# 77% of people who name Meta's superintelligence lab were already at Meta when it formed Current Meta staff whose profile headline or title names Meta Superintelligence Labs, by when they joined Meta, where the newcomers had just worked, and what they do, on the Metix AI Platform on September 22, 2026. ## Findings 936 profiles visible through the Metix AI Platform hold a current Meta job and name the lab in their headline or current title, as superintelligence or as MSL. 722 of them (77.1%) were at Meta when the lab formed: a Meta job that was not an internship began before June 2025 and was still running in May 2025 or later. The other 214 (22.9%) joined the lab from outside Meta: 170 (18.2%) had never had a Meta job before June 2025, 19 (2.0%) had worked at Meta, left, and came back, and 25 (2.7%) had only interned there. Counting every earlier Meta entry, internships included, 766 (81.8%) had been at Meta before June 2025. The split holds under a narrower definition: of the 790 who write superintelligence, 617 (78.1%) were at Meta when the lab formed. Meta put its existing AI teams, FAIR among them, into the lab when it announced it at the end of June 2025, so a large share of long-time staff follows from the memo; the counts show how large, and who came in. Most of the group first joined Meta long before the lab. 433 (46.3%) first joined before 2022 and 547 (58.4%) before 2023, so the typical member first joined in 2022. First joins run at 128 in 2018 and 2019, 190 in 2020 and 2021, 114 in 2022, and 118 in 2024, with only 24 in 2023. The 170 who were new to Meta started in a burst from June 2025, the month the lab formed: 20 in June 2025, 36 in July, 28 in August, and 23 in September, 107 in four months, then 29 from October to December and 34 from January 2026 on. Profiles lag, so the most recent months are undercounted (see Limits). Few of the outside hires came straight from one of five AI labs. 31 of the 214 (14.5%) have a job that ended in 2025 or later at Google DeepMind, OpenAI, Anthropic, xAI, or Thinking Machines Lab: 18 at Google DeepMind, fewer than 10 each at OpenAI and xAI, and none at Anthropic or Thinking Machines Lab. 16 (7.5%) came from Scale AI, 16 from other parts of Google, 10 from Amazon, and 22 (10.3%) from a university; Apple, Microsoft, and NVIDIA account for fewer than 10 each. 101 (47.2%) had just left some other employer or show no job ending in 2025 or later. Each person counts once, at the first of these employers in that order. Across the whole group, earlier jobs at the large technology companies are common: 133 (14.2%) once worked at Google, Google DeepMind included, 110 (11.8%) at Amazon, 92 (9.8%) at Microsoft, and 36 (3.8%) at Apple. Earlier jobs at OpenAI (12, 1.3%) and Anthropic (none) are rare. Outside hires lean toward research. Half of them (108, 50.5%) list Research as their current function, against 30.6% of those who were already at Meta (221 of 722); across the group, 35.1% are in Research, 39.1% in Engineering and Technical, and 6.4% in Product. 37.9% of outside hires list a doctorate (81 of 214), against 29.6% of those already at Meta. Among people whose current function is Research, 67.5% of the lab list a doctorate (222 of 329), against 41.0% of Meta's other research staff (3,670 of 8,955). The lab's research staff are also more often labelled Senior, which here covers senior, staff, and principal titles (15.8% against 10.8%); the share at manager level or above is 6.7% against 9.0%, a gap within chance. Few name any of the four groups searched inside the lab. 63 (6.7%) name FAIR, 49 (5.2%) name infrastructure (11 of the first 15 matches read name MSL Infra), 13 (1.4%) name TBD, 10 of them in TBD Lab or reporting to it, and fewer than 10 name Products and Applied Research; at least 85% name none. 238 more current Meta profiles name FAIR without naming the lab, so people who work in the lab's groups without writing its name fall outside this population. 94.1% list the United States (881): 453 (48.4%) in Bay Area cities, 118 (12.6%) in New York State, 84 (9.0%) in Washington State, 27 elsewhere in California, and 199 (21.3%) elsewhere in the US or with no state; 19 (2.0%) list the United Kingdom. On the day of the snapshot, 30 open Meta postings name the lab, 1.4% of Meta's 2,104 open postings: 23 in Menlo Park, 3 in San Francisco, 2 in New York, and 2 elsewhere. 14 have research in the title, and 7 of the 26 that state an experience requirement ask for 36 months or less. Microsoft, which announced a superintelligence team in November 2025, shows the same pattern at a smaller scale. 