🔍 Read the full analysis: Six Measures Of Whether AI Is Gaining Ground In 2027 on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
ThorstenMeyerAI.com published a watchlist on Nov. 18, 2026, for measuring whether a proposed 2027 AI self-improvement scenario is unfolding. Its six signals cover agent capability, AI use in AI research, computing infrastructure, electricity, robotics and employment; the author stresses that the thresholds are judgments, not forecasts, and that several readings lack scheduled release dates.
ThorstenMeyerAI.com published a six-signal tracker on Nov. 18 for assessing whether a scenario of faster AI-driven AI development is taking shape in 2027. The tracker sets proposed thresholds for agent capabilities, AI participation in research, data-centre construction, electricity supply, industrial robots and employment, while describing them as the author’s judgments rather than official benchmarks or a forecast.
The tracker follows a separate scenario article, titled “Scenario 2027: A Year of AI Progress in Two Months, Until Compute Runs Out?” That article proposes that AI systems could increasingly perform work involved in building later AI systems, accelerating progress until physical limits such as computing capacity, buildings and power constrain it. The new piece focuses on how to monitor that idea, rather than making the case for it again.
One of the central capability measures is METR’s 80% time horizon: the duration of a software task, measured by the time a skilled human needs, that an AI agent is predicted to complete reliably. The last public record cited is 185.9 minutes for an early Claude Mythos Preview in April 2026. The tracker says the scenario gains support if a new task suite measures a horizon of at least 16 hours by mid-2027, but loses support if new suites continue to show only a few hours. METR has cautioned that its current task suite cannot reliably measure beyond 16 hours.
A second signal tracks AI’s role in AI research and formal self-improvement assessments. Anthropic reported that Claude “leads” 26% of its AI research and development work as of August 2026, but the source says no measured part of that work ran fully autonomously. OpenAI’s GPT-6.2 Astra is described in the tracker as its first released model to reach the “High” threshold for AI self-improvement. The remaining indicators examine compute infrastructure, power generation, factory robots and AI-related job cuts.
The Bottlenecks Behind AI Acceleration
The tracker matters because it separates a broad claim about rapid AI progress into measures that can be checked. If AI systems help develop later systems, capability gains could affect research timelines; whether that translates into sustained deployment also depends on infrastructure and labor-market changes. The measures are intended to show whether progress is converging across those areas, rather than relying on a single striking model result.
The figures cited point to different speeds and constraints. The source reports US data-centre construction at an annual rate of $96.5 billion in September 2026, up from $50.7 billion a year earlier. It also cites an expectation that US power generation will grow 2.7% in 2026 and 1.2% in 2027. Those indicators do not establish that AI demand can be met: construction spending and projected generation are not direct measures of available power for AI data centres.
Adoption outside labs appears slower in the tracker’s cited indicators. Factories installed 603,000 robots in 2025, with a 2026 forecast of 655,000, while 2% of US firms reported AI-related job cuts. These numbers provide context, but neither robots nor reported job cuts on their own measure how much work AI systems perform. The tracker says no single reading decides the scenario; the direction of several measures matters.
Products Worth Considering

Logitech MX Keys S Combo Wireless Keyboard and Mouse, Graphite

Brother P-touch EDGE PT-E310BT Industrial Label Printer

Dell 34 Plus USB-C Curved Monitor, S3425DW, 34-Inch WQHD VA, 120Hz
As an affiliate, we earn on qualifying purchases.
How the Six Signals Are Set
The six measures come from the scenario article that Thorsten Meyer published earlier on Nov. 18, 2026. The tracker identifies what to measure, the relevant publisher, the latest reading cited and a threshold at which the scenario would gain or lose support. Its thresholds are the author’s own judgment calls, fixed before future readings arrive; they are not standards set by METR, Anthropic, OpenAI or government agencies.
The six areas are longer AI task horizons; AI’s share of AI research and formal self-improvement assessments; computing capacity; electricity; industrial robotics; and labor-market effects. The source says that, for computing and labor, no publisher defines a specific threshold, so the tracker will follow the direction of the data. It also discloses that the article was researched and drafted with AI assistance, including Anthropic’s Claude, and that Anthropic, OpenAI and METR did not review or sponsor it, despite evidence from those organizations featuring in the analysis.
The timing of updates is uneven. The source lists Dec. 1, 2026, for the next construction figure and Dec. 8 for the next power figures. It gives no calendar for new AI-specific readings from METR or the labs, which publish when they choose. METR’s public time-horizon page was described as no longer actively updated, with its last change dated Sept. 8, 2026.
“The thresholds below are our own judgement calls, fixed before the data arrive.”
— Thorsten Meyer, ThorstenMeyerAI.com
Gaps in the Current Measurements
Several parts of the proposed test remain difficult to assess. METR’s public page has no scheduled next update, and the tracker says its current task suite cannot reliably measure horizons beyond 16 hours. A higher score could also be distorted by how attempts to cheat task evaluations are classified; the source cites sharply different results under different scoring assumptions and says those figures are not robust measurements.
There are also no fixed release calendars for the lab-specific measures. Anthropic and OpenAI have not stated when they will publish further readings, and the tracker notes that no publisher sets a formal threshold for the compute and labor series. The source material provided ends during its discussion of Anthropic’s automation scale, so it does not give the full current breakdown of all six signals or every threshold.
The cited construction, power, robotics and employment figures do not by themselves prove that AI is accelerating its own development. Some are forecasts, some refer to broad economic activity, and their definitions and coverage differ. How well they map to AI-specific capacity or work remains uncertain.
Scheduled Data and Lab Updates
The tracker identifies the next listed releases as the US data-centre construction figure on Dec. 1, 2026, and power figures on Dec. 8. Those dates offer the first scheduled checks after publication. The source does not give release dates for new METR task-horizon results or updates from Anthropic and OpenAI.
Meyer says the tracker will be updated with dated notes as data arrive. Readers will be able to compare those readings with the stated thresholds, while keeping in mind that the thresholds are the author’s judgments and that no single signal is meant to settle the scenario. A meaningful assessment will depend on whether multiple indicators move in the same direction and whether the underlying measures can support that interpretation.
Key Questions
What is the 2027 scenario the tracker is testing?
It is a scenario in which AI systems take on more of the work of developing future AI systems, potentially speeding progress until limits in computing capacity, buildings and power constrain it. The tracker measures indicators that could support or weaken that scenario; it does not present it as a prediction.
What does METR’s 80% time horizon measure?
It estimates the duration of a software task, measured by skilled-human completion time, that an AI agent is predicted to complete four times out of five. The tracker cites 185.9 minutes as the last public record, for an early Claude Mythos Preview in April 2026.
Does Anthropic’s 26% figure mean Claude works independently?
No. In the tracker, “leads” refers to AI completing most of a task end to end while a human supervises. The source says no measured part of Anthropic’s AI research and development work ran fully autonomously.
When are the next measurements due?
The tracker lists Dec. 1, 2026, for the next data-centre construction figure and Dec. 8 for power figures. It gives no scheduled date for the next METR, Anthropic or OpenAI AI-specific readings.
Are the tracker’s thresholds official benchmarks?
No. Thorsten Meyer describes them as his own judgment calls, fixed before future data arrive. The tracker says no single reading decides the scenario and that some areas, including compute and labor, have no publisher-defined threshold.
Source: ThorstenMeyerAI.com
