v11.1 (September 20th 2026). Watch list, Europe: seven entries added at Florin’s request, since the hundred carries fewer Europeans than the continent’s work deserves and the watch list is where the next moves show first: Oriol Vinyals (Discovery Loop), Xavier Amatriain (Expedia Group), Guillaume Lample and Timothée Lacroix (Mistral AI), Thomas Wolf (Hugging Face), Pushmeet Kohli and Lila Ibrahim (Google DeepMind) and Nuria Oliver (ELLIS Alicante), each with a 2026 claim and a fetched source. Vinyals moves there from the “carried by someone else” line. Spanish role labels renamed the same day: Doer is now Constructor and Shaper is Decisor; Hacedor and Configurador read as translations.

v10 (September 19th 2026). The title now names companies. Part 7 lists nineteen companies ranked by humans reached and purchasing power touched, from a three-round model panel published as the Structural Reliance Matrix. Part 1’s top twenty was reordered under the new stance-and-significance rule in the Method, with a fourth badge for stance: Hassabis first, then Huang, Pichai, the Amodeis, Altman, Trump, Xi, Liang Wenfeng, Wei, Nadella, Ternus, Musk, Eddie Wu, Zuckerberg, Fouquet, Bengio, Karpathy, Atallah, Liang Rubo, Jassy. Bengio, Fouquet, Liang Rubo and Jassy enter the twenty; Steinberger, Kavukcuoglu, Mark Chen and Yang Zhilin move to 21 to 27; Pope Leo XIV moves from 77 to 28 on the encyclical, with the signing and publication dates corrected. Each top-20 entry ends with the panel’s reach figures. Two conflicts recorded: Florin sells on Amazon, which touches the Jassy entry; the drafting agent is an Anthropic model, which touches the Amodei entry, and no Anthropic model sat on the four panels of September 19th 2026. Two harm clauses now carry sources (Meta, Reuters November 6th 2025; ByteDance, the state attorneys-general suits of October 2024); a Rufus clause on Amazon was dropped for lack of a source. The Dalai Lama stays off Part 1 until a dated 2026 instrument can be named. Later the same day (v10.1) the By instruments tag was withdrawn: heads of state carry no stance tag, and the 2026 list gained an order control with three views, impact, humans reached and purchasing power touched.

Top-20 validation round (September 19th 2026). Thirteen houses asked, twelve delivered (DeepSeek V4 Pro returned an API error), judged on the declared rule rather than their own. Threshold for returning a placement to Florin was eight DISAGREE; none reached it. Lowest agreement: Ternus 5 AGREE and 4 DISAGREE on tenure, Kavukcuoglu 5 and 4, the Amodeis 6 and 2 with four CANNOT-TELL on 2026 facts past the models’ cutoffs. Most protected: Huang and Hassabis (eight houses each), Xi (seven), Trump (six), Wei and Fouquet (five). Symmetry findings adopted: the ByteDance harm clause needed a source (six houses); Ternus’s tenure is stated in his entry; Musk’s in-progress tag is scoped to Optimus and Grok. Kavukcuoglu’s seat went to Bengio. Recorded, not adopted: Apple’s App Store as an attention footprint, since the rule scores AI conduct. Cost 0.70 USD.

Purchasing-power round (September 19th 2026). Nineteen entities on five money measures, twelve of thirteen houses delivered; Mistral looped and was discarded, Thinking Machines’ Inkling joined. Combined humans-and-money order: Nvidia, Google, Meta, Microsoft, Apple, Amazon, ByteDance, TSMC, OpenAI, Tencent, Alibaba, ASML, Huawei, Anthropic, SpaceX with Tesla and xAI, Xiaomi, DeepSeek, Zhipu, Kimi. TSMC and ASML enter from nowhere on money; ByteDance drops from 4th on humans to 7th; Apple rises on the wealth of its users; Alibaba and Amazon hold rank on the commerce their models recommend. The claim that Apple plus Tesla and X reach 80 percent of the rich holds in the US and Western Europe and fails globally, at 62 to 77 percent.

ByteDance round (September 19th 2026). ByteDance added as a seventh Chinese entity, eleven of twelve houses delivered: 2.45 billion humans, rank 5 of 16; the Chinese cluster 2.65 billion with it, overlap with the US cluster 1.75 billion, China unique 950 million on apps and about 400 million counting silicon; ByteDance inference an estimated 60 percent on Nvidia. The tightest agreement on any entity in any round. Cost 0.91 USD.

