Company teardown · Lila Sciences

Periodic hired OpenAI. Lila hired five university labs.

Periodic Labs was founded by a frontier-lab crew. Lila's chiefs came from MIT, Caltech, Harvard and UCF — and brought students, postdocs and collaborators with them. Three years, $550M and 535 people later, we read 98 of its AI, robotics and science staff to see how a frontier lab gets assembled from faculty — and what it shows academic labs about their own path to the frontier.

Get the full visual brief as a PDF.
Founded
2023 · Cambridge, MA
People
98 read · 535 in all
Raised
$550M · $1.3B valuation
North star
Scientific superintelligence
Sells
Discovery, by partnership

0.1 — Lila Sciences in one page

A biotech incubator's playbook, a robotic lab in its own building and $550M — aimed at automating the scientific method itself.

98: 535
Profiles read, of the company
The AI, robotics and founder subset — roughly two-thirds of LinkedIn's Research facet.
$550M
Raised, at a $1.3B valuation
Two years in stealth, then $200M seed and a $350M Series A in seven months.
61%
Hold PhDs — 60 of 98
More than double Periodic Labs' one in four.
1 · The idea
Science is bottlenecked on running experiments, not on having ideas. Lila pairs a reasoning model with robotic “AI Science Factories” so the model can propose, test and learn from the result.
2 · The business
Flagship's company-building model, pointed at AI: partners get Lila's model, its scientists or its factory time; discoveries are licensed out, not manufactured. First customers since October 2025.
3 · The founders
Ten of them: a Flagship general partner who has founded six companies, a Harvard professor as CTO, MIT and Caltech professors as chiefs, a Navy SEAL as president — and George Church as chief scientist.
4 · The team
535 people in three years; we read 98 of them. Three in five hold PhDs, hired from Genentech, Cohere and Boston's robotics belt rather than frontier labs — professors as chiefs, their students following them in.
5 · The lessons
Hire the chair and the lab follows; lift a bench behind its leader; keep a seat for fresh eyes; source AI-for-biology from pharma — and know this playbook assumes an incubator and $550M.

Everything here is public — LinkedIn, the Greenhouse job board, funding announcements, and four founder appearances with transcripts. The reading is ours.

§ 01 — Part one of five · The idea

What if the bottleneck in science is the experiment, not the idea?

Lila's founders think a model that has read everything still can't discover anything until it can test — so they built the lab first, and the company around it.

1.1 — The problem they're solving

Why a model that has read every paper still can't run one.

Reason oneIdeas are cheap, tests are slowA hypothesis is worth nothing until it meets a bench. “All competitive advantage in science will flow to those who can run the most brilliant next experiment,” says von Maltzahn.
Reason twoPapers don't reproduce“Pick a random Nature paper and want to repeat it — that could take a year, or longer.” The know-how lives in laboratories, not in the text a model was trained on.
Reason threePublic data is nobody's moatBuild on public data and “all a competitor has to do is hire one person that knows the secret.” Only data you generated yourself is defensible — so generate it, at scale.

— The founder's framing

“The ability to conduct the scientific method at a level beyond human intelligence at every step of the process.”

— von Maltzahn, defining scientific superintelligence. IBM Think, November 2025.

— Scientific superintelligence, translated

Not a better predictor: a system that does every step of the method — ask, design, run, measure, decide what's next — better than people do. The lab is part of the model's loop, which is why robotics engineers and thin-film chemists sit beside RL researchers.

1.2 — The idea

A mind for science, and a body to go with it.

A reasoning model proposes and designs; a robotic facility makes and measures; results return as reward. “Building a beautiful, superintelligent mind for science requires a new kind of body for science,” writes von Maltzahn.

01 · Reason

Lila Iris — the model A reasoning model trained with RL “across both simulations and laboratory experiments.” It proposes, and designs the test.

02 · Make & measure

AI Science Factory Instruments under AI control — thin films, cell-free protein synthesis, X-ray characterization, imaging — run by robots.

03 · Learn

Scientific tokens Every result, failures included, becomes training signal. “Hundreds of thousands” of closed-loop experiments in the first factory.

“the outcomes become the reward signal” — Andrew Beam, CTO

AI Science FactoryA building, not a metaphorLila's name for a facility where model, software and custom hardware close the loop. The first sits in Flagship's Cambridge building; more are planned for San Francisco and London. Its staff: 27 of 50 open roles.
Scientific tokensResults as training dataLila's term for experimental results as training data. The stated aim: “more instruments under AI control than any company on earth” — and from them, new scaling laws for scientific intelligence.

