A Year Of AI Progress In Two Months: The 2027 Compute Question
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: A Year Of AI Progress In Two Months: The 2027 Compute Question on ThorstenMeyerAI.com

Age 18–24?Offer from Amazon

Prime made for students and young adults

  • Fast, free delivery for dorm and study essentials
  • Prime Video and Amazon Music included
  • Member-only deals
Try Prime for Young Adults Free trial for eligible 18–24 year olds
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 for a year of progress to arrive in two months in 2027, before compute limits slow the cycle. METR’s public software-task measurements show rapid gains, but its series is no longer actively updated, its task suite has limits, and the cited evidence does not establish a sixfold acceleration. The scenario remains a hypothesis, not a forecast confirmed by the data.

Thorsten Meyer published a 2027 AI scenario on November 18 arguing that AI systems could help develop their successors quickly enough to compress a year of progress into about two months, until physical limits on computing capacity slow the cycle. The article presents this as a hypothesis to test, not a confirmed forecast; the available measurements show progress in selected software tasks, but do not establish the proposed sixfold acceleration.

The scenario rests on a feedback loop: AI systems contribute more to coding, debugging and experiments used to build future systems, which could in turn make the next development cycle faster, as explored in Anthropic’s compute constraints. Meyer’s proposed sequence then reaches a compute ceiling—the supply of chips, data-centre space and electricity—before a slower cycle of physical expansion allows further growth. He calls that later stage “embodied self-improvement,” involving automation in factories and laboratories.

The article uses METR’s 80% time horizon as a public measure of agent capability. It estimates how long a task, judged by human completion time, an AI agent can complete with four-in-five reliability. Selected record-setting results rose from 0.89 minutes for GPT-4 in March 2023 to 185.9 minutes, or about three hours, for an early Claude Mythos Preview result in April 2026. These are observations on METR’s task suite, not projections of general workplace performance.

Meyer’s calculations produce different timelines depending on which measurements are fitted. An exponential fit using data from 2024 doubles about every 3.4 months and reaches a 16-hour horizon in January 2027; a fit from 2019 doubles about every 6.4 months and reaches that point in June 2028. The article says the more dramatic hyperbolic curve performed worst in its backtests. METR marked its public results page as no longer actively updated on September 8, 2026, adding uncertainty to any extrapolation.

At a glance
analysisWhen: Published November 18, 2026; the propos…
The developmentThorsten Meyer published a scenario examining whether AI-driven research could compress a year of progress into two months by 2027 before hardware, power and construction constraints intervene.

The Gap Between Capability and Deployment

If AI-assisted research substantially speeds the creation of more capable systems, the development cycle could become harder to track using annual milestones. But the article’s evidence concerns a narrow class of software tasks; it does not show that AI can autonomously carry out all research, training and deployment work needed to produce the proposed acceleration.

The distinction matters beyond labs. Meyer cites figures 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 capability and broad economic adoption are separate questions. Even rapid gains on a benchmark would not by themselves establish that businesses can deploy systems safely, affordably or at scale.

Physical infrastructure could also shape the pace. The article cites a September 2026 annualized rate of $96.5 billion for US data-centre construction, up from $50.7 billion a year earlier, and reports Epoch AI’s view that power may become the first scaling constraint. A bottleneck in electricity, chips or construction could limit how quickly research improvements translate into usable compute.

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

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

Bose QuietComfort Wireless Noise-Cancelling Headphones, Twilight Blue

Bose QuietComfort Wireless Noise-Cancelling Headphones, Twilight Blue

KitchenAid Artisan Series 5-Quart Tilt-Head Stand Mixer, Juniper (KSM150PS)

KitchenAid Artisan Series 5-Quart Tilt-Head Stand Mixer, Juniper (KSM150PS)

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 the task duration, measured by human expert completion time, at which an agent reaches a specified reliability level. The 80% figure is not a measure of all-purpose intelligence, nor does a three-hour result mean an agent can reliably perform any three-hour job.

METR describes its tasks as primarily involving software engineering, machine learning and cybersecurity, and says they are cleaner than economically valuable work in real settings. The source article also notes that METR considers its task suite unreliable above 16 hours. Of 31 tasks estimated to take humans at least eight hours, it says human completion times were directly measured for only five. These limits make the benchmark useful for tracking a defined capability, but insufficient on its own to settle a broad 2027 forecast.

