Team teardown · Elorian AI

“BAIR, assemble.

The senior researchers behind Gemini's data and Apple's foundation models, reunited for a final battle — against the reasoning gap.

Get the full visual brief as a PDF.
Who answered
20 frontier authors
The mission
Visual reasoning
The funding
$55M seed
Where it points
Physical-world AI

0.1 — The read

Twenty people. Six months. Not Gemini-adjacent — senior contributors to Gemini, PaLM 2 and Apple's foundation models.

This brief reads Elorian as a talent-flow story, in five questions — who they assembled, what they're after, who's writing the checks, where it points, and what the rest of us can steal.

QThe questionThe short answer
Q1Who they assembledA frontier-model program, rebuilt in miniature.
Q2What they're afterThe capability the big labs treat as a side quest.
Q3Who's writing the checks$55M, from the same network that built the team.
Q4Where this pointsNo product moat yet. Only a talent moat.
Q5What others can learnThe network doesn't transfer. The method does.

§ 01 — Question one of five

Who did they assemble?

Twenty people in six months — most with their names on the papers behind systems the industry now runs on.

1.1 — The standouts

Ten hires that set the bar.

01
Andrew Dai
CEO & Co-Founder
Co-wrote the 2015 pre-training paper the GPT line cites; co-led GLaM, PaLM 2 and Gemini's data org.
02
Dustin Tran
Chief Reasoning Architect
Joined from xAI — led post-training for Grok 4. Before that Gemini's eval lead; his Gemini-Exp-0801 hit #1 on LMSYS.
03
Yinfei Yang
Co-Founder · Chief Multimodal Architect
Core author across Apple's multimodal line — ALIGN, MM1, MM1.5, Ferret-UI.
04
Seth Neel
Co-Founder
Left a Harvard professorship to build this. Founded machine unlearning; earlier co-founded a company that raised $41M.
05
Le (Tycho) Xue
Founding Researcher
Technical lead of Salesforce's BLIP-3; first author of ULIP — 3D point-cloud reasoning. Driving roots at NVIDIA.
06
Richard Zhang
Founding Researcher
Gemini post-training and reward modeling at DeepMind; co-creator of Google Vizier.
07
Seojin Bang
Founding Researcher
A core architect on Gemini at DeepMind from 2022 to 2025. CMU PhD; the deep-inside-Gemini hire.
08
Forrest Huang
Founding Researcher
Berkeley BAIR PhD; core contributor to MM1.5, Ferret-UI and AFM at Apple, then Gemini-for-UI. 2,000+ citations.
09
Zhen Xu
Founding Member
Apple Intelligence search RLHF and agentic search; co-built MUFASA with Dai at Google Health.
10
Jihua Huang
Researcher
A decade at SRI on DARPA and Toyota vision programs — the one hire from outside the network.

1.2 — Authorship, not proximity

They didn't just use the last frontier. They helped build it.

01
Andrew Dai
LM pretraining '15 · GLaM · PaLM 2 · Gemini data
The pre-training + fine-tuning recipe the GPT papers cite, then Google's flagship data pipelines.
02
Yinfei Yang
ALIGN · MM1 · MM1.5 · Ferret-UI
Apple's multimodal foundation stack, paper by paper.
03
Dustin Tran
Edward · Gemini evals · Grok 4
The evaluation and reasoning layer of two different frontier labs.
04
Le (Tycho) Xue
BLIP-3 / xGen-MM · ULIP
Open multimodal models, and language aligned to 3D geometry.
05
Richard Zhang
Google Vizier · OptFormer
The optimization tooling a large slice of the field runs on.
06
Seth Neel
Descent-to-Delete
The paper that started machine unlearning as a discipline.

Roughly two dozen named systems, twenty people. Primary authorship at this density is the real moat.

1.3 — A model program, in miniature

Every layer already has a named owner.

