Jeff Dean Leaves Google to Automate Science Itself
August 5 was Jeff Dean's last day at Google, after 27 years. He is leaving with Sanjay Ghemawat, the co-author of MapReduce and half of the most famous pair-programming duo in infrastructure history, plus Quoc Le and Oriol Vinyals. The four are founding Discovery Loop (discoveryloop.com), a public benefit corporation for AI-driven science and engineering. The story took 461 points on Hacker News, and the simultaneous DeepMind reshuffle took another 337.
What Discovery Loop actually does: automate the experimental loop. Propose experiments, implement them, run thousands of evaluations in parallel, learn from the results, iterate. First target is machine learning research itself, the thing all four founders spent their careers doing by hand. Eventually: medicine, clean energy. Google invested in the company, and the departure is described as friendly, which tells you Google would rather own a slice of Dean outside than lose him entirely.
Same day, Demis Hassabis stepped aside as Google DeepMind CEO. He becomes Chair of the unit and Chief Scientist of Alphabet, saying he wants "time and space to focus on the big picture." Koray Kavukcuoglu, DeepMind's CTO, takes over operations as SVP reporting directly to Sundar Pichai. In one day, Google lost its chief scientist to a startup and moved its most famous AI leader out of the day-to-day.
The signal worth paying attention to: the people who built TensorFlow, MapReduce, seq2seq, and AlphaFold-era DeepMind looked at the frontier and decided the leverage is not in training the next model. It is in the loop around the model β the automated propose-run-learn cycle. That is the autoresearch thesis, the same one behind the agent experiments we have covered all year, now endorsed by the four most credentialed systems-and-research people on the planet. When Jeff Dean thinks the next TensorFlow is an experiment loop, that is worth more than any benchmark.
← Back to all articles
What Discovery Loop actually does: automate the experimental loop. Propose experiments, implement them, run thousands of evaluations in parallel, learn from the results, iterate. First target is machine learning research itself, the thing all four founders spent their careers doing by hand. Eventually: medicine, clean energy. Google invested in the company, and the departure is described as friendly, which tells you Google would rather own a slice of Dean outside than lose him entirely.
Same day, Demis Hassabis stepped aside as Google DeepMind CEO. He becomes Chair of the unit and Chief Scientist of Alphabet, saying he wants "time and space to focus on the big picture." Koray Kavukcuoglu, DeepMind's CTO, takes over operations as SVP reporting directly to Sundar Pichai. In one day, Google lost its chief scientist to a startup and moved its most famous AI leader out of the day-to-day.
The signal worth paying attention to: the people who built TensorFlow, MapReduce, seq2seq, and AlphaFold-era DeepMind looked at the frontier and decided the leverage is not in training the next model. It is in the loop around the model β the automated propose-run-learn cycle. That is the autoresearch thesis, the same one behind the agent experiments we have covered all year, now endorsed by the four most credentialed systems-and-research people on the planet. When Jeff Dean thinks the next TensorFlow is an experiment loop, that is worth more than any benchmark.
Comments