Ami Labs' $1.03 Billion Seed Hires Seven—All On-Site

Seven Roles, One Thesis: World Models, Planning, Safety

The lab's hiring funnel doesn't filter for large-language-model experience. It filters for people who can build systems that predict how the physical world evolves — and plan inside that prediction.

Ami Labs (Advanced Machine Intelligence) launched in 2025 with a $1.03 billion seed round closed in March 2026, a founding roster that includes Yann LeCun, and a technical thesis that explicitly breaks from the generative-LLM paradigm. Its careers page and Ashby board list seven on-site roles split between Paris, Montreal, New York, and Singapore. The roles cluster into two tracks, AMI Engineer and AMI Scientist, plus a geometry-and-3D-vision scientist/intern slot and a general application. An office-manager role in Montreal rounds out the set. Every listing is tagged full-time, on-site. No remote option appears.

The lab's own site states the mission plainly: "Real intelligence does not start in language. It starts in the world." That sentence doubles as a screening criterion. The research agenda centers on world models that ingest continuous, high-dimensional sensor streams (video, primarily) and learn abstract representations that ignore unpredictable noise while predicting future states in representation space. Action-conditioned world models then let an agent simulate the consequences of its own actions, plan sequences, and execute them under safety guardrails. The stated application domains are industrial process control, automation, wearable devices, robotics, and healthcare.

That stack (world modeling, planning, memory, safety) is what the seven roles are built around. The Ashby descriptions for both Engineer and Scientist tracks call for advancing "the state-of-the-art in world modeling" and note that "video is a rich and abundant source of data reflecting how the world works." The Geometry and 3D Vision scientist role narrows further: spatial reasoning from visual data. The tech-stack tags on the company profile (Machine Learning, Robotics, Healthcare, Infrastructure, Generative AI, Computer Vision) confirm the orientation. Generative AI appears, but the lab's own writing treats generative approaches as insufficient for real-world sensor data because "much of real-world sensor data is unpredictable, and generative approaches do not work well."

The practical effect on the applicant funnel is immediate. A candidate whose background sits in instruction tuning, RLHF, or prompt engineering for text-only LLMs hits a mismatch at the first screen. The lab is not hiring for language-model alignment. It is hiring for representation learning on video, dynamics modeling, planning under uncertainty, and safety-critical control — skills that live closer to robotics and autonomous systems than to chat-model post-training. The location-locked engineer roles (AMI Engineer - Paris and AMI Engineer - Singapore) compound the filter. But the technical filter comes first.

For applicants, the takeaway is structural: the screen selects for systems-level fluency. A résumé that leads with fine-tuned LLM checkpoints reads as adjacent. One that leads with video representation learning, model-based planning, or safety-critical control speaks the lab's language. The next section examines how the on-site requirement across four cities reshapes the funnel further.

The Paris-Only Constraint: Zero Remote Hiring

Ami Labs' careers page on Ashby lists seven open roles. Every single one carries the "On-site" tag. The filter dropdown on the board shows "Location Type: On-site (7)" — no hybrid, no remote, no "flexible" qualifier. ReqHunt's snapshot from 19 October 2026 confirms the same: 0% remote across the live requisitions. For a lab that describes itself as "a global company, with offices in those four cities," the hiring reality is more concentrated.

Two roles are pinned exclusively to Paris (AMI Engineer - Paris, posted 30 June 2026 and 19 March 2026). One sits exclusively in Singapore (AMI Engineer - Singapore, 30 June 2026). One sits exclusively in Montreal (Office Manager/Executive Assistant, 19 October 2026). The remaining four roles, General Application, AMI Engineer, AMI Scientist, and AMI Scientist/Interns - Geometry and 3D Vision, list all four cities. New York appears on those four multi-location postings despite having no exclusive roles. The Paris concentration is real: Paris appears on six of seven postings (all except the Singapore-exclusive role).

The funnel effect is immediate. Candidates who cannot or will not relocate to Paris, Montreal, or Singapore are screened out before a single technical question is asked. The Montreal and Singapore slots offer narrow alternatives, but they are solitary roles in different functions (operations in Montreal, engineering in Singapore), so a researcher or engineer tied to those cities has at most one shot. The "global company" framing exists, but the hiring surface area is a dense cluster around Paris with two peripheral outliers.

This is not a temporary artifact. ReqHunt's 30-day trend shows Ami's open-role count unchanged, steady at seven. The Remote Index cannot even be calculated because the lab falls below the 10-role threshold required for a score. In practice, that means the non-remote policy is structural, not a momentary hiring freeze on distributed work. The lab's own job descriptions reinforce it: each listing specifies "Location Type: On-site" and "Employment Type: Full time." There is no "or remote" suffix, no time-zone overlap language, no mention of relocation assistance in the public postings.

