Korva is a scroll of the internet’s best ideas.
Hassabis steps back from DeepMind
Hassabis steps back
Hassabis announced this week that he is stepping back from day-to-day leadership of Google DeepMind. He becomes chair of Google DeepMind and remains head of Isomorphic Labs, DeepMind's drug discovery research group. He has been shedding executive responsibilities over the past year, focusing instead on research and AGI strategy.
The change centralizes Google's main AI leadership in California. Kavukcuoglu moved there last year, while Hassabis has stayed in the UK. Bloomberg reported that the leadership changes were driven by a desire to centralize Gemini operations in California.
Jeff Dean's new lab
Jeff Dean is leaving to start a new Alphabet-funded lab called Discovery Loop, alongside senior researchers Sanjay Ghemawat, Quoc Le, and Oriol Vinyals. The next big Gemini model, Gemini 3.5 Pro, has been delayed several months beyond its planned June release.
Compute frustration
CNBC reported that some researchers are frustrated that Google is selling compute to rival companies while they struggle to secure compute for their own research projects. Google's cloud business has grown 82 percent this year, powered by the AI boom.
How did Google Brain and DeepMind end up together?
Google acquired the London-based DeepMind in 2014, and the company stayed a largely separate research group. Google Brain, a lab co-founded by Jeff Dean, Greg Corrado, and Andrew Ng, worked on neural networks inside Google itself. In April 2023, the two were merged into a single unit named Google DeepMind. The merger put the teams behind AlphaGo and behind Google's large language models under one roof.
What does Isomorphic Labs do?
Isomorphic Labs was founded in 2021 as an Alphabet company to apply AI to drug discovery. It grew out of DeepMind's AlphaFold, a system that predicts the three-dimensional shape of a protein from its amino acid sequence. Demis Hassabis and John Jumper of DeepMind shared the 2024 Nobel Prize in Chemistry with David Baker for work on protein structure. The Nobel committee honored Hassabis and Jumper for AlphaFold and Baker for computational protein design.
Why is compute so contested inside AI labs?
Training a frontier model means running thousands of specialized chips continuously for weeks or months. Google builds its own accelerators, called tensor processing units, which it first announced in 2016. The research group Epoch AI has estimated that the compute used to train leading models has grown by roughly four to five times each year. Because the supply of chips is fixed at any moment, every training run competes with other uses for the same hardware.
More from Casey Newton