L1
Data & pretraining
Dai · Das
The base models and the data that feeds them.
L2
Multimodal architecture
Yang · F. Huang · Xue
Vision-language and 3D understanding — the core of the thesis.
L3
Post-training & reasoning
Tran · R. Zhang · Xu
RL, reward modeling and the reasoning layer.
L4
Evaluation
Tran · Das
How you actually know the model got better.
L5
Serving & infra
Kumar · Hu
Turning research models into something that runs — ex-AWS Bedrock, ex-Meta.
L6
Company around the lab
Valentine · Jiang · Ling
First PM (ex-DeepMind), first GTM, first finance lead.

Most seed teams are a research core and a wish list. Elorian has a named owner for every layer — that's the anomaly.

1.4 — Hiring topology

A reunion, not a recruiting search.

The team didn't come off the open market. It grew outward from a few tight networks — three closed pools: Gemini alumni, Apple MM1 alumni, Berkeley BAIR. One open-market hire, deliberate, and exactly once.

Google · 9Apple · 3Amazon · 3Meta · 2ex-OpenAI / Anthropic · 0
i.
Berkeley BAIRThe research core — co-authors and labmates
ii.
Gemini & Apple orbitsFrontier-lab colleagues, one hop out
iii.
Open marketThe one deliberate outside hire
Andrew DaiFounder · CEO
Forrest HuangDustin TranRichard ZhangYinfei YangSeth NeelSeojin BangZhen XuMarcella ValentineJihua Huang

1.5 — The connective tissue

Look past the employers, and it's one lab.

The résumés say Gemini and Apple. The thread underneath is a school: six of the team trace to Berkeley, and the research core to its AI lab, BAIR. The trust predates the paychecks.

01
Forrest Huang
Berkeley BAIR · PhD
Then Apple AFM and MM1.5, then Google. The vision-language line, carried out of the lab.
02
Richard Zhang
Berkeley · Applied-Math PhD
Then Gemini post-training and Vizier at DeepMind. Optimization, lab to frontier.
03
Dustin Tran
Berkeley · BS Math + Stats
Then Gemini evals, then Grok 4 post-training at xAI. Berkeley was the starting line.
04
Michelle Ling
Berkeley · CS + Haas
First finance and ops lead — the network reaches past the research bench, too.

Same pattern as our Eigen AI read: the real map is the lab and the co-authors, not the last logo.

1.6 — The differentiator

Where the thesis stops being a slogan.

A visual-reasoning pitch is cheap. Grounding it — geometry, sensors, deadlines — is not. Two hires make the difference.

Images → spaceLe (Tycho) XueFirst author of ULIP / ULIP-2, aligning language with 3D point clouds, with an autonomous-driving background at NVIDIA and Cadence. The bridge from flat images to geometry and physical constraint.
Research → the real worldJihua HuangTen years at SRI on applied vision for DARPA SemaFor, ARPA-H, and Toyota driver-monitoring. The one senior researcher hired off the open market — a lineage of AI that ships on deadlines.

They imported applied, physical-world computer vision on purpose — a team that hires against its own blind spot.

§ 02 — Question two of five

What are they after?

One capability the big labs treat as a side quest: native visual, spatial and physical reasoning — and the industries waiting on it.

2.1 — The thesis, in one line

A concentrated wager on the one capability the big labs treat as a side quest.

The big labs bolt vision onto language as a feature. Elorian's bet is the inverse: visual, spatial and physical reasoning as the foundation — built for robotics, aerospace, medicine and manufacturing.

“An elementary school kid can beat all the frontier models.”

Andrew Dai, CEO — on visual reasoning · The Neuron, June 2026

On the record, three times over: no model clears even BabyVision's six-year-old tier (it tests ages 3–12). His analogy: text AI is in the iPhone era; visual AI is a Nokia — “64×64 pixels,” ARC-AGI's actual resolution. Elorian pledges to publish its own evals.

2.2 — The method, now on record

Reasoning in the image — not about it.

Two July 2026 founder podcasts lay the technical bet out in public for the first time. Four commitments:

01 · Visual tracesPoint, trace, manipulateToday's models reason about images in text. Elorian's models act on them — pointing to count, tracing a path through a maze or a floor plan — the way a child uses a finger.
02 · One model, one spaceGenerate and edit, nativelyOne model producing text, images and video in a shared embedding space — his stated stepping stone to visual reasoning. The pre-/post-training split? “Arbitrary — the only distinction is scale.”
03 · Data firstDistributions, not examples“Garbage in, garbage out” holds at frontier scale. His Bayesian framing: data is a statistics problem, architecture an optimization problem — the muscles Dai's Gemini data org exercised.
04 · SequencingPost-train first, pre-train laterIt's “enough to post-train first” — early signs it lifts visual reasoning — with pre-training deliberately held back. “No need to rush”: the full-stack bench was built for that moment.

