TL;DR
Thorsten Meyer’s 2027 scenario argues that AI could help develop more capable AI quickly, compressing a year of progress into two months before compute constraints slow the pace. Public evidence shows longer software tasks completed by AI agents and some acceleration in model improvement, but it does not establish a sixfold acceleration or predict when physical limits will bind.
Thorsten Meyer has published a 2027 AI scenario in which AI systems help build improved AI systems quickly enough to compress a year of progress into two months, until limits on chips, buildings and power slow the cycle. The article presents this as a scenario, not a forecast; the public measurements it cites show progress in software task performance, but do not establish that sixfold acceleration will occur.
Get everyday essentials delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
The scenario’s central mechanism is a feedback loop: AI systems take on more coding, debugging and experimentation involved in developing future systems, potentially speeding up each subsequent round. Meyer’s proposed pace—a year of progress in two months—would amount to about six times the earlier rate. The article says each part of this chain could fail, including the pace of AI-assisted development and the ability to translate software improvements into practical deployment.
One public measure cited is METR’s 80% time horizon, which tracks the length of software tasks an AI agent completes successfully four times out of five. METR’s reported results rose from under one 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, limiting how current the series is.
Meyer’s calculations fit different growth curves to METR’s measured points. An exponential fit using data since 2024 doubles every 3.4 months and crosses 16 hours in January 2027; a fit beginning in 2019 doubles every 6.4 months and reaches that point in June 2028. Those are projections from selected data and methods, not METR forecasts. The article says the more dramatic hyperbolic fit performed worst in its backtests.

Compute Could Set the Pace
The scenario matters because it links advances in AI research to constraints in the physical economy. If AI helps accelerate model development, but computing capacity cannot expand as quickly, software progress could slow even as demand for capable systems rises. Meyer’s next proposed phase, called embodied self-improvement, would involve robots, factories and automated labs contributing to further improvements over weeks and months.
The article cites US data-centre construction running at an annual rate of $96.5 billion in September 2026, compared with $50.7 billion a year earlier. It also cites Epoch AI’s view that power may be the first constraint on scaling and that new power supply can take three to five years from the decision to build. These indicators point to potential infrastructure pressures, but they do not establish when a compute ceiling would arrive or how high it would be.
Rapid technical progress would not automatically mean rapid economic change. The source reports that 18% of US firms used AI in a business function in late 2025 and early 2026, while 2% reported AI-related job cuts. Those figures illustrate a gap between capability and adoption; they do not settle how employment or business practices will change in 2027.
Products Worth Considering

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

Amazon Kindle Scribe Colorsoft 64GB with Premium Pen

Jackery Explorer 300 Portable Power Station, 292Wh
As an affiliate, we earn on qualifying purchases.
The Evidence Behind the Scenario
The article distinguishes measured benchmark results from its own projections. METR’s time-horizon measure concerns performance on a particular set of software tasks, not every kind of work or a general measure of intelligence. The source says tasks above 16 hours fall outside the benchmark’s measuring range, so projections beyond that point extend beyond what the task suite directly measures.
Meyer’s analysis also cites METR’s estimate that Anthropic is improving models on well-specified tasks at around twice its earlier rate. METR said it “would not be surprised” by a rate between 1.3 and 3 times the previous pace. That range is below the scenario’s implied sixfold acceleration. The article also says OpenAI’s GPT-6.2 Astra is the first externally released model from the company to meet its “High” threshold for AI self-improvement. These are separate indicators; neither by itself demonstrates the full scenario.
The source describes the article as researched and drafted with AI assistance, including Anthropic’s Claude. It says Anthropic, OpenAI and METR did not review or sponsor the article. Meyer’s use of “ASI,” or artificial superintelligence, is not the subject of a settled definition in the article; its analysis focuses on the proposed development cycle and constraints.
“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
How Fast the Feedback Loop Runs
The available evidence does not show whether AI-assisted research will reach the sixfold acceleration required by the scenario. METR’s cited estimate for Anthropic’s improvement rate is lower, and its public time-horizon page was no longer actively updated after September 8, 2026. The source does not provide new benchmark measurements beyond April 2026.
It is also unclear when power, chips or data-centre capacity would become binding limits, whether progress in software tasks transfers to other kinds of work, or how quickly robots and automated labs could contribute to AI development. The 16-hour benchmark boundary means long-range projections cannot be read as direct measurements. The data-centre spending figures and adoption statistics provide context, not a precise timetable for the scenario.
Signals to Watch in 2027
The scenario will be easier to assess if new public benchmark results extend METR’s time-horizon series beyond April 2026 and clarify whether agent performance continues to rise at the faster rates. Readers can also watch for evidence that AI systems are taking on more of the work involved in building subsequent models, alongside disclosed estimates of how quickly those systems are improving.
On the infrastructure side, data-centre construction, access to chips and power supply will indicate whether computing capacity is expanding fast enough to support more intensive AI development. Business adoption and reported effects on jobs will show whether technical capability is reaching wider use. Until those signals are clearer, the proposed 2027 acceleration and later compute ceiling remain an untested scenario.
Key Questions
Is the 2027 AI acceleration a confirmed forecast?
No. Thorsten Meyer presents it as a scenario and a bet, not as a forecast established by the cited evidence.
What does METR’s 80% time horizon measure?
It measures the length of software tasks an AI agent completes successfully four times out of five. It does not measure all forms of intelligence or work.
Does current evidence show a sixfold acceleration?
The source does not establish that. It cites a METR estimate of roughly twice the earlier rate for Anthropic’s improvement on well-specified tasks, with METR saying it would not be surprised by 1.3 to 3 times the earlier rate.
What could limit AI progress in this scenario?
The proposed constraints are chips, data-centre buildings and power. The source cites a view that power could be the first constraint, but does not give a date when any limit will bind.
What evidence would help test the scenario?
New benchmark measurements, clearer evidence of AI systems contributing to later model development, and data on computing infrastructure and business adoption would help show whether the proposed cycle is taking shape.
Source: Thorsten Meyer AI
