Home › Blog › Inside an AGI Startup Team — Narrow Research Bet
Inside an AGI Startup Team · Lane 3 of 3 🟢

Inside an AGI Startup Team: The Narrow Research Bet

An end-to-end look at how the smallest, most senior team in AI actually works — structure, skills, coordination, cadence, compensation, tools, culture, hierarchy, upskilling, and how researchers use AI tools day to day — anchored by a detailed look at Safe Superintelligence and Goodfire's real team dynamics, including SSI's famously flat, secretive, ~50-person structure.

FL
FrontierAGI Team

What "The Team" Actually Looks Like

If Lane 2 is flat-but-fast and Lane 1 is layered-but-lean, Lane 3 is something else entirely: a team so small and so senior that conventional org-design questions barely apply. Every one of SSI's roughly 50 employees works directly on core research, with no support roles, no management layers, and no commercial product — a structure explicitly designed to remove everything Lane 2 and Lane 1 both consider necessary overhead. This article closes out the "Inside an AGI Startup Team" series by covering the same ground — structure, skills, coordination, cadence, compensation, tools, culture, hierarchy, upskilling, and AI-tool use — for the narrow-research-bet lane, using Safe Superintelligence Inc. (SSI) and Goodfire as throughline case studies, the same companies anchoring this series' original Lane 3 build article.

~50 SSI headcount at a $32B valuation, all working on core research
Zero Product managers, sprints, or external validation cycles at SSI
"Stealth Startup" How many SSI employees list their employer publicly, per reporting
$100K-$1M+ Reported SSI base-plus-equity range, full-time research roles

Team Structure: What the Org Chart Looks Like

SSI's structure is reported as genuinely flat and hierarchy-free — no product managers, no sprints, no external validation cycles, and no layer separating a researcher from the founder. Goodfire, while larger and more product-facing than SSI (it ships the commercial Silico/Ember platform), still organizes around a small senior research core (Chief Scientist Tom McGrath's interpretability team) paired with a much smaller engineering/productization function (CTO Dan Balsam's side) than either Lane 1 or Lane 2 would use at comparable funding.

Founder / Research Lead — sets the thesis, works alongside the team with no management layer between
Senior Research Scientists (thesis owners)
Research Scientists / Engineers — each owns an independent experimental thread with near-total autonomy

Note the missing tier relative to Lanes 1 and 2: there is no dedicated "lead" layer distinct from the senior researchers themselves at SSI's scale — the founder(s) and the senior research core largely are the leadership function. This flat structure is only viable because headcount stays small by design (per this series' Lane 3 build article, roughly five employees per billion dollars of valuation at SSI) and because every hire is expected to operate with extreme self-direction and comfort with ambiguity rather than needing management overhead.

Experience Profiles That Populate This Team

🎓 Nearly Universal: Research-Level Mathematical Maturity
Mastery of probability theory, optimization, linear algebra, and information theory is assumed, not optional, at SSI — a baseline closer to a top PhD program's qualifying exam than a typical engineering hiring bar.
🏆 Elite-Institution Pedigree
Both SSI and Goodfire draw heavily from OpenAI, Google DeepMind, and top PhD programs — Goodfire's team includes experts from these exact institutions, and its Chief Scientist personally founded DeepMind's interpretability team.
🔒 Comfort With Institutional Secrecy
SSI employees are reported to be instructed not to disclose their affiliation publicly, with many listing only "Stealth Startup" on LinkedIn — a genuinely unusual professional norm this lane's hires must be comfortable with.

Skills & Competencies Matrix

Competency TypeWhat It Looks Like HereWhy It Matters More in Lane 3
Hard: Independent research designDesigning, running, and interpreting an entire experimental thread with no management oversightThe flat, no-hierarchy structure at SSI has no layer to catch a researcher who can't self-direct
Hard: PyTorch/JAX + distributed training fluencyFoundational, non-negotiable requirement even at a ~50-person research-only companyLarge-scale distributed training (including on TPUs, per SSI's reported stack) is central even to a "pure research" operation
Soft: Extreme self-directionComfort operating for weeks or months without a manager, sprint board, or external deadlineExplicitly named as required by reporting on SSI's culture — this is not a hidden expectation, it's a stated hiring bar
Soft: Comfort with ambiguity and secrecyWorking on an unpublished thesis for years, often without being able to discuss the work publiclyDirectly at odds with the publication-driven incentives of academia or the usage-metric incentives of Lanes 1-2
Hidden: Genuine belief in the thesisJoining because you believe the specific research question matters, not for compensation or brandReporting on SSI explicitly frames this as the reason researchers work on hard problems here — mission alignment substitutes for management structure