Structural reliance round (September 19th 2026). Fifteen entities ranked by humans reached, twelve houses, Anthropic excluded on purpose. Kept: Nvidia first (ten of eleven), Google above Meta, TSMC and ASML as chokepoints below Nvidia with no audience of their own. Rejected: Nvidia at 6 billion, which exceeds the internet population (4.3 to 5.5 billion kept); Meta AI as the largest consumer AI; Hunyuan as reaching all of China; Anthropic at 100 million and 14th, rejected by every model. Still open: ByteDance’s Nvidia inference share, Llama downstream reach, Bedrock end users, and the substitution-cost matrix that would separate the top three. Cost 1.97 USD; the full matrix and its record are published as their own page.

v8 (September 18th 2026). Florin’s top 20, applied after source checks: Huang first (the Hugging Face acquisition, September 3rd 2026, confirmed), then the Amodeis, Altman, Hassabis, Trump, Musk, Xi, Pichai, Nadella, Liang Wenfeng; John Ternus enters at 11 as Apple’s CEO from September 1st 2026; Karpathy’s entry now reflects his move to Anthropic’s pre-training team on May 19th 2026; Murati, Brockman and Sutskever move to 21 to 23; Richard Socher, Peter Norvig and Johny Srouji enter behind them; Harrison Chase enters in the application tier. Dual tags (Doer+Shaper, Doer+Thinker) were added where one role does not describe the person. Five entries whose organisation is carried by someone above them left the hundred. The watch list gained a section on world models, agents and self-improvement. A Models tab lists one hundred models from OpenRouter’s catalogue at the date. One correction to a source Florin forwarded: AutoGPT was created by Toran Bruce Richards, not David Zhou.

v7 (September 18th 2026). Tone pass on all 100 entries of the 2026 list by a Sonnet agent against the brief “no criticism, no under or overstatements, no predictions, no guru phrases, lower the drama”: 43 entries changed. The top ten were then put to GPT-6 Astra, Grok 4.6 and Claude Opus 5 with the same brief; all three asked for rewrites of all ten and agreed on the causes (unsourced superlatives, ranking arguments inside entries, one prediction, and the TIME comparison in the Hassabis entry). The ten were rewritten in one plain register from the three proposals. Every name on the four lists now links to a page about the person, fetched and checked; where none could be confirmed the name stays plain. The intro and byline were rewritten at Florin’s instruction, the review record moved below the lists, the layout widened, and a hero image added: a stylised collage with no real faces, because likenesses of real people are not something this document generates.

v6.5 (September 18th 2026). Part 3 grew from 21 to 100 advocates and popularisers at Florin’s request, grouped by channel and ordered by fetched reach figures; 85 names came from seven Sonnet research batches with every reach figure and URL fetched, plus Ken Liu (Pantheon), Ezra Klein, Ben Thompson and the Newton and Roose pair. Part 4 grew from 3 to 32 dated claims across six regional batches. The Amodeis now share the first entry, Mira Murati stays at her rank, and the Method carries a “no commander” note.

v6.4 (September 18th 2026). Florin’s ranking call: Dario and Daniela Amodei share the first entry, then Huang, Altman and Hassabis, whose AlphaGenome Atlas (September 8th 2026) is now his key work. “Sets the pace” is gone from Altman’s entry and the Method says there is no commander. Academia check: seven academics with a significant publication in the last twelve months came in (van der Schaar, Bosselut, Gal, Kai Chen, Khapra, Adelani, Florez), each with a fetched source, and seven entries whose 2026 lever was carried by someone else went to the note above. Alex Atallah (OpenRouter, acquired by Stripe for over 7 billion USD, August 2026) added as the routing layer every multi-model workflow, this review included, runs on. Academic titles removed for everyone.

v6.2 (September 18th 2026). Peter Steinberger added at Florin’s request after a source check (OpenClaw, joined OpenAI February 15th 2026, foundation set up for the project); Arvind Krishna moved to the considered-and-not-included note, since IBM’s 2026 influence on what AI is runs through the regulators and hyperscalers already listed.

v6 (September 18th 2026). Every 2026 entry now carries one key work with a link and, for doers, the company’s own page: 100 lookups by five research agents, each link fetched and checked, with the handful that sites blocked marked unverified rather than guessed. The research also caught one stale fact: Noam Shazeer left Google for OpenAI on June 18th 2026, and the entry now says so. The two “considered and not included” notes were rewritten so that each name carries its own reason, grouped by kind; the version history that used to sit there lives here instead.

