Discovery Loop Spun Out; Google Kept a Stake

On August 5, 2026, The New York Times reported that Jeff Dean and three other Google executives had launched Discovery Loop outside Google with seed backing from Radical Ventures, Khosla Ventures, and others—and that Google held an equity stake. Four executives left Google, but the capital link survived.

Taken together, Discovery Loop, Periodic Labs, and Lila Sciences support a limited proposition: companies and investors are funding teams that can choose problems, build models, run experiments, and absorb feedback before the cited reporting establishes repeatable discovery.

Google had already put commercialization in a focused company

In 2024, DeepMind CEO Demis Hassabis described AI breakthroughs in biology as a potential business worth more than $100 billion and pointed to Isomorphic Labs as the commercialization route. DeepMind kept the broader biological research program while Isomorphic Labs pursued commercial applications.

By 2025, Lila Sciences had paired a scientific mandate with a confirmed $200 million seed round. The nonprofit FutureHouse had launched a platform and API with four AI-based research tools. Periodic Labs assembled researchers from Meta, OpenAI, DeepMind, and other companies; Discovery Loop added outside venture capital and equity from the executives’ former employer in 2026.

Researchers still have to close the experimental loop

General-purpose model labs build capabilities that many downstream users can apply. Scientists choose a problem, turn a conjecture into a falsifiable hypothesis, design a test, interpret the result, and use it to select the next test. A model can improve one stage without shortening the entire sequence.

Discovery Loop chose drug discovery and chip design, two domains where hypotheses and designs must survive biological or physical constraints.

Two Nature papers on Google Co-Scientist and FutureHouse reported that the systems could perform drug-retargeting tasks by developing hypotheses; one FutureHouse tool also analyzed some of the data. In a separate randomized study at a corporate laboratory employing more than 1,000 researchers, teams using AI discovered 44% more new materials than teams using standard workflows.

The studies leave an autonomous scientist unproven. Google and FutureHouse tested a narrow drug-retargeting task, while the materials study measured teams using AI within a larger research process. Researchers gained value by inserting models into a sequence of decisions and tests.

Google, Lila, and Periodic financed operating capacity

Discovery Loop paired outside seed capital with Google equity. Radical Ventures, Khosla Ventures, and other investors funded the focused company; Google retained exposure to its financial upside; and the founders gained an organization dedicated to drug discovery and chip design.

Discovery Loop disclosed neither its seed-round size nor its valuation, so readers cannot compare its financial capacity directly with Lila Sciences or Periodic Labs. Its standalone structure allows the founders to make the experimental program the governing mandate. Google’s stake establishes shared financial exposure, while the startup’s operating arrangements remain undisclosed.

Lila Sciences had previously raised a $200 million seed. After a $115 million raise, its confirmed Series A total reached $350 million.

Periodic Labs recruited 20+ researchers from frontier AI companies, including Meta, OpenAI, and DeepMind. It later reportedly sought hundreds of millions of dollars at an approximately $7 billion valuation. Its financing remained reported rather than confirmed. Neither company’s cited reporting identified a product available for customers to buy or use.

Researchers who combine infrastructure knowledge, scientific judgment, data practice, and experimental design form the talent moat these investors are trying to secure.

By August 6, 2026, Discovery Loop had disclosed no scientific result in the reporting cited here. Its founders, domains, investors, and Google relationship document organizational capacity and intent, while discovery performance remains unestablished.

Laboratories keep the failed tests that teach the model

When a molecule fails, a material behaves differently under stress, or a chip design violates a constraint, the result supplies a label tied to the physical world rather than a text corpus. The laboratory that ran the test retains the conditions, the result, and the rationale for choosing the next experiment.

FutureHouse packaged four research tools behind a platform and API and stated an ambition to build an AI scientist. Lila Sciences targets novel drugs and materials, domains where physical evaluation can generate specialized feedback for later models and hypotheses.

An outside model supplier may receive only a selected result, while the laboratory keeps the failed test and the reasoning it changed. The cited reporting does not disclose whether Discovery Loop owns laboratory facilities, contracts experiments, or how it would pay for experimental throughput.

Discovery Loop must show that a gain repeats

Discovery Loop will need evidence that lets researchers reject hypotheses, inspect methods, and reproduce results. For one prospective drug-discovery test, the company could preregister a target set, baseline workflow, success metric, and stopping rule. It could time-stamp the model’s hypotheses, chosen experiments, and rejected paths, then compare validated hits, time, and cost with the baseline. An independent laboratory could repeat the experiment; applying the same protocol across several targets would test whether the gain recurs.

Google still occupies two disclosed positions: it develops Co-Scientist internally and owns part of Discovery Loop externally. Four former executives moved the scientific problem list into a standalone company, while Google preserved financial exposure without any disclosed compute agreement or IP transfer. Google kept a stake; the notebook changed address.