How Different Teams Coordinate

With no management layers and no sprints, coordination at SSI-scale happens through direct peer conversation rather than any formal process — the flat structure means a researcher raises a blocker or shares a finding directly with whoever is relevant, up to and including the founder, without routing through intermediate reporting lines. Goodfire, being larger and split between research and productization, has slightly more formal coordination: the research team (interpretability science) and the engineering team (building Silico/Ember as a shippable platform) maintain a standing translation function, since a research finding only becomes a product feature once engineering can operationalize it.

1
Researcher ↔ Founder: At SSI's flat structure, this is often a direct relationship with no intermediate layer — a researcher's thesis-continuation decision can be discussed straight with leadership.
2
Research ↔ Productization (Goodfire): Interpretability findings from the research core get evaluated by the engineering/product side for whether they're ready to become a Silico/Ember platform feature — a genuine, if lightweight, handoff process absent at SSI.
3
Cross-researcher peer review: Because there's no formal management layer, quality control runs substantially through peer scrutiny of results — researchers challenge each other's interpretations directly rather than through a manager-mediated process.
4
Domain-science partnerships (Goodfire specifically): Collaboration with external domain experts (e.g. Prima Mente for the Alzheimer's biomarker work) requires a distinct coordination skill — translating interpretability findings into a form domain scientists outside ML can use.

Daily, Weekly, Monthly & Release Cadence

Daily
Independent deep work
Individual or small-pod research work — reading, experiment design, running and interpreting results — with far less synchronous meeting time than either other lane, consistent with the "no sprints" structure.
2-3x Weekly
Research syncs / reading groups
The primary coordination mechanism — pods share findings and challenge each other's interpretations, functioning closer to an academic lab meeting than a corporate standup.
Weekly
Founder/lead thesis review
A recurring check on whether each active research thread is still worth funding — the closest thing this lane has to a "project management" ritual.
Rare, Irregular
External signal / release cadence
No fixed release calendar exists — a paper, an open-weight release, or a scientific finding (Goodfire's Alzheimer's biomarker work) ships only when the research genuinely justifies it, sometimes years apart.
Ongoing
Founder-led fundraising & compute partnerships
Runs continuously and almost entirely through the founder(s), since there's no metrics dashboard or usage curve to hand to a dedicated function.
Beta/Continuous (Goodfire)
Platform release cadence (productized labs only)
Goodfire's Silico moved through a private beta in 2026 with continuous internal iteration — a genuinely different, faster cadence than SSI's, reflecting Goodfire's partly-productized structure.

How a Research Result Actually "Ships": A Walk-Through

Take Goodfire's real Alzheimer's biomarker finding as the template. Months 1-3: A research scientist applies Goodfire's interpretability methods to a biology foundation model built by domain partner Prima Mente, initially as an open-ended exploratory thread with no defined deliverable. Months 3-5: Early signals suggesting a novel biomarker class emerge; the researcher shares findings in the recurring research sync, and peers stress-test the interpretation rather than a manager reviewing it. Months 5-7: The finding is validated further, now involving direct collaboration with Prima Mente's domain scientists to confirm the result means something biologically real, not just an interesting activation pattern. Months 7+: Only once the result clears this bar does it become a publishable/announceable finding — the entire arc running 6+ months with no fixed deadline at any stage, contrasted sharply with Lane 2's 2-week feature-ship cycle and Lane 1's 3-month model-release cycle.

There is no roadmap ticket for "discover a novel Alzheimer's biomarker." The entire delivery model in Lane 3 is: fund a small number of people to pursue open-ended questions for as long as it takes, and let peer scrutiny — not a release calendar — decide when a result is real.

Compensation by Role

SSI's own reported compensation structure is unusually wide, reflecting both its research-only focus and its access to $32B-valuation-scale equity.

