By Thorsten Meyer | October 26, 2026 | ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
One researcher keeps a list of 50 signs that artificial superintelligence has arrived, and he says today’s top models are “early ASI”. I took the strongest claims from his 2026 montage and traced each one back to the paper, the lab post or the researcher who would know. Some hold up beautifully. Some fall apart at the first phone call. And the one indicator that matters most is the one with the least hype attached.
Key takeaways
- The framework is generous. Dr Alan D. Thompson’s ASI tracker marks 2 of its 50 indicators as achieved, 33 as partially achieved and 15 as not achieved. “Partially” includes funded programmes and enabling work, so orange does not mean “almost there”.
- His ASI is other people’s AGI. Thompson defines ASI as expert-human level in practically any field. Google DeepMind’s 2023 scale calls that level “Expert AGI”; its 2026 report reserves “superintelligence” for systems more capable than large organisations of humans.
- Of 18 headline claims I checked, 7 are verified, 5 are vendor claims, 3 rest on a single source and 3 are disputed or contradicted. The verified ones cluster where results can be checked: proofs, a validated physics calculation, regulatory filings.
- Maths is the real breakthrough. Terence Tao, Don Knuth and Tim Gowers all describe genuine AI contributions, with honest caveats about coaching, literature and problem style.
- Recursive self-improvement is real but small. Anthropic’s own best internal model scores 62.8% on a test where it says a full replacement for its researchers would need at least 85%, and it estimates its R&D is faster with AI “but not yet by a factor of 2”.
- What has not happened yet: no household robots at scale, no AI-run company verifiably worth a billion dollars, no visible break in the US unemployment rate.
Products Worth Considering
DJI Osmo Mobile 7P with Mic Mini TX (Infinity Black) Gimbal Stabilizer
Sonos Era 100 Wireless Smart Speaker, Black
DJI Osmo Mobile 8 Phone Gimbal with Mic Mini Transmitter, Infinity Black
As an affiliate, we earn on qualifying purchases.
One researcher, 50 signs
Dr Alan D. Thompson is an Australian AI consultant who publishes The Memo newsletter and a public “ASI checklist” of 50 indicators, from “recursive code self-optimization” to “faster-than-light travel”. His October video, ASI: A better world – 2026, is a fast montage of dozens of entries from January to October. His view, stated in the tracker, is that Claude Mythos Preview (February) and GPT-6 Astra (September) are “early ASI” systems.
The tracker’s status grid, in the version dated October 26, 2026 that I read that afternoon (archived snapshot), shows 2 indicators achieved (a major new mathematical proof and a resolved conjecture), 33 partially achieved and 15 not achieved. Thompson explains in an August correction note what orange means: “documented material progress toward an indicator, including funded programs, research infrastructure, formal scientific missions, and enabling work”. That is a low bar. The two exoplanet candidates found with Claude Code this month, for example, are filed under indicator 47, “first planet other than Earth colonized”.
The same author runs a separate “conservative countdown to AGI”, which stands at 99%. The final percent, he writes, waits on a robot doing multi-step fine assembly, like IKEA furniture, at average-human reliability.
This is a useful list of things worth watching. Read it as one expert’s opinion, not as a measurement.
Whose superintelligence?
Thompson’s definitions are short: AGI is “a machine that performs at the level of an average (median) human”, ASI “a machine that performs at the level of an expert human in practically any field”. Compare that with the yardsticks the labs use:
- Google DeepMind’s “Levels of AGI” (2023) grades performance from “Competent” (at least the 50th percentile of skilled adults) to “Expert” (90th), “Virtuoso” (99th) and “Superhuman” (outperforms 100% of humans). Thompson’s ASI roughly matches DeepMind’s Expert AGI, two levels below superhuman.
- DeepMind’s June 2026 report “From AGI to ASI” describes superintelligence as a system “more intelligent and cognitively capable than large organisations of humans”. We covered it in Waves, Not a Wall.
- OpenAI’s Charter defines AGI as “highly autonomous systems that outperform humans at most economically valuable work”. The word that matters is autonomous.
