Why I expect superintelligence sooner rather than later — and why that doesn’t make me a doomer
Let me say up front what I don’t have. I have no insider information. Nobody at a frontier lab texts me. What I have is a feed — the corner of Twitter where researchers, evaluators, forecasters and the occasional overcaffeinated founder argue about AI in real time. I’ve been reading it daily for a long time.
Out of that feed, a feeling has built up. It’s the sense that sometime between 2027 and 2028, we could be looking at artificial superintelligence — systems that outperform the best humans at nearly every cognitive task that matters, including the task of building better AI.
A feeling isn’t a forecast. So this essay does two things. It checks that feeling against what the experts I trust are actually saying. Then it explains why, even after that check, I land where I do: not as a doomer, but as an AI Realist. On my better days, I’m a low-key giddy one.
What the experts actually say
Here’s the awkward part: the two groups I lean on don’t quite say “2027–2028.” One of them never made a date forecast at all. The other has moved its date back.
METR measures; it doesn’t prophesy. METR’s core number is the “time horizon”: the length of a task, in human working time, that an AI agent can finish on its own about half the time. Their January 2026 update, Time Horizon 1.1, found that horizon doubling roughly every seven months across 2019–2025. Since 2024, though, it has doubled about every three months. That’s an acceleration, not a slowdown.
Then the ruler ran out. In March 2026, METR put Claude Mythos Preview at a time horizon of at least 16 hours. That is the ceiling of what their test suite can measure, with a confidence interval from about 8.5 to 55 hours. METR’s own caveat matters here: a 16-hour horizon doesn’t mean the model can do every 16-hour job. The tasks lean heavily toward software, machine learning and cybersecurity.
The AI 2027 team has stretched its own clock. The scenario that named my feeling is AI 2027 from Daniel Kokotajlo and colleagues. It imagined coding agents kicking off an intelligence explosion, with superintelligence around early 2028. Since then, the authors have revised. By early 2026, Kokotajlo’s median for superintelligence was 2031, and Eli Lifland’s was the mid-2030s. Their updated model points to 2034. Kokotajlo has also said their timelines were longer than 2027 even when they published. The year in the title was his single most likely year, not a 50-50 bet.
So if I’m honest, the experts I cite are less aggressive than my gut. The forecasters moved their date out. The measurers are saying, in effect, “we can’t measure this anymore.”
Why the feeling survives the fact-check
I could stop there and say the vibes outran the evidence. I don’t think that’s the full story. Four things keep my gut where it is.
First, the scenario has been right about direction. One scorecard of AI 2027 found its qualitative calls holding up. Agents arrived unreliable, then got better. AI started speeding up AI research. Software jobs felt it first. Where it missed was speed, running roughly 12 to 18 months too fast. Shift its milestones by that much and you’re looking at 2028 or 2029, not 2034.
Second, the instruments are struggling to keep up. When a benchmark built to measure progress for years hits its ceiling within months of its release, that’s a signal in itself. METR says it is working on harder tasks. Until those exist, nobody can see the next part of the curve clearly. That includes the forecasters whose models feed on that data.
Third, “slower” still means soon. Kokotajlo’s revised median of 2031 is five years away. The 2034 horizon is eight. Five years ago, most serious people put superintelligence somewhere between “decades” and “never.” The argument now is between “three years” and “eight,” and that’s a huge shift by itself.
Fourth, “ASI” probably won’t arrive on a single day. Progress is jagged. A system can be superhuman at code and math while being clumsy at physical-world common sense. My 2027–2028 feeling is really a feeling about when the superhuman parts start to dominate the headlines, the economy and the politics. That could easily come before a forecaster would check the box marked “superintelligence.”
I’ll own the limits of this. I’m reading curves and moods, not running models. If the ruler gets longer and the line flattens, I’ll say so in this space.
An AI Realist, not a doomer
The AI conversation tends to split into two camps, and I don’t belong to either one.
- The doomers treat superintelligence as an extinction event with a release date. For them, the only sane move is to stop.
- The dismissers treat the whole thing as hype, a chatbot bubble that will pop on schedule.
I’m an AI Realist. A Realist accepts the trend lines and doesn’t need superintelligence to arrive on a particular Tuesday to take it seriously. A Realist also doesn’t assume that something smarter than us must want us dead. My own p(doom) is low. My odds that ASI changes nearly everything about how we live are close to certain. Those two numbers don’t contradict each other. Most of the danger I see is quieter than Skynet: a slow, comfortable handover of medicine, schooling, love and government to systems we don’t understand and can’t vote out.
Realism has duties, too. If this is coming in years rather than decades, then legislatures, courts and ordinary people need to start having the hard arguments now. Who owns a superintelligence? Who answers for it? What does it owe us, and what do we owe it? I’m pro-AI, and I still want brakes on the car. Wanting brakes isn’t fear. It’s what you think about before you get on the highway.
The low-key giddy part
That’s realism at its gloomiest. On a good day, the same evidence makes me a little giddy, and I don’t think I should apologize for that.
Think about what a system that can work alone on a multi-day research task could actually do, if we point it at the right problems:
- Medicine: drug candidates screened in weeks instead of years, and rare diseases finally worth someone’s attention because the attention is cheap.
- Science: open problems in math, materials and energy taken apart by minds that never get tired or bored.
- Education: a patient, expert tutor for every kid on Earth, whether they’re in Fairfax or a village with one schoolhouse.
- Everyday life: the paperwork, red tape and phone trees that eat our lives handled by someone, or something, that doesn’t mind.
None of that is guaranteed. Each item on that list could go wrong in a way I’ve written about before. But “it could go wrong” is true of fire, flight and the printing press. The cool things are as real as the scary things, and I’d rather be honest about both than practice a dread I don’t actually feel.
Where that leaves me
My prediction comes down to this. The people who measure AI can barely keep a ruler on it. The people who forecast it have pushed their dates out, but only to early-to-mid next decade. My gut says the transformative stretch starts sooner than their medians, around 2027 or 2028. It won’t come as a single moment, but as a run of capabilities that stop feeling like software.
I could be wrong about the year. I doubt I’m wrong about the decade. Either way, it isn’t time to panic or to shrug. It’s time to pay attention, argue in public about the rules, and allow ourselves a little hope.
Sources
AI 2027: How Accurate Has the Superintelligence Forecast Been So Far? — Better Stack, September 2026
Time Horizon 1.1 — METR, January 2026
Claude Mythos Shows 50% Time Horizon of 16+ Hours on METR Benchmark — OfficeChai
AI 2027 — AI Futures Project, April 2025
AI expert delays timeline for its possible destruction of humanity — Irish Examiner
