Metix AI · Discovery Loop Talent Brief

Jeff Dean leaves Google, revealing AI’s next talent gap
AI competition is moving toward a closed loop of models, systems, and science

Jeff Dean, Sanjay Ghemawat, Quoc V. Le, and Oriol Vinyals' public track records point to four capabilities: research systemization, low-level infrastructure, model automation, and agents/evaluation. Metix AI's talent data spans the full depth of these four capabilities — from infrastructure, the deepest pool, to more specialized tracks like Model Serving, AutoML/NAS, Deep RL, and Seq2seq/NLP, all precisely locatable.

Focus Discovery Loop talent stackSupply depth 594 verified · 12 rolesAs of 2026-08-08
Founder stack

01The four founders’ public track records point to four recruitable capabilities

Jeff Dean, Sanjay Ghemawat, Quoc V. Le, and Oriol Vinyals bring public track records across research systems, low-level infrastructure, model automation, and agent evaluation. A Discovery Loop-style organization needs these four capabilities inside the same operating system from research idea to experimental feedback.

4
Four founders, four core capabilities
Spanning 12 hireable roles
594
Verified talent across the project lineage
397 precisely matched
264
Postings at peer companies
21 peer companies · 2026-08-08
3,612
Peer-company talent-record pool
Top four companies hold 50.2%
01The four capabilities already map to a real, verifiable talent pool

Those 594 people sit across 12 concrete roles in six countries, including the U.S., U.K., and Canada — the founders' capability framework already holds up in today's real hiring market.

02Supply concentrates in low-level infrastructure; model automation and agents are thinnest

The 4 low-level infrastructure roles hold 276 people combined, while model automation and agents together hold only 145 — the closer a role sits to training infrastructure, the deeper the hireable pool; the closer to frontier research, the narrower it gets.

03Peer-company hiring centers on infrastructure and domain science

Of 264 postings, Infrastructure/Platform (96) and Domain Science (85) together account for more than half, with seniority concentrated at Mid-Senior level (183) — demand closely mirrors the capability mix the four founders represent.

04Talent supply concentrates in a handful of companies

21 peer companies hold 3,612 talent records combined; Recursion, Lila Sciences, Isomorphic Labs, and Generate:Biomedicines alone account for 50.2% — starting there is more efficient than spreading across all 21 companies.

"Automating discovery to accelerate science and engineering for the world" — Discovery Loop

The four capabilities form a left-to-right chain: research systemization connects models and platforms, low-level infrastructure supports data and scheduling, model automation expands the search space, and agents plus evaluation bring results back into verifiable tasks.

01Research systemizationJeff Dean

Turns research ideas into global-scale AI systems, spanning training through deployment.

Google BrainDistBeliefTensorFlowPathwaysTPUPaLM/GeminiML at ScaleML Platforms
02Low-level infrastructureSanjay Ghemawat

Covers dataflow, storage, scheduling, reliability, and performance.

MapReduceBigtableSpannerPathwaysRPC systemsPerf ToolsData/Storage
03Model automationQuoc V. Le

Scales search and transfer across model architecture, training methods, and reasoning.

Seq2seqNMTNASAutoMLFLANEfficientNetGLaMAlphaGeom
04Agents and evaluationOriol Vinyals

Brings sequence modeling and reinforcement learning into verifiable, multimodal tasks.

Seq2seqAlphaStarAlphaCodedistillationTensorFlowDeep RLMultimodalEval/Bench
Talent lineage

02Project lineage resolves into 12 hireable role profiles

Projects go out of date, capability doesn't. MapReduce and Google's RPC systems belong to the last technology cycle, but the distributed data processing and cross-node communication capability behind them has been repackaged — reappearing today as large-scale training data pipelines and GPU-cluster networking. The four founders' capability lineage resolves into four groups of 12 hireable roles.

Founders, Projects, and Roles at a Glance

Model automation traces back to 11 landmark projects, the most of any group — but most were built in just the last two or three years (Gemini and AlphaGeometry are both post-2023 work), so the field's talent base is still early-stage; the 80 people hireable today are already at the front of it. Low-level infrastructure's project lineage reaches back to GFS in 2003 — two decades of accumulation is what makes its 276-person pool run deep.

