Jeff Dean Founds Discovery Loop to Automate Science

⏱️ Lectura: 9 min

A lab founded by four of the world’s most cited AI engineers, Jeff Dean, Sanjay Ghemawat, Quoc Le, and Oriol Vinyals, aims to fully automate the scientific research cycle: proposing, implementing, running, and evaluating experiments, without relying on manual iteration from a human team.

📑 En este artículo

1. TL;DR
2. Introduction
3. What Happened: The Launch of Discovery Loop
4. Context and Background
5. Technical Details and Performance
6. How to Start Testing It
7. Impact and Analysis
8. What’s Next
9. Frequently Asked Questions
   1. What is Discovery Loop?
   2. Does Discovery Loop already have a public product?
   3. What area will Discovery Loop start working on?
   4. What’s the team’s relationship with Google?
   5. What are the company’s long-term goals?
   6. Has Discovery Loop disclosed funding or valuation?
10. References

The company is called Discovery Loop and will start with machine learning before targeting the world’s grand engineering challenges.

TL;DR

• Discovery Loop is the new lab founded by Jeff Dean, Sanjay Ghemawat, Quoc Le, and Oriol Vinyals, former Google DeepMind leaders.
• It aims to automate the entire research cycle: proposing, implementing, running, and evaluating experiments with AI.
• It will start with machine learning research and engineering before expanding into other sciences.
• Discovery Loop will be its own first customer: it will use its systems to optimize its own technology stack.
• The founding team built Google Search, Ads, GFS, MapReduce, BigTable, Spanner, TensorFlow, TPUs, AlphaFold, and Gemini.
• The ultimate goal is the National Academy of Engineering’s Grand Challenges, such as solar energy and clean water.
• Discovery Loop is recruiting a small, in-person team; it has not disclosed a launch date or funding figures.

Introduction

Discovery Loop starts from a simple diagnosis: the scientific method is one of the most powerful tools that exists, but it runs slowly. Proposing a hypothesis, implementing it, running the experiment, reviewing results, and iterating again is a cycle that in most labs still depends on people working sequentially.

According to the project’s official website, discoveryloop.com, that bottleneck isn’t a talent problem but a scale problem: there are too many scientific questions and too few experimentation cycles running at the same time. Discovery Loop’s bet is to build AI systems capable of running thousands of those cycles in parallel, compressing iteration time and raising both the quantity and quality of scientific and engineering output.

What Happened: The Launch of Discovery Loop

The announcement confirms the formation of an unusual founding team. Jeff Dean and Sanjay Ghemawat previously led much of Google’s infrastructure (Google File System, MapReduce, BigTable, Spanner), while Quoc Le and Oriol Vinyals spearheaded central advances of the modern AI era, from sequence-to-sequence models and chain-of-thought reasoning to AlphaStar and AlphaFold.

Discovery Loop isn’t a product a developer can install today. For now, it’s a mission statement and a hiring call: the company is looking to form a small, in-person team to carry out the vision, according to its own website. There’s no public product launch date, and no funding, valuation, or customer figures have been announced.

Context and Background

The team’s pedigree matters for understanding the bet. Jeff Dean co-authored foundational papers on distributed systems like MapReduce and BigTable, the base on which much of modern data infrastructure was built. Sanjay Ghemawat worked alongside him on those same systems and on Google File System, the design that inspired Hadoop and much of the large-scale data ecosystem of the 2010s.

Quoc Le drove, among other work, sequence-to-sequence models and helped popularize chain-of-thought reasoning, the technique that now lets large language models explain their intermediate steps before answering. Oriol Vinyals led AlphaStar, the system that beat professional StarCraft II players, and took part in developing AlphaFold, the protein structure prediction system whose scientific impact was recognized with the 2024 Nobel Prize in Chemistry.

That track record explains the line Discovery Loop uses to sum up its edge: the company says its relative advantage isn’t just the team’s technical ability, but the unprecedented scale of the systems they built before. Discovery Loop claims to combine full-stack depth, from chips to hardware infrastructure, software infrastructure, AI models, and product, within a single founding team.

💭 Key takeaway: the same team that built Spanner, TensorFlow, and AlphaFold is now betting that infrastructure scale, more than a new algorithm, is what’s missing from scientific automation.

Technical Details and Performance

The mechanism Discovery Loop proposes has three steps that repeat in a loop: proposing an experiment, implementing and running it, and examining the results to refine the next iteration. The novelty isn’t in the loop itself (research has worked this way for centuries), but in who runs it and at what speed.

Approach	Who iterates	Parallelism	Main limitation
Traditional scientific cycle	A researcher or human team	Low, experiments in series	Scales with the number of people available
Discovery Loop’s automated loop	AI systems on large-scale infrastructure	Thousands of simultaneous experiments	Depends on the problem having a clear evaluation metric

flowchart TD
A["Propose experiment"] --> B["Implement and run"]
B --> C["Examine results"]
C --> D["Refine hypothesis"]
D --> A

That closed loop is, in essence, an automated version of the classic scientific method. What changes is the unit of parallelism: instead of a lab running one experiment at a time, Discovery Loop’s proposal is to run thousands simultaneously and let an AI system filter out which ones are worth continuing to explore.

