# Bootstrap prompt

Paste the prompt below into an agent that can reach the Metix AI Platform: one with the [metix-skills](https://github.com/MetixAI-Official/metix-skills) installed, the MCP server connected, or plain REST access with `METIX_KEY` set. Followed end to end, it costs about 500 to 600 API Credits: 353 for the published counts of both stages, about 110 to find and read the audit slices and postings, and 40 to 140 for the counts that test the definitions and size the slices. It stops and asks before 650. Every published number is a count; the reads only check the definition.

```text
Answer two questions with the Metix AI Platform: of the people who name Meta Superintelligence Labs on their public profile, how many were already at Meta when the lab formed in June 2025, and where did the rest come from? And where do Meta's visible AI staff sit, and where does the lab sit among them? Work only through the public Platform (REST at https://mira-api.metix.ai, the MCP server, or the metix-skills) with the key in METIX_KEY, and never print the key.

1. Read before querying. Call GET /contract (free) and build every condition from querySpecByEntity.profile and querySpecByEntity.job; read GET /docs/api/people (free) too. Conditions about one job go inside one has_experience entry, so they describe the same job. Text fields match word by word, in any order, ignoring case; current_function, current_seniority, experience.seniority, and experience.company.type take fixed values with eq or in. A query holds at most 64 conditions and nests at most 6 levels. A count with size 1 costs 1 API Credit and nothing when it finds no one, a search 1 API Credit per 25 IDs, and a detail read 1 API Credit per 5 records, so plan every published number as a count. Send each query once with an extra word that matches nothing first: the Platform checks it against its limits and charges nothing for an empty result. Check the balance with GET /auth/key/status (free) at the start and at the end, and stop and ask before the run passes 650 API Credits.

2. Population. A current job at Meta (experience.company.name match "Meta" and experience.is_current eq true, in one has_experience entry) and a headline or current_title that matches superintelligence or MSL. Check with counts that company.name eq "Meta" gives the same total, so the name catches no other company, and that Facebook is the same employer on the Platform rather than a separate test. An earlier Meta job is an entry at Meta, Facebook, Instagram, WhatsApp, or Oculus. An internship is an entry with experience.seniority eq "Intern" or an experience.title that matches intern, tested in the same entry. Count the narrower definition (superintelligence only) too, and, as context, current Meta profiles that name FAIR but no lab term.

3. Audit before counting. Search ranks results by match quality, so the top of a search looks cleaner than the group. Read whole slices instead: a search with size 25 costs 1 API Credit and returns the total, so a slice of 25 or fewer comes back complete. Read about 60 profiles in five slices (people with no Meta entry before June 2025 in one state, people with one in two cities, the group whose Meta entry has no start date, and people who write MSL but not superintelligence in one state) with POST /entity/v1/profiles/detail-by-id and _source headline, current_title, current_function, and experience company.name, title, start_date, end_date, is_current, and seniority, never a name. Report the share that describe working in the lab, check that the veterans had real Meta jobs rather than only internships, and look for Meta company names the list misses. Read the matches for each sub-team term, a whole slice of the Microsoft group, and every open posting the same way. Keep every record out of every published file.

4. Cohorts. The cut is 2025-06-01. Count the total; people with an earlier Meta entry that started before the cut; the same without internships; and people who stayed, with a non-intern Meta entry that started before the cut and is current or has experience.end_date gte "2025-05-01" (the any of those two inside the entry). Returned is the non-intern count minus stayed, former interns the first count minus the non-intern count, and new to Meta the total minus the first count; the last three together are the outside hires. Count total, earlier entry, and stayed for the narrower definition too, and the four groups for current Microsoft profiles that name superintelligence, with a cut of 2025-11-01 and an end of 2025-10-01.

