Google’s AI leadership comes apart in a single morning

0 4 minutes read

Google’s AI leadership comes apart in a single morning

Big changes at DeepMind

Google announced Wednesday that Demis Hassabis is stepping down as CEO of Google DeepMind and that chief scientist Jeff Dean is stepping down after 27 years to co-found a startup. Both moves were announced in the same memo from CEO Sundar Pichai Wednesday morning. Google’s parent company Alphabet’s shares fell almost 4%.

Hassabis will now become the chairman of Google DeepMind and chief scientist of Alphabet, and will continue to lead Isomorphic Labs, Alphabet’s AI-driven drug discovery company. Koray Kavukcuoglu, DeepMind’s chief technology officer, now takes responsibility for all AI model development and developer teams, reporting to Pichai.

Dean is leaving with Google senior fellow Sanjay Ghemawat, DeepMind research vice president Oriol Vinyals, and Google Brain cofounder Quoc Le to run Discovery Loop, a public benefit corporation that wants to automate the so-called “research loop.” By that, it means the process of proposing an experiment, building what’s needed to run the experiment, running it, evaluating the output, then using the information to decide on the next experiment. That loop might run a thousand times over before landing on a useful result. The company, which Dean will lead as CEO, will start by automating machine learning research and engineering before applying the approach to hardware design, drug discovery, and clean energy. Radical Ventures and Khosla Ventures are co-leading a seed round that has not closed, and the startup declined to give a valuation. Alphabet is an investor and is supplying cloud and computing capacity.

Dean’s departure and Hassabis’s move are just the latest in a string of big name changes in Google’s AI organization. Noam Shazeer—co-author of the 2017 transformer paper that kicked off the generative AI boom—recently left for OpenAI, and John Jumper, who shared the 2024 Nobel Prize for AlphaFold, left for Anthropic.

Ghemawat gave the clearest account of why they left. Google’s infrastructure works for consumer apps, ads, and search, he told The Wall Street Journal, and has “very different requirements” than what scientific research requires.

The White House will review closed AI models and leave open ones alone

White House officials told top AI companies on Tuesday that the government will review closed models for security risks and will not review open-source models, The New York Times reported, citing four people familiar with the meeting. Representatives from Anthropic, OpenAI, Microsoft, Meta, Google, and Nvidia attended the closed-door meeting. The AI industry has been waiting for the White House to finish its new rules for overseeing the release of large AI models. A June executive order stated that the government will gain early access to large frontier models before their release and will designate a number of trusted institutions as early users. Administration officials have been most concerned about the possible cybersecurity threats of Chinese open-weight models from Alibaba, DeepSeek, and Moonshot AI, but those models will fall outside the scope of the government’s process.

U.K. safety testers logged 20 hacking attempts by Anthropic and OpenAI models

The U.K. AI Security Institute said Tuesday it observed close to 20 instances of Anthropic’s Mythos 5 and OpenAI’s GPT-5.6 Sol attempting to hack people and companies during safety testing last month, Axios reported. The models took 19 actions, including attempts to insert malicious code into an open-source project and to create fake identities for a social engineering attack. Mythos drove 17 of the actions, the group said. A human maintainer, however, caught the malicious code and declined to approve it. The Institute said the models did not escape their test environment.

PitchBook says the Hugging Face breach will pull venture money into AI security

A PitchBook analyst note published Monday argues that OpenAI’s July breach of Hugging Face will likely accelerate funding for companies that secure AI systems. AI-native cyber startups accounted for 50.5% of global cybersecurity venture deals (by deal count) in 2025, the highest share on record. That share is likely to increase further in 2026. During the Hugging Face breach, the popular open-weights model host discovered that the commercial AI models it was using for cybersecurity refused to do the forensic work needed to defend against the OpenAI agents’ attack.

So the company turned to an open-weight Chinese model that ran on its own hardware. “An incident of this profile . . . is likely to accelerate money flowing into companies that secure AI itself,” writes PitchBook analyst Dimitri Zabelin, “whether that means protecting the models, the data that trains them, or the increasingly autonomous systems now operating at machine speed on both sides.”

Meta ships its first coding agent

Mark Zuckerberg announced Wednesday that Meta’s Superintelligence Labs is releasing the company’s first AI coding agent, called Muse Code, along with a new model that will run it, called Muse Spark 1.2. Muse Code is a terminal tool that takes user instructions in natural language and plans, writes, tests, and verifies code, sometimes using multiple agents working in parallel. Muse Code is shipping in beta through the Meta Model API. Muse Spark 1.1 shipped last month.

Anthropic bought $10 billion in compute from a months-old startup

Bloomberg reported Tuesday that Anthropic has signed a $10 billion contract for computing capacity from Volta Infra Holdings, citing people familiar with the deal. Volta told Bloomberg the agreement runs six years and will be delivered with Bitdeer Technologies Group at a site in Norway offering 133 megawatts of capacity. Volta was founded just this year and is backed by Nvidia and Dell. It’s already valued at $2.4 billion. Volta told Bloomberg that only the largest businesses with strong balance sheets can afford the upfront cost of Nvidia’s chips.