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

🔍 Read the full analysis: How Compute Could Shape AI’s Rapid Progress In 2027 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

A November 18 scenario from ThorstenMeyerAI.com argues that AI systems could accelerate AI development in 2027, then run into limits imposed by chips, data centers and electricity. The article presents this as a possibility, not a forecast: METR’s software-task measurements show rapid gains but have measurement limits, and other evidence does not establish a sixfold acceleration or broad economic impact.

Thorsten Meyer published a scenario on November 18 arguing that AI could make rapid progress in 2027 by helping build the next generation of AI systems, before running into constraints in chips, data-center capacity and electricity. The piece is an opinion scenario, not a confirmed forecast: its central claim that a year of progress could arrive in two months is not established by the measurements it discusses.

The scenario rests on a software-development feedback loop: AI systems help with coding, debugging and experiments used to create newer systems, which may then do more of that work. Meyer argues that if this loop accelerates sufficiently, progress could compress from annual steps into weeks. He then expects that faster development to meet a physical ceiling, followed by a slower cycle of expansion in factories, laboratories and computing infrastructure. The article calls that later phase embodied self-improvement.

For a public measure of agent capability, the article uses METR’s 80% time horizon: the length of a software task, measured by estimated expert completion time, at which an AI agent is predicted to succeed four times out of five. Selected record-setting results cited in the article rise from 0.89 minutes for GPT-4 in March 2023 to 185.9 minutes for an early Claude Mythos Preview in April 2026. These are measured task results, not a direct measure of general intelligence or progress across the economy.

The article also reports that METR marked its public time-horizon page as no longer actively updated on September 8, 2026. Its account of the data says different fitted trends produce markedly different projections: a fit using observations since 2024 reaches a 16-hour task horizon in January 2027, while one using data from 2019 reaches that point in June 2028. Meyer describes the faster extrapolation as uncertain and says the hyperbolic fit performed worst in backtests. The projections are his analysis of the data, not METR forecasts.

At a glance
analysisWhen: Scenario published November 18, 2026; t…
The developmentThorsten Meyer published a scenario on November 18, 2026, arguing that AI-assisted development could accelerate progress in 2027 until physical compute constraints become more binding.

Compute Could Set the Pace

The scenario matters because it links two questions often discussed separately: whether AI can speed up the work of AI development, and whether companies can supply enough computing capacity to turn that work into deployed systems. If AI-assisted research and engineering produce faster capability gains, the time available to respond to those changes could shrink. But development speed alone would not show that systems are safe, useful across jobs, or adopted widely.

Physical infrastructure could shape the pace even if software improves quickly. The source article 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, and says Epoch AI expects power to be the first scaling constraint. It also says new power supply can take three to five years after a decision. Those figures point to potential bottlenecks, not proof that compute shortages will halt AI progress in 2027.

Economic effects may lag behind technical benchmarks. The article 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. These measures have different meanings: reported business use is not the same as extensive automation, and job cuts do not capture every effect on work. The scenario’s claim of broad labor and business impact remains a possibility rather than an observed outcome.

Products Worth Considering

As an affiliate, we earn on qualifying purchases.

What the Benchmarks Measure

METR is a nonprofit that evaluates frontier AI agents. Its time horizon estimates the task duration at which an agent is predicted to succeed at a specified reliability level. The 80% measure is more demanding than a 50% success threshold for the same task, but it remains a measure of performance on a particular task set.

METR’s tasks primarily cover software engineering, machine learning and cybersecurity. The source article notes that the organization describes them as cleaner and more defined than economically valuable work in real settings. It also reports that METR cautions that measurements above 16 hours are unreliable with its current task suite. Of 31 tasks estimated to take humans at least eight hours, the article says human completion times were measured for only five; the other times were estimates.

