5,000 frontier users would be 38.2% of the core labs' reported staff
Data as of
- Core contributor floor
- 3,913
- Core staff ceiling proxy
- 13,100
- Broader contributor floor
- 4,207
How many people can use frontier AI systems inside the labs? We count 3,913 named model contributors. Dated staff records sum to about 13,100 people. Access for 5,000 people would require 38.2% of that workforce. Access lists remain private.
Andrej Karpathy estimated about 5,000 people. Scope: People with internal access to frontier-grade AI systems, the last step of a funnel that starts with about 6 billion people who have barely used LLMs. Supporting text: “And finally we get to the ~5,000 people (~0.00006%) with access to frontier-grade systems internally.” Our core scope covers Google DeepMind, OpenAI and Anthropic.
Key findings
- We count 3,913 distinct contributor names across reports from 3 labs.
- The latest dated all-staff observations sum to about 13,100 people.
- An access count of 5,000 would cover 38.2% of that recorded workforce.
- The wider set of 9 labs supplies 4,207 contributor names. Its full workforce is unknown.
Distinct names across core model reports: 3,913. Claim reference: 5,000. Dated staff total: 13,100. These measures describe credited work and a workforce ceiling proxy; internal access counts are unknown.
Contributor lists and dated staff sources for Google DeepMind, OpenAI and Anthropic appear in the tables below. Claim: Andrej Karpathy, 11 October 2026. Supporting text: “And finally we get to the ~5,000 people (~0.00006%) with access to frontier-grade systems internally.” Scope: People with internal access to frontier-grade AI systems, the last step of a funnel that starts with about 6 billion people who have barely used LLMs.
Data table
| Measure | People | Meaning |
|---|---|---|
| Named contributor floor | 3,913 | Distinct normalized names across the latest qualifying reports in the source set. |
| Claim reference | 5,000 | Internal frontier access estimate attributed to Andrej Karpathy. People with internal access to frontier-grade AI systems, the last step of a funnel that starts with about 6 billion people who have barely used LLMs. |
| Staff ceiling proxy | 13,100 | Sum of the latest recorded total staff at each core lab; dates differ. |
Start with the people credited on models
A model report gives the first check: people publicly credited with work on a model. The core definition covers developers of Gemini, GPT and Claude. We treat their general-purpose model families as frontier-grade for this comparison. The wider definition below shows the effect of including other model developers.
| Core lab | Latest qualifying report in the source set | Named contributors |
|---|---|---|
| Google DeepMind | Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities (7 July 2025) | 3,434 |
| OpenAI | GPT-5 System Card (19 December 2025) | 486 |
| Anthropic | Claude Opus 4 and Sonnet 4 System Card (2025-05) | No individual byline |
Google DeepMind supplies 3,434 of the 3,920 available per-lab report credits (87.6%), the largest count. This share is calculated before matching names across labs; missing bylines can change the balance.
The per-lab counts sum to 3,920. Matching identical normalized names across labs removes 7 repeat entries, giving 3,913. This is a floor on the credited model work represented by these reports. Bylines can include outside contributors and former staff. Name variants and shared names can affect the count.
We use each lab’s latest qualifying report. Older reports remain in the archive. Anthropic, Meta Superintelligence Labs / Meta AI, xAI and Alibaba Qwen team have no individual byline on the latest qualifying report or model card. A missing byline leaves the contributor floor incomplete.
Staff totals set a ceiling proxy
We next ask how many employees could have access if it extended across the whole recorded workforce. The latest staff sources for the core comparison give a ceiling proxy of 13,100 people. The claim would require 38.2% of that recorded workforce to have access.
| Core lab | Selected all-staff figure, date, scope and evidence |
|---|---|
| Google DeepMind | 5,600 (8 September 2025). Google DeepMind unit; broad unit headcount including Gemini teams. Selected as the latest dated all-staff observation. Source text: “The Google unit now has 5,600 employees.” |
| OpenAI | about 4,500 (21 March 2026). All company employees. Forecasts on the source page are excluded. Selected as the latest dated all-staff observation. Source text: “up from about 4,500 employees.” |
| Anthropic | around 3,000 (18 May 2026). All company employees. Selected as the latest dated all-staff observation. Source text: “Anthropic has around 3000 employees, up from roughly 300 in December 2023.” |
We select each lab’s most recent dated observation of all staff in the source set. This rule excludes hiring targets, research-only totals, team subsets and planned cuts. The table shows the selected figure and its scope for each lab. Older observations remain in the source records below.
