AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

DeepMind's Pushmeet Kohli on AI's Scientific Revolution

· Training Data

Pushmeet Kohli of Google DeepMind joins the Training Data podcast to discuss Alpha Evolve, a new evolutionary AI system that pairs large language models (Gemini Flash and Pro) with automated evaluators to discover entirely new algorithms and tackle long-standing mathematical problems. He frames Alpha Evolve as an evolution of earlier work like FunSearch and AlphaTensor, noting that it removes the template restriction of FunSearch, requires far fewer function evaluations, and can search across large algorithmic spaces in multiple languages including C++, Python, and Verilog for chip design. Kohli situates Alpha Evolve within a broader generator-verifier paradigm that DeepMind has been applying across projects, including the Co-scientist multi-agent system where Gemini plays the roles of hypothesis generator, critic, and reviewer. He argues this architecture mirrors the scientific method itself and that the next bottlenecks are validation (bridging digital outputs to real-world experiments) and accessibility of the technology to working scientists. He uses AlphaFold 2 as a case study, recounting a biologist who solved a decade-old protein structure problem instantly, to illustrate how AI advances, accelerates, and democratizes science while also raising the bar for what problems scientists choose to tackle next. Looking ahead, Kohli says we are in the middle of an AI-accelerated scientific discovery era and predicts human-AI teams will increasingly dominate Nobel-level work, with domain impact spanning energy, materials, healthcare, and chip design. He is bullish on inference-time compute (somewhat), robotics, and humanoids in the medium-to-long term, and flags calibrated uncertainty as a missing capability in current models.

AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.