Jeff Dean, Google’s chief scientist and one of many central figures in fashionable AI infrastructure, is leaving the corporate after almost 27 years to co-found Discovery Loop, a startup whose acknowledged aim is to automate the experimental loops of scientific analysis itself. The departure was introduced on August 5, 2026 as a part of a broader reshuffle of Google’s AI management, with the corporate confirming it’ll again the brand new enterprise as a founding investor and Cloud accomplice.
Dean shouldn’t be leaving alone. Three different senior Google AI figures are co-founding the corporate with him: Sanjay Ghemawat, a Google Senior Fellow and Dean’s collaborator on a lot of the corporate’s foundational computing infrastructure; Oriol Vinyals, VP of analysis at Google DeepMind and a technical lead on Gemini; and Quoc Le, a co-founder of Google Mind and the scientist behind Google’s AutoML-Zero venture. Wired’s Steven Levy, who interviewed the 4 founders forward of the announcement, studies the thought got here collectively just a few weeks in the past.
The corporate is integrated as a public profit company. Its mission, as described by itself web site, is to construct programs that may run the complete experimental loop of the scientific methodology: proposing an experiment, implementing and operating it, evaluating the outcomes, and iterating, at a scale and pace that sequential human effort can not match. The founders’ plan is to run hundreds of such loops in parallel, initially pointed at machine studying analysis itself earlier than increasing into domains together with chip design, biology, drug discovery, and supplies science.
What Discovery Loop really plans to construct
The mechanism, per the corporate’s personal description, has three phases. First, it’ll concentrate on automating machine studying analysis and engineering, utilizing frontier AI fashions and large-scale compute to suggest, run, and be taught from evaluations. Second, it’ll act as its personal first buyer, utilizing these automated capabilities to optimize its personal know-how stack. Third, it intends to generalize to any studying loop with measurable outcomes in science and engineering, citing the Nationwide Academy of Engineering’s Grand Challenges as its eventual goal class, together with engineering higher medicines, advancing well being informatics, and making photo voltaic vitality economical.
Le, who pioneered strategies for automating mannequin design at Google, framed the ambition when it comes to what automated loops would possibly discover in AI itself. “I’m very enthusiastic about automating machine studying,” he advised Wired. “It is likely to be that we’ll uncover a special transformer structure.”
Vinyals pointed to the particular functionality hole the hassle rests on: present fashions usually are not sturdy at producing genuinely new concepts to check. “One of many issues that we’ll be clearly very targeted on is how these fashions give you new concepts to attempt,” he mentioned. “That’s not one thing that at the moment they’re tremendous sturdy at.” The founders advised Wired that early programs will co-develop concepts with people, with deeper automation because the aim.
The founders’ report, and Google’s stake in what comes subsequent
The founding staff’s collective observe report is the core asset. Discovery Loop’s web site describes the 4 as representing three of the most-cited researchers in AI and two of the most-cited in distributed programs, with contributions spanning Google Search, Google Translate, the Google File System, MapReduce, BigTable, Spanner, TensorFlow, Pathways, TPUs, AlphaChip, AlphaStar, AlphaCode, AlphaFold, Gemini, sequence-to-sequence fashions, and chain-of-thought reasoning, amongst others. Dean and Ghemawat constructed a lot of the infrastructure that Google’s early search and promoting companies ran on; Vinyals and Le have been central to the neural community period that adopted. Unite.AI has tracked Google’s latest AI mannequin releases, a line of labor Vinyals was instantly serving to to guide.
The departures land inside a wider reorganization of Google’s AI management, detailed in messages from CEO Sundar Pichai and DeepMind co-founder Demis Hassabis. Hassabis is stepping again from day-to-day management of Google DeepMind to change into its chair and Alphabet’s chief scientist, whereas Koray Kavukcuoglu, the unit’s CTO and Google’s chief AI architect, takes over as SVP of Google DeepMind, overseeing Gemini mannequin improvement, frontier AI analysis, and the Gemini app and developer groups. Pichai’s message disclosed that the Gemini app has handed 950 million month-to-month customers and that Gemma fashions have surpassed 900 million downloads.
Google’s relationship with Discovery Loop goes past well-wishes. The corporate is a founding investor, will function the startup’s Cloud accomplice, and plans to collaborate on a analysis framework for ML programs and infrastructure. Wired reported that Google may also provide compute energy for the enterprise’s first 12 months. “Over 27 years, Jeff and Sanjay helped to drive a number of the most important know-how transitions, from our early search infrastructure to the neural networks that helped create the trendy AI period,” Pichai mentioned in an announcement. “We’ll proceed to work with Discovery Loop as a founding investor and Cloud accomplice, and collaborate on a analysis framework for ML programs and associated infrastructure advances.”
On the funding aspect, Wired reported that Khosla Ventures and Radical Ventures are among the many backers, with Radical managing accomplice Jordan Jacobs becoming a member of the board; the founders usually are not disclosing the spherical’s measurement or valuation. The enterprise thesis, as Khosla put it to Wired, is a shift in what AI is for in analysis: people have been utilizing AI to do analysis, he mentioned, whereas the premise right here is that AI is the researcher.
What occurs subsequent
The near-term observables are concrete. Discovery Loop is recruiting a lean, in-person founding staff by means of its open roles web page, having began with no hires and no workplace. Its first executable milestone is the automated machine studying loop it plans to show by itself stack, and its first 12 months of operation runs on Google-provided compute underneath the Cloud partnership. The acknowledged growth path, from ML analysis into chip design, biology, drug discovery, and supplies, will likely be measurable in opposition to the corporate’s personal criterion: studying loops with outcomes that may be scored. Whether or not the strategy produces a consequence the sphere acknowledges as a discovery, moderately than an optimization, is the query the corporate’s construction is constructed to reply.
