Jeff Dean Left Google for Discovery Loop. The Follow-On Pool Is 27.

On August 5, 2026, Jeff Dean walked out of Google after 27 years to co-found Discovery Loop with Sanjay Ghemawat, Oriol Vinyals, and Quoc Le. Alphabet dropped roughly 5% intraday, an estimated $160 to $200 billion in market cap, and every AI recruiter with a pulse pointed their boolean at "ex-Google research scientist." That is the wrong pool. The right one is smaller, harder to see on LinkedIn, and closes inside two weeks.

Why the Discovery Loop follow-on pool is 27, not 2,000

The first wave of hires at Discovery Loop will come from a closed network of roughly 25 to 30 people the four co-founders have worked with directly for 14 to 30 years, not the broader ex-Google AI researcher market. Dean himself framed the co-founder cluster on X as a decades-old collaboration. That framing matters for sourcing: trust of that vintage does not scale, and the follow-on list is knowable because it is short.

Discovery Loop is a Delaware public benefit corporation based in Palo Alto, the same legal structure OpenAI and Anthropic use. It will start by automating machine learning research and expand into hardware design, drug discovery, and clean energy. Radical Ventures and Khosla Ventures are co-leading the seed. Google itself is a founding investor and will supply compute for at least the first year.

That last fact rewires the defection math. Google is funding the same departure it just suffered. The implication for recruiters is direct:

• Non-competes will not be enforced against Discovery Loop hires.
• "Google alum joining Google-invested startup" is no longer a conflict signal.
• Assume the next senior researcher exits arrive with a Google check attached, not against one.

Sourcing AI research scientists is the wrong frame here

The persona Discovery Loop actually needs is a distributed-systems engineer who happens to know ML, not a pure model researcher. Automating the scientific experimental loop at scale is a systems problem before it is a modeling problem. That is why the founding four include Ghemawat (MapReduce, Bigtable, Spanner) and Dean (TPU, TensorFlow), not four DeepMind theorists.

Here is what that looks like in the index of professional profiles Refolk maintains:

The scarce persona is not "AI researcher." It is the intersection: someone who has shipped distributed systems in production and has TensorFlow or JAX in their commit history. Globally there are only 2,119 senior research scientists carrying both skills, and Google DeepMind is one of the top three concentrators of them.

2,119

Why hybrids ship this product, not model researchers

Discovery Loop's stated first customer is itself: automating ML research end to end. That means orchestrating thousands of experiments, managing accelerator queues, checkpointing state across regions, and closing the loop with evaluation. Every one of those is a systems problem with an ML wrapper. If you staff it with people whose last five years were spent tuning transformer variants, you get papers. If you staff it with people who wrote the schedulers under those papers, you get a product.

Where the follow-on 27 actually live on the internet

Two weeks after signing, the follow-on hires still look like Google employees on LinkedIn. To catch them before the profile change, source off the artifacts that update immediately or never lag: co-author graphs, GitHub commits, and workshop rosters.

The four highest-signal sources:

1. arXiv co-author graphs. Pull co-authors on MapReduce (2004), Bigtable (2006), Spanner, TPU, TensorFlow, seq2seq (2014), WaveNet, and Deep Q-Networks. Intersect that set. The people who appear on two or more of those papers with Dean, Ghemawat, Vinyals, or Le are the follow-on shortlist.
2. GitHub commit history in tensorflow/tensorflow, jax-ml/jax, and the google-deepmind/ org. Look for authors with commits in the last 18 months who also carry google.com email history in the git log.
3. MLSys and the NeurIPS "Systems for ML" workshop speaker and reviewer lists from 2022 to 2026. This is where the hybrid persona presents.
4. Google Brain and DeepMind internal team pages cached on the Wayback Machine. Team composition rarely publishes cleanly, but archived org pages leak names.

LinkedIn updates lag equity signing by 2 to 6 weeks in this cohort. If you wait for the "Now at Discovery Loop" banner to fire, you are sourcing after the shortlist has closed. This is the exact friction Refolk closes for AI-era sourcing: you describe the person in plain English ("senior distributed systems engineer with TensorFlow or JAX contributions, ex-Google Brain or DeepMind, US-based") and get a ranked shortlist that pulls from GitHub, arXiv, and the open web at the same time, not just LinkedIn.

Sanjay Ghemawat's departure changes what "senior" means

Sanjay Ghemawat leaving matters more for sourcing than Dean leaving, because Ghemawat's lineage is the one Discovery Loop will actively poach from. Dean is the public face. Ghemawat's network is the recruiting engine.

Ghemawat's collaborators are the people who wrote the infrastructure other people's ML runs on top of. That is the follow-on wave. If you are a founder or recruiter chasing this cohort, treat "senior" as a lineage question, not a title question:

• Anyone with a Bigtable, Spanner, or Colossus commit history from 2008 to 2018.
• Anyone who shipped a TPU compiler stack change (XLA, MLIR).
• Anyone in the roughly 25-person cluster around the original TensorFlow distributed runtime.

These profiles typically read "Principal Engineer" or "Distinguished Engineer" at Google, and "Staff" or "Senior Staff" everywhere else. The title mapping across companies is inconsistent, which is why keyword sourcing on "Principal" misses the pool. What is consistent is the paper trail.

