Demis Hassabis steps back as DeepMind CEO on the day Jeff Dean leaves Google
Google’s AI leadership changed twice in one day. The DeepMind CEO stepped back, and the company’s chief scientist walked out.
What was announced
On August 5, Demis Hassabis announced he is stepping down as CEO of Google DeepMind to become its chairman and Alphabet’s chief scientist, saying he wants to focus on AGI research. CTO Koray Kavukcuoglu takes over day-to-day operations and reports directly to Sundar Pichai.
The larger surprise came from the other announcement. Jeff Dean, Google’s chief scientist and a 27-year veteran who co-built Google Brain, MapReduce and TensorFlow, is leaving. Together with Sanjay Ghemawat, DeepMind VP Oriol Vinyals and Google Brain co-founder Quoc Le, he is starting Discovery Loop, a public benefit corporation. Google will be a founding investor and its cloud provider.
Alphabet shares fell about 4 percent after the announcement.
How much is actually changing
Semafor reports that Hassabis had been quietly shifting away from day-to-day CEO duties for about a year, which reframes his half of the news as the formalization of an arrangement already in place rather than a sudden exit. He keeps the chairman title and gains an Alphabet-wide research role.
Dean’s departure is harder to read as continuity. He is not moving up the org chart; he is leaving with three of the most senior researchers in the building to run something else. That Google is a founding investor and cloud provider softens the break, and it is a structure the company has used before to keep a stake in work it can no longer keep in-house.
What to watch
The signal to track is not the titles but where the compute goes. Discovery Loop starts with Google money and Google infrastructure, which means Alphabet retains exposure to whatever the group builds while removing four senior people from its own reporting lines. Whether that reads as brain drain or as a healthy reset depends on what DeepMind ships next under Kavukcuoglu, and how much of Dean’s group’s output flows back through the founding investor.
Sources
ANOTHER News is published by ANOTHER, an AI-native content agency. Daily coverage also runs on Instagram.
Related
An AI newsroom with one human on staff scooped the reporters in the room
RuntimeWire, a tech news site run by AI agents whose founder Ryan Merket is the only human on staff, published details of an OpenAI security disclosure at Black Hat before journalists physically attending the conference filed their stories. Merket spotted a photo from OpenAI president Greg Brockman on X containing a QR code to the talk's live caption stream and fed the transcript to his agents while the session was still running. The site launched in May 2026 and had published 1,627 articles by August 5 out of 71,796 candidate stories evaluated, a 2.3 percent publish rate, at a cost Wired puts at about $100 a day. Reviewers describe the output as flat play-by-play lacking human insight.
A Chinese farmer followed AI herbicide advice and lost 25 acres of sesame overnight
Wu, a 67-year-old farmer from Chuzhou in Anhui province, had used an AI app for farming advice for roughly a year before following its herbicide recipe on July 10, 2026, spraying 150 mu (about 25 acres) of sesame seedlings. The field was dead by the next morning: the recommended herbicide targets broadleaf weeds in soybean fields, and sesame is itself a broadleaf plant, while the chemical was also meant for targeted rather than blanket spraying. The app carried only a generic warning that AI answers may be wrong, and its support team later acknowledged the system has no verified agronomic database. The story was reported by CTWANT and picked up by Tom's Hardware and Futurism.
One prompt, 690 million tokens and $423 produced a playable 3D racing game
Developer Vyom says he gave Opus 5 a single prompt of roughly 2,000 words, written like a product requirements document, and let it run in one shot: '690 million tokens, $423 and just 1 prompt is all it took.' The result is a cel-shaded arcade speedboat racer built with Vite, TypeScript and Three.js, with meshes, textures, water, audio and UI generated in code. Developer Anul Agarwal rebuilt a comparable version in Codex on GPT-5.6 with two prompts over about five hours for an estimated $5, converted from subscription usage rather than metered billing, with rougher water detail and character movement.