80 current Microsoft profiles name superintelligence in their headline or title; 61 (76.3%) had a Microsoft entry before November 2025 and 19 (23.8%) joined since. On the same measure (any earlier entry at the company, internships included), 766 (81.8%) of Meta's group had been at Meta before June 2025; the comparable figure is 81.8%, not the 77.1% who stayed, because the Microsoft group is too small to split further without letting a reader subtract down to fewer than 10 people. Of the 61, 21 first joined Microsoft before 2020, 19 from 2020 to 2024, and 21 in the ten months before the team formed. Profiles stop in April 2026, which hides more of Microsoft's window since November 2025 than of Meta's since June 2025, so Microsoft's share is likely biased upward. ## Population and definitions A person counts when one job entry is current and at Meta, and the headline or current title names the lab: superintelligence or MSL, matched word by word and ignoring case. The Platform treats Facebook and Meta as one employer name. An earlier Meta job is an entry at Meta, Facebook, Instagram, WhatsApp, or Oculus; an internship is an entry whose seniority is Intern or whose title says intern. The cut is June 1, 2025: Meta announced the lab at the end of that month. Four groups add up to the whole: stayed (a non-intern Meta job that began before the cut and was current or ended in May 2025 or later), returned (a non-intern Meta job before the cut, none of which lasted into May 2025), former interns (only internships before the cut), and new to Meta (no Meta entry before the cut, including fewer than 10 whose Meta entry has no start date). Outside hires are the last three together. Where an outside hire had just worked is a job entry that ended in 2025 or later, at the first match in a fixed order (OpenAI, Google DeepMind, Anthropic, xAI, Thinking Machines Lab, Apple, Scale AI, Google, Microsoft, Amazon, NVIDIA, then any employer the Platform types as Educational). Internships do not count for companies, so a PhD student's summer at a lab is not where they came from; they do count for universities, because doctoral students' entries often carry the Intern label. The first-join series counts people with a Meta entry that started before each date, so each band is the date of the person's first Meta entry; a second series skips internships. For people who stayed, the date of their newest current Meta entry shows who recorded a new role and when. current_function and current_seniority are the Platform's fixed values for the current job. The baseline is current Meta staff whose function is Research and whose headline and title name neither lab term. Definitions: [`queries/population.json`](queries/population.json) and [`queries/measures.json`](queries/measures.json). ## Method 108 counts, each 1 API Credit or free when it finds no one; no profile or posting is read. The replay runs both stages in one pass, 358 counts for 353 API Credits (`data/receipt.json`), of which stage 1 takes 108. Bands of dates merge with a neighbour wherever one would hold 1 to 9 people. Every count of people from 1 to 9 is published as "<10"; when the hidden cells inside a published sum would add up to 1 to 9, one more cell is withheld; and the narrower definition and the Microsoft group are split only as finely as keeps every difference between published counts at 0 or at least 10. A share is published only when its denominator is at least 30. `fetch.py` writes eight aggregate files for stage 1, ten `map_*` files for stage 2, and the receipt to `data/`, and writes nothing if any count fails or any published cell could be recovered. ## How it was made An AI agent built this study through the public REST API. A capped probe (42 API Credits) found 790 current Meta profiles naming superintelligence and suggested most were Meta veterans. The agent then read `GET /contract` and the people documentation and audited the definition by reading 63 profiles in whole slices, never the top of a search, because Search ranks by match quality: 62 of 63 describe themselves as working in the lab, and the rest recruit for it. The reads changed the definition three ways. 146 more profiles write MSL without superintelligence, and 13 of 14 read use it for the lab, so MSL became a lab term. An earlier Meta job listed under Instagram is not matched by the name Meta, so Instagram, WhatsApp, and Oculus joined Meta and Facebook as earlier employers. Fewer than 10 of the recent hires read hold data labeling or red teaming roles, so the replay counts such titles among outside hires (also fewer than 10). Reads of sub-team matches showed that applied research mostly matches the job title Applied Research Scientist, so the Products and Applied Research row needs its own name, and that 10 of 13 TBD matches are in TBD Lab. A whole slice of 15 Microsoft profiles all named the Microsoft AI superintelligence team; fewer than 10 are recruiters or sourcers for it. The 30 open postings were read to check that every one names the lab. Every query was first sent once with an extra word that matches nothing, which the Platform checks against its limits without charge. The first replay then stopped before writing anything: its check found that the superintelligence-only group, split three ways, would let a reader subtract their way to fewer than 10 people who write MSL and had an earlier Meta job without staying. The narrower group