Drag to rearrange · Hover to highlight · Data: Metix AI
Jeff Dean Sanjay Ghemawat Quoc V. Le Oriol Vinyals Infrastructure Training Systems Model Automation Agents & Eval
Low-level infrastructure4 roles

Storage and networking deepest, performance narrowest

Storage/database systems (109) and networking (103) carry more than half of this group's pool, but performance engineering has only 19 — within the same infrastructure lineage, supply thins out fastest for tacit tuning skills.

Distributed storage & database
10984 strong/25 weak
Training networking (NCCL/RDMA)
10358 strong/45 weak
AI training data pipeline
4528 strong/17 weak
Training/inference performance
1911 strong/8 weak
Data source: Metix AI
Model automation4 roles

The more specific the direction, the cleaner the match

LLM pretraining (87.5% strong) and model efficiency (100% strong) are both small, but they're the two cleanest matches of all 12 roles — the titles themselves are specific enough that little guesswork is needed.

LLM pretraining research
2421 strong/3 weak
Model efficiency research
99 strong/0 weak
Instruction tuning / alignment
3118 strong/13 weak
AI math / geometry reasoning
168 strong/8 weak
Data source: Metix AI
Large-scale training systems2 roles

TPU is the largest, steadiest pool of all 12

AI Accelerator Systems (TPU) totals 117 people at 82% strong-match — the largest and most precisely countable of all 12 roles, because TPU already reads as a distinct, well-understood job title across the industry.

AI accelerator systems (TPU)
11796 strong/21 weak
Large-scale training systems
5642 strong/14 weak
Data source: Metix AI
Agents & evaluation2 roles

RL is the fuzziest label, code reasoning the scarcest

Reinforcement learning (post-training) totals 60 but only 32% strong-match, the least precise of all 12 — RL-adjacent titles get invoked too loosely. Code reasoning/generation is the opposite extreme: only 5 people found worldwide, the scarcest of the 12.

RL (post-training)
6019 strong/41 weak
Code reasoning / generation
53 strong/2 weak
Data source: Metix AI

Every role was verified with the same search method and matching bar, applied iteratively, so the groups stay comparable. Data source: Metix AI (Mira talent search); scope covers currently working people in the United States, United Kingdom, Canada, France, Netherlands, and Japan, as of 2026-08.

Strong-match samples from four representative roles

6 strong-match people from each of four representative roles: TPU maps to Jeff Dean's signature project; storage and databases map to Bigtable / Spanner, co-authored by Jeff Dean and Sanjay Ghemawat; NCCL/RDMA is the literal example used in the project-lineage narrative above — RPC systems evolving into GPU-cluster networking; LLM pretraining research is where Jeff Dean, Oriol Vinyals, and Quoc V. Le all converge.

AI Accelerator Systems (TPU)

T●● S●● TG-01TPU
Senior Software Engineer - TPU Compiler at Google
United StatesTPU CompilerXLA / compiler
Core compiler work on the TPU Compiler team — a direct match to the TPU project lineage.
J●● S●● TG-02TPU
Senior Staff Software Engineer, XLA TPU Compiler
United StatesXLATPU Compiler
Senior staff-level engineering signal on the XLA TPU compiler.
K●● N●● TG-03TPU
SW Engineer, Cloud TPU — Google, AMD, Ex-Meta, Ex-Intel
United StatesCloud TPUGoogle, AMD, Meta, Intel
Cloud TPU and chip engineering experience spanning Google, AMD, Meta, and Intel.
R●● W●● TG-04TPU
RTL Design Engineer for AWS Trainium
United StatesRTL designAWS Trainium
RTL chip design for AWS Trainium — the same AI-accelerator hardware lineage.
S●● K●● TG-05TPU
ASIC Engineer, Infra Silicon at Meta, Ex Google TPU Lead, PhD
United StatesASICEx Google TPU Lead
Former Google TPU team lead, now working on infrastructure silicon at Meta.
A●● N●● TG-06TPU
TPU Chip Lead at Google
United StatesTPU Chip LeadGoogle
Currently TPU chip lead at Google — a direct match to the TPU project lineage.