5. Dates. Count people with an earlier Meta entry that started before each of 2012-01-01, 2014-01-01, 2016-01-01, 2018-01-01, 2020-01-01, 2022-01-01, 2023-01-01, 2024-01-01, 2025-01-01, the cut, 2025-07-01, 2025-08-01, 2025-09-01, 2025-10-01, 2026-01-01, and 2026-04-01; the difference between two dates is a band of first-join dates. Repeat without internships up to the cut. For people who stayed, count those with a current Meta entry that started on or after each of 2025-01-01, 2025-04-01, 2025-06-01, 2025-07-01, 2025-08-01, 2025-09-01, 2025-10-01, 2026-01-01, and 2026-04-01, which places each by their newest role. For Microsoft, count before 2012, 2016, 2020, 2023, 2025-01-01, and its cut. Profiles lag, so count Meta entries by start month across all profiles and say which month is the last complete one.

6. Sources. For the outside hires (the population, not stayed), count a job entry with experience.end_date gte "2025-01-01" that is not an internship, at each employer in this order: OpenAI, DeepMind, Anthropic, xAI, Thinking Machines, Apple, Scale AI, Google, Microsoft, Amazon or AWS, NVIDIA; then experience.company.type eq "Educational", where internships count. Each row leaves out anyone an earlier row matched, and the rest are other employers or none. For the whole group, count anyone with a past non-intern job at Google or DeepMind, OpenAI, Anthropic, Apple, Microsoft, and Amazon or AWS; those rows overlap.

7. Roles, teams, places, and postings. Count current_function eq Research, Engineering and Technical, and Product, and current_function exists, for the group and for the outside hires. For people whose current function is Research, in the lab and in the baseline of current Meta staff with that function and no lab term, count current_seniority in Senior, in Manager or any level above it, and exists, and a Doctorate education entry; count doctorates for the group and the outside hires as well. Count headline or title mentions of FAIR, TBD, infra or infrastructure, and PAR or product applied research. Count location.country United States, location.state California, New York, and Washington, Bay Area cities within California, and the United Kingdom. For open Meta postings (is_open eq true, company.name match "Meta", and a title matching superintelligence or MSL or a description matching superintelligence labs), count the total, research titles, Menlo Park, San Francisco, New York, postings that state min_experience_months and those at 36 or less, the seniority label, postings since 2026-09-01, and every open Meta posting.

8. Outputs. Write aggregate files with "unit" ("profiles" or "jobs"), the snapshot date, and the query files they came from. Every count of people from 1 to 9 is written "<10". A band of dates that would hold 1 to 9 people merges with a neighbour, never across the cut. When the hidden cells of a published sum add up to 1 to 9, withhold one more cell. Split the narrower definition and the Microsoft group only as finely as keeps every difference between published counts at 0 or at least 10. Publish a share only when its denominator is at least 30. Check every difference a reader could form before writing anything, and name no individual in any text you write, including people named in the press; linking to press coverage is fine.

9. Charts. The first-join bands as a timeline with the cut marked, the months after it for people new to Meta, and Microsoft on the same axis; the four groups as one bar, with the narrower definition and Microsoft beneath it; outside hires by where they had just worked, with the five AI labs grouped; current function for people who stayed and outside hires; doctorates and levels of the lab's research staff against the baseline; sub-team mentions; states; open postings by place. Label bars directly, and give each chart a title that states its finding in neutral words.

10. Limits. The population is people who chose to name the lab, not the lab: people who write only FAIR or nothing are missing, and newcomers and long-time staff may name it at different rates. Counts are visible profiles, never a headcount, and since stale profiles count as current, not a guaranteed floor either. Some profiles are out of date, and recent months are undercounted. A date given only as a year reads as January, which moves some people across the cut; say in which direction. A new Meta entry is a lower bound on who moved, since many team changes add none. Sources rest on end dates. The Microsoft group is small and includes recruiters. Postings are one day. Use neutral verbs (join, leave, move, hire) and describe no pay other than the posted base ranges of step 17.