This validates the hiring read: the data and post-training owners weren't incidental hires — they are the method.

2.3 — The lane

Everyone's chasing the physical world. Elorian wants the layer beneath it.

01
3D / world models
World Labs · Fei-Fei Li
Dai's on-record contrast: world models tackle the visual modality alone — robotics, entertainment — where Elorian combines modalities seamlessly in one foundation model.
02
Robot foundation models
Physical Intelligence · Pi
General-purpose robot policies — embodiment-first, where Elorian stays a layer up, model-side.
03
Humanoid robotics
Figure AI · Hardware + AI
Owns the robot itself — the most applied, most capital-heavy end of the same thesis.
04
The incumbents
DeepMind · OpenAI · Meta
Treat vision as a feature bolted onto language. Dai, on record: Gemini leads, Anthropic trails on MMMU / OCR-class benchmarks — and all of them fail his board-game test.

NEA already brackets Elorian with World Labs, Physical Intelligence, Sakana and CuspAI as the “neolab” cohort. His edge claim within it: a specialist can drop the trivia for smaller, cheaper models that win one capability.

§ 03 — Question three of five

Who's writing the checks?

$55M at seed, and a cap table drawn from the same closed research network that built the team.

3.1 — The cap table

The money knows the thesis.

$55M at roughly $300M, April 2026. The signal isn't the number — it's who.

01
Board · lead
Brian Zhan · Striker Ventures
A physical-AI investor whose portfolio runs through Skild AI, Periodic Labs, Voyage AI. His robotics thesis maps directly onto Elorian's.
02
Co-leads
Menlo · Altimeter
Altimeter brings a public-markets, growth lens to the cap table — unusually early for a seed.
03
Strategic
NVIDIA · Jeff Dean
Silicon on the cap table, and a personal angel check from the person who built Google Brain.
04
Angels
Sharon Zhou + others
An AI-native angel bench layered over the institutional leads — relationships more than capital.

Research relationships over growth capital — the same network that built the team.

§ 04 — Question four of five

What future does this point to?

No product yet, a crowded lane — and a first contract that will define the company.

4.1 — The scoreboard

What they've launched: a thesis, and a bench.

No public model, API, paper or demo. There's no product moat yet — only a talent moat. Which is exactly why this is a talent brief: for now, the team is the company.

WhenMilestone
Feb 2026Founded — weeks after Gemini 3 shipped, the classic post-release exit window. Dai “and some friends” from Apple and DeepMind; in-person by design.
Apr 2026Out of stealth. $55M seed at ~$300M; framed as the first lab built around native visual reasoning.
Jun 2026Dustin Tran joins as Chief Reasoning Architect — ex-Grok 4 post-training lead.
CVPR '26Co-hosted a “reasoning gap in visual AI” dinner with NVIDIA, NEA, Twelve Labs, GMI Cloud.
Jul 2026On record: post-training-first strategy; internal wins over Gemini Flash on visual benchmarks; public model late 2026 — plus a pledge to release their own evals.
OngoingAggressive hiring; capital earmarked for compute, team, and early customer pilots.

4.2 — The first contract

Our guess: the first big check reads technical drawings.

Dai has named the early lanes himself — and the first model is due late 2026. Four candidates, ranked:

01
Engineering & technical drawings
Named first on the podcast
Diagrams and technical drawings defeat object detection. His sharpest tell: models “can't even tell what two things a wire is connected to” — said pointedly about data-center buildout.
02
Architecture & design
Floor plans, twice over
Count the meeting rooms, doors, windows; trace a wheelchair path for code compliance. Design iteration over spatial constraints no text-first model can hold.
03
Video understanding
Industries on legacy CV
Stock tracking, monitoring — buyers running traditional computer vision today, not building their own models. He explicitly won't sell to labs that train models.
04
Robotics OEM
Via the cap table (NVIDIA)
His example: a robot at a factory control panel that reasons to pull the safety lever first. NVIDIA is an investor — still the biggest prize, but downstream of the first model.

