🔍 Read the full analysis: AI Progress In 2027: What Happens When Compute Runs Short? on ThorstenMeyerAI.com
Get the little things that make your day 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
A November 18, 2026, scenario article argues that AI progress could accelerate sharply in 2027 as systems help build their successors, then slow when compute capacity becomes a constraint. Public benchmark results show rapid gains on certain software tasks, but the measure has limits and is no longer being actively updated; the 2027 outcome remains uncertain.
Thorsten Meyer published a scenario analysis on November 18, 2026, asking whether AI systems could accelerate the development of successor systems enough to bring roughly a year of progress in two months during 2027, before compute limits slow further gains. The article presents this as a possibility to test, not a forecast established by current evidence.
The scenario has two loops. First, AI systems do more of the coding, debugging and experimentation used to develop newer AI systems, potentially speeding up progress. Then, if available computing capacity becomes a constraint, expansion depends on slower physical processes: manufacturing chips, constructing data centers and bringing new power supply online. Meyer describes this later phase as embodied self-improvement, in which automation extends beyond software into factories and laboratories.
One public measure cited is METR’s 80% time horizon: the duration of a task, based on human expert completion time, at which an AI agent is predicted to succeed four times out of five. In selected record-setting results, the reported 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. Those figures measure performance on a particular task suite; they are not a direct measure of general intelligence or economy-wide productivity.
The article reports that METR marked its public time-horizon page as no longer actively updated on September 8, 2026. It also cites METR’s estimate that Anthropic is improving models on well-specified tasks at around twice its earlier rate, while noting that METR said it would not be surprised by a range from 1.3 to 3 times. Meyer’s proposed pace—one year’s progress in two months—would amount to roughly six times the old rate, a faster acceleration than that estimate alone establishes.
How Compute Could Shape AI Progress
The scenario matters because it links two issues often discussed separately: how quickly AI capabilities improve and how quickly the infrastructure supporting them can expand. If AI tools materially speed up model development, companies could face shorter intervals between capability gains. But even fast software progress cannot by itself produce more chips, data centers or electricity.
The source cites U.S. data-center construction running at a $96.5 billion annual rate in September 2026, compared with $50.7 billion a year earlier. It also says Epoch AI expects power to be the first constraint on scaling and that new power supply can take three to five years from decision to availability. These figures help illustrate why infrastructure could limit growth, but they do not establish when or how tightly any constraint will bind.
Near-term effects on employment are also not demonstrated by capability benchmarks. The article cites survey 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. That gap cautions against treating rapid progress on software tasks as proof of broad deployment or immediate labor-market change.
Products Worth Considering

Sonos Arc Ultra Dolby Atmos Soundbar, Black

Sonos Roam (Gen 1) Portable Smart Speaker, White

Roborock Qrevo S Pro Robot Vacuum and Mop with 18,500 Pa Suction
As an affiliate, we earn on qualifying purchases.
What the Benchmark Measures
METR describes its time horizon as the task duration, measured by human expert completion time, at which an agent is predicted to achieve a specified reliability. The 80% measure is stricter than a 50% success threshold, and the tasks in the cited suite primarily involve software engineering, machine learning and cybersecurity. METR says the work is cleaner and more bounded than real economically valuable labor; a two-hour benchmark task may resemble work assigned to a new hire or freelancer more than the work of an experienced colleague.
The benchmark also has measurement limits. METR says results above 16 hours are unreliable with its current task suite. Of 31 tasks estimated to take humans at least eight hours, it measured human completion times for five and estimated the rest. These qualifications matter because extrapolating from a narrow set of tasks to a precise date for broad AI progress would go beyond what the measurements show.
Meyer’s article says the projected year changes depending on which historical data points are used. His fits to METR measurements produce a doubling time of 3.4 months when using data since 2024, versus 6.4 months when using data since 2019. The former fit crosses a 16-hour horizon in January 2027; the latter does so in June 2028. Those are model-based extrapolations, not observed results. The article also reports that a hyperbolic fit performed worst in its backtests, having already projected an infinite horizon by January 2, 2026.
“Measurements above 16 hrs are unreliable with our current task suite.”
— METR, as quoted in Meyer’s article
Evidence Behind the 2027 Scenario
The central uncertainty is whether AI assistance in AI research will produce the sustained acceleration the scenario requires. The benchmark records gains on a defined software-task suite, but it does not show that the same pace applies to model research as a whole, physical science, or economic output. The source’s estimate of roughly twofold improvement in Anthropic’s work on well-specified tasks is not evidence of a sixfold acceleration across the field.
Other uncertainties include how quickly compute demand grows, whether power or another input becomes the binding constraint, and how quickly new infrastructure can be built and connected. Meyer’s article cites a reported OpenAI release, GPT-6.2 Astra, as the first external model from the company to meet its “High” threshold for AI self-improvement. That is a company-defined classification as presented in the source, not independent proof that an autonomous improvement loop is accelerating on the article’s proposed scale.
The article also discloses that it was researched and drafted with AI assistance, including Anthropic’s Claude, and that Anthropic, OpenAI and METR did not review or sponsor it. Its numerical projections are the author’s own fits to benchmark data. They should not be read as forecasts from METR or the AI companies.
Signals to Watch Through 2027
The scenario will become more testable if updated, independently documented evaluations show whether AI agents keep improving on long-duration tasks. METR’s public page was no longer actively updated as of September 8, 2026, according to Meyer, leaving the cited results without a continuing update from that page. New comparable measurements, with clear methods and task definitions, would help establish whether the trend continues or changes.
Readers can also watch for evidence of faster AI research cycles, rather than relying only on software-task scores; published information about data-center construction, chip availability and power connections; and signs that AI tools are being adopted across workplaces. The employment figures cited in the source offer a baseline, not a prediction of future job effects.
For now, the defensible conclusion is limited: measured AI-agent performance on selected software tasks rose substantially through April 2026, while the available benchmark has stated limitations. Whether those gains lead to a 2027 acceleration, and whether compute constraints then slow it, remains an open question.
Key Questions
Is the article predicting that artificial superintelligence will arrive in 2027?
No. Meyer presents 2027 as a scenario to test, not a fact or settled forecast. The article also notes that artificial superintelligence has no agreed definition.
What does METR’s 80% time horizon measure?
It estimates the length of a task, measured by human expert completion time, at which an AI agent is predicted to succeed four times out of five. The cited tasks mainly cover software engineering, machine learning and cybersecurity.
What evidence supports faster AI progress?
Selected METR results in the source show the 80% time horizon rising from 0.89 minutes for GPT-4 in March 2023 to 185.9 minutes for an early Claude Mythos Preview in April 2026. The source warns that the benchmark is narrow and has reliability limits above 16 hours.
Why could compute become a constraint?
Advanced AI systems require chips, data-center space and electricity. The article cites an estimate that new power supply can take three to five years from a decision to become available, but the timing and severity of any constraint remain uncertain.
What would show that the 2027 scenario is wrong?
Evidence against it could include updated evaluations showing slower or stalled gains, no sustained acceleration in AI-assisted research, or infrastructure expansion that keeps pace with demand. The article does not set a single test date or threshold that would conclusively settle the question.
Source: ThorstenMeyerAI.com