The archive also gives an all-staff observation for every core lab in 2024 and 2025. Taking each lab's latest observation within each year, with open-ended counts excluded, gives about 5,215 staff for 2024 and 10,900 for 2025. That is 109% growth. Observation dates within each year differ, so these sums give an approximate view of workforce change.
This ceiling includes every role. Access could include researchers, engineers, safety teams and other staff who test internal models. That is a possible reason for an access count to exceed a research-only count. The sources do not show who has permission. Staff figures come from different dates. Later hiring could make our total too low. Outside users with access would expand the relevant population. These records provide no current access bound.
Research shares show the size of technical teams
These dated records separate research teams from the wider workforce.
| Lab | Date | Role | Role count | Total staff | Share | Source |
|---|---|---|---|---|---|---|
| Anthropic | 31 December 2024 | research and development | 521 | 915 | 56.9% | source |
| Mistral AI | 8 September 2026 | researchers | 300 | 1,200 | 25% | source |
The 38.2% access share needed in the core comparison falls within the observed research shares. That comparison describes workforce composition. Access policies could produce a different share. These shares describe different jobs and dates. We apply neither share to another lab. Research status alone leaves access permissions unknown.
A wider lab definition adds evidence gaps
The next scope adds Meta Superintelligence Labs / Meta AI and xAI through their Llama and Grok models. The broad scope also admits Mistral AI, DeepSeek, Moonshot AI and Alibaba Qwen team. It covers open-weight and general reasoning model families. This is an inclusive definition of frontier-grade. Thinking Machines Lab and Safe Superintelligence are left out because the supplied sources do not establish a qualifying frontier model report.
Core closed-model labs: 3,913 distinct contributor names; report counts available for 3 labs out of 3; staff ceiling proxy 13,100. Add Llama and Grok: 3,913 distinct contributor names; report counts available for 5 labs out of 5; staff ceiling proxy Unavailable. Broader model labs: 4,207 distinct contributor names; report counts available for 8 labs out of 9; staff ceiling proxy Unavailable.
Latest qualifying model report in the supplied source set for each lab. Links, dates and staff scopes appear in the tables below.
Data table
| Lab definition | Named contributor floor | Report count coverage | Staff ceiling proxy |
|---|---|---|---|
| Core closed-model labs: Google DeepMind, OpenAI and Anthropic | 3,913 | 3 of 3 labs | 13,100 |
| Add Llama and Grok: Google DeepMind, OpenAI, Anthropic, Meta Superintelligence Labs / Meta AI and xAI | 3,913 | 5 of 5 labs | Full ceiling unavailable. Dated staff subtotal: 16,500. xAI: more than 1,200. |
| Broader model labs: Google DeepMind, OpenAI, Anthropic, Meta Superintelligence Labs / Meta AI, xAI, Mistral AI, DeepSeek, Moonshot AI and Alibaba Qwen team | 4,207 | 8 of 9 labs | Full ceiling unavailable. Dated staff subtotal: 18,160. xAI: more than 1,200. No total for Alibaba Qwen team. Contributor counts exceed dated staff figures at DeepSeek. |
| Lab | Latest qualifying report in the source set | Named contributors | Latest recorded staff total, date and scope |
|---|---|---|---|
| Meta Superintelligence Labs / Meta AI | Llama 4 Scout and Maverick Model Card (5 April 2025) | No individual byline | about 3,400 (17 July 2025). Broad Meta AI organization under Meta Superintelligence Labs, including foundation, product, applied AI, and FAIR teams. Selected as the latest dated all-staff observation. Source text: “a staff totaling about 3,400 in the lab.” |
| xAI | Grok 4.20 Model Card (7 April 2026) | No individual byline | more than 1,200 (12 March 2025). All company staff, including about 900 hourly AI tutors. Selected as the latest dated all-staff observation. Source text: “As CEO, Musk oversees a team of more than 1,200 people.” |
| Mistral AI | Magistral (12 June 2025) | 99 | 1,200 (8 September 2026). All company employees; 300 are researchers. Selected as the latest dated all-staff observation. Source text: “le nombre de 300 chercheurs, sur 1 200 employés, reste en croissance.” |
| DeepSeek | DeepSeek-R1 (22 January 2025) | 197 | about 160 (14 March 2025). All company employees. Selected as the latest dated all-staff observation. Source text: “Deepseek has “about 160” employees.” |
| Moonshot AI | Kimi K3 (27 July 2026) | Count unavailable | 300 (11 August 2026). All company staff. Selected as the latest dated all-staff observation. Source text: “这家300人的公司没有部门。” |
| Alibaba Qwen team | Qwen3.5-Omni Technical Report (17 April 2026) | No individual byline | No lab-specific total found |
Magistral is the latest qualifying general reasoning report in that lab's source set. Small-model and coding reports remain in the source archive. The latest qualifying report count is unavailable for Moonshot AI. Kimi K3 (27 July 2026) has no cached individual byline available for this run. Earlier reports supply no replacement count in the floor.