1.3 — The wager

The wager isn't the loop. It's the scale of it.

300 → 535
People — March to September 2026.
10
Order of closed-loop experiments in the first factory.
3
Hubs — Cambridge, San Francisco, London.
50
Open roles — 27 of them for the factory.

Every rival runs a version of the loop. Lila's bet is that the winner runs the most experiments — so it is building the largest team in the category by a factor of five and organising it like a special-forces network: “small teams applying a common platform… across a very wide diversity of fields.”

— The bet, in his words

“The leader in this pursuit will be the entity that runs the scientific method at the largest scale, speed and intelligence.”

— Geoffrey von Maltzahn, launch announcement, March 2025.

— The operating model

Team of Teams — the doctrine of president Chris Fussell, ex-SEAL Team 6 — is the operating model. The delivery anecdotes come in threes and fours: “a team of about four did that in less than a few months”; “the team of three has run over 600,000 games.”

§ 02 — Part two of five · The business

Flagship's playbook, run on a factory.

Lila doesn't plan to make drugs or materials. It plans to sell the finding of them — as a model, as a service, as factory time — the way its incubator sells companies.

2.1 — The product, as described

Three ways in — and a proof asset, Flagship-style.

01
Who buys
R&D organisations in pharma, materials, energy — and government
Partners “who require extraordinary fidelity of their data” and security. IQT and Analog Devices sit on the cap table; the Department of Energy arrived via Genesis Mission.
02
What they get
Iris · Lila's scientists · factory time
The reasoning model “in the hands of your own scientists”; Lila's scientists on the problem; or factory access with “a proprietary learning loop that compounds with every experiment.”
03
How it's sold
Partnerships and licences, not self-serve
Multi-year discovery agreements; results licensed out. The commercial org is already hired: a Chief Revenue & Product Officer, a VP of Life Science Product, two product directors.
04
What it isn't
A drug company
“We don't plan to manufacture drugs or commercialize clean energy solutions ourselves.” Yet a CAR-T therapy designed on the platform beat a marketed one preclinically. The proof, not the product.
05
Today
First customers · three DOE awards
A “first cohort of customers” since October 2025; Genesis Mission Phase I awards with Caltech and Berkeley Lab, Northwestern and Argonne, and NREL; a Bill Gates visit this summer.

A tenth of the company is in Business Development — and a product director just became Senior Director, Capital Markets.

2.2 — The money

Two years in stealth, then $550M in seven months.

$200M
Seed — out of stealth, March 2025.
$350M
Series A in two closes — September to October 2025.
$550M
Raised in total.
$1.3B
Valuation — CNBC Disruptor 50, May 2026.
I1
Flagship Pioneering
The incubator · Noubar Afeyan, chairman
Moderna's incubator merged two prototypes — FL96 (materials) and FL97 (life science) — into Lila. The first factory sits in Flagship's Cambridge building.
I2
Braidwell · Collective Global
The Series A leads · $235M first close, Sep 2025
A healthcare crossover fund and a sovereign-backed growth investor; Braidwell's co-founder advises the company.
I3
NVentures · Analog Devices · IQT
The strategics · $115M extension, Oct 2025
Nvidia's venture arm, a Flagship industrial partner and the intelligence community's fund — the last two explain July's AI-security architect.
I4
General Catalyst · March · ARK
The institutions · State of Michigan · ADIA
The seed syndicate; later Dauntless, Catalio, Pennant and members of Peter Diamandis's Abundance network.

Unlike Periodic's “primarily a compute cost,” Lila's bill is buildings, instruments and a 535-person payroll.

2.3 — Who else is trying

Five neighbours — and Lila trades people with two of them.

