Richard Sutton's reinforcement-learning group is where it starts. David Silver and Marc Lanctot took their PhDs here. Aja Huang did his postdoc here, years before AlphaGo.
The talent pool behind River AI, and who is hiring from it now.
We had already mapped its founder. Their lineage runs through the network around two landmark papers, and that network has produced labs worth more than $37 billion on closed rounds.
645
Researchers mapped
302
Still at Google DeepMind
45
Founders or chief scientists
$37B+
Value across three labs
In one line
A network that starts companies as readily as it staffs them.
In August 2026, River AI closed a $1.1 billion round two months after launch. Its founder sits in a network this report has been tracking. The people who came out of that network did not join the next company. They started it, and the labs they started now hire from the same pool you do.
The network runs back to 2017, when a group of DeepMind researchers published a method called Policy-Space Response Oracles. It was the game-theoretic idea behind AlphaStar. Two years later AlphaStar reached the top 0.2% of human StarCraft players. We traced the people behind that work and what they have built since.
The network at a glance645
RESEARCHERS
RESEARCHERS
302 remain at Google DeepMind. 45 now hold a founder or chief-scientist title.
$37B+Closed-round value across three founded labs
84Researchers in academia
34Researchers at frontier labs
120Researchers in play
01
The bench you are hiring from is still in one building.
If you are hiring multi-agent or reinforcement learning researchers, this pool is not spread across the market. It is largely one building.
Most of the network is still inside DeepMind. The movement that matters is small in headcount and large in consequence: a thin stream into frontier labs, a return to academia, and a founding wave.
| Destination | People |
|---|---|
| Google DeepMind (stayed) | 302 |
| Academia (Oxford, EPFL, CMU, Mila, UC Irvine) | 84 |
| Frontier labs (OpenAI, Anthropic, xAI, Cohere, Sakana) | 34 |
| Meta FAIR | 9 |
| H Company | 9 |
| Microsoft AI | 8 |
| Ineffable Intelligence | 5 |
| Reflection AI, General Agents and other new labs | ~30 |
We could confirm a current employer for 524 of the 645. The rows name concentrated destinations. The final row is an estimate covering several small new labs together. The rest of the network is spread thinly across other employers.
02
The people who left are now bidding against you.
If you are hiring at senior level, your competition is not the labs you expect. It is people who left this network with a round already closed.
This is the part a flat talent list misses. Each lab below was started or co-founded by someone in this lineage, most in the last two years. This network starts AI companies as readily as it staffs them.
| Lab | Value or raise | What it is | Founders |
|---|---|---|---|
| Ineffable Intelligence | $5.1B | Learning without human data. The $1.1B seed was the largest in European history. | David Silver |
| Reflection AI | $27.5B | Autonomous coding agents and an open frontier model. | Ioannis Antonoglou, a core author on AlphaGo, and Misha Laskin |
| General Agents | Acquired (reported) | Video-language-action agents. Reported acquired by Jeff Bezos's Project Prometheus in 2025, terms undisclosed. | Sherjil Ozair and William Guss |
| H Company | $220M raised | Paris agentic-AI startup focused on computer-use models. Then-largest EU AI seed, May 2024. | Laurent Sifre, Charles Kantor and DeepMind alumni |
| River AI | ~$5B | Personalised, user-owned AI. Closed a $1.1B round in August 2026, two months after launch. | Igor Babuschkin, an xAI co-founder |
03
Where to look before the labs do.
If you are hiring early career, there are two groups worth knowing, and the labs are already watching them.
The senior people are founding companies or already committed elsewhere. The pipeline that keeps refilling is a short list of PhD groups. This is where a hiring manager should look upstream, before the labs do.
The pipeline into DeepMind. Silver taught the reinforcement-learning course a generation learned from, and DeepMind funds PhD places through it.
Yoshua Bengio's institute. Caglar Gulcehre and a cluster of others trained here before London.
Jakob Foerster did his PhD here. His FLAIR group is winding down into BOLD, the UK AI research lab Oxford announced in June 2026, which he now leads while splitting his time with Meta FAIR.
Caglar Gulcehre's CLAIRE group, a live multi-agent and reinforcement learning bench of its own.
04
If you are not in London, you are hiring against six in ten.
If you are hiring outside London, you are hiring against six in ten of this pool.