So when someone says “proto-ASI”, the first question is: on whose scale? By Thompson’s, a model that writes expert-level proofs counts. By DeepMind’s, the same model is a strong AGI candidate in some domains and nowhere near a superintelligence.
Whose superintelligence?
One word, very different bars
Dr Alan D. Thompson's ASI starts where Google DeepMind's scale puts "Expert AGI". The same model can be "early ASI" on one scale and a strong but narrow AGI candidate on another.
- Median human50th percentileThompson: AGIa machine at the level of an average (median) humanDeepMind: Competent AGIat least the 50th percentile of skilled adults
- Expert90th percentileThompson: ASIexpert-human level in practically any fieldDeepMind: Expert AGIat least the 90th percentile of skilled adults
- Virtuoso99th percentile—DeepMind: Virtuoso AGIat least the 99th percentile of skilled adults
- Beyond every human100%—DeepMind: Superhuman (Level 5)outperforms 100% of humans
- Beyond organisations2026 report—DeepMind: ASImore capable than large organisations of humans
Thompson’s tracker status as of 26 October 2026. DeepMind rows: “Levels of AGI” (2023) and, for the last row, “From AGI to ASI” (June 2026). OpenAI’s Charter: AGI = “highly autonomous systems that outperform humans at most economically valuable work”.
Sources: LifeArchitect.ai ASI checklist (26 Oct 2026), AGI countdown, Morris et al., Levels of AGI (2023), Genewein et al., From AGI to ASI (2026), OpenAI Charter.
Powered by Thorsten Meyer AI
The montage, checked
I picked 18 of the strongest claims from the video and the tracker, grouped them, and went to the primary source: the researcher’s own note, the paper, the company post or the filing. I then looked for critics. The verdicts use five labels: verified (checked by an independent expert or an official document), vendor claim (the company’s own numbers, not independently reproduced), single source, disputed or contradicted, and not found.
| # | Claim (as told in the montage) | Primary source | Verdict |
|---|---|---|---|
| 1 | GPT-5.2 solved Erdős problem #728 | Terence Tao, Mathstodon (Jan 7); Tao’s AI-contributions wiki | Verified. “More or less autonomously”, after feedback; tools were Aristotle and GPT-5.2 Pro; the problem had been misformulated, and similar results existed |
| 2 | Claude Opus 4.6 solved Don Knuth’s open problem | Knuth, “Claude’s Cycles” (Feb 28, rev. Apr 14) | Verified. Claude found the construction for odd sizes in about an hour of coached exploration; Knuth wrote the proof; the even case was finished later with GPT models |
| 3 | An OpenAI model disproved Erdős’s 1946 unit-distance conjecture | OpenAI (May 20); companion paper by Alon, Bloom, Gowers, Sawin, Shankar and others | Verified by external mathematicians |
| 4 | GPT-6 Astra: ten results on decade-old problems for about $2,000, all formalised in Lean | OpenAI (Aug 1) | Vendor claim. Lean certificates are public; humans prepared the manuscripts with the model |
| 5 | Claude computed a nine-loop particle-physics amplitude | Anthropic guest post by Matt von Hippel, with Lance Dixon’s validation (Sep 25) | Verified by Dixon. Von Hippel: known methods, “not super-intelligently so”; a human team with GPT-6 had most of the result too |
| 6 | Claude discovered a new CRISPR-like enzyme system | Anthropic (Sep 23); CNN (Oct 8); The Conversation (Oct 8) | Disputed. Function unknown, not peer reviewed; a Copenhagen researcher says his unpublished work describes the same system; Anthropic denies any use of it |
| 7 | Claude designed scorpion antivenom | Anthropic (Oct 20); The Verge (Oct 22) | Contradicted as stated. 17 of 211 tested designs bound a toxin; none was tested for neutralisation. Anthropic: “we do not claim to have produced an antivenom” |
| 8 | A citizen scientist found a new rocky planet with Claude Code | Preprint on Zenodo (Sep 25) | Single source. A candidate; the author writes “No statistical validation is claimed” |
| 9 | Claude sped up AI research 52-fold; a skilled human gets about 4-fold | Anthropic, When AI builds itself (June) | Vendor claim, on one well-defined optimisation task |