Nemotron 3 Ultra, free slug, fourth attempt (September 18th 2026, 0 USD). The failure was never cost and never the model: with reasoning switched off in the request it returned a full review in 283 seconds and said ACCEPT WITH FIXES. Applied: Chervonenkis now shares the Vapnik entry (three houses asked), Yann LeCun is tagged Doer for the same reason Fei-Fei Li is (he is building a lab), and the SK hynix entry says ADR listing rather than “took public”. Its other factual objections (the SpaceX and xAI valuation, SoftBank passing Toyota) are confirmed by 2026 sources it could not see. Its regional proposals repeat names other houses gave (Matsuo, Anandkumar, Finn, already in) plus several the record does not support.

v5 (September 18th 2026, after the Claude Opus 5 and Kimi K3 retries came back). Both accepted with fixes. Consensus acted on: Tegmark and Ng to the advocates list (four houses), Jumper’s 2026 entry removed as double counting (two), a diffusion entry added to the all-time list (two), Yejin Choi and Chelsea Finn added (two each), the A* credit completed, Valiant’s origin fixed, two promotional superlatives softened. Florin’s request for Chinese, Japanese and European thinkers: Andrew Yao, Zhou Bowen, Yutaka Matsuo, Bernhard Schölkopf and Sepp Hochreiter added, plus Zhuang Rongwen, the first Chinese regulator on the list, named by Opus 5. Single-house proposals held for next time: Tri Dao, Luis von Ahn, the fair-use judges (Alsup, Chhabria), Amin Vahdat, Alexey Ivakhnenko, Cynthia Dwork, Barbara Grosz, Lila Ibrahim’s restoration. Nemotron 3 Ultra spent 101K reasoning tokens and returned nothing, twice.

Third pass, the six retries (September 18th 2026, 1.1 USD). Claude Opus 5, Kimi K3, Qwen 3.8 Max, Inkling and GLM 5.3 all returned substance and all five said ACCEPT WITH FIXES; Meta’s Muse Spark returned nothing twice and Muse Glimmer answered every claim CANNOT-TELL. Nemotron 3 Ultra spent 101K reasoning tokens and returned nothing. Consensus fixes are in v5. Inkling questioned the SpaceX IPO; it is confirmed (Nasdaq, SPCX, June 12th 2026, 75 billion USD raised). GLM questioned Ian Hogarth’s title; no source confirms a step-down, so it stands. GLM’s SHRDLU date is fixed. Grace Hopper (GLM), Zoubin Ghahramani (Mistral, Inkling), Rob Fergus (Inkling) and Nathaniel Rochester (Qwen) are noted as single-house candidates.

v4 fixes from the second round. The country-label convention said birthplace first and the entries did not follow it (Musk, Pichai, Nadella, Bengio, Sutskever, Buolamwini, von der Leyen, Grey Walter); the convention now says base first and the entries match. Bellman, Spärck Jones and Cortes were filed in the wrong eras and are now in the era of their work. ASML no longer prints “every” advanced chip, since SMIC prints Huawei’s on DUV. Wu Wenjun is no longer “the founder of Chinese AI research”. Tags: Fei-Fei Li is a Doer in 2026 (World Labs), Masayoshi Son a Shaper (capital), Andrew Ng a Doer (AI Fund). Kavukcuoglu moved above Brockman because he runs DeepMind. Anima Anandkumar restored and Emily Bender added, each named by two houses; Álvaro Soto (Latam-GPT) and Ha Jung-woo (Korea’s AI council) close the two regional gaps the first round left open. All-time: Colmerauer, Kingma and Ronald Williams in; Lenat, Yudkowsky and Murati out.

v3 changes after the first review were Florin’s editorial decisions, not panel findings: six political leaders added, six entries removed to make room, category tags on every entry, and a parallel advocates list.

Second round, v3, twelve houses (September 18th 2026, 1.62 USD plus retries). Asked for one review from each of OpenAI, Anthropic, Google, Moonshot, Alibaba, DeepSeek, Nvidia, Mistral, Thinking Machines, Z.ai, Meta and Tencent. Five answered on the first pass (GPT-6 Astra, DeepSeek V4 Pro, Gemini 3.8 Flash, Mistral Medium 3.5, Hunyuan A13B); Meta’s model returned HTTP 403 on OpenRouter; Claude Opus 5, Kimi K3, Nemotron 3 Ultra, Qwen 3.8 Max, Inkling and GLM 5.3 returned empty bodies at a 14K token budget, which is the reasoning-model failure the skill file documents, and were re-run with a larger budget (see below).