is now split as finely as its differences allow, and the corrected `fetch.py` then ran again from the start (108 counts, 105 API Credits) and wrote stage 1's files and a receipt, which the replays of both stages later replaced (see stage 2). Stage 1's exploration cost 206 API Credits before that run: the 42-Credit probe, 59 for the agent's counts and reads, and 105 for the first replay that stopped before writing. ## Limits Counts are of profiles visible through the Metix AI Platform, not of staff: people in the lab who do not name it are missing and profiles left out of date remain, so a count is neither the lab's headcount nor a guaranteed floor (current Meta profiles outnumber Meta's reported headcount 1.7 to 1). The population is people who chose to name the lab, which is not the same as the lab: people in its groups who write only FAIR (238 profiles) or nothing are missing, and if newcomers name the lab more readily than long-time staff, or less, the 77.1% moves with them. A few profiles are out of date: reads found fewer than 10 that say they have left the lab while the Meta entry is still current. Profiles lag: across all Meta job entries on the Platform, 3,988 start in 2026, 1,104 of them in March and April and fewer than 10 in May or later, so start dates stop in April 2026 and the last months are probably undercounted. Dates come from the profile, where some entries give only a year, which the Platform reads as January: a newcomer whose Meta entry says only 2025 counts as at Meta before the cut, which raises the 77.1%, and an entry whose end date says only 2025 counts as ending before May 2025, which can move someone who stayed into returned. Many people change teams inside Meta without adding an entry, so the dates of newest roles are a lower bound on who moved. Where an outside hire came from rests on end dates in 2025 or later, and 47.2% show none of the listed employers. The Microsoft group is small, includes recruiters, and its history names do not include subsidiaries such as LinkedIn or GitHub. One day of postings, all posted in September 2026 and all labelled Not Applicable for seniority. No individual is named or shown, and nothing here describes pay. ## Stage 2: the map Stage 2 maps Meta's visible AI staff and places the lab among them, on the same day and in the same run as stage 1. The definitions, term lists, unit order, and audit results are in [`queries/mapping.json`](queries/mapping.json); every number below is in one of the `data/map_*.json` files. ### Findings 8,257 current Meta profiles visible through the Metix AI Platform describe AI work in their headline or current title; call them Meta's visible AI staff (P2). 6,888 of them (83.4%) list the United States, and the lab's 936 are 11.3% of them. For scale, 128,970 current profiles hold a job at the Meta that the Platform records with more than 10,000 employees, 60,588 in the US and 68,382 elsewhere or with no country (`map_size.json`). Meta reported a headcount of 75,472 as of June 30, 2026 ([SEC filing](https://www.sec.gov/Archives/edgar/data/1326801/000162828026050596/meta-06302026xexhibit991.htm), July 29, 2026), so current Meta profiles outnumber its headcount 1.7 to 1. Visible profiles include contractors and people who left without updating their profile, so every count here is of profiles, not of staff. Most of Meta's visible AI staff name none of the nine units searched. Each person counts at the first unit they name, in a fixed order, and 5,188 (62.8%) name none of them (`map_units.json`). The rest: the lab 936 (11.3%); GenAI, the name of Meta's former generative AI organization, 506 (6.1%); Reality Labs and AR, VR, and wearables work 400 (4.8%); AI infrastructure 390 (4.7%); ads and monetization 314 (3.8%); FAIR outside the lab 215 (2.6%); ranking and recommendations 152 (1.8%); Instagram, WhatsApp, Messenger, and Threads 104 (1.3%); integrity 52 (0.6%). FAIR outside the lab is the most research-heavy unit: 69.8% list the Research function and 57.2% a doctorate, against 19.8% and 29.2% across P2; the lab is next on research (35.1%), and on doctorates ranking and recommendations (36.2%) is ahead of the lab (31.5%). AI infrastructure is engineering (74.4% Engineering and Technical, 3.8% Research), and ads has the largest share at Manager or above (34.4%, against 14.2% across P2). 94.1% of the lab lists the US, against 66.0% of FAIR outside it. The lab is more research-titled and somewhat more senior than the rest of P2. 35.1% of the lab lists the Research function, against 17.8% of the rest (`map_functions.json`). The comparison is uneven by construction: the lab counts everyone who names it, while the rest counts only profiles that name an AI term, which leaves out about 2,200 research-titled profiles with no AI term; that lowers the rest's research share and raises its machine learning engineer share. 25.3% of the lab hold a research scientist title, against 10.4%; machine learning engineer, the most common title in the rest (24.2%), appears in 3.8% of the lab's titles (`map_titles.json`). 