Meyer also cites a METR estimate that Anthropic is improving models on well-specified tasks at about twice its earlier rate, while METR said it would not be surprised by a range of 1.3 to 3 times. That estimate is not the same as the scenario’s required sixfold acceleration. The article says OpenAI’s GPT-6.2 Astra met the company’s “High” threshold for AI self-improvement, but that designation is a company assessment and does not independently establish the speed or scope of a recursive development loop.

“Measurements above 16 hrs are unreliable with our current task suite.”

— METR, as quoted in Meyer’s article

The Sixfold Acceleration Test

The key unresolved issue is whether AI systems will accelerate AI research enough to produce the scenario’s sixfold pace of progress. The measurements summarized in the article do not demonstrate that rate. Their fitted timelines vary depending on the starting year, and the fastest curve performed poorly in the article’s backtests.

It is also unclear how well software-task gains transfer to broader research or economically valuable work. The source does not establish how much of the cited model-development work is currently done by AI, how the OpenAI self-improvement threshold was assessed, or whether its results are comparable with METR’s measures. The eventual limits from electricity, chips and data-centre construction—and the timing of any response—remain uncertain as well.

The article discloses that it was researched and drafted with AI assistance, including Anthropic’s Claude. It says Anthropic, OpenAI and METR did not review or sponsor it. Its figures and interpretations should be read as the author’s scenario analysis, with claims attributed to the organizations where stated.

Signals to Watch Through 2027

Meyer frames the scenario as testable through future measurements: whether agents sustain gains on comparable long-horizon tasks, whether AI contributes more directly to building successor systems, and whether independently reported evidence shows development cycles shortening toward the proposed two-month pace. The end of active updates to METR’s public page means readers will need to check for new or alternative evaluations before treating its historical curve as a current measure.

Infrastructure indicators provide a separate test. Data-centre construction, chip availability and power connections could show whether added capability can be matched by more compute. The article says Epoch AI expects new power supply to take three to five years from the decision; that estimate points to a possible lag, not a guarantee that power will be the first or binding constraint in every case.

For now, the proposed 2027 acceleration remains a scenario. Evidence of faster AI-assisted development would strengthen one link in the argument, while slower benchmark gains, poor transfer to real work or persistent infrastructure limits would weaken it. The source material does not give a confirmed next milestone or establish when the forecast could be resolved.

Key Questions

Is a year of AI progress in two months confirmed?

No. It is Thorsten Meyer’s 2027 scenario, not a measured result or confirmed forecast. The cited evidence does not show that AI progress has accelerated sixfold.

What does METR’s 80% time horizon measure?

It estimates the duration of tasks, based on human completion times, that an AI agent can complete with 80% reliability. METR’s tasks focus mainly on software engineering, machine learning and cybersecurity, so the measure does not cover all types of work.

Why do the projected dates differ?

The article fits exponential trends to different slices of METR’s historical measurements. A fit beginning in 2024 reaches a 16-hour horizon in January 2027, while one beginning in 2019 reaches it in June 2028. The dates depend on the chosen data and model, and are projections rather than observations.

What could limit AI progress even if systems improve?

The scenario points to chips, data-centre construction and electricity as physical constraints. How binding those limits will be, and when, is not established by the article.

Source: ThorstenMeyerAI.com

You May Also Like

OpenAI’s AI Mathematics Has 722 Proofs. What Could They Lead To?

OpenAI released 722 manuscripts from an unnamed model. Outside mathematicians have not yet confirmed the results or their wider significance.

How Claude Fable 5.2 And Gemini 4 Could Shape Your Workday

Independent tests show Claude Fable 5.2 leading overall, while Gemini 4 models win selected tasks and cost less. Reliability and pricing differ.

Navier-Stokes – Tristan Buckmaster [Pdf]

Analysis of Tristan Buckmaster’s recent PDF on Navier-Stokes equations highlights ongoing research and unresolved questions in fluid dynamics.

The Secret Management Test That Exposes AI’s Working Style

A new live experiment exposes how different AI models handle management crises, highlighting strengths and weaknesses in decision execution.