01
Periodic Labs
Same bet, bits-first · $605M · Menlo Park
A frontier-lab crew and one robotic chemistry lab, founded two years later. Hired Lila's lab-automation principal Sam Cross — talent flows out here.
02
FutureHouse
Same bet, non-profit · Schmidt-backed · San Francisco
AI agents for science, no factory. Lila hired its AI tech lead, Geemi Wellawatte (Aug 2026), and a former intern — talent flows in.
03
Isomorphic Labs
Models, no factory · Alphabet · drug design
The model-first neighbour in biology: AlphaFold's lineage, aimed at drug design, no factory. No talent moves either way in our roster — a peer to watch, not a feeder or a drain.
04
Emerald Cloud Lab
The older model · lab-as-a-service
Remote-controlled instruments by the hour — the pre-AI version of “factory time,” and proof the hard part was never the robot arm.
05
The A-Lab · the national labs
The origin, and the partners · Berkeley Lab · Argonne · NREL
Genesis Mission partners are also feeders: Berkeley Lab ×4, Oak Ridge ×2, Argonne, Los Alamos and Sandia alumni sit on the roster.

For a recruiter this is one market — and a silent one: our 30-day scan found no public discourse for any of these names. This roster is reached by referral, not by feed.

Talent brief

Want the full visual breakdown?

Download the PDF version of this teardown — the seven-team org chart, ten names to know, the build order month by month, the two lineages that arrived in groups, and six sourcing patterns.

§ 03 — Part three of five · The founders

Not two founders. A founding company.

Ten names on the founders' list — a Flagship general partner, three professors, a Navy SEAL, a chess prodigy — and George Church as chief scientist.

3.1 — The two who built it

A company-builder and a professor, both out of Generate.

— Geoffrey von Maltzahn · Co-founder & CEO · the company half

MIT
Chemical engineering, then a PhD in medical physics; 200+ patents.
Flagship
General partner since 2009 — Moderna's incubator builds companies.
Six companies
Founding CEO of Generate, Tessera, Quotient, Indigo; co-founder of Sana, Seres.
Lila
CEO from Dec 2022; prototypes FL96 and FL97 merged into one company.

Brings the incubator's method — chiefs first, then teams — the Generate protein-model lineage, and a chairman named Afeyan.

— Andrew Beam · CTO & Head of AI Research · the AI half

Harvard
Postdoc, then professor from 2019; adjunct since August 2026.
Generate
Co-founder, founding Head of ML (2018–22): the protein language models.
NEJM AI
Co-founder and deputy editor of the medical-AI journal.
Lila
CTO since Aug 2024; hired Stanley within six months; Ray Summit keynote.

Brings the Harvard–Generate ML bench — co-founder Ben Kompa is a collaborator of a decade — and an AI org modelled on frontier labs, staffed from pharma.

3.2 — The professors

Professors as chiefs — some left the chair, some kept it.

01
John Gregoire
Caltech → Chief Autonomous Science Officer
Fifteen years at Caltech in high-throughput experimentation; SVP from March 2024, chief since October 2025. The factory's scientist.
02
Julie Shah
MIT AeroAstro head → Chief Robotics Officer
Head of MIT's aeronautics department; “joined us recently,” said the CEO in March 2026. Her former student had started a month before.
03
Rafael Gómez-Bombarelli
MIT DMSE → CSO of Materials
Associate professor of materials science, on the founding team, still listed at MIT. The chair, kept.
04
Andrew Beam
Harvard → CTO
Associate professor for his first two years as CTO; adjunct since August. The chair, surrendered slowly.
05
Kenneth Stanley
UCF → Uber → OpenAI → SVP
No longer a professor, but recruits like one: five of his nine researchers share his Central Florida history.
06
Hector Corrada Bravo
Maryland → Genentech → VP
Ten years a professor, six at Genentech; arrived April 2026 with seven Genentech colleagues behind him.

Above them, George Church as Chief Scientist; beside them, Taylor Killian — Principal Scientist since March and a BYU assistant professor since August, both at once. The lever is access, not pay: professors join because “this platform is going to make science a lot more fun.” Around them, Flagship installed a full operating C-suite — an Axon CFO, a Waymo CLO, Moderna's CISO and a Navy Chief Government Officer.

3.3 — How they organise and hire · Upfront Summit · ARK · company blog

Small teams, a high bar, and always somebody with fresh eyes.