Distributed Storage & Database Systems

K●● A●● TG-07Storage/DB
Software Engineer @ IBM Ceph Team
United KingdomCephDistributed storage
Distributed-storage engineer on IBM's Ceph team.
B●● H●● TG-08Storage/DB
Software Engineer | Distributed Systems | TiDB
United StatesTiDBDistributed Systems
Systems engineer working on the TiDB distributed database.
s●● r●● TG-09Storage/DB
Principal Software Engineer, Database Internals and C++
CanadaDatabase internalsC++
Principal-level engineer working on database internals in C++.
R●● V●● TG-10Storage/DB
Staff Software Engineer working on Cloud Spanner at Google
United StatesCloud SpannerGoogle
Currently on Google's Cloud Spanner team — a direct match to Jeff Dean's Spanner lineage.
M●● L●● TG-11Storage/DB
Software Engineer - Distributed Database at Huawei Canada
CanadaDistributed DatabaseHuawei Canada
Distributed-database developer at Huawei Canada.
A●● G●● TG-12Storage/DB
Original creator of Apache DataFusion, Apache Arrow & DataFusion PMC Member
United StatesDataFusionApache Arrow PMC
Original creator of Apache DataFusion — an open-source leadership signal in distributed database engines.

Training Networking (NCCL/RDMA)

S●● W●● A●● TG-13NCCL/RDMA
Cloud Network Engineer, HPC & AI Supercomputing, Azure, InfiniBand/RoCE
United StatesInfiniBand/RoCEAzure
InfiniBand/RoCE high-performance networking on Azure — the RPC-systems lineage reappearing as AI-cluster networking.
X●● L●● TG-14NCCL/RDMA
Senior HPC Performance Engineer, Networking - RDMA / GPU Communication - NCCL
United StatesNCCLGPU communication
Senior HPC performance engineer focused on NCCL/GPU communication — a direct match to the RPC-systems lineage.
H●● M●● TG-15NCCL/RDMA
Senior Research Engineer, NTT Network Innovation Laboratories, RDMA/Infiniband Developer
JapanRDMAInfiniband
RDMA/Infiniband developer at NTT's Network Innovation Laboratories.
R●● T●● TG-16NCCL/RDMA
RDMA Developer at China Telecom Corporation Limited
United StatesRDMAChina Telecom
RDMA developer at China Telecom.
M●● G●● TG-17NCCL/RDMA
Lead Systems Engineer (HPC), NVIDIA InfiniBand RDMA Expert - NVL72 (NVLink)
United StatesNVLinkNVL72
NVIDIA InfiniBand RDMA expert covering the latest NVLink/NVL72 GPU-interconnect generation.
N●● M●● TG-18NCCL/RDMA
Senior Linux System Software Developer, Infiniband at SUSE
FranceInfinibandSUSE
Senior Infiniband systems-software developer at SUSE.

LLM Pretraining Research

Q●● L●● TG-19LLM Pretraining
Senior Applied Scientist, Amazon AGI
United StatesApplied ScientistAmazon AGI
Senior applied scientist at Amazon AGI.
H●● Y●● TG-20LLM Pretraining
Research Scientist @ Meta, (Multimodal) LLM for Recommendation
United StatesMultimodal LLMMeta
Research scientist at Meta working on multimodal LLMs.
S●● V●● TG-21LLM Pretraining
RE @ GDM, RE @ Character AI, Senior MLE @ Square, AI Resident @ Facebook AI
United StatesGoogle DeepMindCharacter AI
Research-engineering experience across Google DeepMind, Character AI, and other frontier-model organizations.
Z●● B●● TG-22LLM Pretraining
Research Scientist @ Meta FAIR, LLM pretraining @ Amazon AGI/AWS
United StatesLLM pretrainingMeta FAIR
LLM pretraining research spanning both Meta FAIR and Amazon AGI/AWS.
T●● R●● TG-23LLM Pretraining
LLM pretraining research at the Allen Institute for AI
United StatesLLM pretrainingAI2
LLM pretraining researcher at the Allen Institute for AI.
H●● P●● TG-24LLM Pretraining
Research Scientist/Engineer II (Senior), Adobe Research, Foundation Model Pretraining
United StatesFoundation Model PretrainingAdobe Research
Senior research scientist/engineer at Adobe Research working on foundation-model pretraining.
Job-posting demand