RoleTotal Comp Range (2026)Source
SSI — Full-time research roles$100K-$500K base + equity; senior researchers reportedly exceed $1M total compPin.com AI Compensation Benchmarks, 2026
Frontier-lab median (OpenAI/Anthropic/DeepMind/MSL research scientist)$600K-$795K median; 90th percentile $1.28M+Pin.com, 2026
Goodfire research scientists (Series B stage)Broadly comparable to mid-tier frontier lab bands ($280K-$500K+), plus meaningful equity given $1.25B valuationIndustry benchmark comparison; exact figures not publicly disclosed
SSI research interns (2027-2028 cycle)Competitive with full-time mid-tier roles, per SSI's own internship guideExtern — SSI Internship Guide

Note: With only ~50 employees against an $8B raise, SSI's per-researcher equity stake — while not publicly disclosed in exact percentages — is understood to be substantial relative to typical Series A/B norms, consistent with a deliberate strategy of using equity concentration, not just cash, to retain a very small elite team.

Tools, Software, Hardware & Logistics

PyTorch / JAX Large-scale distributed training (TPUs, per SSI's reported stack) Interpretability tooling (sparse autoencoders, probes) Goodfire's Silico / Ember platform Internal research wikis / lightweight documentation arXiv / Hugging Face for eventual publication

On hardware and logistics: SSI's compute is provisioned via its $5B Nvidia partnership rather than owned infrastructure built up front — compute is matched to validation-stage needs, not pre-built for eventual scale, per this series' Lane 3 build article. Goodfire similarly relies on external compute partnerships while investing more heavily than SSI in productized internal tooling, since Silico itself must run reliably as a customer-facing platform, not just an internal research aid.

Team Culture & How the Hierarchy Works

SSI's culture is described directly as flat and hierarchy-free, requiring extreme self-direction and comfort with ambiguity — no product managers, no sprints, no external validation cycles, and a pervasive culture of secrecy reminiscent of elite research institutions rather than a typical startup. Employees are reported to avoid disclosing their affiliation publicly. This is a genuinely distinct culture from both other lanes: Lane 2's flatness is about shipping speed, while SSI's flatness is about removing anything that could distract from long-horizon research focus.

🔇
Deliberate Secrecy as Culture
SSI's opaque public profile isn't accidental — it's a stated part of insulating the research from short-term commercial and reputational pressure, extending even to how individual employees describe their own jobs.
🤝
Peer-Driven, Not Manager-Driven
With no management layers, quality control and prioritization both run through direct peer interaction — a genuinely flat structure in the literal sense, not just a marketing description of a normal company.
🧬
Mission-Driven Retention (Goodfire)
Goodfire bridges pure research culture with product accountability — its team explicitly frames itself as "the first mechanistic interpretability frontier lab," a mission identity that does real retention work alongside compensation.

How the Founder Actually Works With the Team

At SSI's flat structure, the founder relationship is reported to be genuinely direct — there's no management layer for Ilya Sutskever or Daniel Levy to work through, meaning founder involvement in individual research threads can be closer and more frequent than at either other lane, even as the founder's public-facing role (fundraising, the Nvidia partnership) stays highly visible externally. At Goodfire, CEO Eric Ho (product/GTM background) and Chief Scientist Tom McGrath (pure research background) split founder involvement along exactly the research/productization line described in the team-structure section — Ho engages more with the engineering and business side, McGrath more directly with the research core, an explicit division of founder labor uncommon in the other two lanes.

Team Upskilling Programs

Formal upskilling programs are essentially absent in this lane — the hiring bar already assumes near-complete technical maturity (research-level mathematical fluency is "assumed, not optional" per reporting on SSI), so there is little the company would even need to teach a new hire on fundamentals. What exists instead is informal: peer learning through the recurring research syncs and reading groups, and — increasingly, per Goodfire's own reporting — learning to work effectively alongside AI research agents like Silico, which is itself becoming a new kind of internal upskilling curve unique to this moment in the field.

