🔍 Read the full analysis: How Much Progress Has AI Made Toward AGI And ASI? 50 Signs Checked on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
An October 26 review of 18 prominent claims from researcher Alan D. Thompson’s 50-item ASI tracker classified seven as verified, five as vendor claims, three as based on a single source and three as disputed or contradicted. Thompson marks two indicators achieved and 33 partially achieved, but his broad definition of ASI and low threshold for partial credit do not establish that superintelligence has arrived.
A review published October 26 of 18 prominent claims in researcher Alan D. Thompson’s 50-sign artificial superintelligence tracker found seven verified, five supported only by company claims, three resting on a single source and three disputed or contradicted. Thompson’s tracker labels two of its 50 indicators achieved and 33 partially achieved, but the review says those statuses are not a measurement that superintelligence has arrived.
Thompson, an Australian AI consultant who publishes The Memo newsletter, describes artificial general intelligence as a machine performing at the level of an average human and artificial superintelligence as a machine performing at the level of an expert human in practically any field. He considers Claude Mythos Preview and GPT-6 Astra, which he identifies as 2026 models, examples of “early ASI.” Those are Thompson’s judgments, not a shared scientific finding.
The tracker’s categories require care. In an August correction note, Thompson said “partially achieved” includes documented progress such as funded programs, research infrastructure, formal scientific missions and enabling work. The review says that standard can place work far from the stated outcome in the same category as direct progress. It also notes a separate countdown by Thompson that puts progress toward AGI at 99%, with the remaining test involving a robot completing multi-step fine assembly at average-human reliability.
The strongest independently checked examples in the review involve mathematics and science. It reports that mathematicians documented AI contributions to open problems, though the cases included coaching, human-written proofs, or qualifications about how a problem was stated. Other claims did not have the same level of independent confirmation. The review describes an OpenAI claim about GPT-6 Astra producing ten formalized results on longstanding problems for roughly $2,000 as a vendor claim, even though Lean certificates were made public; the source material does not provide enough of the full claim-by-claim table to report every verdict or its supporting detail.
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
Why the Checklist Needs Scrutiny
The tracker brings attention to developments worth monitoring, but its headline totals can be mistaken for evidence that a widely accepted threshold has been crossed. Partial credit is especially broad: a funded project or enabling technology may be relevant to a future capability without showing that the capability exists now. Readers comparing totals should distinguish research activity from demonstrated performance.
The review also finds that the most compelling evidence tends to be in areas where results can be checked, including formal mathematics, a validated physics calculation and regulatory filings. By contrast, claims about broad autonomy, economic disruption or machines operating in the physical world require evidence beyond a model’s benchmark score or its developer’s announcement. The review says there is no visible break in the US unemployment rate, no household robots operating at scale, and no verifiably billion-dollar AI-run company among the developments it examined. Those observations do not settle what AI may do next, but they limit what can be claimed about its present effects.
For readers, the central issue is not whether models are improving; the cited examples show progress in specific tasks. It is whether that progress justifies labels such as AGI or ASI, which depend on definitions, reliability and autonomy. A system that can help solve a difficult proof is evidence of a useful capability, not by itself proof of expert performance across practically any field.
Popular right now
TCL 85 Designer Series QD-Mini LED TV (Model 85A400U) with Google Smart TV
Breville Bambino Espresso Machine, Brushed Stainless Steel
De'Longhi Dinamica Plus Fully Automatic Espresso Machine
As an affiliate, we earn on qualifying purchases.
Different Scales, Different Labels
There is no single definition of AGI or ASI in the source material. Thompson defines ASI around expert-human performance across practically any field. Google DeepMind’s 2023 “Levels of AGI” framework instead separates levels such as Competent, Expert, Virtuoso and Superhuman. Under that scale, Expert corresponds to performance at the 90th percentile of skilled adults, while Superhuman means outperforming all humans. The review argues that Thompson’s ASI is closer to DeepMind’s “Expert AGI” than to its later description of superintelligence.
