admin The pair are joined by Google DeepMind research leaders Oriol Vinyals and Quoc Le, creating a founding team whose work has shaped internet-scale computing, neural networks and Google’s Gemini models. Dean will serve as chief executive of the Palo Alto-based venture.
Discovery Loop has been incorporated as a public benefit corporation, a structure requiring its directors to balance shareholder returns with a stated public purpose. The company will initially focus on building AI systems capable of proposing, conducting and evaluating experiments with limited human intervention.
Alphabet will be a founding investor and Google Cloud will provide computing resources during the company’s first year. The arrangement signals that the departures were negotiated on cooperative terms, despite removing four prominent researchers from Google during an intensifying contest for advanced AI talent.
Khosla Ventures and Radical Ventures are also backing the company. Funding figures and its valuation have not been disclosed. Radical Ventures managing partner Jordan Jacobs is expected to join the board.
Discovery Loop’s first research target will be machine learning itself. Its founders plan to develop automated experimental cycles that can test model designs, analyse results and generate improved approaches. The resulting technology could then be applied to areas including chip design, biology, medicines and new materials.
The concept extends the use of AI beyond assisting researchers with individual tasks. Discovery Loop wants its systems to perform much of the complete scientific process, from forming hypotheses and writing experimental code to assessing outcomes and determining the next test.
Dean outlined a similar vision at Y Combinator’s Startup School in San Francisco on July 25. He described systems that could run thousands of experimental loops, potentially accelerating work that would otherwise require large teams and long development periods.
The venture remains at an early stage. The founders had not assembled a wider workforce or secured permanent office space when the company was unveiled on Wednesday. Their immediate challenge will be translating a broad research ambition into dependable systems capable of producing verifiable discoveries.
Dean joined Google in 1999 as one of its first employees and became one of the company’s most influential engineers. Working closely with Ghemawat, he helped create infrastructure that allowed Google’s services to process vast quantities of information across large networks of computers.
Their collaborations included MapReduce, a programming model for processing extensive datasets, and Bigtable, the distributed storage system that supported products including Search and Google Earth. They also contributed to the Google File System and Spanner database technology.
Those systems helped establish the technical foundations for cloud computing and large-scale data analysis. Their principles were later adopted across the technology industry, influencing open-source platforms and the infrastructure used by internet companies, banks and research institutions.
Dean subsequently co-founded Google Brain and played a central role in expanding deep learning across the company. He became chief scientist for Google Research and Google DeepMind and served as a co-technical lead for Gemini, Alphabet’s principal family of generative AI models.
Ghemawat held the rank of Google senior fellow, the company’s highest technical designation. Vinyals was vice-president of research at Google DeepMind and a technical leader on Gemini, while Le co-founded Google Brain and contributed to advances in neural architecture search and automated machine learning.
Their departure creates a significant leadership and knowledge gap as Alphabet increases spending on data centres, custom chips and AI models. Google is competing with OpenAI, Anthropic, Meta, Microsoft and a growing group of heavily financed start-ups for researchers capable of developing frontier systems.
Alphabet chief executive Sundar Pichai held several discussions aimed at persuading the four researchers to remain with the company. Google ultimately agreed to invest in Discovery Loop and collaborate with it on machine-learning systems and infrastructure research.
The company’s launch also reflects a widening shift towards AI-led scientific discovery. Google DeepMind’s AlphaFold demonstrated that machine learning could predict protein structures at unprecedented scale, while Isomorphic Labs is applying related technology to pharmaceutical development. Other laboratories are testing AI agents that design molecules, generate mathematical proofs and optimise computer code.
Discovery Loop’s founders contend that a dedicated organisation can move faster than a large corporate research division. Smaller teams using automated experimentation could potentially investigate more ideas and abandon unsuccessful approaches more quickly, although validating machine-generated findings will remain essential.
Scientific work often depends on physical laboratories, specialised equipment and carefully controlled trials that cannot be fully reproduced through software. AI systems can also generate plausible but inaccurate results, making independent testing, transparent methods and human oversight critical to the company’s credibility.