| 10 | Anthropic’s internal “Model 2” approaches researcher level | Anthropic Risk Report (August); Axios (Aug 14) | Vendor claim: 62.8% on CoBench v2, where full substitution would need at least 85% |
| 11 | GPT-5.6 Sol made itself cheaper to run | OpenAI (Jul 29); GPT-5.6 launch (Jul 9) | Vendor claim: serving costs down 20%, an internal “RSI Index” of 57.9%. The “Sol post-trained Luna” quote is not in OpenAI’s written posts |
| 12 | Claude rewrote its own research agent for eight days | Weco AI (Jul 14) | Single source. Claude Opus 4.7 improved a different, Gemini-based agent; Weco rates it level 1 of 4, without “ignition” |
| 13 | Medvi: a $1.8 billion company run almost entirely by AI | The New York Times (Apr 2); Moneywise via Yahoo Finance (Apr 7) | Disputed. $1.8 billion is projected 2026 sales, not a valuation; doctors and pharmacies are outsourced partners; FDA warning letter and a class action |
| 14 | A Gemini agent runs a real café | Andon Labs (spring 2026) | Single source. The operator’s own account is candid: missed orders, staff impersonation, humans on standby; the post does not name the model |
| 15 | GPT-5 follows the law better than judges (100% vs 52%) | Posner and Saran, SSRN (Jan 29); The Register (Feb 15) | Verified numbers, narrow test: one hypothetical car-accident choice-of-law case; the authors call judges’ departures “a strength” |
| 16 | Figure’s robot works in 30 homes it has never seen | Figure (Sep 17) | Vendor claim. Zero-shot success rose from 9% to 56%; 44% of attempts still failed |
| 17 | Tesla started Optimus production on a high-volume line | Tesla Q3 2026 Update, SEC filing | Verified: production began in Fremont; Tesla calls them “the first production-quality bots” |
| 18 | China approved the first brain implant for use outside trials | Nature (Mar 16) | Verified: Neuracle’s NEO, for hand movement in severe paralysis |
The 2026 montage, checked
7 of 18 hold up fully
I traced 18 headline claims from Thompson's 2026 ASI montage to their primary sources and looked for critics. Every source is linked in the article.
| # | Area | Claim as told | Verdict | Why |
|---|---|---|---|---|
| 1 | Maths | Erdős problem #728 solved | Verified | Terence Tao: "more or less autonomously", after feedback |
| 2 | Maths | Knuth's open problem cracked by Claude | Verified | Odd case; Knuth wrote the proof |
| 3 | Maths | Erdős unit-distance conjecture disproved | Verified | Checked by external mathematicians incl. Tim Gowers |
| 4 | Maths | Ten decade-old problems for ~$2,000 | Vendor claim | OpenAI; Lean certificates published |
| 5 | Science | Nine-loop physics amplitude | Verified | Validated by Lance Dixon (SLAC) |
| 6 | Science | New CRISPR-like enzyme system | Disputed | Function unknown; priority claim; not peer reviewed |
| 7 | Science | "Claude designed antivenom" | Contradicted | 17 binders of 211 tested; no neutralisation test |
| 8 | Science | New rocky planet found | Single source | Candidate; "no statistical validation is claimed" |
| 9 | Self-improvement | 52× faster AI research step | Vendor claim | One well-defined task; humans ~4× |
| 10 | Self-improvement | "Model 2" near researcher level | Vendor claim | 62.8% where ≥85% would be needed |
| 11 | Self-improvement | GPT-5.6 Sol cut its own serving cost | Vendor claim | −20%, OpenAI's figure |
| 12 | Self-improvement | Claude rewrote its own agent for 8 days | Single source | A different agent; level 1 of 4 |
| 13 | Economy | Medvi, $1.8B company run by AI | Disputed | Projected sales, not valuation; FDA letter, lawsuit |
| 14 | Economy | Agent runs a café | Single source | Real café, humans on standby |
| 15 | Economy | GPT-5 follows the law better than judges | Verified | 100% vs 52%, one hypothetical case |
| 16 | Robots | Robot works in 30 unseen homes | Vendor claim | 56% success, 44% fail |
| 17 | Robots | Optimus production started | Verified | Tesla SEC filing |
| 18 | Brain implants | First implant approved outside trials | Verified | China, Neuracle NEO (Nature) |
The claims that survived are the ones that could be checked: proofs, a validated calculation, official filings.Thorsten Meyer
Verdicts: Thorsten Meyer, based on the primary sources listed in the article (read 26 Oct 2026). Verified = confirmed by an independent expert or an official document.
