AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: AI In 2027: Two Months Of Progress, With Compute In Question on ThorstenMeyerAI.com

TL;DR

Thorsten Meyer’s November 18 scenario argues that AI could accelerate AI research enough for a year’s progress to arrive in about two months in 2027, before physical compute limits slow the cycle. The supporting evidence is mixed: METR’s software-task measure rose sharply through April 2026 but has not been actively updated since September, and the projected timeline changes substantially with the trend line selected.

Thorsten Meyer published a scenario for 2027 in which AI systems help build more capable successors quickly enough that a year of progress could arrive in roughly two months, before limits on chips, data-centre capacity and electricity slow development. The article, dated November 18, 2026, presents this as a conditional argument, not a forecast established by current measurements.

The scenario depends on a feedback loop: AI systems take on more coding, debugging and experimentation used to develop later systems, which could speed the next round of work. Meyer says the loop would eventually meet a physical constraint: available compute. He then proposes a slower phase of improvement through automation in factories and laboratories, which he calls embodied self-improvement. Each link is a condition in the scenario, not a confirmed outcome.

The clearest public measure cited is METR’s 80% time horizon: the software-task duration at which an AI agent is predicted to succeed four times in five. METR’s recorded results rose from 0.89 minutes for GPT-4 in March 2023 to 185.9 minutes for an early Claude Mythos Preview in April 2026. METR marked its public page as no longer actively updated on September 8, 2026, leaving the cited series without a newer public endpoint in this account.

Meyer’s calculations produce very different timelines depending on which measurements are included. A fit using data since 2024 yields a doubling time of about 3.4 months and reaches a 16-hour horizon in January 2027; a fit starting in 2019 doubles about every 6.4 months and reaches that mark in June 2028. These are the author’s fits to METR points, not METR forecasts. A hyperbolic fit reaches an infinite value in November 2027, but Meyer says it performed worst in his backtests and had already predicted an infinite value by January 2, 2026, when fitted to models released before July 2025.

At a glance
analysisWhen: Published November 18, 2026; describes…
The developmentThorsten Meyer published a scenario for 2027 in which AI-assisted AI development accelerates progress before compute constraints become a bottleneck.

Why a Faster AI Cycle Matters

If AI systems materially speed up the research and engineering that produce later systems, development could become less dependent on the pace of human-led work. That possibility matters to companies, workers and policymakers because a rapid capability shift could affect the timing of investment, workplace adoption and infrastructure demand. The scenario’s proposed two-month compression, however, is much faster than the acceleration described in the cited METR analysis.

Meyer reports that METR estimates Anthropic is improving its models on well-specified tasks at around twice its earlier rate, while saying it would not be surprised by a rate between 1.3 and 3 times as fast. The scenario’s claim that a year’s progress arrives in two months implies roughly six times the old pace. The gap matters: evidence of some acceleration does not by itself establish the much larger rate needed for the 2027 timeline.

There is also a difference between technical capability and broad economic use. The source reports that 18% of US firms used AI in a business function in late 2025 and early 2026, while 2% reported AI-related job cuts. Those figures suggest adoption and employment effects should be tracked separately from benchmark performance; they do not show that more capable systems will automatically be deployed widely or quickly.

Vitamix Nano Personal Blender with SmoothiePilot Technology, Black

Vitamix Nano Personal Blender with SmoothiePilot Technology, Black

As an affiliate, we earn on qualifying purchases.

What the Benchmarks Can Show

METR describes its time horizon as the task duration, based on human expert completion time, at which an agent is predicted to succeed at a specified reliability. The 80% measure is not a general score of intelligence or a direct measure of economic productivity. METR’s tasks primarily involve software engineering, machine learning and cybersecurity, and the organization says they are cleaner and more defined than real economically valuable work.

The measure also has limits at longer durations. METR says its current task suite cannot reliably measure horizons above 16 hours. Of 31 tasks estimated to take humans at least eight hours, the source says METR measured the human completion time for only five; the remaining times are estimates. The threshold is therefore a useful way to track performance on a particular class of tasks, but it cannot alone establish that AI can complete a full workday of varied, real-world responsibilities.

