Frontier AI lab researchers on X — self-asserted handles only

No, not as a discovery mechanism — we deliberately rejected that approach — searching X bios for lab names is noisy and low-precision (see docs/ai-org-intelligence-product-plan-2026-06.md). Instead this is roster-first: we build a list of real researchers/staff from ground-truth academic and organizational records (OpenAlex co-authorship, arXiv paper author lists, Wikidata employer statements), then check whether that specific, already-identified person has self-asserted an X handle — primarily via their own ORCID researcher-url, their own Wikidata P2002 claim, or their own GitHub profile's twitter_username field. A small minority (labeled "Bio (lower confidence)" wherever shown) is instead resolved via an explicit, current-tense employer statement in the person's own X bio — see the next FAQ for why this exists and how it's bounded. X is used to resolve and enrich, never to blindly discover.

31 listings on this page (out of 152 people) are resolved via bio_self_declared: the person's own X bio explicitly and currently states their lab affiliation in plain words (e.g. "Researcher @XiaomiMiMo", "Tech Lead of Multimodal RL at @TencentHunyuan"), but they didn't resolve via ORCID, Wikidata, or a GitHub org member's profile field. This is NOT a reopening of the rejected "search X bios for lab mentions" discovery axis (see the first FAQ) — every one of these people was identified individually, not surfaced by a blind keyword sweep of an unrelated population. Most already sit inside an independently-sampled academic roster (e.g. four Xiaomi MiMo listings are all named authors on the MiMo-VL Technical Report; their bios were only checked because they were already on that roster and didn't resolve any other way). We still treat this as a strictly lower-confidence tier than the other three, because bio text isn't a structured identity-registry field — it can be stale, aspirational, or ambiguous in ways an ORCID researcher-url or a GitHub profile field isn't. It's labeled "Bio (lower confidence)" everywhere it's shown, including the full roster table and the CSV export, so it's never presented at the same evidentiary weight as the other three.

3 listings — Sam Altman, Andrej Karpathy, and Timothée Lacroix — are included even though their CURRENT X bio does not state the affiliation, which every other signal on this page requires. Sam Altman's bio reads "AI is cool i guess" (though the account does carry X's own "OpenAI" affiliate badge); Andrej Karpathy's reads "I like training large deep neural nets" despite his widely-reported May 2026 move to Anthropic; Timothée Lacroix's still reads "PhD student at Facebook AI Research Paris," a leftover from before he co-founded Mistral in 2023. These three were added at the site maintainer's explicit direction, overriding the automated methodology used everywhere else — not discovered or qualified by any of the other four signals. This is the least rigorous tier on the page by design: it trades the "structured evidence only" discipline for well-known, uncontroversial public facts, and it's labeled "Public record (override)" everywhere so it's never confused with a self-asserted or registry-backed signal.

No, and it undercounts substantially on purpose — this is a floor, not a census. 1.6% of the OpenAlex-sampled academic roster and 23.1% of the Wikidata-employer roster had a self-asserted X handle; GitHub org public members ran higher at 35.3%. Most working researchers never link their X account anywhere machine-readable, so true X presence is far higher than what shows up here. The roster also skews toward people with an academic publishing history or public GitHub org membership — not every employee has either.

Every person on this page carries one or more of five signals, shown in the "source" column: their own ORCID researcher-url pointing at x.com/twitter.com, their own Wikidata P2002 (X username) claim, their own GitHub profile's twitter_username field discovered via public membership in that lab's GitHub org, an explicit current-employer statement in their own X bio (a small, clearly-labeled minority), or — for a handful of extremely well-known people whose bio doesn't state it — a maintainer-confirmed public record (see the dedicated "Public record (override)" FAQ, the least rigorous of the five). None of the first four are inferred from a name match or a blind bio-keyword sweep — they are all first-party, self-declared links, held to progressively lower confidence tiers from ORCID/Wikidata/GitHub down through bio text to the override tier. One known-bad Wikidata mapping surfaced during research (a mismatched handle for a DeepMind-affiliated academic) and was excluded rather than shipped.

None of the four roster sources (OpenAlex, arXiv, Wikidata, GitHub) carries a reliable employment date, so a resolved signal can point at a lab someone has since left — an old co-authored paper, a GitHub org membership nobody revoked, a Wikidata employer claim nobody updated. As a one-time snapshot pass on August 11, 2026, every listing with a live X bio was cross-checked: does the bio's current-tense self-description support the assigned lab, or does it explicitly contradict it (past-tense "previously/ex-/formerly <lab>", or a bio naming a different current employer with zero mention of the assigned lab)? 15 lab-listings failed that check and were removed rather than shipped as a false "current" claim. Listings with no bio available were left as-is (an absence of corroboration, not a contradiction). This is a point-in-time check, not a live feed — bios can go stale again after this pass, the same way the underlying roster signals did.

None of Qwen, Tencent Hunyuan, ByteDance Seed, Xiaomi MiMo, or MiniMax are their own OpenAlex institution (Qwen, Hunyuan, Seed, MiMo — they're model teams inside much larger companies: Alibaba, Tencent, ByteDance, Xiaomi), and using the parent company as the roster source would reproduce the exact false-positive problem this whole methodology exists to avoid. MiniMax does have its own Wikidata organization and a real technical report (MiniMax-M1, 109 authors via a pinned OpenAlex work id — "MiniMax" alone isn't distinctive enough for an affiliation-text search), but the one ORCID/Wikidata hit it produced was excluded: that person's own X bio names a different current employer with zero mention of MiniMax, an explicit bio-contradiction, not a real match (same bar as the "people who've since left the lab" FAQ). None of the roughly 750 people surfaced across all five — including 30-74% with a known ORCID — had a self-asserted X handle via ORCID, Wikidata, or a GitHub twitter_username field that survived that check; a plausible read is that X is blocked in mainland China, so self-declaring a handle through an academic identity registry is rare there. Tencent Hunyuan, ByteDance Seed, and Xiaomi MiMo show resolved people anyway, because a follow-up pass individually identified specific researchers (several already on the sampled roster above) whose own X bio states the affiliation directly — see the "Bio (lower confidence)" FAQ. Qwen and MiniMax haven't had that same follow-up pass complete yet, which is why they're still at zero — that's a gap in this pass, not a confirmed true negative.

Yes — the full roster (name, lab, X handle, source signal, followers, verified, role, previous affiliations) is published as a CSV under CC BY 4.0 at https://crawlora.net/ai-researchers-index/download. No email addresses or other non-public contact info are collected or published.

Same source as everything else self-declared on this page: the person's own X bio. Where a bio states a current title ("Research Engineer", "CEO & Co-Founder") we surface it near-verbatim in "Role"; where it uses explicit past-tense language ("previously/prev/formerly/ex-") naming a prior employer, we surface that in "Previously at". Unlike the lab-affiliation signal (ORCID, Wikidata, GitHub — all structured records), these two fields are extracted from free-text bio content, so they're held to a stricter bar: ambiguous, terse, or unclear bio text is left blank rather than guessed. A blank cell means the bio didn't clearly state one, not that we don't know.

That earlier study measured how often AI-lab names appear as TEXT inside bios across the X Users dataset — an aggregate keyword count with no named individuals and no employment verification. This page is the opposite approach: named individuals, each verified through a real employment/affiliation signal (OpenAlex, arXiv, Wikidata) and a self-asserted handle, not a bio-text match.