🔍 Read the full analysis: When AI Progress Speeds Up: A 2027 Scenario About Compute Limits on ThorstenMeyerAI.com
Get smart everyday buys delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Thorsten Meyer’s November 18, 2026 article presents a possible 2027 acceleration in AI development as a scenario, not a forecast established by evidence. METR’s software-task measurements show rapid gains but do not establish a sixfold acceleration, while data-center construction and power supply point to physical limits that could slow expansion.
Thorsten Meyer published a 2027 AI acceleration scenario on November 18, arguing that AI systems could help build more capable successors quickly enough to compress a year of progress into about two months, before available computing capacity becomes a bottleneck. The article presents this as a possibility to test, not a confirmed forecast, and says the evidence does not yet establish that pace.
The scenario depends on a software feedback loop: AI systems take on more coding, debugging and experimentation involved in developing later systems. Meyer argues that, if this loop tightens, progress could accelerate until it runs into limits involving chips, data-center buildings and electricity. A slower, physical feedback loop—what he calls embodied self-improvement—could then involve automated laboratories, factories and robotics. Those steps are part of the proposed scenario, not outcomes shown to have occurred.
The article uses METR’s measure of the software-task duration an AI agent can complete reliably. Its 80% time horizon rose from 0.89 minutes for GPT-4 in March 2023 to 185.9 minutes for an early Claude Mythos Preview in April 2026. But the trend line depends on which historical period is fitted: Meyer’s calculation using data since 2024 reaches a 16-hour horizon in January 2027, while a fit starting in 2019 reaches it in June 2028. The article says the more dramatic hyperbolic fit performed worst in its backtests.
The reported acceleration rate is also short of the scenario’s central claim. METR estimates that Anthropic has improved its models on well-specified tasks at about twice the earlier rate, while saying it would not be surprised by a rate between 1.3 and three times. Compressing a year of progress into two months would imply about six times the earlier pace. METR’s public page was marked as no longer actively updated on September 8, 2026, according to the article.
The Gap Between Capability and Deployment
The scenario matters because a faster AI-development cycle could affect how quickly new capabilities reach businesses and workers. But a rapid improvement in benchmarked software tasks would not by itself show that AI can perform a broad range of paid work or that companies are deploying it at the same speed. Meyer 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 describe different measures of adoption and employment effects; they do not prove that future impacts will remain limited.
Physical infrastructure could also shape the pace. The article reports that U.S. data-center construction was running at a $96.5 billion annual rate in September 2026, compared with $50.7 billion a year earlier. It also cites Epoch AI’s expectation that power will be the first constraint on scaling and that new power supply can take three to five years from a decision. If demand for compute rises faster than infrastructure can be built and connected, capability gains may not translate into an equally fast rollout.
Popular right now

Bose QuietComfort Wireless Noise-Cancelling Headphones, Moonlight Grey

Roborock Q7 M5+ Robot Vacuum and Mop with Self-Emptying Dock

Bose QuietComfort Ultra Wireless Noise Cancelling Earbuds, 2nd Gen, with 20W 2-Port Wall Charger
As an affiliate, we earn on qualifying purchases.
What METR’s Horizon Measures
METR defines its time horizon by the length of a task, measured by the time a human expert would need, at which an AI agent is predicted to succeed with a specified reliability. The 80% figure refers to tasks the agent completes four times out of five. The listed results are measurements, not projections, but the measure covers a limited part of AI capability.
METR’s tasks primarily involve software engineering, machine learning and cybersecurity. The organization has cautioned that they are cleaner and more clearly specified than many economically valuable jobs. Meyer also notes limits in longer-task estimates: METR says measurements above 16 hours are unreliable with its current task suite, and human completion times were measured for only five of 31 tasks estimated to take at least eight hours.
The article also describes OpenAI’s GPT-6.2 Astra as its first externally released model to meet the company’s “High” threshold for AI self-improvement. That label is a company assessment, not independent proof that the full feedback loop in Meyer’s scenario is operating. Meyer disclosed that the article was researched and drafted with AI assistance, including Anthropic’s Claude, and said Anthropic, OpenAI and METR did not review or sponsor it.
“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 Sixfold Acceleration Is Unproven
The available task-horizon data do not establish that progress will accelerate to six times its earlier pace, arrive on a 2027 timetable or produce artificial superintelligence. Meyer notes that the projected date changes substantially depending on the data period used, and that the most dramatic fitted curve failed his backtests. METR’s benchmark is limited to particular software tasks, with stated reliability and measurement limits; it is not a direct measure of economy-wide capability.
It is also unclear how quickly new models will improve their own development, whether the reported self-improvement threshold will lead to a sustained feedback loop, and how much additional compute can be supplied. The article’s infrastructure figures and cited power-delay estimate indicate possible constraints, but do not establish precisely when they would bind. Meyer says ASI has no agreed definition, so the scenario does not settle when that label would apply.
Signals to Watch Through 2027
The scenario can be tested against future evidence: updated, comparable measurements of AI agents on longer tasks; independently checkable evidence that AI systems are speeding up model research and development; and new information about data-center capacity, chip availability and power connections. Meyer’s article says METR’s public page stopped being actively updated in September, making continued access to comparable measures a particular uncertainty.
Readers should also distinguish a model’s performance from its adoption in workplaces. Business-use and employment data could show whether capability gains are reaching firms, though those indicators may move on different timelines. For now, the article offers a framework for evaluating a possible 2027 development. It does not establish that the acceleration will happen, when compute limits will bite, or how broadly any resulting systems will be deployed.
Key Questions
Is the 2027 acceleration a confirmed forecast?
No. Meyer presents it as a scenario and tests evidence for and against it; he does not establish that it will happen.
What does METR’s 80% time horizon measure?
It estimates the length of a task, based on human expert completion time, that an AI agent can finish with 80% reliability. The task set is focused mainly on software engineering, machine learning and cybersecurity.
Does the benchmark show AI progress is accelerating sixfold?
No. Meyer says METR estimates improvement in Anthropic’s models on specified tasks at about twice the earlier rate. A sixfold pace is part of the scenario’s proposed compression, not a result established by that estimate.
What could limit further AI development?
The article points to compute infrastructure, including chips, data-center buildings and connected electricity supply. It cites an estimate that adding power supply can take three to five years from the decision, but the timing and severity of any constraint remain uncertain.
Source: ThorstenMeyerAI.com