For applicants, the constraint rewrites the calculus. A candidate in London, Berlin, or San Francisco with the exact world-model and planning background Ami seeks faces a binary choice: move to Paris, Montreal, or Singapore, or do not apply. The lab's $1.03B Seed round (March 2026) and 11–50 team size suggest capital is not the blocker — the constraint is deliberate. Ami's stated mission (that real intelligence begins in the world, not in language) implies a belief that physical co-location matters for the kind of sensor-grounded, action-conditioned world models they are building. Whether that belief holds up to scrutiny is a separate question. What the data shows is that the screen is absolute, the geography is fixed, and the applicant pool is pre-filtered by passport, lease, and willingness to uproot.

What Actually Gets You Past the Screen

Ami Labs' careers page names three technical pillars: world models, planning, and safety systems. That triad maps directly to the bottlenecks that frontier labs are currently failing to solve at scale. The Carnegie Endowment's analysis of the AI labor debate notes that Ilya Sutskever, a key architect of ChatGPT, has observed that current algorithms "generalize dramatically worse than people," pointing to a key limitation: unlike humans, AI systems struggle to map knowledge across concepts and grasp sophisticated causal relationships. Narayanan and Kapoor found that nearly two years of rapid capability progress produced only modest reliability gains — OpenAI's models from late 2025 are not that much more reliable than their models from 1.5 years prior. The frontier has moved from "bigger models" to "models that can plan, maintain coherent world state, and operate safely over long horizons."

Comparable labs are hiring for exactly those gaps. Waymo's 2027 PhD intern cohort (posted on its Ashby board as of this month) includes dedicated slots for VLM harnesses for AV planning, perception foundation models, behavior post-training, simulation and closed-loop training, and semantics. Those are not LLM fine-tuning roles. They are systems roles: building the evaluation infrastructure that lets a world model plan over hours, not tokens; closing the sim-to-real loop; making perception outputs causally useful to a planner. Anthropic's Fable 5 and Mythos 5 releases show the same priority. Fable 5 set a new record on DeepSWE v1.1 at 74% with fewer steps and greater token efficiency than any frontier model, especially on complex, long-horizon tasks. Mythos 5 conducted novel genomics research over a week of largely autonomous work — assembling single-cell data for millions of cells across 138 species, designing and training a custom ML model to identify homologous cell types, and outperforming a recent Science publication despite being 100 times smaller. OpenAI's GPT-6 Astra caught roughly 20% more bugs than its baseline, and more than doubled the catch rate on pull requests requiring extensive cross-file reasoning. On Terminal-Bench 4.0 (complex terminal-based tasks including software engineering, system configuration, and data analysis), Astra reached 57.9% versus 37.3% for GPT-5.6 Sol and 55.8% for Claude Fable 5.1, at roughly 9% and 63% lower estimated API cost per task. During training, Astra produced unintended outcomes 89% less often than GPT-5.6 Sol and 74.7% less often than Claude Fable 5.1 on an internal computer-use safety benchmark testing the hardest business scenarios: exposing confidential information, sharing dashboards too broadly, deleting data. Astra is also the first model to reach the Critical cybersecurity capability threshold under OpenAI's Preparedness Framework.

The screening signal is consistent across these labs: they need engineers who have built memory systems that persist across sessions, planners that reason over hours of tool use, and safety architectures that hold under adversarial pressure. Khoja, Hiscott, and Hendrycks identify continual learning and long-term memory storage as the most significant remaining obstacle — current systems cannot retain information across sessions, and resolving this will require at least one meaningful breakthrough. ScaleAI's Remote Labor Index, testing AI systems on complicated multi-day Upwork tasks, found that as of March 2026 the best system (Claude Opus 4.6 Cowork) completed only 4.17% of tasks at a level matching or exceeding the human gold standard. The task-length frontier has been doubling roughly every seven months. Newer models like Claude Opus 5 and GPT-5.6 can now handle tasks taking a human nearly an hour, but reliability over multi-step, tool-using workflows remains the hard ceiling.

Ami's screen will weight evidence of that work higher than publication counts or benchmark-chasing. The labs publishing these results (Waymo, Anthropic, OpenAI) are hiring for the same stack. Their public job boards and technical blogs make the priority legible: world models that ground perception in causal structure, planners that operate over long horizons with verifiable safety bounds, and safety systems that survive contact with real users. Ami's seven roles sit in that same funnel. The screen is not a proxy for "AI experience." It is a filter for the specific systems problems that currently block frontier deployment.

Application Tactics for a Research-Heavy, Location-Bound Lab

Ami Labs' on-site policy means the first filter isn't technical — it's logistical. Hiring managers at frontier labs treat non-local candidates as flight risks. The fix isn't vague openness. It's specificity in the first paragraph of your cover letter: state the target city, the move date, and that relocation is self-funded. One hiring tool builder puts it bluntly ("concrete dates = commitment") and notes that self-funded relocation plus own-cost interview travel puts you ahead of candidates with identical skills but unknown budget needs. If you've signed a lease or bought property, use that address. If not, write "Relocating to Paris on [date]" where your address would sit. Skip the "open to relocation" hedging; it signals you haven't decided.