The two-week window, mechanically

You have roughly 10 to 14 business days from August 5 before the follow-on wave signs and locks. Here is what happens inside it:

• Days 1 to 3. Founders reach into their trusted network. This is the closed 25 to 30. If you are not already introduced, you are not in this batch.
• Days 4 to 10. Warm intros from the founding four to second-degree collaborators. This is where competitive offers actually land, and it is the window where a fast recruiter can intercept.
• Days 11 to 20. Verbal accepts convert to signed offers. LinkedIn is still silent.
• Days 21 to 42. Profiles update. By the time your saved search fires, the person is gone.

If you are a founder trying to hire against Discovery Loop, or a recruiter placing into NVIDIA, Amazon, Isomorphic Labs, Recursion Pharma, Protillion Bio, or Relation Therapeutics (the named competitors for this exact talent), your intercept window is days 4 through 10. Miss it and you are recruiting from the runner-up pool.

LinkedIn updates lag equity signing by two to six weeks. If you wait for the banner, you are sourcing the runner-ups.

The Vinyals second wave nobody is sourcing yet

Oriol Vinyals's DeepMind lineage means Discovery Loop will pull reinforcement learning and generative-agent researchers, not just LLM people, and almost no one has that in their boolean yet. Vinyals oversaw foundational work on WaveNet and Deep Q-Networks. His follow-on network is RL-flavored.

The recruiters chasing "LLM researcher" boolean strings will miss this second wave entirely. The signal to pull on:

• Co-authors on the original DQN paper and its 2015 through 2019 descendants.
• AlphaStar and AlphaFold contributors who are not on the biology track.
• RLHF and agent-eval infrastructure authors from 2023 to 2026.

Quoc Le's Google Brain lineage adds a third overlapping cluster: neural architecture search, AutoML, and sequence-to-sequence originators. Between the four founders you are looking at three distinct-but-overlapping networks. The people at the intersection of two or more are the highest-probability follow-on hires.

The retention lever slowing the wake

Koray Kavukcuoglu taking over Google DeepMind as SVP reporting to Sundar Pichai, with Demis Hassabis moving to Alphabet Chief Scientist, is a real retention move that will slow but not stop the follow-on wave. That reorganization landed alongside the Dean announcement for a reason.

The people it retains are the ones who wanted a bigger platform inside Google, not the ones who wanted to build. The founders Discovery Loop is targeting are the second group. Kavukcuoglu's promotion does not change their calculus, especially given that Google is funding the exit and supplying compute. If anything, the "keep the compute contract, forfeit the org chart" model makes leaving cheaper, because the researcher does not have to give up the accelerator budget they are used to.

The recent comparable: Nobel laureate John Jumper left Google DeepMind for Anthropic in June 2026. Dean is the largest name to leave, and the first to leave for his own startup rather than a rival lab. The pattern is now two data points and a funded template. Assume more.

What to actually do this week

Five moves, in order, for founders and recruiters working the wake:

1. Build the founder co-author graph. Pull every paper Dean, Ghemawat, Vinyals, and Le have co-authored since 2004. Extract co-authors. Deduplicate. That is your universe.
2. Filter to current Google or DeepMind employment. Cross-reference the co-author list against current employer. The remaining set is roughly 200 to 400 people.
3. Rank by tie strength. Number of co-authored papers, recency, and whether they appear in more than one founder's network. The top 27 to 40 is your intercept list.
4. Source before the signal. Reach out via GitHub, arXiv contact emails, and warm intros. Do not wait for LinkedIn. The whole point is that LinkedIn is the trailing indicator.
5. Have an offer that beats "keep the compute contract." If your pitch is "come do ML at a smaller company," you lose. If your pitch is a specific technical problem the person has spent a decade wanting to solve, you have a shot.

The founders who win the next 30 days will not be the ones with the biggest recruiter budget. They will be the ones who understood that ex-Google AI researcher sourcing, done from LinkedIn, is already too late for this cohort.

FAQ

Who left Google with Jeff Dean to found Discovery Loop?

Sanjay Ghemawat, Oriol Vinyals from DeepMind, and Quoc Le from Google Brain. All four are co-founders. Dean is CEO. The four have worked together for 14 to 30 years, which is why the trusted follow-on network is small enough to enumerate rather than search broadly. Google is a founding investor and will supply compute for at least the first year, and Radical Ventures and Khosla Ventures are co-leading the seed.

Why is the "real" follow-on pool 27 when there are 597 senior TensorFlow researchers in the US?

Because Discovery Loop's first wave will hire from the founders' direct collaboration network, not the broader ex-Google AI researcher market. With 14 to 30 years of co-tenure, that network is roughly 25 to 30 people who have co-authored on the MapReduce, Bigtable, TPU, TensorFlow, seq2seq, or Deep Q-Networks lineages. The 597 and 2,119 numbers describe the pool competitors will fish from once the first wave is closed.

How do I source ex-Google researchers before their LinkedIn updates?

Work from artifacts that update in real time or never lag: arXiv co-author graphs, GitHub commit history in tensorflow/tensorflow, jax-ml/jax, and google-deepmind/, and MLSys or NeurIPS Systems for ML workshop rosters. LinkedIn lags equity signing by 2 to 6 weeks in this cohort, so any saved search that depends on the current employer field is trailing by roughly a month. Refolk pulls from GitHub, LinkedIn, and the open web simultaneously, so the arXiv and GitHub signals surface before the profile change.

Will Google enforce non-competes against Discovery Loop hires?

Almost certainly not, given that Google is a founding investor and compute provider. The Discovery Loop deal structure signals a new operating model where Google keeps the commercial relationship and forfeits the org chart. That reduces legal risk for both the departing researcher and the hiring startup, and it should be assumed as the default posture for future senior Google Brain alumni hiring moves until Google signals otherwise.