluators filter them, and evolutionary search preserves diverse solutions that can improve existing Google infrastructure or tackle long-standing scientific problems.The approach makes precise evaluation the key unlock and bottleneck, with math and computer science offering cheap scoring while biology needs simulators; a 23% training-infrastructure speedup suggests limited self-improvement, but whether gains accumulate remains open.
Google AlphaEvolve - Discovering new science (exclusive interview)
AlphaEvolve found a 48-multiplication algorithm for general 4×4 matrix multiplication, surpassing Strassen’s 49 and showing that evolutionary search can compound verified partial improvements rather than merely sample an LLM more often.At Google, evolved scheduling recovered an average 0.7% of fleet-wide compute, while another change sped training of the next Gemini version by 1%.Scaling remains bounded by evaluator costs, fixed time budgets, specialized domain knowledge, and an unclosed reinforcement-learning loop for distilling evolutionary discoveries into future base models.