Company teardown · Goodfire
Two recruiters built a frontier AI lab that top researchers rush to join.
Eric Ho and Daniel Balsam spent the better part of a decade building an AI recruiting company. Now their lab hires researchers who, by Ho's account, turn down Anthropic, DeepMind and OpenAI to come — a Meta research scientist we interviewed pointed us to this startup. We read 66 of 72 profiles to see how: an idea researchers can't get anywhere else, and a fellowship that doubles as the interview.
0.1 — Goodfire in one page
A lab that reads the insides of neural networks and rewrites them — built by recruiters, and hired the way recruiters would.
Everything here is public — LinkedIn, GitHub, the Greenhouse job board, funding announcements, and eight founder and staff appearances with transcripts. The reading is ours.
§ 01 — Part one of five · The idea
We train AI we can't read. What if we could?
Goodfire's founders think the most important unsolved problem in AI is what's happening inside the model — and that solving it turns training into engineering.
1.1 — The problem they're solving
Why a model that works still can't be trusted, fixed or designed.
— The founder's framing
“Never before has there been such a big gap between how widely a technology is being deployed and our understanding of it.”
— Eric Ho, in Lightspeed's Investment Memo, February 2026.
— Mechanistic interpretability, translated
Reverse-engineering a trained network: finding the internal features and circuits that produce a behaviour, then changing the behaviour by changing them. A science, which is why Goodfire hires neuroscientists and physicists beside ML engineers.
1.2 — The idea
Understand it, learn from it, then design it.
Three verbs — “understand, learn from, and design AI systems.” Ho's history of the field runs the same way: circuits, then SAEs, then “intentional design” — a bonsai, pruned as it grows rather than fought afterwards.
01 · Decode
02 · Understand
03 · Design
↺ what you can read, you can change — and what you change, you can test
1.3 — The wager
A research lab that bets on engineering.
Most interpretability lives in academic groups and one team at each frontier lab. Goodfire's bet is that the science scales like software — so it hires more than two engineers for every researcher, and ships its research as open weights, SDKs and tools.
— The bet, in his words
“It's hard to overestimate just how important like good engineering skills are.”
— Eric Ho, on hiring. Sequoia Training Data, July 2025.
— The count, and who is counting
“Around 40 people, growing very, very quickly” was Ho's own count at the Series B in February 2026; 72 is LinkedIn's count today. The seed memo had planned for ten in year one.
§ 02 — Part two of five · The business
A microscope for models, sold as a platform.
Goodfire sells access to what it can see inside a model — as a hosted product, as research partnerships, and in life sciences as a way to learn biology from biology models.
2.1 — The product, as described
A platform, partnerships — and open research as the shop window.
Silico appears only in an employee profile; the partner work is “under NDA” — the public demos are the tip of the iceberg, by their own account.
2.2 — The money
A $7M seed, a $50M A nine months later — then $150M inside two years.
2.3 — Who else is trying
A small field — and Goodfire hires from its rivals.
The pool is tiny — “maybe a few hundred researchers focused on this full-time,” by the seed investor's count — and loud: 37 AI-safety openings logged in a single day in September.
Talent brief
Want the full visual breakdown?
Download the PDF version of this teardown — the seven-group org chart, ten names to know, the build order month by month, the two pools that never touch, and six sourcing patterns.
§ 03 — Part three of five · The founders
Two recruiters and one scientist.
The CEO and CTO built an AI recruiting company together for years. The Chief Scientist co-founded DeepMind's interpretability team. The company is hired in that shape.
3.1 — The three who built it
A recruiting company's leaders, and DeepMind's interpretability co-founder.
The OpenAI credit is Nick Cammarata, an early member absent from LinkedIn; Sharkey introduced Ho to McGrath.
3.2 — The leadership layer
PIs from rivals, operators from Glean — and one seat already empty.
Also on the list: Nathan Rourke, Head of Finance (Dimensional Fund Advisors). No VP layer; the org runs PIs, leads and two flat titles.
3.3 — How they hire · Cognitive Revolution ×3 · Sequoia · MLST · Lightspeed · Latent Space
Engineers first, good people attract good people, and freedom as the offer.
§ 04 — Part four of five · The team
Sixty-six of 72, hired from two pools that never touch.
Operators from the founders' last company, researchers from the AI-safety ecosystem and academia — and a fellowship that turns visitors into staff.
4.1 — The org chart, from the bets · inferred from profiles
Seven groups — each hired from a different pool.
4.2 — Ten names to know
The hires that set the bar.
4.3 — The shape of the team · 66 profiles
No juniors, and no ladder.