The broader documented floor remains 793 below the claim. Missing bylines and reports leave room for more contributors. The staff side also has gaps. xAI has an open-ended total of more than 1,200 (12 March 2025). A lab-specific staff total is missing for Alibaba Qwen team. The broader dated staff subtotal is 18,160. It excludes open-ended totals and labs with no staff figure. It supplies no full ceiling.
DeepSeek's latest qualifying report names 197 contributors. Its staff source reports about 160 (14 March 2025). The contributor count exceeds the staff figure. Different dates, outside contributors or different organizational scopes are possible reasons. The records do not resolve the mismatch, so this pair cannot define a valid staff range.
What this means for the access estimate
We estimate about 13,100 staff across the core labs from their latest dated all-staff records. Access for 5,000 people would reach 38.2% of that workforce. The scale is compatible with the recorded total. Contributor lists and workforce totals leave current access unmeasured.
A direct count needs current access rules and user lists from each lab. Missing public records leave our comparison incomplete.
Sources
Evidence through 14 October 2026. We use the reports and staff records listed below; later model releases and hiring can change this comparison.
All dated staff sources
| Lab | Date | Reported figure | Scope | Selection | Source text | Source |
|---|---|---|---|---|---|---|
| Google DeepMind | 17 May 2024 | about 2,600 | Google DeepMind after Google Brain merged into it; broad unit headcount. | Earlier all-staff observation | “oversees about 2,600 employees, according to internal data seen by BI.” | Entrepreneur, republished from Business Insider |
| Google DeepMind | 7 August 2025 | around 6,000 | Google DeepMind workforce; broad unit headcount. | Earlier all-staff observation | “Google’s DeepMind workforce has swelled to around 6,000 employees.” | Mint, republished from The Wall Street Journal |
| Google DeepMind | 8 September 2025 | 5,600 | Google DeepMind unit; broad unit headcount including Gemini teams. | Selected: latest dated all-staff observation | “The Google unit now has 5,600 employees.” | The Information |
| OpenAI | 2024 (exact month not stated) | 1,000 | All company staff; rounded historical figure. | Earlier all-staff observation | “OpenAI grew from 1,000 to 3,000 people in the year he was there.” | TechCrunch |
| OpenAI | 2 October 2024 | about 1,700 | All company employees. | Earlier all-staff observation | “company has gone from about 700 to 1,700 employees over the past year.” | The Information |
| OpenAI | 18 November 2024 | more than 1,700 | All company employees. | Earlier all-staff observation | “more than 1,700 employees.” | The Information |
| OpenAI | 15 July 2025 | about 3,000 | All company staff. | Earlier all-staff observation | “OpenAI grew from 1,000 to 3,000 people in the year he was there.” | TechCrunch |
| OpenAI | 21 March 2026 | about 4,500 | All company employees. Forecasts on the source page are excluded. | Selected: latest dated all-staff observation | “up from about 4,500 employees.” | The Information |
| Anthropic | 8 October 2024 | roughly 800 | All company employees. | Earlier all-staff observation | “Anthropic, with roughly 800 employees” | The Information |
| Anthropic | 2024-10 | about 1,000 | All company staff; approximate figure reported during late 2024. | Earlier all-staff observation | “expanded in less than a year from nearly 200 people to about 1,000.” | WIRED |
| Anthropic | 31 December 2024 | 915 | All employees at year end. | Earlier all-staff observation | “Anthropic ended 2024 with 915 employees, 521 of whom were in research and development.” | The Information |
| Anthropic | 31 December 2024 | 521 | Research and development employees at year end; subset of the 915 total employees. | Excluded: research | “521 of whom were in research and development.” | The Information |
| Anthropic | 4 December 2025 | about 2,300 | All company employees. | Earlier all-staff observation | “from around 500 employees in late 2023 to about 2,300 today.” | Fortune |
| Anthropic | 18 May 2026 | around 3,000 | All company employees. | Selected: latest dated all-staff observation | “Anthropic has around 3000 employees, up from roughly 300 in December 2023.” | The Information |