01 · The orgTeam of TeamsChris Fussell — SEAL Team 6, co-author of Team of Teams — is President. “Small teams applying a common platform… across a very wide diversity of fields.” Delivery comes in threes and fours.
02 · The centreFour orgs build the machine“At the center of Lila are AI, software, robotics, and hardware teams building the scientific method machine.” Around them, science runs as small vertical teams, each with its own VP or director.
03 · The barLeaders in their field“A great and difficult mission with an okay team is not a good recipe.” A bar, not a screen: people who ran the field elsewhere — professors, a Genentech director, Intel Labs' AI-for-science lead.
04 · The leverFun, not payProfessors join “based on the inference that this platform is going to make science a lot more fun.” Access to the factory is the offer — and concurrent professorships say it allows a foot in both worlds.
05 · Stanley's screenAlways somebody with fresh eyes“There should always be somebody who has fresh eyes.” Diversity hired the way he searches — quality-diversity, applied to a team: an Artbreeder founder and a diamond-growth data lead beside four UCF alumni.
06 · The readThe moat is people.“If it's all public data, all a competitor has to do is hire one person that knows the secret.” A theory of why proprietary data matters is also a theory of why staff get poached — and why Lila hires whole lineages, not individuals.

§ 04 — Part four of five · The team

Ninety-eight of 535, hired the way the founders said they would be.

Chiefs from universities, benches from pharma and Boston's robotics belt, small teams under many VPs — and, in 2026, a hard turn toward the factory.

4.1 — The org chart, from the bets

Seven small teams — each one a bet, staffed from a different pool.

T1
AI Research · RL
The bet: outcomes as reward · Beam · Killian · Kyro · Banbury · Jain · J. Lee
RL from Mila, Apple and Microsoft; a Cohere MTS. Kyro: the agent “learns from proprietary laboratory data.”
T2
Applied AI
The bet: models against real labs · Kalyanpur · Zhao · Shaan D · H. Yu · Balestri
Cohere ×3, Writer, Amazon AGI, Genentech — enterprise-LLM production people, hiring more in public.
T3
Life Sciences AI
The bet: the pharma bench · Corrada Bravo · Hofmann · Gunsalus · Jorstad
Genentech ×5 on this row alone; CZI Biohub, Novo Nordisk. Cell biology, sequence and biomolecule design.
T4
Physical Sciences AI
The bet: materials, catalysts, films · Miret · Pan · Jagriti S. · Baibakova · Cruse
Intel Labs, Meta's FAIR Chemistry, Oak Ridge, Berkeley Lab, Imperial. All did AI-for-materials first.
T5
Open-Endedness
The bet: fresh eyes, by design · Stanley · Lehman · D'Ambrosio · Fontaine · Simon
UCF ×5, Google DeepMind, SynthLabs, Sakana — and an Artbreeder founder. Nine, “and growing.”
T6
Autonomous Science Platform
The bet: dexterous robots · Shah · Banfi · Shen Li · A. Wang · Shukla · Mistry
MIT's robotics groups, Amazon Robotics, Berkshire Grey ×2, Boston Dynamics, Symbotic. Built in 2026.
T7
The bench · science
The bet: make and measure · Gregoire · Page · McCourt · Dey · M. Zhang · Erdosy
Caltech, Argonne, Los Alamos, Harvard: thin films, X-ray characterization, HTE, imaging, cell-free protein.

4.2 — Ten names to know

The hires that set the bar.

01
Geoffrey von Maltzahn
Co-Founder & CEO
Flagship general partner; founding CEO of Generate, Tessera, Quotient and Indigo; 200+ patents. CEO since December 2022.
02
Andrew Beam
CTO & Head of AI Research
Harvard professor, adjunct since August; Generate's founding Head of ML. Built the AI org from August 2024.
03
Kenneth Stanley
SVP, Open-Endedness
NEAT, novelty search, Why Greatness Cannot Be Planned; led OpenAI's Open-Endedness team. Nine researchers, five from UCF.
04
John Gregoire
Chief Autonomous Science Officer
Fifteen years at Caltech pioneering high-throughput experimentation; joined March 2024, chief since October 2025.
05
Julie Shah
Chief Robotics Officer
Head of MIT AeroAstro, director of its Interactive Robotics Group; joined early 2026 with a former student already on staff.
06
Hector Corrada Bravo
VP, AI for Cell Biology
Genentech's Director of ML for Genomics; ten years a Maryland professor. April 2026 — seven Genentech colleagues followed.
07
Santiago Miret
Senior Director, Materials Intelligence
Intel Labs' AI-for-science research lead, with Aspuru-Guzik and Mila collaborations; Berkeley PhD. June 2025.
08
Trushant Kalyanpur
Senior Principal MLE, Applied AI
Founding Senior Staff at Cohere, custom enterprise LLMs; June 2026. Hiring the Applied AI org in public.
09
Atul Mistry
Chief Software Architect
A decade as Symbotic's Chief Software Architect (warehouse robots). Feb 2026: the factory's software spine.
10
Ben Kompa
Co-Founder · Head of AI Lab Innovation
Harvard PhD; MIT Technology Review 35 Under 35; “the connective tissue between our AI and the lab.”

