Discovery Loop's Next 20 Hires: A Pool of 412, Mapped

On Aug 5-6, 2026, Jeff Dean and Sanjay Ghemawat left Google after 27 years to found Discovery Loop with Quoc Le and Oriol Vinyals. Alphabet dropped roughly 4% on the news, adding to the $270 billion already shaved off since June's exodus. If you are a founder or a sourcer, the question is not whether Discovery Loop will hire; it is which people they will call first, and who you should be calling before they do.

The pool Discovery Loop will actually recruit from is tiny

Discovery Loop's next 20 hires will come from a US+UK bench of senior researchers who have shipped Gemini, Pathways, AlphaFold, or TPU-scale RL, drawn out of a wider pool of 412 senior/director-level ML research scientists and engineers with reinforcement-learning skills. That is the entire addressable market for a founding team whose thesis is automating the experimental loop.

In Refolk's index of professional profiles, the senior RL bench across the US and UK is 412 people. Not thousands. Four hundred and twelve. That is the ceiling on Discovery Loop's realistic first-100 recruit pipeline, and it is the same pool that Anthropic, OpenAI, Meta FAIR, Laurent Sifre and Karl Tuyls' new "Holistic" startup, and every AI-for-science lab is fishing in right now.

412

Discovery Loop is targeting roughly 20 hires in year one. To hit that, they need to convert about 5% of the entire senior RL bench, or, put more sharply, poach nearly every Google DeepMind and OpenAI RL senior on the list.

Where the 412 actually work

The senior RL pool is bunched into a handful of employers, and Meta FAIR (not Google) is the single biggest bucket. Below is the shape of the bench as it sits in Refolk's index today.

Two things jump out. First, Meta FAIR employs more of this bench than Google DeepMind does, which means competing founders should be raiding FAIR before the Discovery Loop pipeline consolidates. Second, the SF Bay concentration is 32% of the sample, which is why Discovery Loop planted its flag in Palo Alto. Geography is a moat here, not a preference.

The research-engineer bottleneck nobody's pricing

The scarce hire is not the researcher. It is the research engineer who can wire a training loop across a TPU pod without babysitting it. In Refolk's 25-person sample of the senior RL bench, Research Scientists outnumber Research Engineers 16 to 9, roughly 1.8 to 1. Discovery Loop's whole thesis, "automate ML, science, and engineering research," is engineer-heavy. The person who can propose the experiment is not scarce. The person who can run 10,000 of them unattended is.

If you are hiring against Discovery Loop, do not benchmark yourself on scientist headcount. Benchmark on the ratio of engineers to scientists on your team. If it is below 1:1, you are staffed like a lab, not a discovery loop.

Why Google is funding its own talent leak

Google is Discovery Loop's founding investor and year-one cloud partner, alongside Radical Ventures and Khosla, which means Google is literally underwriting the departure of its own top scientists. That is not sentiment. It is a hedge against the compute grievance that pushed Dean and Ghemawat out in the first place.

Insiders have said researchers grew frustrated over access to the computing capacity they need while Google Cloud sold TPUs to outside customers, including Anthropic. If the top ICs are leaving anyway, better to have them leave into a company you own equity in and rent TPUs to.

For competing founders, this has two consequences:

• Expect a soft non-poach on the 5 Google DeepMind seniors in the pool. Discovery Loop will hunt ex-Googlers who are already at Anthropic, OpenAI, and Meta first, because they are structurally cleaner to close.
• The winning pitch to a DeepMind IC in 2026 is guaranteed TPU quota, not salary. Discovery Loop can offer that on day one. You probably cannot. Adjust your outreach accordingly.

Compute grievance beats comp. Guaranteed TPU quota is the offer letter now.

Who Discovery Loop will call first, by name and by pattern

Discovery Loop's first 20 calls will follow the founders' rolodex, which means ex-Google Brain and ex-DeepMind engineers who have already moved once (to Anthropic, OpenAI, or a stealth AI-for-science shop) and are cleaner to re-recruit than currently seated Googlers. The archetypes are already public:

• Noam Shazeer returned to OpenAI in June 2026, less than two years after Google paid nearly $3 billion in an acquihire to bring him back. All eight authors of "Attention Is All You Need" have now left Google. That is the IC-1 archetype Discovery Loop will try to clone.
• John Jumper, AlphaFold co-creator and 2024 Nobel laureate, now at Anthropic. The AI-for-science archetype.
• Jonas Adler and Alexander Pritzel, Gemini researchers who left for Anthropic. The IC-2 gradient competing founders should exploit: DeepMind to Anthropic is a well-worn path with lower friction than DeepMind to a brand-new PBC.
• Laurent Sifre and Karl Tuyls, reportedly leaving DeepMind for "Holistic," a rumored competing model lab. Same pool, opposite direction.

If you are sourcing against Discovery Loop, the pattern to target is: ex-Google Brain or DeepMind, currently at Anthropic/OpenAI/Meta FAIR, joined their current employer in 2023-2025, has co-authored on RL-from-experience or scientific-domain benchmarks. That is a specific enough query that you can run it in plain English against a good index and get a real list back, which is the exact gap Refolk closes: describe the person in one sentence and get a ranked shortlist across GitHub, LinkedIn, and the open web.

