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

🔍 Read the full analysis: What Might Slow AI In 2027 After A Burst Of Progress? on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on monitors, keyboards and dev gear

  • 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 scenario article argues that AI-assisted AI development could accelerate progress in 2027, but it is not a forecast and several links in its argument remain unproven. METR’s task-duration results show rapid gains through April 2026, while measurement limits, data-center power constraints and slow business adoption could slow or complicate the trend.

Thorsten Meyer published a 2027 AI scenario on November 18 that tests whether AI systems could speed up the development of later systems, then run into limits in computing hardware, electricity and deployment. The article is explicitly an opinion scenario, not a forecast presented as fact; its central claim is that a rapid software-development loop could accelerate progress, while physical infrastructure could slow it.

The scenario starts from a measurable but narrow trend: AI agents have improved on certain software tasks. METR defines its 80% time horizon as the length of a task, measured by human expert completion time, at which an agent is predicted to succeed four times out of five. In the results cited by Meyer, the measure 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 in April 2026.

Those numbers describe observed task performance, not a direct measure of general intelligence or economic productivity. METR says its tasks focus mainly on software engineering, machine learning and cybersecurity, and are cleaner than real economically valuable work. The organization also says measurements above 16 hours are unreliable with its current task set. METR marked its public page as no longer actively updated on September 8, 2026, according to the source article.

Meyer’s analysis finds that the projected timing changes sharply depending on which historical measurements are used. Its fit to METR data from 2024 onward doubles the measured horizon every 3.4 months and crosses 16 hours in January 2027; a fit starting in 2019 doubles every 6.4 months and reaches that point in June 2028. The article says a hyperbolic fit, which reaches infinity in November 2027, performed worst in its backtests. These are the author’s extrapolations, not METR predictions.

At a glance
analysisWhen: Published November 18, 2026; the 2027 o…
The developmentThorsten Meyer published a scenario on November 18, 2026, examining what could accelerate or slow AI progress in 2027, including compute, power and deployment constraints.

The Limits Behind Faster AI

The scenario matters because it connects two questions that are often discussed separately: whether AI can accelerate AI research, and whether the infrastructure needed to train and run models can keep pace. If AI tools take on more coding, testing and experimentation, they could help shorten development cycles. But faster software work does not automatically produce more chips, data centers or electricity.

Meyer cites an estimate from METR that Anthropic is improving its models on well-specified tasks at about twice the earlier rate, while noting that the organization would not be surprised by a range of 1.3 to 3 times. The article’s proposed pace—one year of progress in two months—would imply sixfold acceleration, well beyond that cited estimate. That gap makes the scale and durability of the feedback loop a central uncertainty.

Even if model capabilities rise quickly, the effects on workers and businesses may arrive more slowly. The source cites figures showing 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 what will happen in 2027, but they illustrate why capability benchmarks and broad economic adoption should not be treated as the same thing.

Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation

Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation

Ninja Luxe Café Premier 3-in-1 Espresso Machine, Drip Coffee & Cold Brew

Ninja Luxe Café Premier 3-in-1 Espresso Machine, Drip Coffee & Cold Brew

As an affiliate, we earn on qualifying purchases.

From Model Gains to Infrastructure

The proposed chain has four steps: AI systems assist in creating later systems; that assistance speeds up research; progress eventually meets a compute ceiling; and improvements to physical processes such as robotics, factories or automated labs could later expand capacity. Meyer calls the last stage embodied self-improvement. The article presents it as a possibility, not as a demonstrated operating cycle.

The physical constraint is already visible in investment and planning, though those facts do not prove a specific bottleneck date. Meyer cites U.S. data-center construction running at an annual rate of $96.5 billion in September 2026, compared with $50.7 billion a year earlier. The article also attributes to Epoch AI the view that power could be the first constraint on scaling, and says new power supply can take three to five years from a decision.

