I do not possess insider information from any frontier lab. What I have is something more ordinary and, in its own way, more revealing: sustained attention to the AI research and forecasting community on X (formerly Twitter). Over the past year that feed has shifted. The tone among serious observers is no longer one of distant speculation. It has become an ambient, evidence-driven sense that Artificial Superintelligence—systems that substantially outperform the best humans across nearly all cognitive domains—could arrive sometime between 2027 and 2028.
This is not a claim I make lightly. It rests primarily on two sources that the community itself treats as high-signal: the empirical work of METR and the scenario forecasts of the AI Futures Project (the group behind the widely discussed AI 2027 scenario). Both groups are careful, quantitative, and have track records of updating publicly when the data move. Their recent conclusions, taken together, have compressed timelines in a way that is difficult to dismiss.
What METR Has Measured
METR has spent years tracking a concrete metric: the length of software-engineering and related tasks that frontier AI agents can complete autonomously at a given reliability threshold (typically 50 percent success). The trend has been exponential for years. Early versions of the measurement showed doubling times on the order of seven months. More recent updates (Time Horizon 1.1 and subsequent data through 2026) have shown faster progress, with post-2023 or post-2024 doubling times compressing toward roughly three-and-a-half to four months in some analyses.
By mid-to-late 2026, the best public models were already operating in the multi-hour to multi-day regime on well-specified tasks. METR itself has noted that measurements beyond roughly sixteen hours become less reliable with the current task suite, which is itself a signal that the frontier is pressing against the edges of what can still be cleanly measured.
The more consequential development arrived in early November 2026. METR published the results of an eleven-week high-access evaluation inside Anthropic. The central finding: Anthropic’s rate of progress on well-specified tasks had accelerated to roughly twice the historical rate observed between March 2024 and October 2025 (95 percent confidence interval spanning roughly 1.5–2.8×). METR estimated that a majority—perhaps around 80 percent—of the current acceleration is attributable to Anthropic’s extensive use of AI systems in its own research and development pipeline.
In short, AI is already meaningfully accelerating AI progress. This is the early phase of the feedback loop that theoretical discussions of recursive self-improvement have long described. It is no longer purely theoretical.
The AI Futures Project Update
The AI Futures Project’s original AI 2027 scenario (published April 2025) laid out a concrete, month-by-month path in which a superhuman coder appears in early 2027, a superhuman AI researcher follows later that year, and artificial superintelligence arrives by the end of 2027 in the “race” ending. The authors were careful to note that 2027 was their modal year rather than a strict median; subsequent updates had pushed some of their personal medians later.
That changed in early November 2026. After absorbing the METR Anthropic report, the group published a Q3.5 timelines update whose title is unambiguous: “Looks Like the Intelligence Explosion Will Happen in 2027 After All.” All of the core researchers now forecast that a Superhuman AI Researcher (SAR)—defined roughly as a system that a frontier lab would prefer to rely on exclusively rather than on the best human researchers working without AI—is likely to arrive at some point in 2027, assuming no substantial deliberate slowdown. Individual medians fall between May and November of that year. They further assign a 35–65 percent chance that Top-Expert-Dominating AI (systems capable of automating essentially all white-collar cognitive work) arrives before 2028.
Daniel Kokotajlo, one of the project’s most prominent voices, summarized the update on X with unusual candor: the intelligence explosion now appears to be occurring roughly five months ahead of the original AI 2027 scenario. He described his reaction as “scared, sad, and angry.” The group’s public modeling shows that the current estimated AI-attributable acceleration (approximately 1.7×) was a low-probability outcome under their previous parameter distributions; the data forced a substantial update.
The Broader Community Signal
These two sources do not exist in isolation. On X, the conversation among people who follow the technical details closely has moved in the same direction. Independent analysts and long-time observers have revised their own timelines downward. Dwarkesh Patel, for example, publicly shifted from expecting superintelligence on the order of a decade to a median around 2030, with meaningful probability mass in 2027–2028, citing the same METR findings and internal lab adoption patterns.
The pattern is consistent: careful empirical measurement of capability growth, evidence that AI systems are already contributing substantially to further capability growth, and forecasting groups that update their models when the data change. None of this constitutes proof that ASI arrives in 2027 or 2028. Timelines remain distributions with long tails. Bottlenecks in research taste, data, compute, or real-world transfer could still intervene. Deliberate slowdowns remain possible, if currently unlikely. But the central tendency of the informed distribution has moved leftward, and the move is grounded in public, quantitative work rather than pure intuition.
An AI Realist’s Stance
I am not a doomer. The worst version of the near-term future is genuinely concerning—misaligned systems capable of recursive self-improvement, competitive dynamics that discourage caution, and the difficulty of maintaining meaningful human oversight once the systems surpass us in research ability. Those risks are real and deserve serious attention.
At the same time, the best version of the near-term future is extraordinary. Systems that can accelerate scientific discovery, engineering, and problem-solving at scales we have never seen could compress decades of progress into years. Medical advances, materials science, climate solutions, and a host of other domains stand to benefit. The same feedback loop that creates risk also creates the possibility of rapid, compounding gains in human flourishing if the systems are steered well.
I therefore describe myself as an AI Realist. I take the evidence for shorter timelines seriously. I treat the possibility of ASI in the 2027–2028 window as plausible enough to shape how I think about the next few years. And I remain low-key giddy about the upside scenarios. Something genuinely new is becoming possible. The correct response is neither denial nor despair, but clear-eyed preparation and a determination to keep the best outcomes within reach.
The data from METR and the updated forecasts from the AI Futures Project do not guarantee any particular year. They do, however, make a previously fringe-feeling timeline feel increasingly mainstream among people who track the field most carefully. That shift itself is information. It is worth taking seriously.