03Peer-company hiring is filling Infrastructure / Platform and Domain Science

As of 2026-08-08, there are 264 job postings, concentrated in Infrastructure / Platform and Domain Science, with seniority centered on Mid-Senior level. Demand skews toward mature talent that can build platforms, run experiments, and move domain problems forward.

Hiring themes

Infrastructure / Platform and Domain Science lead, with Lab Automation third at 39 postings — the main hiring surface for this peer set.

Infrastructure / Platform
96
Domain Science
85
Lab Automation
39
Research Engineering
25
AI / ML
9
Product / Ops / Business
5
Other / Unknown
5
Data source: Metix AI (2026-08-08)

Seniority mix in hiring

Mid-Senior level has 183 postings, Director has 25, and Entry level has 16. The hiring center of gravity sits in the experience band that can directly own systems, experiments, and project execution.

Mid-Senior level
183
Director
25
Entry level
16
Internship
13
Associate
12
Executive
9
Not Applicable
6
Data source: Metix AI (2026-08-08)
Company map

04The company map shows talent-record concentration in a few AI for Science companies

The 21 peer companies span five lanes, with 16 in AI for Science Platform. Recursion (758), Lila Sciences (368), Isomorphic Labs (361), Generate:Biomedicines (325) rank as the top four talent-record pools and form the first layer of the company map.

Peer-company talent-pool scale

21peer companies
3,612talent-record scale
264job postings
AI for Science Platform16 companies

Models, experimental data, and scientific problems sit in one R&D line.

Autonomous Lab / Robotic Experiment Platform2 companies

Lab equipment, automation scheduling, and result feedback are the main capability signals in this lane.

Foundation Model Research Infrastructure1 companies

Research engineering, training platforms, model serving, and evaluation platforms are the main capability signals in this lane.

AI-designed Hardware / Materials Discovery1 companies

Materials, chips, accelerators, and simulation form the main capability signals in this lane.

Automated AI Research Loop1 companies

Productized research tasks, evaluation, and iteration mechanisms are the main capability signals in this lane.

Talent-record scale in peer companies

The company map concentrates automated discovery talent search into a small set of high-density organizations and reduces noise from broad AI-company pools.

Recursion
758 United States
Lila Sciences
368 United States
Isomorphic Labs
361 United Kingdom
Generate:Biomedicines
325 United States
Insilico Medicine
285 United States
Xaira Therapeutics
202 United States
Absci
175 United States
Sakana AI
173 Japan
Genesis Therapeutics
155 United States
Iambic Therapeutics
144 United States
Data source: Metix AI
FAQ

Questions this report answers

How large is the verified Discovery Loop talent pool in this Jeff Dean report?
Those 594 people sit across 12 concrete roles in six countries, including the U.S., U.K., and Canada — the founders' capability framework already holds up in today's real hiring market.
How many peer-company postings sit next to that 594-person pool?
Of 264 postings, Infrastructure/Platform (96) and Domain Science (85) together account for more than half, with seniority concentrated at Mid-Senior level (183).
How should this report be cited?
Metix AI Talent Intelligence, 2026-08-08. Jeff Dean leaves Google, revealing AI’s next talent gap | Metix AI. https://metix.ai/reports/mapping/ai-next-talent-gap-2026

Automated discovery talent lists can go deeper

Metix AI can split talent records by hiring keyword, company, region, seniority, and technical layer to form a verifiable talent-search list.

Start a Mira trial
Company, job, stack, and candidate-fit views can be expanded.
中文版