Jeff Dean leaves Google to start Discovery Loop
This is one of the clearest signals that the next frontier for frontier-model veterans is not only building models, but automating the experimental loops behind ML, science, and engineering itself.
Links: Original source · Shared link · Related link · Related link 1 · Related link 2
Logged at IST: 2026-08-06 10:21 IST
What it is: Jeff Dean’s public farewell note from Google, plus the new Discovery Loop homepage for the public benefit corporation he is starting with Sanjay Ghemawat, Oriol Vinyals, and Quoc Le.
Gist: Dean says he is leaving Google after 27 years, having watched it grow from 25 people to more than 190,000. His internal farewell note frames the work as a shared accomplishment across consumer products, large-scale infrastructure, research, hardware, and AI systems: Search, Ads, News, Translate, MapReduce, BigTable, Spanner, DistBelief, TensorFlow, Pathways, TPUs, Google Brain, Gemini, Gemma, model distillation, mixture-of-experts architectures, word2vec, neural architecture search, and multiple generations of LLMs.
The new company is Discovery Loop. Its site describes the mission as automating the experimental loops of science and engineering: propose an experiment, implement and run it, examine results, and iterate. The initial focus is machine-learning research and engineering, using frontier AI models and large-scale computational infrastructure to run many experiments in parallel and compress iteration time.
The interesting part is not just the personnel move. It is the thesis: the same people who built Google-scale compute, data systems, ML infrastructure, and foundation-model products are now aiming that stack at the process of discovery itself. Discovery Loop’s own language is explicitly full-stack: chips, hardware infrastructure, software infrastructure, ML models, and products, applied first to ML automation and later to broader National Academy of Engineering-style grand challenges.
Newsletter angle: Good marker for a larger shift from “AI helps researchers write code” to “AI systems operate the experimental loop.” The talent history makes it worth tracking, but the durable idea is the productization of discovery infrastructure.