In a June 2026 report, Google DeepMind described superintelligence as a system “more intelligent and cognitively capable than large organisations of humans.” OpenAI’s charter defines AGI in relation to highly autonomous systems that outperform humans at most economically valuable work. These definitions put different weight on breadth, autonomy and comparison with human abilities. So a claim that a model is “early ASI” cannot be assessed without asking which definition and evidence standard are being used.
The tracker is best read as one researcher’s selection of possible indicators, not a standardized scientific instrument. Its 50 items range from concrete technical milestones to distant outcomes. A status label can show how Thompson interprets a development, but it does not independently establish that the corresponding capability has been achieved.
“documented material progress toward an indicator, including funded programs, research infrastructure, formal scientific missions, and enabling work”
— Alan D. Thompson, in an August correction note about the tracker
What the Evidence Does Not Establish
The review’s figures cover 18 selected claims, not all 50 tracker indicators, and the source material does not identify the full evidence behind every verdict. Seven claims were classed as verified, but that does not mean seven indicators of ASI were achieved: verification of an individual result is different from proof of broad, general capability. The full basis for the remaining claims, and how the review’s selection affects the overall picture, is not detailed in the supplied material.
It is also unclear how well results from narrow tasks generalize to unfamiliar fields or sustained real-world work. The cited mathematical examples include human guidance or proof-writing, while some headline results rely on companies’ own reports. The source material does not establish that current models can independently perform expert work across practically any field, run organizations, or improve their own research capabilities at a scale that changes development speed.
On self-improvement, the material cites an internal Anthropic assessment: the company’s best model scored 62.8% on a test, while Anthropic said a full replacement for its researchers would require at least 85%. It also estimated that AI makes its research and development faster, “but not yet by a factor of 2.” These are company-reported figures; the supplied account does not describe independent replication or the test’s full methodology. They indicate a gap under Anthropic’s stated standard, not a general measure of every lab’s progress.
Evidence to Watch in Future Updates
The next useful test is whether current claims can be independently reproduced and whether systems perform reliably beyond curated examples. For mathematics, that means examining not just a correct answer or formal certificate but also the role of human prompting, problem selection and proof construction. For scientific and economic claims, readers will need source documents and methods that allow outside experts to check the result.
Physical-world capability remains another unresolved benchmark in the material. Thompson’s AGI countdown waits for a robot to complete multi-step fine assembly, such as putting together flat-pack furniture, at average-human reliability. The source does not give a date for that milestone. Until it is demonstrated, the countdown is a forecast framework rather than confirmation that general-purpose robotics has reached that level.
Future tracker revisions may change the status of individual signs, while new lab reports and independent evaluations may strengthen or weaken specific claims. For now, the review’s clearest conclusion is limited: there is credible progress in some tightly defined tasks, but the checked evidence does not establish that ASI has arrived. Whether the next developments meet a shared standard for broad, autonomous capability remains open.
Key Questions
How many ASI tracker indicators does Thompson mark as achieved?
His tracker listed two achieved, 33 partially achieved and 15 not achieved as of October 26, 2026. These are Thompson’s classifications, not an independent scientific consensus.
What did the review find when it checked 18 claims?
It classified seven as verified, five as vendor claims, three as based on a single source and three as disputed or contradicted. The review examined selected claims, not every item on the 50-sign list.
Does “partially achieved” mean a sign is nearly complete?
No. Thompson says the category includes funded programs, research infrastructure, formal missions and other enabling work. A partial status can reflect progress toward a goal without showing that the goal itself has been met.
Has the review established that AI has reached ASI?
No. It found evidence for specific advances, including mathematical contributions, but says the tracker’s definitions and partial-credit rules do not establish that AI can perform expert work across practically any field.
What major capabilities remain unconfirmed in the source material?
The account says household robots are not operating at scale, it found no verifiably billion-dollar AI-run company, and it reports no visible break in the US unemployment rate. It also describes limits in an Anthropic-reported test of AI-assisted research. These points do not prove what future systems will or will not achieve.
Source: ThorstenMeyerAI.com
