The new public benefit corporation will develop AI systems designed to automate machine learning, scientific research and engineering experiments at large scale.
Published: August 6, 2026
Jeff Dean, Google’s longtime chief scientist and one of the company’s most influential engineers, is leaving after approximately 27 years to establish an artificial intelligence startup called Discovery Loop.
Dean is launching the company with three other senior Google researchers: Sanjay Ghemawat, Quoc Le and Oriol Vinyals. The four founders have worked across Google’s large-scale computing infrastructure, Google Brain, Google DeepMind, Gemini and automated machine-learning research.
Discovery Loop has been formed as a public benefit corporation. Its stated mission is to automate machine learning, science and engineering by building AI systems capable of managing complete experimental cycles.
Dean is expected to serve as the startup’s chief executive.
Discovery Loop brings together researchers who played major roles in Google’s computing and artificial intelligence programs.
Jeff Dean joined Google in 1999. His work included contributions to the distributed computing systems that supported Google Search, as well as leadership roles in machine learning and AI research. He co-founded the Google Brain project in 2011 and later became Google’s chief scientist.
Sanjay Ghemawat worked closely with Dean on several of Google’s foundational distributed systems. Their work contributed to technologies used for processing, storing and managing information across large computing clusters.
Quoc Le was a founding member of Google Brain and worked on deep learning, neural architecture development and automated machine-learning systems.
Oriol Vinyals held a senior research role at Google DeepMind and contributed to major AI projects, including the development of Gemini models.
| Founder | Principal areas of work at Google |
|---|---|
| Jeff Dean | Distributed systems, Google Brain, machine learning and AI leadership |
| Sanjay Ghemawat | Large-scale computing infrastructure and distributed systems |
| Quoc Le | Google Brain, deep learning and automated machine learning |
| Oriol Vinyals | Google DeepMind, advanced model research and Gemini |
| Total founders | 4 |
| Expected CEO | Jeff Dean |
Discovery Loop intends to create AI systems that can perform multiple stages of an experimental process.
The planned system would generate possible ideas, design experiments, run or coordinate those experiments, evaluate the results and use the findings to determine what should be tested next.
Instead of researchers conducting one experiment at a time, the company wants to use large-scale computing resources to operate thousands of experimental loops simultaneously.
The founders believe this approach could increase both the number of experiments performed and the amount of useful information obtained from them.
The startup is initially expected to use its technology internally. Its first application will be improving machine-learning algorithms and AI architectures through repeated automated experimentation.
Potential future research areas identified by the founders include:
| Research area | Intended application of automated experiments |
|---|---|
| Machine learning | Testing and improving AI algorithms and model architectures |
| Biology | Evaluating biological hypotheses and experimental results |
| Drug discovery | Exploring possible compounds and research directions |
| Chip design | Testing computing architecture and hardware design options |
| Materials science | Searching for materials with specific physical properties |
| Engineering | Comparing designs and optimizing technical systems |
The company’s proposed research cycle can be represented as a continuous process:
1. Generate a hypothesis
The AI system identifies a research question or proposes a possible solution.
2. Design an experiment
It selects the method, variables and evaluation criteria required to test the idea.
3. Execute the experiment
The system runs a digital experiment or coordinates the required computing and laboratory tools.
4. Evaluate the results
It measures the outcome and determines whether the hypothesis was supported.
5. Update the next experiment
The findings are used to modify the hypothesis, model or experimental design.
6. Repeat at scale
Multiple versions of the loop run simultaneously to explore a larger number of possible solutions.
The company also plans to investigate recursive AI improvement. Under this approach, AI systems would help design and evaluate improved versions of other AI systems. Discovery Loop has not announced a finished product or demonstrated a commercially available version of this technology.
Google’s parent company, Alphabet, is supporting Discovery Loop as a founding investor. Google will also act as a cloud partner and provide computing resources for the startup’s early research.
The relationship means the founders are leaving their positions at Google while maintaining a formal commercial and research connection with the company.
Google and Discovery Loop also plan to collaborate on a research framework involving machine-learning systems and related computing infrastructure.
The startup’s initial funding round is being co-led by Radical Ventures and Khosla Ventures. Kleiner Perkins, Lightspeed, Doerr Capital and Alphabet are also named as participants.
The amount raised and the company’s valuation have not been publicly disclosed.
Discovery Loop was publicly announced before establishing a large operating organization.
At the time of its launch, the company had not disclosed a completed product, customer list, headquarters or public launch schedule. Reporting around the announcement indicated that the founders had not yet assembled a broader team or secured permanent office space.
The immediate focus is expected to be the development of the underlying AI and computing infrastructure required to run automated experimental loops.
| Confirmed information | Information not publicly disclosed |
|---|---|
| Company name is Discovery Loop | Total funding amount |
| Four co-founders | Company valuation |
| Structured as a public benefit corporation | Headquarters |
| Jeff Dean expected to serve as CEO | Initial headcount |
| Focused on automated research loops | Product release date |
| Alphabet is a founding investor | First external customers |
| Google is a cloud partner | Revenue model |
| Six named initial investors | Commercial pricing |
Dean joined Google when the company was still in its early stage and remained through its expansion into search, cloud computing, mobile software and artificial intelligence.
His work covered Google’s early search infrastructure, large-scale data processing, machine learning systems and the formation of Google Brain. He later helped direct AI research across Google Research and Google DeepMind.
Ghemawat also spent much of his career building the large-scale systems that supported Google’s products. Le and Vinyals became prominent researchers during Google’s expansion into deep learning, automated AI development and multimodal models.
| Year | Development |
|---|---|
| 1999 | Jeff Dean joins Google |
| 2011 | Dean co-founds the Google Brain project |
| 2023 | Google introduces the first Gemini model family, with Dean, Le and Vinyals among the listed contributors |
| 2026 | Dean, Ghemawat, Le and Vinyals announce Discovery Loop |
| 2026 | Alphabet becomes a founding investor and Google becomes the startup’s cloud partner |
Discovery Loop’s founding team has extensive experience developing large-scale computing and AI systems, but the company’s central technology remains under development.
The startup has not yet published performance results showing that its systems can independently produce scientific discoveries or operate complete experimental programs without substantial human direction.
Its early work will therefore focus on testing whether automated loops can reliably generate, execute and evaluate experiments at the scale described by the founders.
The company’s first measurable results are expected to come from its internal machine-learning research, where experiments can be conducted through software and computing infrastructure before the system is applied to laboratory-based scientific fields.
Share your thoughts about this article.
Be the first to post a comment!