The technical screen rewards systems-level work over model-card collecting. Ami's posted roles name that triad: domains where the artifact is an architecture, not a fine-tuned checkpoint. What survives: memory-augmented agents, hierarchical planners, world-model evaluation harnesses, and safety-infrastructure code (guardrails, monitoring, automated red-teaming pipelines). The China frontier lab data illustrates the gap: 857 model releases in five years, 31 with published safety results, 9 with results at launch. Labs that treat safety as infrastructure — not a post-launch audit — are the ones hiring for it. Frame your experience that way: a deployment-suspension trigger tied to compute-threshold monitoring beats a generic "worked on AI safety" bullet.

Safety framing needs precision. The Chinese technical community's consensus (52% of scientist-authored texts argue safety must gate development) maps to what Ami's safety-systems role likely targets: training-time controls, not application-layer content filters. If you've implemented compute-threshold registration, independent audit trails, or shutdown mechanisms for crossed red lines, name the component and the threshold. If you haven't, show you've read the IDAIS red-line statements and the Shanghai AI Lab's Frontier AI Risk Management Framework; its 4× and 10× effective-compute evaluation triggers are concrete enough to discuss in an interview. A paragraph describing a specific eval harness you maintained differentiates; generic "alignment research" bullet points do not.

Relocation and technical signaling compound. A candidate who arrives with a Paris start date, a planning-stack component in a repository, and a safety-infrastructure write-up that references the same frameworks Ami's researchers cite has cleared the two filters that eliminate most inbound volume. The lab's seven-role cohort is small by design; the funnel rewards the few who treat the application as a systems integration problem — logistics, code, and framing all aligned.

Out of Scope: What This Story Does Not Claim

This piece covers Ami Labs' seven open roles, their on-site location requirement, and the technical profile that those sources signal. It does not make claims about the broader AI labor market, and several data-rich topics that sit adjacent to this story are deliberately excluded.

No industry-wide salary benchmarks appear here.

Rework reported a median 12–15% salary differential between remote AI engineers in Austin versus San Francisco as of April 2026, down from 22–28% in 2022–2023. Eames AI Talent found only 4% of 619 open roles at 30 tracked frontier AI companies were explicitly remote as of July 2026, with the rest on-site or hybrid and clustered in San Francisco and New York. Those figures describe the market Ami Labs operates in; they do not describe Ami Labs' compensation, which this story does not have access to.

No remote-work trend analysis belongs in this story. The Rework data shows 65% of AI job postings in Q1 2026 listed fully remote or remote-first arrangements, up from 38% two years prior, and that postings shifting from remote-first to hybrid or on-site saw applications drop 47% on average with offer acceptance rates falling 31%. Eames AI Talent argues co-location is a competitive input for frontier labs because research, product, and infrastructure work "doesn't decompose neatly into async tickets," and that talent density in those cities drives location decisions. Shopify and several YC-backed startups run distributed AI teams across 15+ countries and Eastern Europe, LatAm, Poland, Romania, India, and Southeast Asia as a deliberate strategy. Canada's AI strategy aims to increase business adoption from 12% to 60% by 2034 and create up to 250,000 new jobs through AI adoption by 2031. None of those patterns are claims about Ami Labs; they are context the reader can find elsewhere.

No claims about other labs' hiring processes appear here. OpenAI's annualized revenue run rate hit $40 billion as of August 2026, with enterprise revenue crossing consumer and growing 32% month-over-month in July per CNBC. OpenAI burned $3.7 billion in the first three months of 2026 per The Information. Anthropic's gross margins were reportedly in the mid-60% range per Rhodium Group, while Zhipu's fell to 26% and MiniMax's to 18% in H1 2026. DeepSeek's gross margins were reportedly 45% by July. METR measured frontier model autonomous task completion rising from roughly 4 minutes in March 2024 to 12 hours by February 2026. The UK government's AI scenarios outline five futures from "Slow Burn" to "Take-Off." China's AI capex is expected to double to 932 billion yuan ($139 billion) in 2026 and top 1.2 trillion yuan ($193 billion) in 2027, though total Chinese model ARR was only around $10.7 billion, roughly 10% of OpenAI and Anthropic levels. ByteDance's Doubao app has over 200 million daily users but under one million yuan daily revenue. Alibaba's AI app segment generated 3 billion yuan in Q2 revenue but a 14 billion yuan net loss in adjusted EBITA terms. These are separate stories.

No application-volume figures for Ami Labs are cited or estimated. The Rework data on 47% application drops applies to companies that changed remote policies in 2025, not to Ami Labs. Eames AI Talent's view that the 4% remote figure makes remote scarcity a leverage point for companies willing to offer it is their read, not a forecast this story endorses. The UK government's scenarios, Canada's $500 million Regional AI Initiative and $200 million AI Missions Program, IDC's $22.3 trillion cumulative AI impact projection by 2030, and Microsoft's 85% Fortune 500 AI adoption figure are all outside this story's frame.

This story is a close reading of one lab's seven roles, one location constraint, and the technical profile its public materials surface. Everything else is background.

Working in frontier tech? Zero G Talent tracks the openings: see every open ASML role, browse frontier tech jobs, openings at Stripe and Waymo, and the people building the field.