Two Months Of AI Progress In 2027: Will Compute Keep Up?
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: Two Months Of AI Progress In 2027: Will Compute Keep Up? on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

TL;DR

A November 18, 2026 scenario from Thorsten Meyer argues that AI could accelerate AI development enough to deliver a year’s progress in two months during 2027, before physical compute limits slow the cycle. Public evidence supports faster AI performance on some software tasks, but does not establish that sixfold acceleration, artificial superintelligence, or broad economic change will arrive on that schedule.

Thorsten Meyer published a 2027 AI scenario on November 18 arguing that AI systems could help build successors quickly enough to compress a year of progress into two months, before infrastructure limits slow development. The article presents this as a conditional bet, not an established forecast; available measurements show rapid gains on selected software tasks, but do not confirm the proposed sixfold acceleration or a 2027 arrival of artificial superintelligence.

The scenario links four developments: AI systems do more work on AI research and engineering; that feedback loop speeds progress; available chips, data centers and electricity eventually constrain it; and a later, slower cycle of automated factories and laboratories expands capacity. Meyer calls that physical-stage process “embodied self-improvement.” Each link could fail, and the article does not establish that the full sequence is underway.

The public capability measure at the center of the argument comes from METR, a nonprofit that tests frontier AI agents. Its 80% time horizon estimates the length of a task an agent can complete successfully four times out of five, using human expert completion time as the reference. In the selected record-setting results cited by Meyer, the measure 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 task results, not projections of general intelligence.

Meyer’s own curve-fitting produces very different timelines depending on which measurements are included. An exponential fit using data since 2024 doubles the horizon about every 3.4 months and reaches 16 hours in January 2027. A fit beginning in 2019 doubles about every 6.4 months and reaches that point in June 2028. A hyperbolic fit reaches infinity in November 2027, but Meyer says it performed worst in backtesting, having already projected infinity by January 2, 2026, when fitted to earlier data.

At a glance
analysisWhen: Published November 18, 2026; the 2027 t…
The developmentThorsten Meyer published a scenario examining whether AI-assisted AI development could sharply accelerate progress in 2027 before chips, data centers and power become limiting constraints.

Why Compute Sets the Pace

The scenario matters because it connects AI capability gains to the physical infrastructure needed to train and run systems. If automated research accelerates, but data centers cannot be built or powered quickly enough, software progress may run into a material bottleneck. Meyer cites U.S. data-center construction running at a $96.5 billion annual rate in September 2026, compared with $50.7 billion a year earlier. That increase shows investment activity, not proof that enough usable compute or electricity will be available.

The possible effects on work also need to be separated from benchmark results. The source cites estimates that 18% of U.S. firms used AI in a business function in late 2025 and early 2026, while 2% reported AI-related job cuts. Those figures do not establish future employment effects or show that benchmark improvements translate directly into widespread deployment. For readers, the central issue is whether capability, infrastructure and adoption can advance together—or whether one becomes the limiting factor.

Products Worth Considering

As an affiliate, we earn on qualifying purchases.

What the Benchmarks Measure

METR’s time horizon is a measure of agent performance on tasks, not a countdown to a particular AI milestone. Its task set primarily covers software engineering, machine learning and cybersecurity. METR has cautioned that these tasks are cleaner and more defined than economically valuable work in everyday settings, and that a measured two-hour task is closer to work a new hire or freelancer could complete than to an experienced colleague’s work.

The measure also has limits at longer durations. METR says results above 16 hours are unreliable with its current task suite. Of 31 tasks estimated to take humans at least eight hours, it measured human completion times for only five; the rest are estimates. METR marked its public page as no longer actively updated on September 8, 2026, according to Meyer. That status means the cited series may not provide a continuing public readout of the trend.

There is evidence of AI contributing to AI development, but the scale is disputed. Meyer reports that METR estimates Anthropic is improving models on well-specified tasks at around twice the earlier rate, while saying it would not be surprised by a range of 1.3 to 3 times. A year of progress in two months would imply roughly sixfold acceleration, a substantially stronger claim than that estimate. Meyer also says OpenAI’s GPT-6.2 Astra is its first externally released model to meet the company’s “High” threshold for AI self-improvement; that is a company classification, not independent evidence that the broader scenario is occurring.

