Jeff Dean Launches Discovery Loop to Automate the AI Experimental Cycle

Founding Team and Mission

Jeff Dean announced his departure from Google to co‑found Discovery Loop with three longtime collaborators—Sanjay Ghemawat, Oriol Vinyals, and Quoc Le. The company is registered as a Public Benefit Corporation with the mission to automate machine learning, science, and engineering, thereby accelerating discovery and progress.

Core Idea: Automating the Experimental Loop

The pitch deck presents a single diagram that captures the universal scientific cycle: Propose Experiment → Implement and Run → Evaluate → Return to start . The bottleneck, according to the authors, is the manual execution of this loop, which is slow, costly, and unscalable.

Discovery Loop’s solution is expressed in four bullet points:

Drive down iteration time – reduce the duration of each loop iteration.

Propose and automatically run many experiments – generate and execute experiments without human intervention.

Learn from each evaluation – feed evaluation results back into the design of the next experiment.

Scale quality and volume → breakthrough results – increase both the quantity and quality of experiments to achieve qualitative breakthroughs.

The company’s website states that by combining cutting‑edge AI models with massive compute, the system can “quickly propose, run, and learn from evaluations,” running thousands of experiments in parallel and compressing iteration time by several orders of magnitude.

Roadmap: Three‑Step Plan

The public roadmap is deliberately modest and consists of three stages:

Start with Machine Learning – first automate ML research and engineering, the area where the founders have deepest expertise and the cleanest evaluation signals (loss reduction, benchmark improvements).

Act as Our Own First Customer – use the automated ML capability to improve their own tech stack (“eat their own dog food”) before external rollout.

Grand Challenges – extend the approach to any domain with measurable learning loops, such as drug discovery, materials science, chip design, and energy.

The second step is highlighted as a compliant expression of recursive self‑improvement. In an interview with Wired, Quoc Le expressed excitement about automated ML potentially uncovering new Transformer architectures, while Vinyals noted that current models struggle to propose truly novel ideas, suggesting early systems will co‑generate ideas with humans before deeper automation.

Team Track Record

The deck’s most information‑dense slide lists the founders’ achievements across four categories: Products, Infrastructure, AI Research, and AI Applications. A subsequent slide shows a “team‑building” diagram that maps the founders’ management experience and enumerates AI company founders who emerged from their teams, including Dario Amodei, Chris Olah, Tom Brown (Anthropic), Ilya Sutskever (OpenAI/SSI), and others.

Future Timeline

2024‑2025 – Coding Agents achieve the propose‑execute‑evaluate loop on code.

2025‑2026 – Loop Engineering, Agentic RL, and self‑evolving skills extend the loop to agent behavior; AlphaEvolve and AI Scientist initiatives begin scientific research.

August 2026 – Jeff Dean aims to turn the experimental loop itself into infrastructure, starting with ML research and ultimately targeting recursive self‑improvement.

https://x.com/JeffDean/status/2085083442669318443
https://www.discoveryloop.com/