The source also points to compute investment as a possible counterweight to faster software progress. It reports US data-centre construction running at an annual rate of $96.5 billion in September 2026, compared with $50.7 billion a year earlier. It cites Epoch AI’s view that power may be the first constraint on scaling, with new power supply taking three to five years from a decision. These figures frame a potential infrastructure bottleneck; they do not show exactly when it would bind or how much capacity future AI systems will require.

„We reach ASI either in 2027 or shortly after. The effect: in two months we see the progress of a whole year, and then further progress until we hit a ceiling in available compute.“

— Thorsten Meyer, in his November 18, 2026 article

The Missing Measurements and Bottlenecks

There is no agreed definition of ASI, and the source does not use the label as a testable benchmark. More directly, the latest METR result cited is from April 2026, while its public page was marked no longer actively updated in September. The material does not provide later comparable measurements that could confirm whether the observed pace continued, slowed or changed.

The 2027 date depends heavily on choosing the faster, more recent trend fit. The longer historical fit points to mid-2028 for the same 16-hour benchmark, and the hyperbolic extrapolation failed Meyer’s cited backtest. None of these curve fits proves what future performance will be. The source does not provide enough information to independently assess every modeling choice behind the fits.

Other parts of the chain remain open as well: how much AI contributes to the development of its successors, whether reported acceleration transfers beyond well-specified tasks, when electricity or chip supply would constrain growth, and whether improved systems would be affordable and reliable enough for widespread use. Meyer discloses that the article was researched and drafted with AI assistance, including Anthropic’s Claude, and says Anthropic, OpenAI and METR neither reviewed nor sponsored it.

Signals That Could Test the Scenario

The next useful evidence would be updated, comparable task measurements that extend beyond the April 2026 METR result and explain any changes in task mix or evaluation methods. A sustained rise across independent measures would offer stronger support for faster software-task performance than a single extrapolated curve; a flattening or slower rise would weaken the proposed compression.

Readers can also watch for concrete evidence that AI systems are contributing more of the coding and research work behind later models, alongside disclosures about model-development timelines. On the infrastructure side, data-centre construction, grid connections, power availability and chip supply will help show whether physical capacity is becoming a binding constraint. Those indicators would test separate parts of the scenario, rather than confirm it all at once.

For now, Meyer’s article is best read as a scenario to monitor, not a settled 2027 prediction. The timing, the scale of any acceleration and the route from stronger software agents to broader workplace effects remain unresolved.

Key Questions

Is the article predicting that ASI will arrive in 2027?

No. Meyer presents a possible scenario for 2027 or shortly after, not a forecast established as fact. The article says ASI has no agreed definition and tests the proposed acceleration mechanism instead.

What does METR’s 80% time horizon measure?

It estimates the duration of software tasks, measured against human expert completion time, at which an AI agent is predicted to succeed four times in five. The cited tasks are mainly software engineering, machine learning and cybersecurity, not a broad measure of job performance.

Why do the projections give different dates?

The calculations use different portions of METR’s historical measurements. Meyer’s fit from 2024 doubles about every 3.4 months, while a fit from 2019 doubles about every 6.4 months. The resulting dates are extrapolations, not METR forecasts.

What could limit AI progress even if models improve quickly?

The scenario identifies chips, data-centre buildings and electricity as possible constraints. The source cites an estimate that new power supply can take three to five years from a decision, but it does not establish when or whether those limits will halt AI scaling.

Source: ThorstenMeyerAI.com

This content is for general information only and is not financial, tax or legal advice. Consult a qualified professional for decisions about your money.
You May Also Like

Introducing Forezai · TradingAgents — a committee of LLMs decides paper-trades

Forezai · TradingAgents introduces a multi-LLM system that autonomously runs paper-trades using a structured agent framework, advancing AI-driven trading research.

AI Tools Face Exposure From Stolen Sessions And Leaked Logs

Three October reports describe AI session theft, exposed agent traces and a flaw researchers say exposed hidden model reasoning.

Analyzing NTSB’s Investigation Into Miami B-767 Runway Excursion

The NTSB has released an investigative update on a runway excursion involving a B-767 at Miami International Airport, highlighting safety concerns.

Phone-based injury-risk movement screening for hiring

A new phone-based movement screening tool for industrial hiring is being tested, aiming to reduce injury risks and costs through remote assessment.