17.6% of the lab are at Manager or above against 13.7% of the rest, and 5.3% at Head or above against 3.3% (`map_levels.json`, `map_units.json`). The lab names foundation models and LLMs more often (28.7% against 18.5%, in headline, title, or skills) and ranking and recommendations less often (8.2% against 17.9%) (`map_directions.json`). More of the lab once worked at Google and Microsoft. 14.2% of the lab once worked at Google, Google DeepMind included, against 8.3% of the rest, and Microsoft 9.8% against 7.7%; OpenAI is 12 people (1.3%) against 19 (0.3%), and the Amazon gap (11.8% against 10.2%) is within chance (`map_sources.json`). Carnegie Mellon and Berkeley appear in the education of 5.4% and 5.7% of the lab, against 3.8% and 3.9% of the rest; the Stanford gap (5.1% against 3.9%) is within chance. The China-educated share does not differ: 19.8% of lab members with a Bachelor entry studied at a mainland-China institution (144 of 727), and 19.6% of the rest (1,083 of 5,514). In the rest, the share is 21.8% among individual contributors (765 of 3,509) and 15.6% at Manager and above (117 of 752); the lab is too small to split by level. More people moved from Meta to the AI labs than the other way (`map_flows.json`, every role at Meta, moves that began in June 2025 or later). 191 people who left Meta hold a new job at OpenAI, while 37 joined Meta from OpenAI; Anthropic 81 out and fewer than 10 in; xAI 33 and 11; Thinking Machines Lab 16 and fewer than 10; Google DeepMind 58 and 41. With Amazon, Microsoft, and Apple it runs the other way: 691 joined Meta from Amazon and 153 went there; 343 joined Meta from Microsoft and 154 went there; 158 joined Meta from Apple and 134 went there. NVIDIA is the exception, with 82 out and 20 in, and Google outside DeepMind is about even (317 in, 308 out). Across all employers, 7,787 joined Meta and 8,054 left in the window. 218 open Meta postings posted in the last 180 days carry an AI title, 10.4% of Meta's 2,104 open postings (`map_postings.json`, `postings.json`). 107 (49.1%) are machine learning or software engineer titles, 38 (17.4%) research scientist, and 14 (6.4%) research engineer. 46.3% name ranking or recommendations in the title or description, 35.8% LLMs or foundation models, 34.9% AI infrastructure, and 17.4% agents. 197 US postings state a base range in dollars (`map_pay.json`): 120 (60.9%) top out at $250,000 or more and 8 (4.1%) at $340,000 or more, and the ranges fall on a few fixed bands. Research scientist postings (33) do not state higher ranges than machine learning and software engineer postings (99): 54.5% and 67.7% of them top out at $250,000 or more. Research engineer postings (12) are too few to publish. ### Definitions and method P2 is a current Meta job, the stage-1 entry, and an AI term in the headline or current title: machine learning, ML, deep learning, computer vision, NLP, LLM, large language models, reinforcement learning, generative AI, GenAI, artificial intelligence, superintelligence, MSL, or FAIR (the full list is in `queries/mapping.json`), or AI in the current title alone. Everyone outside the lab also needs a current job at the Meta recorded with more than 10,000 employees, a current function outside the non-technical ones (human resources, sales, marketing, and the like), and a current title without non-technical words (recruiter, marketing, and the like). The lab's 936 are in P2 by construction, so the lab is exactly the stage-1 group, 38 of them in non-technical roles. Units are the first of lab, FAIR, Reality Labs and AR or VR words, infrastructure, ads, ranking, integrity, the apps, and GenAI named in the headline or current title. GenAI comes after the product areas because a third of the people who write it use it for the topic, often inside ads, integrity, or an app. China-educated uses the institution list of the China-educated study. The definition was audited by reading 365 profiles in whole slices (one state or country and a narrow range of total experience, every match read), never the top of a search. The reads removed three traps: AI alone in a headline is mostly a buzzword, plain research scientist and research engineer titles at Meta also cover people research, survey science, photonics, and software testing, and outside the US the name Meta also matches other companies. Of the 150 read profiles that fall in the final P2, in eight whole slices, 138 (92.0%) hold AI roles; in the last slice, read after the definition was fixed, 35 of 39 (89.7%) when two borderline roles count as misses. Separate whole slices showed that GenAI names the former org in 16 of 24 reads, that generative AI without it is mostly the topic (18 of 23, so it is not a unit term), that AR, VR, and wearables words describe Reality Labs products in 12 of 16, and that infra means AI infrastructure in 26 of 28. 