Jeff Dean's First Post-Google Act: Automating the Scientific Method

Key takeaways from Asian American Scholar Forum fireside chat (Stanford Frontier & Pioneer Symposium) with Discovery Loop co-founder and CEO Jeff Dean, in his first public talk since leaving Google after 27 years:

The mission. Discovery Loop is a public benefit corporation set up to automate machine learning science and engineering — building AI that proposes experiments, implements them, evaluates results and iterates, running thousands of research cycles in parallel to accelerate discovery across fields.

Why a PBC. Dean said the company may make decisions that are not in its financial interest but serve the broader societal good, aiming to distribute scientific discoveries as widely as possible rather than lock them up.

The four founders. Dean is building with Sanjay Ghemawat, Oriol Vinyals and Quoc Le — collaborators of 14 to 30 years who together shaped MapReduce, Bigtable, Spanner, TensorFlow, model distillation and neural architecture search. Initial team target: roughly ten people in Palo Alto.

Focused start, broad ambition. The company will begin with a narrow set of domains — starting with ML research itself — before expanding, betting that reusable infrastructure and PhD-level, multi-domain model expertise let it orchestrate multi-agent systems on hard scientific sub-problems.

Why small beats large here. Cloud compute lets a small, focused team raise capital and lean on providers like Google for infrastructure; Dean argued a ten-person company "focused exclusively" on science automation cuts through the distractions of a large organisation.

Mixture-of-experts, in hindsight. Dean revisited his 2017 sparsely-gated MoE work — now underpinning most frontier models — explaining the intuition of huge-capacity models that activate only the relevant expert per token, which showed ~10x better compute-to-quality ratios than dense models.

TensorFlow lessons. He credited its computational-graph abstraction and hardware-agnostic design, but named two mistakes: lacking eager execution (later added, popularised by PyTorch/JAX) and the sprawling contrib directory that left "ten ways of doing everything."

Recursive self-improvement. Dean traced the lineage from neural architecture search and the evolved transformer (~30% more efficient) to Discovery Loop's thesis: automating the full loop of data, evals and architecture to compress experiment cycles from days to minutes.

Read more: Discovery Loop · Tae Kim on X