| Meta Superintelligence Labs / Meta AI | 17 July 2025 | about 3,400 | Broad Meta AI organization under Meta Superintelligence Labs, including foundation, product, applied AI, and FAIR teams. | Selected: latest dated all-staff observation | “a staff totaling about 3,400 in the lab.” | The Information |
| Meta Superintelligence Labs / Meta AI | 22 October 2025 | roughly 600 | AI organization positions to be cut. This figure counts planned job cuts. | Excluded: reductions | “cutting roughly 600 positions out of the several thousand roles within Meta’s superintelligence lab.” | Axios |
| xAI | 2024-07 | 97 | All company employees; PitchBook estimate cited by Observer. | Earlier all-staff observation | “xAI’s current employee headcount measures at 97, according to PitchBook.” | Observer |
| xAI | 12 March 2025 | more than 1,200 | All company staff, including about 900 hourly AI tutors. | Selected: latest dated all-staff observation | “As CEO, Musk oversees a team of more than 1,200 people.” | Business Insider, republished by Yahoo Finance |
| xAI | 15 September 2025 | 1,500 | Data annotation team, including full-time and contract staff. | Excluded: subset | “the xAI’s data annotation team was one of the largest, with 1,500 full-time and contract staff members.” | Los Angeles Times |
| Mistral AI | 11 June 2024 | 60 | All company employees. | Earlier all-staff observation | “Avec 60 employés, Mistral AI souhaite également développer son équipe commerciale.” | Le Monde |
| Mistral AI | 9 September 2025 | 350 | All company employees across six offices. | Earlier all-staff observation | “La start-up revendique 350 employés au sein de six bureaux dans le monde.” | Le Monde |
| Mistral AI | 28 May 2026 | 1,000 | All company employees; 300 work directly with businesses. | Earlier all-staff observation | “300 de ses 1 000 salariés travaillent en contact direct avec les entreprises.” | Le Monde |
| Mistral AI | 8 September 2026 | 1,200 | All company employees; 300 are researchers. | Selected: latest dated all-staff observation | “le nombre de 300 chercheurs, sur 1 200 employés, reste en croissance.” | Le Monde |
| DeepSeek | 14 March 2025 | about 160 | All company employees. | Selected: latest dated all-staff observation | “Deepseek has “about 160” employees.” | Tom’s Hardware, citing the Financial Times |
| Moonshot AI | 20 February 2024 | more than 80 | All company employees. | Earlier all-staff observation | “Moonshot AI had raised USD300 million in financing before this round and employs more than 80 people.” | Yicai Global |
| Moonshot AI | 11 August 2026 | 300 | All company staff. | Selected: latest dated all-staff observation | “这家300人的公司没有部门。” | Sina Finance, republished from 企业家杂志 |
| Thinking Machines Lab | 18 February 2025 | about 30 | All company employees. | Earlier all-staff observation | “Thinking Machines Lab says it has about 30 employees.” | Axios |
| Thinking Machines Lab | 24 April 2026 | around 140 | All company employees. | Selected: latest dated all-staff observation | “The startup’s headcount now stands at around 140.” | TechCrunch |
| Safe Superintelligence | 4 September 2024 | 10 | All company employees. | Selected: latest dated all-staff observation | “SSI, which currently has 10 employees” | Reuters, hosted by Investing.com |
All model report sources
| Lab | Report date | Report and source | Distinct contributors | Count method | Date in source notes |
|---|---|---|---|---|---|
| Google DeepMind | 8 March 2024 | Gemini 1.5 Technical Report | 1,134 | Distinct normalized individual names in cached arXiv citation metadata; collective labels excluded. | Same date |
| Google DeepMind | 7 July 2025 | Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities | 3,434 | Distinct normalized individual names in cached arXiv citation metadata; collective labels excluded. | Same date |
| OpenAI | 25 October 2024 | GPT-4o System Card | 417 | Distinct normalized individual names in cached arXiv citation metadata; collective labels excluded. | Same date |
| OpenAI | 5 December 2024 | OpenAI o1 System Card | Unavailable | Individual names could not be checked from the cache. | Same date |