By the primary city of each researcher's current employer, London alone holds 307 of the 524 researchers with a confirmed employer.
| Location | People |
|---|---|
| London (DeepMind, Ineffable) | 307 |
| San Francisco Bay Area (frontier labs) | 43 |
| Oxford | 15 |
| Lausanne (EPFL) | 14 |
| Montréal (Mila) | 12 |
| Paris (H Company) | 8 |
| Cambridge | 6 |
| Princeton | 5 |
Smaller locations are not listed, so the rows above do not add to 524.
London is where this pool sits. DeepMind is there, and so are the spin-out labs around it. If you want this talent, you need a London presence or a credible London-remote proposition. The other doors are the San Francisco frontier labs and European university cities.
05
Match the specialism, not the brand.
If you are hiring for reasoning and search, you are fishing in the largest specialism here. If you are hiring multi-agent, the original core is smaller than the headline suggests.
From two multi-agent reinforcement learning papers, the co-author network spread across the 2026 frontier. Here are the specialisms it moved into, by how often they show up in recent work.
| Specialism | Mentions |
|---|---|
| Reasoning and search | 113 |
| Alignment and safety | 76 |
| Multi-agent and game theory | 46 |
| World models and open-endedness | 46 |
| Representation and theory | 29 |
| Sequence models and architecture | 16 |
The original multi-agent specialism now sits mid-table. The network's centre of gravity moved to reasoning and search, with alignment close behind. That is where its alumni took it.
Interactive map
The lineage behind the headline.
The most prominent public researchers in the lineage, from the source papers to the institutions they started or now lead.
06
Intelligence for a hiring manager.
A talent map is only useful if it tells you what to do. These are the six reads of this network for anyone hiring from it.
01
The most senior people are founding, not job-hunting.
45 of this network hold a founder or chief-scientist title, and people at that level raise rounds rather than answer job specs. The depth of this network is still on the working bench they came from, the 300-plus at DeepMind.
Move: Offer a co-founder-level mandate, not a job spec.
02
Hire for the capability, not the brand.
The network split into specialisms. Reasoning and search talent sits with the MuZero and AlphaProof people. World-models depth is in DeepMind's open-endedness work. Multi-agent expertise is the original core. Alignment moved to Anthropic.
Move: Match the specialism to the role.
03
The depth regroups when a team breaks up.
xAI's founding research team has broken up. H Company's leadership has turned over. A founder-scale exit at DeepMind pulled a cohort with it. Each time, the talent reassembles somewhere new.
Move: Read team break-ups as a signal about where the work is going.
04
The early-career pipeline runs through two PhD groups.
Oxford's BOLD, which Jakob Foerster leads as FLAIR winds down, and EPFL's CLAIRE under Caglar Gulcehre. Both are run by alumni of this network who went back into academia, Foerster while also splitting his time with Meta FAIR. Both produce the multi-agent and RL researchers this network is made of.
Move: Recruit upstream, before the labs do.
05
They move towards a problem.
Every major exit in this network went to a founding role or a frontier problem. We can see where they went, not what they were paid, so read that as destination and not as motive.
Move: Lead with the mandate, and assume it still has to survive a competing offer.
06
You are bidding against the alumni.
The labs this network founded recruit from exactly the same pool you do. Ineffable and Reflection AI are among them. The people you want are being courted by the people who left.
Move: Know who else is fishing here before you start.
This network was rebuilt from two papers. Every role you are hiring for sits inside a network like it, with its own bench, upstream groups, and people already bidding for it. The difference is whether anyone has mapped yours.
Methodology
How the network was built. Fuze maps technical talent through public publication and professional data. For this report we built the co-author network rooted in the PSRO and AlphaStar work, traced each researcher to their current role, then grouped the network by destination, training, location, and research theme.
What we could confirm. A current employer for 524 of the 645 mapped researchers.
How to read the locations. Location is taken from each person's current employer. It shows where people cluster, not where they live.
Funding figures. From public reporting. The full mapped pool is larger than the names shown in the interactive map.
Network data current to 16 June 2026. The employer figures describe the network as it stood then, and some will have moved since. Funding and valuation figures were re-checked against public reporting on 13 August 2026.
645 researchers. 120 are in play.
We know
which 120.
Do you?
Tell us the role, specialism, seniority, and location flexibility. We will show you the researchers in your hiring zone, the alumni founders already recruiting them, and the adjacent bench worth approaching first.
Book an intelligence briefing.
See the specific talent network around your role before you commit to a search.