“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

Where the 2027 Case Could Break

The largest uncertainty is whether the task benchmark predicts broader AI research or economic capability. METR’s own cautions about task scope, low-context work and long-duration measurement make the 16-hour threshold a weak basis for a precise arrival date. The curve fits also show that changing the starting period shifts the projected date by more than a year, while the most dramatic fit failed in Meyer’s cited backtest.

It is also unclear how much AI systems are currently speeding up frontier model development, how the cited estimate for Anthropic was derived, and whether other labs would report comparable effects. The supplied material does not provide independent validation of OpenAI’s “High” designation for GPT-6.2 Astra or evidence that it has produced a sixfold acceleration. Nor does it establish when additional power and computing capacity will become available; Meyer cites Epoch AI’s view that power may be the first scaling constraint and that new supply can take three to five years after a decision.

Finally, ASI has no agreed definition, and capability progress does not automatically mean deployment across workplaces. The article was researched and drafted with AI assistance, including Anthropic’s Claude, and says Anthropic, OpenAI and METR neither reviewed nor sponsored it. Those disclosures provide relevant context for assessing the analysis; they do not, by themselves, validate or disprove its claims.

Signals to Watch in 2027

The scenario will be easier to test if new, independently described measurements track agents on longer tasks and clarify their real-world relevance. Readers can watch whether METR resumes or replaces its public time-horizon series, whether other evaluators publish comparable results, and whether reported AI contributions to model research are backed by transparent methods rather than company labels alone.

Infrastructure is another test. Data-center construction spending does not show how much capacity is connected, available to AI workloads or supported by dependable power. Updates on electricity supply, construction timelines and compute availability would help show whether physical constraints are tightening or easing. Meyer’s article offers no confirmed milestone showing that this bottleneck has already been reached.

For now, the defensible conclusion is narrower than the headline scenario: AI agents have improved substantially on a particular set of software tasks, and some evidence points to AI helping with AI development. Whether those gains compound fast enough to compress a year into two months in 2027 remains an open question. The next useful evidence will be repeatable capability measurements, clearer reporting on AI-assisted research, and actual—not projected—changes in compute and power capacity.

Key Questions

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

No. Meyer describes a scenario and personal bet, not a confirmed forecast. The source also notes that ASI has no agreed definition.

What does METR’s 80% time horizon measure?

It estimates the length of a task, measured by human expert completion time, that an AI agent can complete successfully four times out of five. The cited tasks mainly concern software engineering, machine learning and cybersecurity.

Does the benchmark establish that AI progress is accelerating sixfold?

No. The benchmark records task performance over time, while the sixfold figure is what Meyer says would be required for a year of progress to arrive in two months. The cited METR estimate for AI-assisted model improvement is around twice the earlier rate, with a possible range of 1.3 to 3 times.

What could limit AI progress even if models improve quickly?

The scenario identifies chips, data-center buildings and electricity as physical constraints. The source says power may be the first limit and that new power supply can take three to five years after a decision, but the timing and scale of any constraint remain uncertain.

What evidence should readers watch next?

Useful signals include updated, comparable evaluations of longer AI tasks; transparent evidence about AI’s contribution to model development; and data on connected compute capacity and power supply. These would help distinguish measured progress from extrapolation.

Source: ThorstenMeyerAI.com

You May Also Like

US lifts curbs on Anthropic’s Fable, Mythos AI models

The US government has lifted restrictions on Anthropic’s Fable and Mythos AI models, allowing broader deployment and use amid ongoing AI regulation debates.

Anthropic’s $100 Million Bet On Training The Next Wave Of AI Engineers

CNBC reports a $100 million Anthropic investment to train AI engineers. Program structure, timing, eligibility and expected outcomes are not yet specified.

Bitcoin Battles Unfold In Live Warzone Visualization

A new web-based visualization transforms real-time Bitcoin trades into a cinematic battlefield, highlighting market dynamics visually.

Price Per 1M Tokens Is Meaningless

Experts confirm that pricing models based solely on 1 million tokens are misleading and do not reflect true value or cost of AI services.