4.3 — The shape of the team · 98 scoped profiles

A PhD bench on a pharma ladder.

12.6
Years average experience — a senior bench, not a graduate one.
61%
Hold PhDs — 60 of 98.
54%
Senior level or above — a different cut from the career-stage bar below.
28%
Big-tech background.
48%
In Greater Boston — 23% in San Francisco.

— Career stage at join · 96 of the 98 scoped profiles classified

Titles run on a ladder — Scientist I and II, Senior, Staff, Principal, Senior Principal, Director, VP, SVP, Chief — the convention of the pharma companies the bench came from. Underneath it, 59 joined at senior level and 27 as veterans. Three in five hold PhDs — more than double Periodic Labs' one in four.

Schools: Harvard ×7, UC Berkeley ×5, Stanford ×5, CMU ×3; MIT is a prior employer for 13. A second title system is creeping in: nine Members of Technical Staff, all 2025–26 AI hires — the frontier-lab convention inside the pharma ladder.

4.4 — The build order · first Lila role, 98 scoped profiles

Chiefs in 2024. Teams in 2025. The factory in 2026.

Dec 2022 – 2024
Inside Flagship, then the chiefs. Von Maltzahn as CEO; the first software engineer (Nov 2023). In 2024: Gregoire from Caltech (Mar), Hennek as CPO (May), Beam as CTO (Aug); ML engineer #1 in July — “ML team < 6”; thin-film, protein and HTE scientists by autumn.
Jan – Mar 2025
Out of stealth with $200M. Stanley as SVP Open-Endedness (Feb), Lehman as his first hire; Kyro (Yale, ex-Pfizer) and Teufel (Novo Nordisk) into AI research; a lab-automation scientist from Syensqo.
Apr – Aug 2025
Product and materials. Malleo as VP for the factories, Leung as VP Life Science Product; Miret from Intel Labs; Shaan D, the first Cohere alumnus; an Apple GenAI product manager as Chief of Staff, AI.
Sep – Dec 2025
Series A and the Open-Endedness wave. $550M closed; Simon, D'Ambrosio, Gabriel and Ivy Zhang join Stanley; Jain (Mila RL); two Directors of Product; Azure Quantum's Zheng, Argonne's McCourt, Microsoft's Banbury.
Jan – Apr 2026
The robotics turn. Shah's student Shen Li and Symbotic's Mistry (Feb), Banfi (Mar), Rohrs and Snyder (Apr); Shah named Chief Robotics Officer. Corrada Bravo and Hütter open the Genentech lift (Apr).
May – Sep 2026
The lift completes; the pace doubles. Six more from Genentech (Jun–Aug); Kalyanpur and Zhao from Cohere; the Berkshire Grey pair; FutureHouse's Wellawatte; an AI-security architect; a Talent Partner from the agency.

Fifty-two of 98 joined in the first nine months of 2026 — about six a month in this subset alone. The company went from “about 300” to 535 associated members between March and September.

4.5 — Where they came from

Benches, not résumés.

The first wave came from the incubator — Flagship, Generate and the founders' Harvard and MIT labs. The 2026 wave came in benches: eight from Genentech in five months, four from Cohere across a year, five from Kenneth Stanley's UCF lineage. Frontier-lab logos barely register — OpenAI ×1 — though the unscanned compute team carries frontier-grade credentials.

Genentech · 8MIT · 13Harvard · 11Central Florida · 5Cohere · 4Berkeley Lab · 4Intel · 3Berkshire Grey · 2FutureHouse · 2
The Genentech lift · Apr – Aug 2026
Hector Corrada BravoJan-Christian HütterAmy W.Laura GunsalusNikolas JorstadJennifer HofmannPaula PlutaHarry Yu
A VP in April · seven follow by August
both
in San Francisco
Stanley's lineage
Kenneth StanleyJoel LehmanDavid D'AmbrosioM. FontaineIvy Zhang
UCF → Uber → OpenAI → Lila

Two more groups arrived the same way: Shaan D · Lee · Kalyanpur · Zhao from Cohere ×4 between June 2025 and June 2026, and Wellawatte · Ponnapati hired out of FutureHouse. The Genentech lift and Stanley's lineage both sit in San Francisco — the AI org is the West-Coast half of a Cambridge company.