The source also says OpenAI released GPT-6.2 Astra as its first externally released model to meet the company’s “High” threshold for AI self-improvement. That is a company-defined category, not independent proof that AI systems are now broadly improving themselves or that development will accelerate at the rate proposed. The article says neither OpenAI, Anthropic nor METR reviewed or sponsored its analysis, and discloses that it was researched and drafted with AI assistance, including Anthropic’s Claude.

“An automated flywheel pushes forward.”

— Thorsten Meyer, describing embodied self-improvement

Where the 2027 Scenario Could Break

The available evidence does not establish that AI progress will accelerate sixfold, that the 16-hour task threshold will be crossed on a particular date, or that artificial general or superintelligent systems will arrive in 2027. The date estimates depend on which data points are fitted, and the underlying benchmark has limitations in task selection, human-time estimates and measurement beyond 16 hours. METR’s public page was no longer actively updated as of September 8, 2026, according to the source, leaving the later trend harder to assess from that series.

Other links in the scenario are also unresolved. The source does not provide evidence that AI-developed models will consistently make the next generation faster at the proposed rate, or that automated labs and factories will improve themselves on a comparable cycle. The cited construction spending and power lead times indicate infrastructure demands, but they do not specify when electricity or computing capacity will constrain any particular AI project. The article’s adoption figures likewise cannot determine how quickly future capabilities will spread through businesses or affect employment.

Signals to Watch Through 2027

The scenario will be easier to test if new, comparable evaluations continue to track how long AI agents can complete tasks reliably. Readers should look for published methods and results that clarify whether performance gains persist across task types, rather than relying on a single extrapolated curve. Updated METR measurements or another transparent benchmark could show whether the trend continues beyond the limits of the current task suite.

Infrastructure indicators will provide a separate test: data-center construction, access to electricity, and the time needed to bring new power supply online. Evidence that those constraints are delaying projects would support the scenario’s proposed physical ceiling; evidence of timely capacity growth would weaken that part of it. The source gives no confirmed milestone for when such a ceiling will arrive.

Finally, the key question is whether benchmark advances translate into sustained use in workplaces and measurable changes in work. Company announcements about AI-assisted research should be separated from independently measured results, and capability claims should be compared with deployment data. As of November 18, 2026, the 2027 outcome remains open: the article lays out signals to watch, not a settled prediction.

Key Questions

Is rapid AI progress in 2027 confirmed?

No. Meyer’s article labels the 2027 outlook a scenario and opinion, not a forecast established by current evidence.

What does METR’s 80% time horizon measure?

It estimates the length of a task, based on human expert completion time, at which an AI agent is predicted to succeed four times out of five. The cited tasks are mainly in software engineering, machine learning and cybersecurity.

Why might computing power or electricity slow AI development?

Training and operating AI systems requires chips, data-center space and electricity. Meyer cites a view from Epoch AI that power could be the first scaling constraint, while noting that new supply may take three to five years from a decision.

Does the benchmark show that AI can replace workers?

No. METR’s task-duration measure is not a direct measure of employment effects or broad workplace productivity. The source separately cites U.S. figures in which 18% of firms reported AI use in a business function and 2% reported AI-related job cuts in late 2025 and early 2026.

Source: ThorstenMeyerAI.com

You May Also Like

10 Mini PCs For Local AI Models 2027

A source-based guide to 10 mini PCs for local AI, with key tradeoffs in memory, cooling and expandability—and important gaps in the available specs.

iPhone 18 Pro Could Be Noticeably Thicker Than iPhone 17 Pro

Leaked reports suggest the iPhone 18 Pro could be noticeably thicker than its predecessor, raising questions about design and features.

Glasspane: When Transparency Itself Becomes the Product

Glasspane introduces role-aware dashboards and AI-driven insights, transforming infrastructure transparency for enterprise and MSPs. Latest features expand its capabilities.

Quantinuum Reports A 10X Reduction In Logical Error Rate On Helios

Quantinuum says a color-code test on its Helios system reached 99.9995% logical fidelity, with errors falling as code distance increased.