Khosla backs Jeff Dean's plan to automate science, starting with AI research

Why it matters

Discovery Loop is testing how far founder pedigree can carry an AI venture before public benchmarks or commercial results. Its machine-learning-first approach offers a faster path to proving automated experimentation than starting in a physical laboratory.

Khosla Ventures Managing Director Samir Kaul used an August 21st Bloomberg Technology interview to explain why Khosla Ventures moved quickly to back Jeff Dean (@JeffDean) and Discovery Loop, which Dean formed with three longtime Google collaborators to automate scientific experimentation.

Dean, Sanjay Ghemawat, Quoc Le and Oriol Vinyals publicly launched Discovery Loop on August 5th after careers spent building much of Google's computing and AI stack. Dean and Ghemawat joined Google in 1999. Le helped found Google Brain, while Vinyals held a senior research role at Google DeepMind and worked on systems including AlphaStar and Gemini.

TechCrunch reported that Alphabet invested in Discovery Loop. Discovery Loop has announced an undisclosed initial funding round co-led by Khosla Ventures and Radical Ventures, with participation from Kleiner Perkins, Lightspeed and Doerr Capital.

Kaul's appearance adds the investor's case for a deal whose core terms remain private. Discovery Loop has not published a round size or valuation. Axios reported on August 6th that a source described the financing as worth "hundreds of millions" of dollars.

The financing terms will show how much of Discovery Loop investors bought for that risk. The public record establishes the investor group and Alphabet's equity participation while leaving the capital structure unpriced.

A founder bet at seed scale

Discovery Loop gives venture investors a concentrated version of the defining AI wager: price the founders' technical history before operating results exist.

The four founders have worked together for decades, according to Dean's launch account. Their shared record spans Google Search, MapReduce, BigTable, Spanner, TensorFlow, TPUs, Pathways, AlphaFold and Gemini. Discovery Loop argues that this experience across chips, distributed infrastructure, models and consumer-scale products is the full stack needed to automate experimentation.

That pedigree explains the speed of Khosla Ventures' decision. It also explains why the financing discussion has outgrown the usual meaning of a seed round. Discovery Loop is still assembling an in-person Palo Alto operation and hiring its founding group.

Start where experiments stay inside the computer

Discovery Loop's first product decision is disciplined. Dean and his co-founders plan to begin with machine-learning research and engineering, where experiments can be proposed, executed and evaluated computationally. Discovery Loop says its eventual system will run thousands of experiments in parallel, learn from the results and select the next tests without waiting for researchers to manage each iteration.

Discovery Loop intends to serve as its own first customer. Dean wrote that rapid internal feedback would guide the founders as they build the infrastructure, models and systems. A useful automated-research system could improve the technology that powers the next version of the same system, creating a tight development loop before Discovery Loop enters medicine, materials, clean energy or other fields where experiments require laboratories, instruments and physical supplies.

That sequencing separates Discovery Loop from autonomous-science businesses that began closer to the wet lab. Lila Sciences, for example, has raised $550 million for scientific models and physical "AI Science Factories" spanning life science, chemistry and materials. Automata raised a $45 million Series C in January to build robotics, software and data infrastructure for automated laboratories. Autoscience raised a $14 million seed round in March around autonomous machine-learning experimentation.

Discovery Loop can initially avoid the robotics and instrument-integration work those approaches demand. Discovery Loop still must show that automated experiments produce reproducible improvements rather than large volumes of cheap, inconclusive trials. Discovery Loop's public materials currently describe the architecture and hiring plan rather than a released system or technical benchmark.

Khosla's science investor

Kaul's enthusiasm for Discovery Loop's science thesis comes with operating history. According to Khosla Ventures' biography of Kaul, he worked on the Arabidopsis Genome Initiative at The Institute for Genomic Research and published research on the first complete genome of a flowering plant. He later helped co-found sequencing venture Helicos BioSciences and served as founding CEO of synthetic-biology venture Codon Devices.

Kaul's background and Khosla Ventures' history in technical businesses fit Khosla Ventures' interest in Discovery Loop's science thesis. The founders remain the decisive asset. Alphabet's investment gives Discovery Loop access to computing infrastructure while preserving an economic link between Google and four researchers who helped build its most important systems. Khosla Ventures and Radical Ventures bring outside capital while betting that the founders can turn their internal research practices into a repeatable product.

Discovery Loop's next proof point is a working loop that produces measurable research gains. Dean, Ghemawat, Le and Vinyals spent their careers building systems used by other engineers and researchers. Discovery Loop is betting those systems can begin doing a larger share of the research themselves.