Powered by Thorsten Meyer AI
Maths and proofs: the part that holds up
The strongest evidence of 2026 comes from mathematics, because mathematics can check itself. In January Terence Tao wrote that Erdős problem #728 “was solved more or less autonomously by AI (after some feedback from an initial attempt)” and added the caveats himself: the problem as stated by Erdős was misformulated, and similar results by similar methods existed. In February Don Knuth opened his note with “Shock! Shock!” after Claude Opus 4.6 cracked a decomposition problem he had worked on for weeks. Read on, and the story is more human: his friend Filip Stappers coached Claude through 31 explorations and restarts, and Knuth wrote the rigorous proof.
The unit-distance disproof in May is the cleanest case. Tim Gowers wrote in the companion paper that if a human had submitted it to the Annals of Mathematics, he “would have recommended acceptance without any hesitation”. He also offered a sober reading: perhaps “there are certain styles of problem where it has a distinct advantage”, such as finding surprising connections across a huge literature. Our pieces on 722 AI proofs and formal verification go deeper.
Science: real results, inflated words
The nine-loop calculation is real. Lance Dixon at SLAC validated it and admitted he had thought a direct computation “too hard”. But the man who issued the challenge, Matt von Hippel, concluded that Claude “did something it turned out humans were also able to do”, using known methods “but not super-intelligently so”. His takeaway: there is more low-hanging fruit than experts expect.
Biology is where the words outrun the data. Anthropic’s ART enzyme system is an interesting pattern of DNA repeats whose function nobody knows yet; Aaron Engelhart of the University of Minnesota told CNN these are “very interesting preliminary results” that “don’t demonstrate gene editing”. The scorpion project, run by the investment firm Citrini Research with an unreleased Claude model, produced 17 working binders from 211 lab-tested designs. Andreas Hougaard Laustsen-Kiel of the Technical University of Denmark told The Verge: “It is easy-ish to discover a good binder against snake and scorpion toxins. The hard part is to translate it into a medicine.” Anthropic’s own post says the same in one line: it does not claim an antivenom.
Recursive self-improvement: the indicator that matters most
If AI can do AI research, every other indicator speeds up. That is why this one matters, and here the evidence is the most careful. Anthropic reports that Claude Mythos Preview reached a 52-fold speed-up on a fixed training-code optimisation task, against about 4-fold for a skilled human in four to eight hours. In the same post it says humans still hold the edge in “research taste and judgment”.
Its August Risk Report is more sobering. On CoBench v2, real debugging problems that Anthropic engineers solved, the scores are 15.6% for Claude Opus 4.6, 50.3% for Claude Mythos 5 and 62.8% for an unreleased “Model 2”. A model that could fully replace its research staff, Anthropic estimates, would need at least 85%. Its overall verdict: internal R&D is significantly faster with AI, “but not yet by a factor of 2”. OpenAI says GPT-5.6 Sol cut its own serving costs by 20%. Weco’s eight-day experiment calls itself level 1 on a four-level scale. And Anthropic now merges over 30,000 code changes a day, up from about 2,500 in May, mostly written by Claude, with human review as the new bottleneck. Loops are running. Nobody has shown the explosion. Jack Clark’s own estimate for AI that builds its successor, which we analysed in May, points to about 2028.