Satellite appears as wildfire detection, not defense. Factory automation folds into engineering. Long-run he still wants a broad model: “the only way to recoup the compute.”

4.3 — Staying power

How they plan to survive being easy to leave.

APIs cut both ways — easy to adopt, easy to churn off. Asked directly, Dai gave a three-step answer.

01
The wedge
Marquee names, few verticals
“A few key verticals and a few marquee enterprises” — relationship-led proof the model wins on real visual-reasoning work, not benchmarks.
02
The opening
Incumbents weakest here
Where current models fail outright, there's no loyalty to churn against — “less incentive to stick with an API” if a rival is priced right and accurate.
03
The moat
API → tools → ecosystem
The retention plan is explicitly the Anthropic route: platforms and tooling around the API once the first model ships.

Recruiting signal: expect platform, DX and forward-deployed roles to open within two quarters of the model launch — a different hiring market than the research bench.

§ 05 — Question five of five

What can the rest of us learn?

You can't copy the network — it took a decade to earn. But the method underneath it transfers.

5.1 — The lessons

Five moves worth stealing.

01 · ScreeningHire authors, not alumni“Worked on Gemini” describes thousands of people. Named authorship describes a handful. Screen for the citation, not the logo on the last job.
02 · SourcingRecruit the trust graphThe fastest search isn't a search. Map co-authors, labs and shared codebases — people who've already built together align in weeks, not quarters.
03 · SequencingName an owner per layerBy hire twenty, every layer of a model program had a named owner. Hire against the org you'll need, not the résumés you happen to like.
04 · The exceptionBuy your blind spotThe one open-market hire was the skill the network lacked. Know precisely which hire your friends can't supply — and go outside for exactly that one.
05 · The pitchSell the problem, not the packageA Harvard chair and a Grok 4 lead didn't move for comp. They moved for one unsolved problem — and full-stack scope: at big labs “if you work in post-training, you don't look at pre-training.”
The catchReunions must be earned.This playbook assumes a decade of relationships to mine. Without one, the map has to be built deliberately — which is a craft of its own.

5.2 — The culture transplant

He isn't just hiring the people. He's rebuilding the room.

“Google Brain was the Bell Labs of this era.”

Andrew Dai — on the diaspora behind OpenAI, Anthropic, SSI… and Elorian

01 · OsmosisIn-person, deliberatelyResearch taste transfers by proximity — corridor talk, micro-kitchen debates, early wrong results. Elorian is in-office by design to recreate it.
02 · ScreeningThe residency filterBrain's residency screened for unusual backgrounds and intense curiosity, not GPA — and produced a generation of founders. Expect the same filter here.
03 · SafetyComfortable being wrongShow results early, wrong ones included; tell anyone “this isn't the right direction.” The conviction on top is Sutskever's old line: “success is guaranteed.”

Recruiting signal: his own Brain interns — Fedus, Ha, Gu — now run their own labs. The Brain diaspora is both Elorian's talent pool and its competition.

Base to Base · Recruiting

The strongest early AI teams aren't hired off the open market. They're reassembled from a trust graph — co-authorship, shared labs, and a decade of working together.

Read that graph and you can see where a company's value is forming before the market narrative catches up.

— The takeaway

The team is the moat —
and it was assembled
one relationship at a time.

Brief
Elorian AI · Talent Brief
Prepared by
Base to Base · Recruiting

Methodology & limitations

How this read was assembled.

— Sources

  • Public LinkedIn profiles and work history.
  • Google Scholar, citations, and published research.
  • Funding disclosures and cap-table reporting.
  • Founder interviews and July 2026 podcasts.

— Limitations

  • Team of ~20; a few roles remain approximate.
  • This is not a complete org chart.
  • The goal is a pattern-level read on how the team was formed — not a roster. Teardowns like this are how our searches begin; this one's on us.