11. The map's population. Meta's visible AI staff (P2) are people with the stage-1 current Meta entry and an AI term in the headline or current title: machine learning, ML, deep learning, computer vision, NLP, natural language processing, LLM, LLMs, large language models, reinforcement learning, generative AI, GenAI, artificial intelligence, superintelligence, MSL, or FAIR, or AI in the current title alone. Everyone outside the lab also needs a current Meta entry with experience.company.size eq "10,001+" in the same has_experience, a current_function outside Human Resources, Sales, Marketing, Administrative, Finance & Accounting, Legal, Customer Service, and Real Estate, and a current title without marketing, sales, recruiter, recruiting, sourcer, sourcing, talent, administrative, assistant, counsel, attorney, paralegal, accountant, communications, or partnerships. Put each of those three filters in an any beside the two lab-term leaves, so every stage-1 lab member stays in P2 and the lab is the same group in both stages. Send a list of terms with in on a text field: it matches like one match leaf per term and keeps each query under 64 conditions.

12. Audit the map. Read at least 100 profiles in whole slices of P2 (one state or country and a narrow range of total_experience_months, every match read) with headline, current_title, current_function, summary, and the current entry's title, description, company name, and size, never a name, and report the share in AI roles; it must reach 90%. Test the traps the reads found before settling the terms: AI alone in a headline is mostly a buzzword; research scientist and research engineer match word by word and cover people research, survey science, photonics, and software engineers in Reality Labs Research; outside the US the name Meta also matches other companies. Read whole slices of the unit terms as well (GenAI, generative AI, AR and VR words, infra) and decide from them whether each names a team or a topic, and in which order the units go.

13. Units and size. Count current Meta profiles in the US, and current entries at the 10,001+ Meta in the US and elsewhere, so that no count comes back as 100000+. Put each P2 member in the first unit whose terms appear in the headline or current title, in this order: the lab (superintelligence, MSL); FAIR; Reality Labs, with AR, VR, XR, augmented reality, virtual reality, mixed reality, wearables, smart glasses, and Oculus; infra or infrastructure; ads, monetization, or advertising; ranking, recommendation, recommendations, recommender, or recsys; integrity; Instagram, WhatsApp, Messenger, or Threads; GenAI; and none of these. For each unit and for P2, count the total, the Research and the Engineering and Technical functions, manager and above, a doctorate, the US, a Bachelor entry, and a Bachelor entry at a mainland-China institution on the list in cases/china-educated-ai-talent-2026/queries/institutions.json. The lab's total, functions, doctorates, and US count are the stage-1 counts of the same group.

14. Functions, levels, and titles. For the lab and for P2, count data titles (data, annotation, annotator, labeling, prompt, knowledge expert, rater, evaluator) outside Research, then Engineering and Technical, Product with Design and Project Management, and no function, each without a data title; research is the Research count of step 13 and other is what is left. Count current_seniority Intern, Specialist, Senior, Manager, and Head or Director; Vice President and above is manager and above minus Manager, Head, and Director. Count current titles containing machine learning engineer, software engineer, research scientist, research engineer, engineering manager, product manager, program manager, data scientist, data engineer, prompt engineer, production engineer, and director, and keep the ten largest.

15. Directions. For foundation models and LLMs, post-training and alignment, multimodal and perception, AI infrastructure, agents, ranking and recommendations, AR, VR, and devices, and safety and evaluation, count the lab and P2 by a phrase in the headline or current title or a single word in skills, and Meta AI postings (step 17) by a phrase in the title or a single word in the description. Every Meta description carries a paragraph on augmented and virtual reality and most mention responsible AI, and a phrase matches its words anywhere in a long text, so read a whole slice of postings before choosing the description words; the list is in queries/mapping.json.