| OpenAI | 19 December 2025 | GPT-5 System Card | 486 | Distinct normalized individual names in cached arXiv citation metadata; collective labels excluded. | Same date |
| Anthropic | 2025-02 | Claude Sonnet 3.7 System Card | 0 | The official system card PDF has no named individual author or contributor list. | Same date |
| Anthropic | 2025-05 | Claude Opus 4 and Sonnet 4 System Card | 0 | The official system card PDF has no named individual author or contributor list. | Same date |
| Meta Superintelligence Labs / Meta AI | 31 July 2024 | The Llama 3 Herd of Models | 559 | Distinct normalized individual names in cached arXiv citation metadata; collective labels excluded. | Same date |
| Meta Superintelligence Labs / Meta AI | 5 April 2025 | Llama 4 Scout and Maverick Model Card | 0 | The official model card names Meta as model developer and has no named individual author list. | Same date |
| xAI | 20 August 2025 | Grok 4 Model Card | 0 | The model card names xAI as provider and has no named individual author list. | Same date |
| xAI | 7 April 2026 | Grok 4.20 Model Card | 0 | The model card names xAI as provider and has no named individual author list. | Same date |
| Mistral AI | 8 January 2024 | Mixtral of Experts | 26 | Distinct normalized individual names in cached arXiv citation metadata; collective labels excluded. | Same date |
| Mistral AI | 12 June 2025 | Magistral | 99 | Distinct normalized individual names in cached arXiv citation metadata; collective labels excluded. | Same date |
| Mistral AI | 8 August 2025 | Devstral: Fine-tuning Language Models for Coding Agent Applications | 103 | Distinct normalized individual names in cached arXiv citation metadata; collective labels excluded. | Same date |
| Mistral AI | 13 January 2026 | Ministral 3 | 119 | Distinct normalized individual names in cached arXiv citation metadata; collective labels excluded. | Same date |
| DeepSeek | 7 May 2024 | DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model | Unavailable | Individual names could not be checked from the cache. | Same date |
| DeepSeek | 27 December 2024 | DeepSeek-V3 Technical Report | 197 | Distinct normalized individual names in cached arXiv citation metadata; collective labels excluded. | Same date |
| DeepSeek | 22 January 2025 | DeepSeek-R1 | 197 | Distinct normalized individual names in cached arXiv citation metadata; collective labels excluded. | Same date |
| Moonshot AI | 10 April 2025 | Kimi-VL | Unavailable | Individual names could not be checked from the cache. | Same date |
| Moonshot AI | 25 April 2025 | Kimi-Audio | Unavailable | Individual names could not be checked from the cache. | Same date |
| Moonshot AI | 28 July 2025 | Kimi K2 | 197 | Distinct normalized individual names in cached arXiv citation metadata; collective labels excluded. | Same date |
| Moonshot AI | 2 February 2026 | Kimi K2.5 | 335 | Distinct normalized individual names in cached arXiv citation metadata; collective labels excluded. | Same date |
| Moonshot AI | 27 July 2026 | Kimi K3 | Unavailable | Individual names could not be checked from the cache. | Same date |
| Alibaba Qwen team | 15 July 2024 | Qwen2 | 62 | Distinct normalized individual names in cached arXiv citation metadata; collective labels excluded. | Same date |
| Alibaba Qwen team | 19 December 2024 | Qwen2.5 | 42 | Distinct normalized individual names in cached arXiv citation metadata; collective labels excluded. | Same date |
| Alibaba Qwen team | 26 January 2025 | Qwen2.5-1M | 28 | Distinct normalized individual names in cached arXiv citation metadata; collective labels excluded. | Same date |
| Alibaba Qwen team | 14 May 2025 | Qwen3 Technical Report | 60 | Distinct normalized individual names in cached arXiv citation metadata; collective labels excluded. | Same date |
| Alibaba Qwen team | 17 April 2026 | Qwen3.5-Omni Technical Report | 0 | Distinct normalized individual names in cached arXiv citation metadata; collective labels excluded. | Same date |
| Thinking Machines Lab | 15 July 2026 | Inkling Model Card | 0 | The model card has a Thinking Machines Lab provider credit and no named individual author list. | Same date |
| Thinking Machines Lab | 30 July 2026 | Inkling-Small Model Card | 0 | The model card has a Thinking Machines Lab provider credit and no named individual author list. | Same date |