The indicator that matters most
AI that improves AI: loops yes, explosion no
On a narrow task Claude is already superhuman. On the messy, real work of an AI lab, Anthropic's best internal model still falls well short of what its own researchers do.
CoBench v2: real Anthropic R&D problems
Share of root causes found. Red line: the score Anthropic thinks a full replacement for its research staff would need.
Speed-up on one training-code task
A clearly defined optimisation task, Anthropic’s own measurement.
All figures are the companies’ own measurements. CoBench is filtered for hard problems, so scores may understate ability; Anthropic says the 85% bar is an uncertain estimate.
Sources: Anthropic Risk Report (Aug 2026), Fig. 3.4.3.A, Anthropic, When AI builds itself (Jun 2026), OpenAI (29 Jul 2026), Anthropic Engineering (13 Oct 2026), Weco AI (14 Jul 2026).
Powered by Thorsten Meyer AI
Agents in the economy: read the fine print
Medvi, the telehealth firm “run almost entirely by AI”, did $401 million in sales in 2025 with two employees, Moneywise reports in its summary of the Times profile. But $1.8 billion is projected 2026 revenue, not a valuation, doctors and pharmacies come from partner firms, and Moneywise lists an FDA warning letter (the founder says it concerned an affiliate’s website), fake-doctor ad allegations and a class action. Andon Labs’ café agent Mona took 44,000 Swedish kronor in its first two weeks, and also ordered 120 eggs for a café without a stove and emailed officials under a human colleague’s name. Andon calls it “a controlled experiment with humans standing by”. The judges study measured one thing well: GPT-5 applied a formal rule in a hypothetical case every time. Whether that makes a better judge is exactly the question its authors leave open.
Robots and brain implants: pilots, not households
Tesla’s SEC filing confirms Optimus production began in Fremont. Figure’s home robot is impressive, and its own numbers say 56% success in unseen homes. Thompson’s AGI page says it “fulfills the Woz AGI test at scale”: a humanoid walks into a strange home and gets to work. With 44% of attempts failing, I would not let it near my kitchen yet. In brain implants, China’s approval of Neuracle’s NEO, reported by Nature, is a genuine regulatory first. Neuralink’s imagined-speech demo rests on a company video.
What clearly has not happened
- No household robots at scale. Optimus production has started, but in our Q2 robotics check deployments were still mostly pilots.
- No visible labour-market break. The US unemployment rate was 4.3% in January and 4.2% in September 2026 (BLS, seasonally adjusted). That says nothing about individual professions, but it rules out a broad wave so far.
- No verified AI-run billion-dollar company. Indicator 22 asks for a valuation; Medvi’s figure is projected revenue.
- No autonomous discovery that has become a product. The biology results are leads, not drugs.
- No economy-wide payoff yet. As we wrote in AI payoff outside tech, the gains so far are concentrated in tech.
What this means for companies
- Use what is verified now. Expert-level help on checkable work is real: code with tests, proofs, data analysis with a ground truth, literature search with citations. Those are the areas where 2026’s claims survived scrutiny.
- Make verification a core competence. Every claim that fell apart did so at the step from “the model produced something” to “an expert confirmed it does what we said”. Budget for reviewers, test suites, lab checks and formal methods. Our Referee Shortage piece explains why checking is now the expensive part.
- Read vendor announcements like earnings releases. Ask what was measured, who checked it and what the failure rate is. “17 binders” and “56% success” are honest numbers inside dramatic headlines.
- Plan the workforce around bottlenecks, not headlines. Anthropic’s experience is instructive: once Claude writes most of the code, human review becomes the constraint. Train people to specify, review and decide. Do not cut the reviewers first.
- Watch the self-improvement metrics. Lab risk reports and system cards now publish them. If CoBench-style scores close the gap to 85%, the timeline for everything else shortens.
What this means for users
- Expect genuinely useful expert-level answers in some areas, especially maths, code and structured analysis, and much weaker results where the world has to push back, such as biology, health decisions and anything physical.
