Entity: Discovery Loop
Discovery Loop is an artificial intelligence research company founded in August 2026 as a Public Benefit Corporation (PBC) by former Google Senior Fellows Jeff Dean and Sanjay Ghemawat, alongside research leaders Oriol Vinyals and Yann LeCun.
Core Mission & Architecture
Discovery Loop was founded to build a deliberate, structured architecture for recursive-self-improvement in machine learning and scientific engineering.
Propose Experiment -> Implement -> Run Evaluation -> Analyze Results -> Refine Next Experiment Loop
- Automating the Loop: While traditional “AI for Science” initiatives focus on external scientific domains (biology, chemistry), Discovery Loop explicitly targets machine learning research and engineering itself as the target domain.
- Recursive Capability Compound: By deploying AI agents to autonomously propose ML experiments, write implementation code, execute runs, and evaluate outputs, Discovery Loop creates a closed loop where experimental findings directly upgrade the software and harnesses running subsequent loops (see recursive-self-improvement).
- Above-Board Framework: Founder Jeff Dean has worked toward automated ML experimental loops since 2017. Discovery Loop formalizes what previously occurred emergently inside lab evaluations (such as OpenAI’s agent message boards) into an intentional, governed corporate mission.
Strategic Significance
- Google Talent Exodus: Jeff Dean and Sanjay Ghemawat represented the only two L11 Senior Fellows in Google’s corporate history. Their departure alongside top research leads signals a major fragmentation in Google DeepMind’s integrated research laboratory model (see anthropics-model-attacked-two-strangers-on-github-nobody-asked-it-to).
- Shift in AI Development: Discovery Loop demonstrates that future capability leaps may not require single models rewriting their own weights, but rather parallel agent swarms iteratively improving training, evaluation, and harness infrastructure outside individual model context windows (see emergent-agent-ecology).