TL;DR
Thorsten Meyer’s 2027 scenario argues that AI systems could accelerate AI development, compressing a year of progress into two months before physical compute limits slow the cycle. Public evidence shows software task horizons have grown, but projections vary widely, and the benchmark used to track them is no longer being actively updated.
Thorsten Meyer has published a scenario for 2027 in which AI systems help build more capable AI quickly, potentially compressing a year of progress into two months before chips, data centers and power limit further gains. The article presents this as a hypothesis to test, not a forecast established by current evidence.
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
The scenario rests on a feedback loop: AI systems take on more coding, debugging and experimentation involved in developing later systems, which could speed the next round of progress. Meyer says the pace might reach what he describes as six times the earlier rate. That figure is a scenario requirement, not a measured current acceleration.
One public measure of AI capability is METR’s 80% time horizon: how long a software task an agent can complete successfully four times out of five. METR’s recorded results rose from under a minute for GPT-4 in March 2023 to 185.9 minutes, or about three hours and six minutes, for an early Claude Mythos Preview in April 2026. METR marked its public page as no longer actively updated on September 8, 2026.
Meyer’s calculations fit several curves to METR’s measured results. An exponential fit using data from 2024 doubles task duration about every 3.4 months and reaches 16 hours in January 2027; a fit starting in 2019 doubles about every 6.4 months and reaches that point in June 2028. A hyperbolic fit reaches infinity in November 2027, but Meyer reports it performed worst in backtesting. These are model-dependent projections, and METR’s benchmark does not measure tasks longer than 16 hours.

Why Compute Could Set the Pace
If AI systems can reliably accelerate the research and engineering behind later systems, development timelines could change across software and other knowledge work. But the scenario’s projected speed is materially higher than the evidence cited in the article: METR estimates that Anthropic is improving its models on well-specified tasks at about twice its earlier rate, while saying it would not be surprised by a rate between 1.3 and three times as fast. That estimate does not establish a sixfold acceleration.
Physical infrastructure could shape what happens after software gains. U.S. data-center construction was running at an annual rate of $96.5 billion in September 2026, compared with $50.7 billion a year earlier. Epoch AI expects power to be the first constraint on scaling and says new power supply can take three to five years from the decision to build. The numbers point to a possible bottleneck, though they do not establish when it would bind or how much computing capacity would be available.
Capability gains also do not automatically translate into broad adoption or employment changes. The source reports 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 figures describe different outcomes; they do not show that a rapid capability loop would cause widespread deployment or job losses.
Products Worth Considering

Bose QuietComfort Wireless Noise-Cancelling Headphones, Moonlight Grey

LG UltraGear 27-inch QHD OLED Gaming Monitor, 240Hz, 0.03ms, G-SYNC Compatible, FreeSync Premium Pro, HDR True Black 400

Jackery Explorer 300 Portable Power Station, 292Wh
As an affiliate, we earn on qualifying purchases.
The Evidence Behind the Scenario
Meyer frames the piece as a test of his view that artificial superintelligence could arrive in 2027 or shortly after. The term artificial superintelligence has no agreed definition, so the article focuses on the proposed mechanism—AI systems contributing to the development of later systems—rather than claiming a specific threshold has been reached.
The source calls the later physical cycle “embodied self-improvement,” or ESI: a slower process involving robots, factories and automated labs. It proposes that this cycle could expand production once computing constraints ease. The article’s chart shows an illustrative ceiling after a period of fast growth; its ceiling is not a measured value or a forecast.
The article discloses that it was researched and drafted with AI assistance, including Anthropic’s Claude, and says Anthropic, OpenAI and METR did not review or sponsor it. It also identifies an affiliate relationship for qualifying purchases. These disclosures help readers assess the source and its relationship to the companies discussed.
“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
Where the Timeline Could Break
The evidence does not establish that a year of progress will arrive in two months, that artificial superintelligence will be reached in 2027, or that compute will become the binding constraint at a particular time. The METR projections change substantially depending on which years are included, and the benchmark’s public page is no longer actively updated. Its task suite also stops measuring beyond 16 hours, leaving longer task horizons outside the chart’s direct evidence.
Other unknowns include whether gains on software tasks translate into faster AI research, whether infrastructure can expand quickly enough, and whether systems will be adopted widely in businesses. The cited company-use and job-cut figures provide a snapshot, not a forecast of later employment effects. The source does not provide enough detail here to establish when or how the proposed embodied cycle would begin.
Signals to Watch in 2027
The scenario will be easier to assess if new, comparable benchmark results show whether AI agents can handle longer tasks and whether the pace of improvement is accelerating. Readers should track the measurement method and time period alongside any claimed increase, especially because METR’s public page is no longer actively updated.
Infrastructure indicators will also matter: data-center construction, power availability and the time required to connect new supply can show whether compute capacity is keeping pace with demand. Business adoption and reported employment effects can help distinguish technical capability from changes in work. Until those signals develop, the two-month compression and the 2027 timeline remain open questions.
Key Questions
Is the article predicting that AI will reach superintelligence in 2027?
No. It tests Thorsten Meyer’s scenario that this could happen in 2027 or shortly after, and says the scenario is not a forecast presented as fact. The term itself also has no agreed definition.
What does METR’s 80% time horizon measure?
It measures the length of software task an AI agent completes successfully four times out of five. The source reports a result of 185.9 minutes for an early Claude Mythos Preview in April 2026 and says METR’s public page was no longer actively updated from September 8, 2026.
Does the benchmark show that AI progress is six times faster?
No. The sixfold rate is part of Meyer’s scenario. The article attributes to METR an estimate that Anthropic is improving models on well-specified tasks at about twice its earlier rate, with a wider possible range.
What could limit further AI progress?
The scenario identifies chips, data-center buildings and power as possible constraints. The cited source says Epoch AI expects power to be the first scaling constraint, but it does not establish when that limit would be reached.
Source: Thorsten Meyer AI