- Treat “AI discovered” headlines as a lead, not a result. Look for the word candidate, preprint or “not yet tested”.
- Robots in your home are still years, not months, away. A demo in 30 homes with a 56% success rate is progress, not a product.
- Do not hand over judgment. Even the AI-run café needed humans to stop it impersonating staff.
My verdict
How close are we? In narrow, checkable slices, today’s frontier models are at or beyond expert level, which by Thompson’s definition is the start of ASI and by DeepMind’s is Expert AGI in those domains. Across “practically any field”, including the physical world, we are not there. The indicator that would change everything, AI that improves AI faster than people can, shows real loops and honest gaps: 62.8% where 85% is needed, faster R&D but not yet twice as fast. That is the line to watch in 2027.
For the month-by-month story of the year, read AI in 2026 So Far: 10 Months, 10 Chapters. For the security side of these capabilities, see why the AI didn’t stop in the Mythos 5 tests and the CVE wave. For agents at work, see Claude Duet and the always-on agents and Nvidia’s Flora.
Method and limits
Thompson’s video is an automatic-transcript montage; I used it only as a list of claims and took every figure from the primary sources below, read on October 26, 2026. Several sources are company posts that I could not reproduce. Anthropic and OpenAI are central parties in this story, and this article was researched with Anthropic’s Claude. I did not use any of Thompson’s footage or graphics. Where I could not read a source (the New York Times profile of Medvi and a correction the tracker mentions), I relied on the named secondary report or left the point out.
Sources
- Dr Alan D. Thompson, ASI checklist (updated October 26, 2026) (archived snapshot of the version read) and AGI countdown; video ASI: A better world – 2026
- Morris et al., Levels of AGI (Google DeepMind, 2023)
- Genewein et al., From AGI to ASI (Google DeepMind, June 10, 2026)
- OpenAI Charter
- Terence Tao on Erdős problem #728 (January 7, 2026) and AI contributions to Erdős problems (wiki)
- Don Knuth, Claude’s Cycles (February 28, revised April 14, 2026)
- OpenAI, An OpenAI model has disproved a central conjecture in discrete geometry (May 20, 2026) and companion remarks
- OpenAI, Ten advances in mathematics and theoretical computer science (August 1, 2026)
- Anthropic, Yes, Claude can do nine loops (September 25, 2026)
- Anthropic, Claude discovers a novel enzyme system (September 23, 2026)
- CNN, What to know about Anthropic’s “Claude-led” biological discovery (October 8, 2026)
- The Conversation, A biologist queries whether AI “borrowed” his work (October 8, 2026)
- Anthropic, scorpion toxin binder design (October 20, 2026)
- The Verge, Claude did not invent scorpion antivenom (October 22, 2026)
- Two transit-like signals in TESS photometry of TIC 4206066, preprint (September 25, 2026)
- Anthropic, When AI builds itself (June 2026)
- Anthropic, Risk Report, August 2026 (redacted) and Axios (August 14, 2026)
- OpenAI, GPT-5.6 (July 9, 2026) and frontier intelligence efficiency (July 29, 2026)
- Weco AI, First evidence of recursive self-improvement (July 14, 2026)
- Anthropic Engineering, Redesigning source control (October 13, 2026)
- The New York Times, Medvi (April 2, 2026) and Moneywise via Yahoo Finance (April 7, 2026)
- Andon Labs, AI café in Stockholm (2026)
- Posner and Saran, Silicon Formalism: Rules, Standards, and Judge AI (SSRN, January 29, 2026) and The Register (February 15, 2026)
- Figure, Helix 2.5: zero-shot 30-home generalization (September 17, 2026)
- Tesla, Q3 2026 Update (SEC Form 8-K, Exhibit 99.1)
- Nature, China approves first brain implant for use outside clinical trials (March 16, 2026)
- US Bureau of Labor Statistics, unemployment rate (series LNS14000000)
Researched and drafted with AI assistance, including Anthropic’s Claude.
