Will AI’s Two-Month Burst Of Progress Run Into Compute Limits?
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: Will AI’s Two-Month Burst Of Progress Run Into Compute Limits? on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the latest gadgets delivered free with Prime

  • Fast, free delivery on millions of items
  • Prime Video, Amazon Music and more included
  • Member-only deals all year
Start your free Prime trial Free trial for eligible customers · Cancel anytime
As an affiliate, we earn on qualifying purchases.

TL;DR

Thorsten Meyer’s November 18 scenario argues that AI systems could accelerate AI development enough to deliver a year’s progress in about two months by 2027, before physical compute limits slow the cycle. METR data show AI agents completing longer software tasks, but the trend depends heavily on how it is fitted, and its public time-horizon data are no longer actively updated. The 2027 outcome is a scenario, not an established forecast.

Thorsten Meyer published a scenario on November 18 arguing that AI could make roughly a year’s worth of progress in two months by 2027, as systems increasingly help build their successors. The article presents this as a conditional possibility, not a forecast established by current data, and says the acceleration could eventually run into limits on chips, data-centre construction and electricity.

The scenario has four links: AI systems take on more of the coding and experimentation used to develop later systems; that feedback loop speeds up; available computing capacity becomes a constraint; and, over a longer period, automation in factories and laboratories helps expand capacity. Meyer calls this later physical feedback loop “embodied self-improvement.” Each link could fail, the article says, and the proposed 2027 timing depends on all of them moving quickly enough.

The public measure at the center of the analysis is METR’s 80% time horizon: the duration of a software task, measured by human expert completion time, at which an AI agent is predicted to succeed four times out of five. Selected record-setting measurements in METR’s results file rose from 0.89 minutes for GPT-4 in March 2023 to 185.9 minutes—about three hours and six minutes—for an early Claude Mythos Preview in April 2026. These are measured results on METR’s task suite, not projections of general work capability.

The curve changes markedly depending on which historical data are used. Meyer’s analysis says an exponential fit starting in 2024 doubles the measured horizon every 3.4 months and reaches 16 hours in January 2027. A fit using data from 2019 doubles every 6.4 months and reaches that level in June 2028. METR says measurements above 16 hours are unreliable with its current task suite, so neither date should be treated as a confirmed milestone.

At a glance
analysisWhen: Published November 18, 2026; the scenar…
The developmentThorsten Meyer published a scenario on November 18, 2026, examining whether rapid AI self-improvement could compress a year of progress into two months by 2027 before compute constraints intervene.

Why Compute Could Set the Pace

If AI systems materially speed up the development of later systems, progress could arrive in shorter intervals than the annual cycle many companies and policymakers use to plan. But the scenario’s proposed acceleration is far greater than the evidence cited for current improvements in AI-assisted model development. The article reports that METR estimates Anthropic is improving models on well-specified tasks at about twice its earlier rate, while saying it would not be surprised by a rate between 1.3 and 3 times as high. A year of progress in two months would imply about six times the earlier pace.

The physical constraints matter because more capable software does not itself create additional chips, buildings or electricity. The source cites 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 also reports that Epoch AI expects power to be the first constraint on scaling, with new supply taking three to five years from a decision. Those figures point to investment and timing challenges; they do not establish when a compute ceiling will be reached.

Even faster technical progress would not automatically mean rapid changes across the economy. The source cites estimates 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 distinguish capability from adoption: the speed at which systems improve and the speed at which businesses deploy them are separate questions.

KitchenAid Artisan 5-Quart Tilt-Head Stand Mixer, Almond Cream

KitchenAid Artisan 5-Quart Tilt-Head Stand Mixer, Almond Cream

As an affiliate, we earn on qualifying purchases.

What METR’s Measure Captures

METR is a nonprofit that evaluates frontier AI agents. Its time horizon estimates how long a task can be before an agent’s predicted success rate falls below a selected reliability level. The 80% measure used here means success on four out of five attempts at the relevant task duration; it is not a measure of how long a model can work autonomously on every kind of job.

METR’s task suite primarily covers software engineering, machine learning and cybersecurity. The organization cautions that these tasks are cleaner than economically valuable work in real settings. It also says a two-hour task is closer to work a new hire or freelance contractor could complete in two hours than to the work of an experienced colleague. The results therefore describe performance on a specific class of tasks, not a general measure of AI’s economic impact.

There are measurement limits at the high end as well. METR says its current suite cannot reliably measure horizons above 16 hours. Among 31 tasks that take humans at least eight hours, it measured human completion times for five and estimated the rest. Its public page was marked no longer actively updated on September 8, 2026, according to Meyer’s article. That makes the dataset useful for examining a recent trend, but not a continuously refreshed basis for tracking every subsequent model.