91 postings read in two whole cities set the posting rules: every Meta description carries a paragraph on augmented and virtual reality and most mention responsible AI, so descriptions are searched only for distinctive single words. Stage 2 adds 250 counts to the replay and reads nothing. Beyond the small-cell rule, `fetch.py` lists every sum a reader can form, including the stage-1 cells that contain a stage-2 cell, solves them together by exact elimination, and withholds a count (null) until no group of 1 to 9 people can be worked out. That withholds the research, doctorate, US, and China-educated counts of the units-that-name-none row and the US count of the apps, merges the lab's other and no-function rows, publishes levels in five groups, and leaves the lab's China-educated split by level unpublished. The first full replay (353 API Credits) wrote files that a check written afterwards found would let a reader work out the lab's interns (fewer than 10) from two sums taken together, which the first check, solving one sum at a time, had missed; the check now solves all sums at once. A second full replay (353 API Credits) reproduced every number. A fact check then found that moves matched Meta by name alone, which outside the US also matches other companies: moves now take Meta as the employer of P2 (the name with employer size 10,001+, and Facebook, Instagram, WhatsApp, and Oculus by name), and anyone with a current entry under the name Meta at any size is not counted as a leaver. Two cheap probes (9 API Credits) sized the change, a third full replay (353 API Credits) tested it, and the fourth, with the final definition, wrote the published files and `data/receipt.json`; only the moves changed. Stage 2's exploration cost 1,318 API Credits: the 32-Credit probe, 218 for the agent's counts and reads (including 27 counts the first output had hidden), three superseded full replays (1,059), and the 9 of probes. With stage 1's 311 (206 before its replay, and its own replay of 105, which the combined replay replaced), the case cost 1,629 API Credits to explore. ### Limits Units are what people write, not Meta's organization chart, and 62.8% write none of the nine units searched. P2 is neither a floor nor a ceiling on Meta's AI staff: stale profiles and contractors who list Meta are in it, while research scientists who name no AI word (about 2,200 profiles) are left out because the reads found many of them outside AI, and the audit put the share in AI roles at about 92%, with stale profiles and generic engineers who list AI words among the misses. Skills lists are noisy, and descriptions are longer than headlines, so directions compare within profiles or within postings, not across them. Postings are one day's open roles (every one read was posted within three weeks of the snapshot), and a role posted in several cities counts once per city; the jobs index has no employer size, so a few postings at other companies named Meta may be included. Moves count every role, not only AI roles; they rest on profile dates that stop in April 2026 and fall across the October 2025 reductions in Meta's AI organization ([SiliconANGLE](https://siliconangle.com/2025/10/22/meta-lays-off-600-ai-workers-looks-streamline-superintelligence-labs/), October 22, 2025) and before the May 2026 reduction, which profile dates do not reach; and few people at the AI labs carry AI words in their titles (fewer than 10 of the 191 who moved to OpenAI, none of the 16 at Thinking Machines Lab), so the AI-titled subset says little there. Pay is the posted base range only, not bonus, equity, or what anyone is paid. How hard these roles are to fill needs a comparison across employers and is out of scope for one company. ## Sources Meta announced the lab in an internal memo on June 30, 2025, which [CNBC published](https://www.cnbc.com/2025/06/30/mark-zuckerberg-creating-meta-superintelligence-labs-read-the-memo.html). Meta's investment in Scale AI was [reported by Axios](https://axios.com/2025/06/13/meta-scale-ai-deal) on June 12, 2025. Microsoft formed its superintelligence team in November 2025, [reported by CNBC](https://www.cnbc.com/2025/11/06/microsoft-forms-superintelligence-team-under-ai-head-mustafa-suleyman-.html). Meta split the lab into four teams (TBD Lab, FAIR, Products and Applied Research, and MSL Infra) in August 2025, [reported by Built In](https://builtin.com/artificial-intelligence/meta-superintelligence-reorg) on August 27, 2025, and cut more than 600 roles in three of them in October 2025, leaving TBD Lab untouched, [reported by SiliconANGLE](https://siliconangle.com/2025/10/22/meta-lays-off-600-ai-workers-looks-streamline-superintelligence-labs/) on October 22, 2025. Meta's headcount of 75,472 on June 30, 2026 is from its [results release filed with the SEC](https://www.sec.gov/Archives/edgar/data/1326801/000162828026050596/meta-06302026xexhibit991.htm) on July 29, 2026. ## Rerun ```bash export METIX_KEY=metix_xxxxxxxxxxxx # create one at https://platform.metix.ai/api-keys python3 cases/meta-superintelligence-labs-2026/fetch.py ``` To make it with an agent instead, use [`PROMPT.md`](PROMPT.md).