Archived OpenAlex links and exclusion reasons
| Lab | Institution links | Archived queries and dates | Why counts are omitted |
|---|---|---|---|
| Google DeepMind | 2023 archived query; through 14 October 2026. 2024 archived query; through 14 October 2026. 2025 archived query; through 14 October 2026. 2026 archived query; through 14 October 2026. | Author counts are omitted because the archived queries include coauthors from other institutions. | |
| OpenAI | 2023 archived query; through 14 October 2026. 2024 archived query; through 14 October 2026. 2025 archived query; through 14 October 2026. 2026 archived query; through 14 October 2026. | Author counts are omitted because the archived queries include coauthors from other institutions. | |
| Anthropic | 2023 archived query; through 14 October 2026. 2024 archived query; through 14 October 2026. 2025 archived query; through 14 October 2026. 2026 archived query; through 14 October 2026. | Author counts are omitted because the archived queries include coauthors from other institutions. | |
| Meta Superintelligence Labs / Meta AI | 2023 archived query; through 14 October 2026. 2024 archived query; through 14 October 2026. 2025 archived query; through 14 October 2026. 2026 archived query; through 14 October 2026. | Author counts are omitted because the archived queries include coauthors from other institutions. The affiliation IDs also cover a wider company or research unit. | |
| xAI | 2023 archived query; through 14 October 2026. 2024 archived query; through 14 October 2026. 2025 archived query; through 14 October 2026. 2026 archived query; through 14 October 2026. | Author counts are omitted because the archived queries include coauthors from other institutions. | |
| Mistral AI | 2023 archived query; through 14 October 2026. 2024 archived query; through 14 October 2026. 2025 archived query; through 14 October 2026. 2026 archived query; through 14 October 2026. | Author counts are omitted because the archived queries include coauthors from other institutions. | |
| DeepSeek | 2023 archived query; through 14 October 2026. 2024 archived query; through 14 October 2026. 2025 archived query; through 14 October 2026. 2026 archived query; through 14 October 2026. | Author counts are omitted because the archived queries include coauthors from other institutions. | |
| Moonshot AI | 2023 archived query; through 14 October 2026. 2024 archived query; through 14 October 2026. 2025 archived query; through 14 October 2026. 2026 archived query; through 14 October 2026. | Author counts are omitted because the archived queries include coauthors from other institutions. | |
| Alibaba Qwen team | 2023 archived query; through 14 October 2026. 2024 archived query; through 14 October 2026. 2025 archived query; through 14 October 2026. 2026 archived query; through 14 October 2026. | Author counts are omitted because the archived queries include coauthors from other institutions. The affiliation IDs also cover a wider company or research unit. | |
| Thinking Machines Lab | No lab-specific record verified in the source notes. | Author counts are omitted because the archived queries include coauthors from other institutions. | |
| Safe Superintelligence | No lab-specific record verified in the source notes. | Author counts are omitted because the archived queries include coauthors from other institutions. |
We omit OpenAlex author counts because the archived queries include coauthors from other institutions. Several institution mappings remain unverified. Meta and Alibaba IDs cover wider organizations.
Institution searches: Google DeepMind; OpenAI; Anthropic; Meta AI; Meta FAIR; Meta Superintelligence Labs; xAI; Mistral AI; DeepSeek; Moonshot AI; Alibaba Qwen; Thinking Machines Lab; Safe Superintelligence; Meta Platforms; Facebook AI Research; Fundamental AI Research; Meta Superintelligence; Meta GenAI; Alibaba DAMO Academy; Qwen; Thinking Machines; Safe Superintelligence Inc.
Query method: OpenAlex grouping documentation.
Career data comes from the biographies on this site and from OpenAlex author records, as of 15 October 2026. People whose OpenAlex record may be incomplete or mixed are left out of citation figures.