4.6 — Who builds the factory

The factory is being built by people who build factories.

Nine robotics and automation hires from February to August 2026, under a Chief Robotics Officer who ran MIT's aeronautics department — and 27 of 50 open roles for the Autonomous Science Platform, 41 of the 50 in Cambridge. Lila's own overview claims more than LinkedIn shows: Tesla's Gigafactory Berlin automation design, SpaceX launch-pad control, rollouts at Rivian, Bayer, Ginkgo and 10x, and 70+ workcells at HighRes BioSolutions — none of them in our 98.

01
Shen Li · Banfi · A. Wang
The Shah lineage · MIT IRG · MIT postdoc · MIT PSFC
Shah's own student; an MIT postdoc via Amazon Robotics; a fusion-control PhD from Boston Dynamics. Three MIT roboticists in four months.
02
Shukla · F. Lin · Mistry
The manufacturing belt · Berkshire Grey ×2 · Symbotic
Two Berkshire Grey R&D engineers a month apart; Symbotic's chief architect. Where Periodic bought liquid handlers, Lila hires dexterous manipulation.
03
Layilla W. · E. Brewer · Rohrs · Snyder
The automation engineers · Syensqo · Kimball Physics
A robotics lab built “from the ground up” at Syensqo; an Oxford physics PhD; a technician staging protocols “for training limited-context robotic models.”
04
Ng · Zyto · Malkomes · K. Brewer
The software spine · First engineer · Intel · Seqera
Lila's first software engineer (Nov 2023) built the closed loop; a staff backend engineer; Intel's accelerator lead, now on “autonomous discovery.”
05
Page · McCourt · H. Zhang · Dey · M. Zhang
The bench scientists · Argonne · Los Alamos · Berkeley Lab
Thin films, X-ray characterization, catalyst QC, autonomous high-throughput experimentation, imaging. The instruments have owners.

4.7 — Who runs the GPUs · Ray Summit, Aug 2026

Fifteen people run the compute — a team our 98-profile scan never reached.

Two platform engineers, not the CTO, gave Lila's second Ray Summit talk: GPU training made self-service for scientists. It ended with “we're hiring.” The credentials here are frontier-grade: the company credits unnamed staff with co-creating the Pile, training on four of the TOP500's ten largest machines, and a Gordon Bell finalist for breaking the exaflop barrier.

The team
≈15 people · AI Platform · Tyler Titsworth, lead. Small project teams that re-form per problem, plus a user-support group — Team of Teams applied to infrastructure.
The orchestrator
Flyte v1 · Ray · Anyscale. Flyte runs DAGs across clusters, clouds and regions; Ray fans out training and inference inside them. Ray clusters are per team and ephemeral.
The guardrails
Kueue · Kyverno · OPA. Per-team GPU quotas, topology-aware scheduling, policy on every pod — the boundary between a vibe-coded notebook and a 120-GPU run.
The doorway
Chariot · Hydra. An internal CLI — “if you want GPUs, use Chariot” — is the one door to compute; a single Hydra config is the source of truth for every run.
The plumbing
Argo CD · Crossplane · W&B. GitOps for the infrastructure, Weights & Biases for model artifacts — on AWS plus a neocloud, multi-cluster and multi-region.

For a recruiter: the screening vocabulary for the platform and AI postings, and a pool to source from — Anyscale, Union.ai, the neoclouds, frontier-lab infra teams. Silent on social; it speaks from conference stages.

4.8 — The census and the job board

Engineering first, research second — and the factory is what's posted.

41%
Engineering, by LinkedIn's census
219 of 535.
10%
Business development
55 of 535.
27 of 50
Open roles in the factory
The Autonomous Science Platform.

— The company · LinkedIn's function census, 535 members · Sep 2026, self-reported · the five below sum to 541, more than the 535

— The job board · 50 open roles · Cambridge 41 · San Francisco 17 · London 1 · those three hub counts sum to 59, more than the 50

Two in five build software; a tenth sell — a sales floor for a “first cohort” of customers. The job board says where the next hundred go: the factory out-posts AI 27 to 20, Cambridge out-posts San Francisco 41 to 17, and London has one open role.

§ 05 — Part five of five · What transfers

What the rest of us can steal.

