Jeff Dean, Google's chief scientist and employee number 30, is leaving the company after 27 years to launch Discovery Loop, an AI startup focused on automating scientific research. He will serve as CEO. CNBC WSJ

Dean co-founded Discovery Loop with Sanjay Ghemawat, Quoc Le, and Oriol Vinyals, all senior Google AI researchers. The company is structured as a public benefit corporation, a legal form that requires it to balance shareholder returns with a defined public good, in this case accelerating scientific research. The goal is to use AI to speed up scientific discovery by running thousands of experiments at the same time and partially automating the research process. TechCrunch

Discovery Loop also plans to explore recursive self-improvement, the idea of using AI systems to build more capable AI systems. The initial funding round is co-led by Radical Ventures and Khosla Ventures, with Kleiner Perkins, Lightspeed, and Doerr Capital participating. Alphabet, Google's parent company, has also provided financial backing. TechCrunch

Dean confirmed the founding on his official X account. The startup operates the handle @DiscoLoopAI. X / Jeff Dean

Google's CEO acknowledged the departure in a blog post, confirming that Dean is leaving after 27 years and stating that he wants to try something new and that Google is excited to support him. Google Blog

The departure is part of a broader AI leadership reshuffle at Google. On the same day, Demis Hassabis stepped down as DeepMind CEO to take on a chief scientist role. Reuters CNBC

Dean joined Google in 1999 as the company's 30th employee. His technical contributions span the architecture of Google Search itself, the crawling and indexing system and the query-serving infrastructure that powered it, through to the multimodal models behind Google Gemini. He played a leading role in Google's early AI research, helping build the foundation that evolved into Google DeepMind's current model lineup. TechCrunch

His interest in AI-driven scientific discovery predates Discovery Loop. Dean was an angel investor in Profluent, a company using AI for drug discovery, backed in 2024. He also angel-backed Periodic Labs, a startup founded by OpenAI and Google Brain researchers that attracted a $300M venture round in 2025. Earlier, in 2025, Dean was linked to a $100M fund for AI researchers, pledged by Databricks and Perplexity co-founders, that would award $3 million annually over five years, including for AI applied to scientific discovery. TechCrunch (Profluent) TechCrunch (Periodic Labs) TechCrunch ($100M fund)

Dean also co-authored a 2022 study that pushed back against the framing of AI as an inherently environmentally costly endeavor. TechCrunch

The co-founding team brings complementary strengths. Ghemawat, a longtime Google distributed systems engineer, collaborated with Dean on core infrastructure going back to the early search systems. Quoc Le was a principal contributor to sequence-to-sequence learning, a technique where a neural network takes an input sequence and produces an output sequence, which underpins modern neural machine translation. Oriol Vinyals contributed foundational research in sequence models and reinforcement learning, a type of AI training where systems learn through trial and error with rewards, during his tenure at DeepMind. Together, the group covers the systems, modeling, and applied research layers that Discovery Loop's vision of parallelized automated experimentation would require. TechCrunch

The technical ambition is substantial. Running thousands of concurrent experiments under algorithmic control implies a pipeline that can generate hypotheses, design experimental protocols, execute them (likely through automated lab interfaces or computational surrogates), evaluate results, and feed those results back into model improvement, all at machine speed and scale. The recursive self-improvement component adds a second loop: the AI systems improving the AI systems that drive the research. If realized, this architecture would compress iteration cycles that currently take researchers months or years into potentially hours.

The broader context here is a shift in where frontier AI research leadership is choosing to work. Dean's departure from Google, combined with the simultaneous DeepMind leadership reconfiguration, leaves Google's AI apparatus in a materially different shape than it was a week ago. Alphabet's decision to back Discovery Loop financially signals that Google sees more value in maintaining a relationship with this team as an external entity than in retaining them internally. That calculation may reflect the constraints large organizations face in pursuing research directions with uncertain commercial timelines, or the difficulty of justifying the compute and infrastructure costs of large-scale parallelized experimentation within a publicly held company's financial statements.

For the AI research community, Discovery Loop's public benefit corporation structure is a detail worth noting. It places a legal obligation on the company to balance shareholder returns with a defined public benefit. Whether that structure meaningfully constrains decision-making when capital-intensive infrastructure and investor expectations come into play is an open question. But it does place Discovery Loop in a different governance category than the standard venture-backed startup, and aligns with a pattern several AI safety and research organizations have adopted.

The funding lineup, Radical Ventures, Khosla Ventures, Kleiner Perkins, Lightspeed, Doerr Capital, plus Alphabet, represents a notably wide coalition for a company at founding stage. Alphabet's participation is the detail that stands out: a company losing its chief scientist to a startup, and then investing in that same startup, is an unusual sequence. It suggests Google's leadership concluded that Discovery Loop's research direction was better pursued outside the company's walls, with Google positioned as a financial stakeholder rather than an employer.

Dean's track record at Google covers the full arc of modern computing infrastructure: from the systems that scaled web search to planetary size, through the TensorFlow era that made deep learning accessible to practitioners, to the Gemini multimodal models currently competing in the frontier model landscape. The question now is whether that systems-building instinct, applied to scientific research automation rather than information retrieval, produces a comparable leap.