How the Team Uses AI Tools Day to Day

This is where Lane 3 diverges most sharply from the other two lanes: at Goodfire specifically, AI tools aren't just an engineering convenience — they've become a research collaborator in their own right. Every one of Goodfire's researchers and engineers reportedly uses AI tools daily, and the company has built internal infrastructure specifically for agentic research workflows, culminating in Silico — an autonomous agent that plans and runs its own interpretability experiments, monitors progress across computing nodes, and returns inspectable results without human hand-holding.

1
AI as a research collaborator, not just a coding tool: Goodfire's own blog post "You and Your Research Agent" documents lessons learned from using agents specifically for interpretability research — a genuinely different relationship than a Lane 2 engineer using Claude Code to write product code.
2
Silico runs experiments end-to-end: Researchers describe an investigation in plain language, and the platform autonomously builds an experimental plan, dispatches parallel agents to execute sub-tasks (training sparse autoencoders, mapping neural geometry, testing causal hypotheses), and returns inspectable results.
3
Standard AI coding tools for infrastructure work: Where research engineers do write conventional code (platform infrastructure, tooling), they use AI coding assistants similarly to Lane 1's infra engineers.
4
SSI's own AI-tool usage is far less publicly documented — consistent with its secrecy culture — but its reliance on PyTorch/JAX and large-scale distributed training implies conventional engineering tool use at minimum, without the same publicly visible agent-research-collaborator pattern Goodfire has shown.

Case Study: SSI & Goodfire's Real Team Dynamics

Safe Superintelligence & Goodfire Founded June 2024
SSI: ~50 employees, $32B valuation, zero hierarchy Goodfire: $1.25B valuation, Silico research-agent platform Both: no sprints, mission-driven retention

SSI's team structure is the purest real-world demonstration of Lane 3's core organizational bet: that removing management overhead entirely — no product managers, no sprints, no external validation cycles — lets a small, elite research team move faster on a single thesis than any conventionally-managed organization could. Reporting describes SSI's culture as reminiscent of elite research institutions, complete with a genuine secrecy norm where employees avoid disclosing their affiliation, extending the company's broader strategy of insulating research from external pressure all the way down to individual employees' public profiles.

Goodfire occupies an instructive middle ground: founded by Eric Ho (CEO, product/GTM background from RippleMatch), Dan Balsam (CTO, engineering background), and Tom McGrath (Chief Scientist, founder of Google DeepMind's interpretability team), it pairs SSI-style deep research culture with an actual shippable product — the Silico platform, which opened private beta in 2026 and represents a genuinely novel team-dynamics artifact: an AI agent that functions as a standing research team member, planning and executing its own interpretability experiments. Goodfire's own reporting states plainly that every researcher and engineer at the company uses AI tools daily, with dedicated internal infrastructure built around agentic research workflows.

The contrast between the two companies shows Lane 3's real range: SSI represents the "purest" version of the no-hierarchy, no-product research bet, while Goodfire shows what happens when that same research-first culture is deliberately paired with productization — a second, smaller engineering culture layered on top of, but never replacing, the research core's autonomy.

The team-dynamics lesson: Both companies prove the same underlying principle from opposite ends — SSI shows that removing all management structure is viable when headcount stays small and mission alignment is total, while Goodfire shows that even a partly-productized research lab can preserve that same flat, peer-driven research culture as long as productization stays a clearly separate function rather than absorbing the research core into a normal corporate hierarchy.

🎥 Recommended Videos on Team Dynamics in This Lane

🧭 Closing — What Makes a Lane 3 Team Actually Work

🎯 The Bottom Line
A Lane 3 team's defining structural trait is the near-total removal of organizational overhead in favor of extreme individual autonomy, mission alignment, and peer-driven quality control — SSI's real, reported ~50-person, zero-hierarchy, secrecy-cultured structure at a $32B valuation is the clearest existing proof that this model can be funded and sustained at genuine scale. Goodfire's parallel structure — pairing that same research-first culture with a real product (Silico) and, remarkably, an AI research agent that now functions as a de facto team member — shows the model's range: it holds even once a research lab starts shipping something the world can use. For a first-time reader, the throughline across all three articles in this series: team structure, cadence, culture, and compensation are never independent design choices — in every lane, they are the direct organizational consequence of what that lane is actually betting on, whether that's shipping velocity, a training run's compute budget, or a single research thesis's freedom from every kind of external pressure.