You can't copy the incubator, the professors or the $550M. The lifts, the lineages and the screens transfer.

5.1 — Sourcing patterns

Six patterns — most of them arrived in groups.

01 · The leader-then-bench liftGenentech, eight in five monthsA VP in April; Hütter the same month; Amy W. and Gunsalus in June; Jorstad in July; Hofmann, Pluta and Harry Yu in August. One company's computational groups, moved almost whole.
02 · The professor's lineageStanley's nine, five from UCFLehman and D'Ambrosio were his PhD students; Fontaine taught at UCF; Ivy Zhang came from its evolutionary-computation lab. A lineage reconvened after Uber and OpenAI — then salted with outsiders.
03 · The manufacturing beltFactory builders, re-aimedBerkshire Grey ×2 a month apart, Amazon Robotics, Symbotic, Dexterity, Boston Dynamics — and, per the company, Tesla and SpaceX automation leads. People who have built a factory before are on no lab-automation recruiter's list — yet.
04 · Peer flows, both waysIn from FutureHouse, out to PeriodicFutureHouse's AI tech lead and an ex-intern came in; a lab-automation principal left for Periodic's founding team. The AI-for-science set trades people — and everyone in it knows the pool by name.
05 · The enterprise-LLM shopsCohere ×4, Writer, SynthLabsPost-training and custom-model people from applied-LLM companies — OpenAI ×1 on the roster. Where frontier alumni are unaffordable, the enterprise shops are the bench.
06 · The recruiterPoached from the agency.June 2026: a Talent Partner (AI Research) hired from Strativ Group, the UK agency that staffs frontier-AI teams. When the agency's consultant joins the client, the client means to compete for frontier talent directly.

5.2 — The lessons

Five moves worth stealing.

01 · SourcingHire the leader; the bench followsOne Genentech VP in April became eight Genentech hires by August. Before an offer closes, map the leader's last org chart — the next seven hires are on it, and they already trust each other.
02 · ScreeningAlways somebody with fresh eyesStanley's rule, applied to a team: four UCF alumni, then an Artbreeder founder and a diamond-growth data lead hired for “outsider perspective.” Reserve a seat on every team for the outsider.
03 · SequencingChiefs first, then their labsGregoire (2024), Shah (2026), Corrada Bravo (2026): the chief arrives; students and colleagues follow within months. Hire the field's chair and you've opened the field's pipeline.
04 · The poolAI-for-biology lives in pharmaGenentech ×8, Cohere ×4, OpenAI ×1. When the domain is biology the frontier-lab bidding war is optional: pharma's computational groups are the bench, and they are under-courted.
05 · Units of hireLift pairs and lineagesBerkshire Grey ×2 in a month; Cohere ×4 across a year; UCF ×5 over eighteen months. Recruit relationships — a pair skips the forming stage; a lineage skips the interview.
The catchThis is Flagship's model — $550M and 535 people.Team of Teams needs a bench of chiefs, and an incubator supplied them. Without one, start with one professor and one lineage — the lift works at any scale; the org chart doesn't.

— The takeaway

They hired the chairs.
The labs followed them in —
and the factory is what's hiring now.

Brief
Lila Sciences · Talent Brief
Prepared by
Base to Base · Recruiting
Written for
Founders building AI-for-science teams

Methodology & limitations

A scoped read — honestly counted.

— Sources

  • The LinkedIn company page (535 associated members; function, school and location census) with full profile histories for 98 scoped staff — the AI, robotics/autonomy and founder/CTO subset.
  • The Greenhouse job board (50 roles), and a 30-day public-signal scan (X · Hacker News · Reddit · GitHub · YouTube, Aug 16 – Sep 15).
  • Four founder appearances with transcripts (Upfront Summit 2026, ARK ×2) and the company's post on its Open-Endedness team.
  • Funding announcements, CNBC Disruptor 50 and press.

— Limitations

  • The 98 are a scoped subset, not the company — roughly two-thirds of LinkedIn's Research facet. Every ratio here describes that subset, not all 535.
  • The AI Platform team of ~15 was never reached by the scan; it is described from a conference talk, which is why its credentials appear without names.
  • The public-signal scan returned no on-target discourse for the category — that absence is recorded as a finding, not a gap.
  • Two failed scrapes were recovered by hand; two Open-Endedness members were collected after the company named them. Collected September 2026. Teardowns like this are how our searches begin — this one's on us.