— Career stage at join · all 66 profiles classified
A Stanford professor, a Meta research director and a first-year PhD student can all carry the same title: Member of Technical Staff, the frontier-lab convention. Below the executives and a handful of PIs and leads, there is no Senior, Staff or Principal. Nobody was hired early-career.
Schools, by LinkedIn's census of all 72 members: Stanford 11, Yale 6, UCLA 4, MIT 3, Brown 3. The roster we read is a slightly different base — Stanford appears 9 times across the 66.
4.4 — The build order · first Goodfire role, 66 profiles
Operators in 2024. A rival's lab in 2025. Professors in 2026.
Thirty of 66 joined in 2026 — June alone brought ten. The fellows arrive like an academic intake, not a trickle.
4.5 — Where they came from
Two pools, zero overlap.
Eight people carry RippleMatch — the founders' last company — and they run operations, people and platform. Fifteen carry the AI-safety ecosystem — Apollo, MATS, Conjecture, EleutherAI, FAR.AI, DeepMind, Anthropic — and they do research. Not one person is in both.
These count prior affiliations — employers, schools and programmes — across the 66 profiles we read. OpenAI shows zero because its one link here, Nick Cammarata, is absent from LinkedIn and so outside the 66 — not because no one came from it.
people in both
Everyone else — 43 of the 66 — comes from academia, big tech, Glean and South Park Commons. And the RippleMatch pull is still running: 2024, 2025 and twice in 2026.
4.6 — The fellowship funnel
The Research Fellowship is the interview.
Eleven people have a profile showing a Research Fellow role converting to staff, usually within three to six months — and a twelfth is documented only by a podcast. Nine more are fellows now — mostly PhD students on leave, from a generation where, as the Head of Product put it, “every incoming PhD student wants to study interpretability.”
Nine named fellows sit on the roster today; the title census above counts eleven under the Fellow title — the deck does not reconcile the two, and neither do we.
4.7 — Who studies minds
Professors keep their chairs — and cognitive scientists join the bench.
Three professors hold Goodfire roles alongside their chairs, and a fourth academic keeps a Harvard research affiliation. Around them, a thread the tool never tagged: people trained to study brains, now studying networks — because, as the Head of Product put it, “we have unfettered access to this artificial mind. You can run as many ablations as you want.”
4.8 — The census and the job board
Engineering first — and the next hires aren't researchers.
— The company · LinkedIn's function census, 72 members · Sep 2026, self-reported · the five below sum to 74, more than the 72
— The job board · 28 open roles · San Francisco 25 · New York 1 · London 1 · those three hub counts sum to 27, one short of the 28
McGrath's “most urgent” roles: engineering, life-sciences research, research engineering — “and we're going to be growing a go-to-market and product function soon.” The board agrees: 19 of 28 roles are GTM, People or general.
§ 05 — Part five of five · What transfers
What the rest of us can steal.
You can't copy a DeepMind co-founder or a $1.25B valuation. The fellowship, the lift and the flat title transfer.
5.1 — Sourcing patterns
Six patterns — most of them are programmes, not people.
5.2 — The lessons
Five moves worth stealing.
— The takeaway
Two recruiters sold researchers
an idea worth a pay cut —
then let a fellowship do the interviewing.
Methodology & limitations
A near-complete read — honestly counted.
— Sources
- The LinkedIn company page (72 associated members, 7 of them investors; function, school and location census) with profile histories for 66 staff — four of whom were found only via GitHub and founder media.
- The GitHub org goodfire-ai (24 repos) and the Greenhouse job board (28 roles).
- Eight founder and staff appearances with transcripts — Cognitive Revolution ×3, Sequoia Training Data, MLST, Lightspeed's Investment Memo, Latent Space, and a McGrath talk at Founders You Should Know.
- A 30-day public-signal scan (X · Hacker News · Reddit · GitHub · YouTube, Aug 21 – Sep 20) and funding announcements.
— Limitations
- The PhD rate is a floor: education pages returned no content, so degrees were inferred from roles and headlines. Read 47% as “at least.”
- The 72 is LinkedIn's associated-member count, not a staff count — the company page lists 7 investors among them, so every “of 72” ratio here has a denominator that isn't purely staff. The 66 we read are staff profiles, four of them found outside LinkedIn entirely. The source gives no staff-only total, so we quote the count it gives rather than compute one it doesn't.
- LinkedIn's function census sums to 74 across a 72-member page, and the job board's hub counts sum to 27 against 28 roles. Both are reported as found.
- The org chart on 4.1 is inferred from profiles, not published by the company — groups are our reading of who works on what.
- Headcount is a moving target: Ho said “around 40” in February 2026 and LinkedIn shows 72 today. Collected September 2026. Teardowns like this are how our searches begin — this one's on us.