The source also attributes a separate estimate to METR: Anthropic may be improving its models on well-specified tasks at about twice the earlier rate. METR is quoted as saying it “would not be surprised” by a range of 1.3 to 3 times. That estimate does not establish the sixfold acceleration required for a year of prior progress to arrive in two months. The source says OpenAI’s GPT-6.2 Astra is its first externally released model to meet the company’s “High” threshold for AI self-improvement; this is OpenAI’s classification, not independent confirmation of a recursive development loop.

“an automated flywheel pushes forward. As long as the compute limits are then removed, this becomes a big rollout of AI that affects far more labour and business.”

— Thorsten Meyer, describing the proposed later phase

Forecasts Outrun the Evidence

The central unknown is whether AI systems will speed up AI research and engineering enough to produce the proposed sixfold acceleration. The cited benchmark trend measures specific software tasks, and its projections shift substantially depending on which years are included. METR’s reported limits on long-task measurements also make extrapolations toward multi-hour work uncertain.

The source does not establish that an AI system has independently built a substantially more capable successor, or that the reported model-development rate translates into a comparable acceleration of overall AI progress. OpenAI’s “High” self-improvement threshold is a company classification. The article says neither OpenAI, Anthropic nor METR reviewed or sponsored it, and discloses that it was researched and drafted with AI assistance, including Anthropic’s Claude.

It is also unclear when compute or electricity would become a binding constraint, how quickly new data centers and power supply could be added, or whether embodied automation would speed up research and deployment. The source presents no confirmed 2027 outcome. Its forecasts should be read as conditional claims whose links can fail, not as established events.

Signals to Watch in 2027

The scenario can be tested against new, comparable measurements of AI performance on longer software tasks. A useful signal would be sustained improvement across independent evaluations, with clearly reported task definitions, success rates and measurement methods. Because METR’s public page was no longer actively updated as of September 8, 2026, readers would also need to watch for updated benchmark data or other transparent evaluations before treating its older trend as a current trajectory.

Other signals include evidence that AI tools are materially contributing to model research, documented changes in the time or cost needed to develop new systems, and data on construction, grid connections and power availability. Adoption and employment figures could help distinguish capability gains from real-world effects. Until those indicators show a consistent pattern, the proposed 2027 acceleration remains a scenario rather than a settled forecast.

Key Questions

Is rapid AI progress in 2027 confirmed?

No. The source is an opinion scenario published by Thorsten Meyer on November 18, 2026. It outlines a possible path and the evidence it uses, but does not establish that the projected acceleration will happen.

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 mainly involve software engineering, machine learning and cybersecurity, so the result is not a general measure of workplace capability.

Why might compute limit AI progress?

AI systems run on chips in data centers that need buildings, equipment and electricity. The article cites data-center construction and power-supply timelines as signs that infrastructure could constrain scaling, but it does not establish when or whether such limits will bind.

Does AI already accelerate the development of other AI systems?

The source describes AI-assisted development as a potential feedback loop and attributes to METR an estimate of roughly twice the earlier rate of improvement for Anthropic models on well-specified tasks. That estimate does not prove a self-sustaining loop or the sixfold acceleration proposed in the scenario.

Source: ThorstenMeyerAI.com

You May Also Like

Terminal-RSI And The Limits Of AI Self-Improvement Claims

A Terminal-RSI preprint reports benchmark gains from a human-designed training pipeline, not an AI system improving itself without oversight.

SenseTime’s SenseCore Maps Out Clusters For Agentic AI Inference

A Pandaily headline describes separate prefill and decode clusters for SenseCore agentic inference, but deployment and performance details are unavailable.

Will You Be Among The First To Test SpaceXAI’s Grok Bot? Here’s How

SpaceXAI has launched Grok, its conversational AI, into early beta. Limited users can now test its capabilities; broader release details remain unknown.

SpaceXAI Puts $1.5 Million In Compute Toward Omarchy’s Linux Distro

A Verge headline says SpaceXAI is backing Linux distribution Omarchy with $1.5 million in compute, but the terms and timeline are not clear.