Rejected from the second round. Gemini’s and Mistral’s factual objections were all knowledge-cutoff artefacts again (Pope Leo XIV, Alexandr Wang, Jumper, Sutton and Barto’s Turing Award, the Anthropic order), each re-checked against a primary source. Mistral proposed replacements who do not exist or hold no such role (an “Adebayo Adewole”, a “Thiago Ventura”, a “Woojay Han”), which is the fabrication risk the skill file warns about, and its ten additions were discarded whole. Tencent’s Hunyuan review contradicted verified facts on Sacks and the encyclical and was discarded. Gemini’s objection that heads of state confuse state power with AI agency is recorded as a dissent; Florin’s decision stands. Single-model proposals without a second vote (Watkins, Lowe, Quinlan, Sohl-Dickstein, Dwork, Chervonenkis as a separate entry, Kowalski, Bajcsy) are noted here as the strongest next candidates for the all-time list.

Panel run through OpenRouter, five models, 0.91 USD, about five minutes. Each model was asked to falsify nine numbered claims, list factual errors, propose additions and removals, and give a protect-list.

Panel behaviour. GPT-6 Astra and DeepSeek V4 Pro gave usable, specific reviews. Gemini 3.8 Flash (self-reported cutoff March 2026) and Nemotron 3.5 Lightning (cutoff June 2024) flagged verified 2026 facts as “hallucinations”: Alexandr Wang at Meta, Yann LeCun’s AMI Labs, Pope Leo XIV, Bezos’s Prometheus, Sutton and Barto’s Turing Award. Every one of those was checked against a primary source today and kept. Kimi K3 returned an empty body, its fourth such failure on this panel.

Confirmed and fixed (from the panel). Deep Blue beat Kasparov in a game in 1996 and in a match in 1997. ChatGPT reached 100 million users in two months, not a billion. Backpropagation (1986) was Rumelhart, Hinton and Williams. MYCIN was built by Shortliffe, not Feigenbaum. Falcon is TII’s, not G42’s. Several “first” claims (Turing 1950, Shannon 1950, Grey Walter, the perceptron, MENACE) were softened to what the record supports. Lample was a senior author of Llama, not its lead author. The Pope contradicted the head-of-state exclusion; the exception is now declared. Country labels are now defined. The entry count is now disclosed honestly.

Confirmed and fixed (from the agent’s own source check, before the panel returned). Hassabis became Chair of DeepMind and Chief Scientist of Alphabet on August 5th 2026, with Kavukcuoglu running DeepMind. Jeff Dean and Oriol Vinyals left Google the same day to found Discovery Loop. John Jumper left DeepMind for Anthropic (June 2026). Fidji Simo stepped down from OpenAI (July 2026) and was replaced on the list by Bret Taylor. SpaceX absorbed xAI (Feb 2026). Jan Leike stepped back from running alignment (May 2026).

Ranking changes on panel consensus (three of four working models). C.C. Wei moved from 43 to 10. Daniela Amodei moved from 14 to 30. Ilya Sutskever moved from 10 to 16. Kavukcuoglu moved into the top 15 because he now runs DeepMind. Lisa Su moved to the top of the compute tier.

Additions on two-model consensus. Robin Li (Baidu) for Joseph Gordon-Levitt. Nandan Nilekani for Daniel Kokotajlo. Christophe Fouquet (ASML) for Pat Grady and Alfred Lin. Meredith Whittaker for Lila Ibrahim. All-time: Karen Spärck Jones for Ray Kurzweil, Richard Bellman for Yuval Noah Harari, Corinna Cortes for Mustafa Suleyman. Chervonenkis credited alongside Vapnik.

Rejected, with reason. Splitting the transformer authors into eight entries (would cost seven slots for one paper; the grouping is disclosed instead). Kurt Gödel and Thomas Bayes (upstream mathematics, consistent with the stated rule). Replacing Llion Jones, Jay Y. Lee, Michael Intrator or Pelonomi Moiloa with alternates (proposed by one model each, no consensus, and each proposal traded a verified 2026 lever for a plausible one). Nemotron’s proposal to remove ranks 1 through 33 (it argued every entry reflected fame; that is a rejection of ranking, not a finding).

Still open. Latin America is represented by one US-based founder; no model could name a Latin America-based person with comparable AI influence, and neither could the agent. Korea is represented only through memory chips. Both are stated limits, not oversights. The calls above are Florin’s; the drafting, the research runs and the review rounds are his Claude agent’s, and the trail records both.