Where competing founders should actually be sourcing

If you are building against Discovery Loop, do not fight for the same shortlist. Fight for the adjacent bench Discovery Loop cannot or will not touch. Four places to look:

1. Meta FAIR. The largest single employer bucket in the senior RL pool (7 of the 25 sampled). FAIR ICs are structurally poachable: no PBC restriction, no Google investor circularity, and Meta's Llama roadmap has left research-heavy engineers looking for more open-ended problems.
2. Isomorphic Labs. Demis Hassabis has stepped back to Chair of Google DeepMind while shifting focus toward AGI and Isomorphic, which makes Isomorphic Discovery Loop's closest sibling on the AI-for-science axis. It is also its most natural talent-swap partner, which cuts both ways: you can raid Isomorphic before the swap starts.
3. Recursion, Insilico Medicine, and academic groups including MIT Jameel Clinic, Stanford SNAP, and CMU's ML department in Pittsburgh (where 2 of the 412 sit). The intersection of "senior RL" and "wet-lab or physical-sciences fluency" is a small subset of the 412, and the domain-fluent tail lives here rather than at FAANG.
4. Waymo and other applied-RL teams. Waymo has 2 seniors in the pool. Applied-RL engineers who have shipped safety-critical systems are exactly the profile Discovery Loop needs to run automated experiment loops without human babysitting.

Running these searches by hand across LinkedIn, GitHub, and lab pages takes weeks. Running them as one plain-English query against a unified index is the difference between shipping a shortlist on Monday and shipping it in November. "Ex-DeepMind engineers at Isomorphic or Recursion with wet-lab publications" is a valid Refolk query, not a research project.

The pre-IPO equity problem Discovery Loop cannot solve

Discovery Loop is a public benefit corporation, which means its equity is mission-aligned but liquidity-constrained, and that structurally loses the 30-40 age cohort to Anthropic and OpenAI. Both of those companies are approaching IPOs and can offer pre-IPO equity that neither Alphabet (a $2 trillion public company) nor a fresh PBC can match on paper.

Where Discovery Loop wins:

• The 45+ cohort. Dean's peers. People who already have liquidity and are motivated by mission and compute access.
• The IC-1 layer where compute quota matters more than upside.
• Researchers who want to work directly with Dean, Ghemawat, Le, or Vinyals. That rolodex is a comp component in itself.

Where Discovery Loop loses:

• The 30-40 IC layer looking to compound wealth before a family-liquidity event.
• Anyone whose partner also works in tech and is doing the household math on Anthropic secondaries.
• Researchers whose ambition is domain-specific (drug discovery, materials) rather than the general "automate ML" mission.

$270B

If you are competing for the 30-40 IC layer, price your equity against Anthropic's last secondary, not against Google's RSU grant, and definitely not against Discovery Loop's PBC common. This is the cohort where a well-run outbound campaign, powered by a plain-English index like Refolk, still beats a founder rolodex, because Dean's rolodex skews older.

What to do this week

Three moves, in order:

1. Pull the Meta FAIR list. Seven of the 412 are there. That is your single densest, cleanest target list. Reach out before Discovery Loop's recruiter pipeline goes live.
2. Map the ex-Google-Brain diaspora at Anthropic and OpenAI. These are the people Discovery Loop will call first. Get on their calendars before the founders do.
3. Build a second list: domain-fluent RL engineers at Isomorphic, Recursion, and academic AI-for-science labs. This is where the real AI-for-science pool lives, and it is not on Discovery Loop's shortlist yet.

The 412 will not grow this quarter. Sourcing right now is a zero-sum game against Dean, Ghemawat, Sifre, Tuyls, and every recruiter at Anthropic. Move first.

FAQ

How many people can Discovery Loop realistically hire in year one?

Around 20, based on the founding team's stated "lean team" positioning and a founder rolodex that concentrates on a small slice of the 412. To hit even that, they need to convert about 5% of the entire US+UK senior RL bench, which is why the pool is effectively spoken for the moment the round closes.

Why is Meta FAIR a better target than Google DeepMind right now?

Meta FAIR is the single largest employer of senior RL researchers in the US+UK pool (7 of the 412), and its ICs are not subject to the soft non-poach that Discovery Loop will observe with its own investor, Google. DeepMind's 5 seniors are effectively locked; FAIR's 7 are not.

What signals matter more than title when sourcing this pool?

Compute access, publication cadence on RL or scientific-domain benchmarks, and whether the person is a research engineer or a research scientist. The scientist-to-engineer ratio in the senior RL pool is 1.8 to 1, so engineers are the actual bottleneck. Look at GitHub contribution patterns on distributed-training or experiment-orchestration code rather than paper counts.

Does Discovery Loop's Google Cloud partnership actually matter to candidates?

Yes, materially. The internal push factor that drove Dean and Ghemawat out was compute grievance, specifically that Google Cloud was selling TPUs to Anthropic while internal researchers waited in queue. A year-one guaranteed TPU commitment is a real, structural offer component that most competing startups cannot match, and it will close candidates that salary alone would not.