🔍 Read the full analysis: AI’s 2027 Progress: A Fast Start And An Uncertain Compute Supply on ThorstenMeyerAI.com
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
TL;DR
Thorsten Meyer’s November 18, 2026 article presents a scenario—not a forecast—that AI could compress a year of progress into two months in 2027 before running into compute limits. METR’s measured software-task results show rapid gains through April 2026, but its public tracking page is no longer actively updated, and the trend depends heavily on which historical period is fitted.
Thorsten Meyer published a scenario for 2027 in which AI systems help build better AI systems quickly enough to compress a year of progress into about two months, before growth encounters limits in chips, data centres and power. The article, published November 18 on ThorstenMeyerAI.com, presents the idea as a possibility to test, not a forecast established by current evidence.
The scenario rests on a feedback loop: AI systems take on more coding, debugging and experimentation involved in developing later systems, potentially speeding up each generation. Meyer’s proposed sequence then reaches a physical compute ceiling, followed by a slower phase in which improvements to factories, laboratories and other infrastructure could expand computing capacity and the use of AI. He calls that later process “embodied self-improvement.”
The article uses METR’s software-task time horizon as one measure of progress. METR defines the 80% horizon as the human-expert task duration at which an agent is predicted to succeed four times in five. In the selected record-setting results cited by Meyer, that measure rises from 0.89 minutes for GPT-4 in March 2023 to 185.9 minutes for an early Claude Mythos Preview in April 2026. METR’s public page was marked no longer actively updated on September 8, 2026, according to the article.
Meyer’s own curve fits produce sharply different timelines. A fit using data since 2024 doubles the measured horizon about every 3.4 months and reaches a 16-hour task horizon in January 2027; a fit using data from 2019 doubles about every 6.4 months and reaches that level in June 2028. The article says a hyperbolic fit reaches infinity in November 2027, but also reports that this was the worst-performing of five fits in its backtests. These are extrapolations by the author, not observed results or METR forecasts.
The Race Between Models and Infrastructure
If the software-development loop accelerates, it could shorten the time between AI capability gains and increase pressure on companies, workers and infrastructure providers to adapt. But a faster benchmark trend would not by itself establish that AI can perform a comparable range of economically valuable work. METR’s tasks mainly cover software engineering, machine learning and cybersecurity, and the organization says they are cleaner and more defined than real-world work.
The proposed compute bottleneck also matters for the timing of any gains. Meyer cites U.S. data-centre construction at a September 2026 annual rate of $96.5 billion, compared with $50.7 billion a year earlier. The article says Epoch AI expects power to be the first constraint on scaling and estimates that new power supply takes three to five years after a decision. Those figures point to a gap between rapid software improvement and the slower construction and connection of physical infrastructure; they do not establish exactly when a shortage will bind.
Deployment is another separate question. 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. Those measurements suggest that capability progress and broad workplace adoption are not interchangeable. The 2027 scenario therefore involves several distinct claims: faster research, a compute ceiling, later infrastructure growth and wider economic use.
Products Worth Considering

Logitech MX Creative Console with 9 Customizable Keys, Graphite

DJI Mini 5 Pro Fly More Combo with DJI RC 2, 1-Inch CMOS

Bose TV Speaker Compact Soundbar with Bluetooth and HDMI-ARC, Black
As an affiliate, we earn on qualifying purchases.
What METR’s Time Horizon Measures
METR’s time horizon is a benchmark estimate, not a direct measure of an AI system’s general intelligence or economic value. The 80% figure represents the task length at which a model is predicted to succeed at an 80% rate. The task duration is based on how long human experts take, but METR says a two-hour task is closer to work a new hire or freelance contractor might complete in two hours than work by an experienced colleague.
The measurement has limits that matter for the forecast. METR says its tasks are concentrated in software-related fields and are less messy than economically valuable work. It also 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 METR measured human completion times for only five; the remaining times are estimates.
The source also distinguishes evidence of AI helping improve AI from the much faster acceleration assumed in the scenario. It reports that METR estimates Anthropic is improving models on well-specified tasks at around twice its earlier rate, while saying it would not be surprised by a rate between 1.3 and 3 times. Compressing a year of progress into two months would imply roughly six times the earlier pace. The article 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 an OpenAI classification, not independent confirmation of the full feedback-loop scenario.
“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, as quoted in his November 18 article
Which Forecast Assumptions Hold?
The central uncertainty is whether the AI-development feedback loop can deliver the acceleration Meyer assumes. The benchmark record supports a rise in measured software-task horizons through April 2026, but the author’s projections vary substantially depending on which years of data are included. The most dramatic curve, the hyperbolic fit, performed worst in the article’s reported backtests, so its November 2027 endpoint is not reliable evidence that progress will reach a singularity then.
It is also unclear whether the benchmark can track progress beyond its current range. METR’s page is no longer actively updated, and the organization’s own cautions identify limits in task coverage, human-time estimates and measurements above 16 hours. The supplied source does not provide independent validation of Meyer’s curve fitting or enough detail to assess the uncertainty around each projected date.
Other links in the scenario remain unresolved: whether AI can contribute meaningfully to building its successors, whether compute or power becomes the binding constraint and when new capacity can come online. The source’s employment and adoption figures describe earlier periods; they do not establish how deployment or job effects will change in 2027. Meyer discloses that AI, including Anthropic’s Claude, assisted with research and drafting, and says Anthropic, OpenAI and METR did not review or sponsor the article.
Signals to Watch Through 2027
Meyer’s article says the scenario should be tested against future evidence rather than accepted on the basis of its most aggressive curve. Relevant signals include new, independently documented software-task measurements; whether AI systems take on a larger role in AI research and development; and whether progress continues at a rate closer to the recent trend or to the sixfold acceleration the scenario requires.
Readers will also need to watch whether benchmark providers resume or replace public tracking, and whether task suites can measure longer and more representative work reliably. On the infrastructure side, data-centre construction, grid connections and added power supply could indicate whether physical capacity is catching up with demand. The source gives no confirmed date for a new METR update, a power-supply milestone or a 2027 capability threshold. The proposed year remains a scenario whose key tests are still ahead.
Key Questions
Is the article predicting that AI will reach ASI in 2027?
No. Meyer describes a scenario and personal bet, not a forecast presented as fact. The article also says ASI has no agreed definition and concentrates on the proposed mechanism for faster progress.
What does METR’s 80% time horizon measure?
It estimates the duration of a human-expert task 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, so the measure does not cover all work.
Why do the projected dates differ?
Meyer’s projections use different historical periods. The fit starting in 2024 implies faster growth than the fit starting in 2019, producing January 2027 and June 2028 dates, respectively, for reaching a 16-hour horizon. Both dates are author-generated extrapolations.
What could limit AI progress in the scenario?
The proposed limits are physical: available chips, data-centre buildings and connected power. The article cites an estimate that new power supply takes three to five years after a decision, but it does not establish when or where a binding shortage will occur.
Does the benchmark prove AI is already replacing workers at scale?
No. The benchmark measures performance on defined tasks, not economy-wide job replacement. The article separately cites U.S. business-use and job-cut figures, which describe limited periods and do not establish future employment effects.
Source: ThorstenMeyerAI.com