How to Start Testing It

Discovery Loop hasn’t yet published an API, a CLI, or an open repository: for now it’s a mission page and a hiring call, not a tool a developer can install today. What you can do is replicate the same pattern (propose, run, evaluate, iterate) with open source tools that already exist for automatic hyperparameter search.

A minimal example with Optuna, a Python library for automatic optimization, illustrates the loop’s core idea. Installation is the same on Windows, macOS, and Linux, since Optuna is distributed as a pip package:

pip install optuna

import optuna

def objective(trial):
    x = trial.suggest_float("x", -10, 10)
    return (x - 2) ** 2

study = optuna.create_study(direction="minimize")
study.optimize(objective, n_trials=50)
print(study.best_params)

This minimal loop tests 50 values of x and tries to minimize (x - 2) ** 2 using Bayesian search instead of brute force. Optuna decides which experiment to try next based on previous results, the same propose-and-refine logic that Discovery Loop describes, though on a much smaller scale.

study = optuna.create_study(
    study_name="tuning_modelo_clasificacion",
    storage="sqlite:///discovery_loop_demo.db",
    direction="maximize",
    load_if_exists=True,
)

study.optimize(objective_clasificacion, n_trials=200, n_jobs=8)

With storage pointing to a shared SQLite database, multiple processes (n_jobs=8) can run experiments in parallel and write their results to the same study. It’s a minimal, desktop-scale version of the massive parallelism that Discovery Loop plans to bring to thousands of simultaneous experiments on its own infrastructure. To confirm the trials ran in parallel, study.trials_dataframe() returns a table with the start and end timestamps of each attempt; if several overlap, the parallelism is genuinely working.

💡 Tip: tools like Optuna, Ray Tune, or Vertex AI Vizier already apply the same propose-and-evaluate pattern automatically, though on a much smaller scale than what Discovery Loop promises.

Impact and Analysis

Discovery Loop’s bet has an honest limit: the automated loop works best when a clear evaluation metric exists, like a model’s accuracy or training time. Scientific questions without an objective metric, much of systems biology or social research, are far harder to fit into that cycle.

That limit isn’t new: it’s the same reason AlphaFold was able to solve protein structure prediction, a task with a very clear success metric (structural match with crystallographic data), but similar tools still don’t solve scientific problems with ambiguous evaluation.

Another limitation the company itself acknowledges is one of scope. Discovery Loop says it will start with machine learning research and engineering, and will use that capability to optimize its own technical stack before expanding into other domains. That means that, at this early stage, the project is mostly a meta-optimization bet: using AI to improve the AI that Discovery Loop builds, not yet medicine, energy, or clean water, the long-term goals the company mentions.

📌 Note: Discovery Loop plans to eventually take on the National Academy of Engineering’s Grand Challenges: better medicines, affordable solar energy, accessible clean water, and cybersecurity, among others.

What’s Next

Discovery Loop hasn’t published a roadmap with dates or verifiable technical goals beyond its own mission page. The only concrete thing for now is the hiring call for a small, in-person team.

The next real indicator for evaluating the project won’t be a mission announcement but a public technical result: a model trained with this method, a paper with reproducible results, or a demonstration that the automated loop beats a human team at a concrete machine learning task. Until that happens, the most honest comparison is with the early announcements of other AI labs: stated ambition tends to run years ahead of public evidence.

📖 Summary on Telegram: View summary

If you want to see the same propose-and-evaluate pattern working today, install Optuna with pip install optuna and run the example above on your own laptop.

Frequently Asked Questions

What is Discovery Loop?

It’s a lab founded by Jeff Dean, Sanjay Ghemawat, Quoc Le, and Oriol Vinyals that aims to automate the entire scientific and engineering research cycle using large-scale AI systems.

Does Discovery Loop already have a public product?

No. For now the company has only published its mission and a hiring call for a small, in-person team. There’s no API, repository, or announced launch date.

What area will Discovery Loop start working on?

The company says it will start with machine learning research and engineering, and will use that capability to optimize its own technology stack before expanding into other domains.

What’s the team’s relationship with Google?

The four founders have long track records at Google and Google DeepMind, where they took part in projects like Google Search, Google File System, MapReduce, BigTable, Spanner, TensorFlow, TPUs, AlphaStar, AlphaFold, and Gemini, among others.

What are the company’s long-term goals?

Discovery Loop mentions the National Academy of Engineering’s Grand Challenges: better medicines, affordable solar energy, accessible clean water, and cybersecurity, as the problems it aspires to solve in the long run.

Has Discovery Loop disclosed funding or valuation?

No. The company’s official website doesn’t mention investment figures, valuation, or customers; the announcement is limited to the project’s mission and the hiring call.

References

• Discovery Loop: official mission page, founding team, and hiring call.
• Wikipedia: Jeff Dean: professional profile and career of the co-founder at Google.
• Wikipedia: AlphaFold: background on the protein structure prediction system recognized with the 2024 Nobel Prize in Chemistry.
• Optuna on GitHub: open source library used in this article’s code examples.

📱 Do you like this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day.

Imagen destacada: Foto de Omar:. Lopez-Rincon en Unsplash