16. Sources and moves. For the lab and P2, count a past non-intern job at Google or DeepMind, OpenAI, Anthropic, Apple, Microsoft, Amazon or AWS, NVIDIA, and ByteDance or TikTok, and any education entry at Stanford, Carnegie Mellon, Berkeley, the Massachusetts Institute of Technology, Urbana, and an Indian Institute of Technology. Split the China-educated share by level (Intern, Specialist, and Senior against manager and above) only where every cell holds 20 or more. For moves since 2025-06-01, with the same window on both sides and Meta the same employer as in P2 (the name with employer size 10,001+; Facebook, Instagram, WhatsApp, and Oculus by name): out to X is a current non-intern entry at X that started on or after 2025-06-01, a non-intern Meta entry that ended on or after 2025-05-01, and no current entry under the name Meta at any size; in from X is a current non-intern Meta entry that started on or after 2025-06-01, a non-intern entry at X that ended on or after 2025-05-01, and no current entry at X. Count OpenAI, DeepMind, Anthropic, xAI, Thinking Machines, Microsoft, Apple, Amazon or AWS, NVIDIA, Google without DeepMind in the same entry, and any employer outside the Meta names, and each again with the AI terms where the row holds 10 or more.

17. Postings and pay. Take Meta postings with posted_date gte "now-180d" whose title holds a P2 AI term or AI and none of the non-technical title words, and count them by family, first match in order: research scientist; research engineer; software engineer, machine learning engineer, ML engineer, or AI engineer; data titles; product manager, product management, program manager, or designer; other. For US postings in USD that state both salary.annual_min and salary.annual_max, count all of them and each of the first three families, and for each with 15 or more, count annual_min at or above 150,000, 180,000, 215,000, and 265,000 and annual_max at or above 215,000, 250,000, 295,000, and 340,000. That is the posted base range, not pay.

18. Map outputs. Write one aggregate file per measure: data/map_size, map_units, map_functions, map_levels, map_titles, map_directions, map_sources, map_flows, map_postings, and map_pay. Every count of people from 1 to 9 is "<10" and a share needs a base of 30. List every sum a reader can form (the units add up to P2, each table row to its total, a unit's US staff sit inside its count, a stage-2 lab cell sits inside the stage-1 cell that holds it, a move row's AI-titled part sits inside the row), and withhold the smallest cell until no group of 1 to 9 people can be worked out. Stage 1 published the lab's functions, so when splitting out data titles would leave such a difference, merge the lab's other and none. Write nothing if any check fails.

19. Map charts and limits. The org map as a tree: P2, its units, and the lab with its stage-1 groups beneath it; then units by function and level, directions as profiles against postings, sources, moves as paired bars, and posted pay bands by family. Say that units are what people write, not Meta's org chart; that P2 counts profiles, not staff, with the accuracy the audit measured; that postings are one day's open roles and a role posted in several cities counts once per city; that moves rest on profiles that lag to April 2026 and fall across the October 2025 reductions in Meta's AI organization; and that hiring difficulty is out of scope for one company.
```

## Adapt it

| To change | Edit | For example |
| --- | --- | --- |
| The lab | The employer and lab terms in step 2, audited again in step 3 | Google DeepMind staff who name Gemini, or Microsoft AI |
| The date the lab formed | The cut in steps 4 and 5 | 2025-11-01 for Microsoft's superintelligence team |
| Where newcomers came from | The employer order in step 6 | Add Mistral AI or Cohere |
| The map's units | The order and terms in step 13, audited again in step 12 | A unit for WhatsApp alone, or Llama |
| The API Credit ceiling | Steps 1, 3, and 12 | Run steps 1 to 10 alone for about 200 API Credits |

## What to ask before running it

When someone brings a looser version of this question, settle these first. Each one changes the query or the cost:

1. Which words mark someone as part of the lab, and does any of them also mean something else?
2. Which company names count as the same employer, now and in the past?
3. What date separates people who were there from people who joined?
4. Do internships count as having been there?
5. How many API Credits may the audit reads spend?
6. Which words in a headline name a team, and which only a topic?