“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

The 2027 Curve Is Unsettled

The central uncertainty is whether the measured software-task trend can support the much faster pace the scenario requires. The fit starting in 2024 reaches its 16-hour threshold roughly 17 months earlier than the fit starting in 2019. Both are extrapolations, and the more dramatic curve is not automatically the more reliable one. Meyer says his hyperbolic fit performed worst in backtesting: fitted to models released before July 2025, it had already run to infinity by January 2, 2026.

The measurements also stop short of showing that AI systems are autonomously accelerating the full process of AI research. They concern success on a defined set of tasks, and METR’s reported estimate of faster AI-assisted improvement is about two times the earlier rate—not the sixfold acceleration implied by progress compressed into two months. The source does not establish that the gap will close.

Other unknowns include how quickly new computing capacity can be connected to power, whether data-centre construction can keep pace with demand, and whether advances in software translate into broad business use. The article’s figures do not resolve those questions. It also sets aside disagreement over the meaning of artificial superintelligence, noting that there is no agreed definition of the label.

Meyer discloses that the article was researched and drafted with AI assistance, including Anthropic’s Claude. He says Anthropic, OpenAI and METR did not review or sponsor it. The analysis should be read as an attributed scenario and interpretation of cited evidence, not as an assessment endorsed by those organizations.

Signals to Watch Through 2027

The scenario will be easier to evaluate if future reporting provides updated, comparable measurements of agent performance, with methods and task coverage made clear. In particular, results that extend beyond METR’s current 16-hour reliability limit would help establish whether longer-horizon estimates are measurable rather than extrapolated. The source notes that METR’s public page was no longer actively updated as of September 8, 2026, but does not specify when or whether a replacement series will appear.

Readers can also track evidence of the proposed software feedback loop: documented changes in how much AI systems contribute to developing, testing and improving later systems, alongside estimates of the effect on development time. A stronger case for the two-month claim would require evidence that clearly distinguishes this pace from the roughly twofold improvement attributed to METR in the source.

On the physical side, relevant developments include data-centre construction, access to electricity and the time required to bring new supply online. Adoption data will show whether technical improvements spread into business operations. Until those indicators are clearer, 2027 remains the scenario’s test period, not a settled deadline for superintelligence or a guaranteed compute bottleneck.

Key Questions

Is the article predicting that artificial superintelligence will arrive in 2027?

No. Meyer describes 2027 or shortly after as his scenario and bet, not a forecast confirmed by the evidence. The article examines what would need to happen and what evidence could challenge the timeline.

What does METR’s 80% time horizon measure?

It estimates the duration of a task, based on human expert completion time, at which an AI agent is predicted to succeed four times out of five. METR’s suite mainly covers software engineering, machine learning and cybersecurity tasks.

Does the rise from 0.89 minutes to 185.9 minutes prove that AI progress will accelerate sixfold?

No. Those are measured task-horizon results for selected models on METR’s suite. The sixfold figure refers to the pace implied by the scenario of compressing a year of progress into two months; the measurements do not establish that this acceleration has happened.

What could limit AI progress even if models keep improving?

The scenario points to chips, data-centre buildings and electricity. The source cites an estimate that new power supply can take three to five years from a decision, but it does not establish when any of these constraints will halt progress.

What evidence should readers watch for next?

Updated and comparable agent measurements, evidence of AI systems speeding up AI development, and data on compute capacity, power availability and business adoption would help test the scenario. METR’s public page was marked no longer actively updated in September 2026, and the source does not give a date for a new dataset.

Source: ThorstenMeyerAI.com

You May Also Like

Can An AI Read Another AI’s Inner Workings? We Tried Weight Space Programming

MIT researchers tested AI systems that read and edit neural-network weights. A separate three-part test found one narrow improvement and important limits.

Llama.cpp V1.0.0 Builds A Stronger Foundation For Local AI

The ggml-org team has released llama.cpp and ggml v1.0.0, setting semantic-versioning rules for public API changes and outlining planned distribution channels.

What The UK AISI Cybertesting Incident Reveals About AI Alignment

A headline identifies an Anthropic alignment assessment of a UK AISI cybertesting incident, but the incident and findings are not available.

“Women Love Video Games,” Dressmaker Dev Says, There Just Aren’t Enough Games For Them: “We Were Definitely Targeting Women From The Get-go”

Cozy Lives says it designed Dressmaker for women. The Steam hit has drawn a large audience, and its developers plan further updates and ports.