Two jagged minds
We keep asking whether AI can do everything we can. That question hides most of what it can do.
The chart, the evidence table and the source list need JavaScript, which isn't running in this browser. The text still makes the argument on its own.
How we usually see it
Each spoke is one ability. Blue is the best AI reported at each date. AGI tests ask if AI matches a capable adult (amber) at everything, so nothing past amber counts.
Take the ceiling off
Same data, no cap. Each ring is now a rung on a ladder of people: capable adult, professional, best human. The hatched outer band is ability no person has.
It's hard to appreciate an ability you don't have, so AI's biggest leads are the easiest to discount.
Count what the test leaves out
AGI tests skip some abilities by design. Add four back, two where AI leads (languages, coding) and two where people do (physical skill, long projects). The average barely moves, because the gains and losses cancel out. But both minds get more jagged.
Nobody is a circle
The capable-adult ring is an average, and nobody is average at everything. Real people have spikes too. Pick a sketch:
Flip the ruler
Whoever you measure against looks like a perfect circle. So make AI the ruler. It's the blue circle; each ring in or out is one rung.
Now people turn jagged: ahead on learning and physical skill, far behind on speed, knowledge and languages.
Neither picture is the true one. They're two different kinds of mind.
Where it still breaks
A total can hide a fatal flaw: an AI that can't learn on the job fails at its weakest point. The thin blue ring is AI's score. How much should weak spots count?
Superhuman things we stopped noticing
Machines have been passing people for 80 years. Each time, the feat stopped counting as intelligence. In 2022, professional forecasters gave AI a 2.3% chance of Olympiad gold by 2025. It happened in July 2025.
“Intelligence is whatever machines haven't done yet.”
The strongest objection: maybe the weakest link is what matters
When the steps of a job depend on each other, one weak step can sink it. Economists call this the O-ring problem, after the seal that failed on the Challenger. Judged task by task, AI now matches professionals most of the time. Judged on whole projects, it still mostly falls short.
of single, well-specified work tasks where GPT-5.2 matched or beat experienced professionals in blind grading (Dec 2025, reported by OpenAI).
of whole real freelance projects the best AI agent completed acceptably in Sept 2026, up from 2.5% a year earlier.
The longer the task, the less reliable. The best measured AI succeeds half the time on software tasks that take a skilled person about 17 hours. To succeed 80% of the time, the tasks have to be about 3 hours or shorter.
The edge is invisible from inside. Consultants using AI on a task just outside its abilities were 19 percentage points less likely to get it right than those working without AI.
A mind comparable to ours, shaped very differently
Two jagged minds · Ivan Lay · Evidence as of September 2026
So, is it AGI? By the AGI test, not yet. AGI asks whether AI can stand in for a person across the board, and 2026 AI still can't, mostly because it can't learn from experience or stay reliable over long, messy work.
But judge it by the whole shape of what it can do, not by its weakest point, and AI is already a mind comparable to ours. In four years it went from fumbling multiplication to research mathematics, and from reading nothing to reading a novel in seconds. Even Google DeepMind's 2026 AGI framework notes that AI “already possess[es] some capabilities not found in humans.”
So “is it as good as us yet?” is the wrong question. Ask what it's good at, and what we're good at. The answer changes every few months.
Easy for you, hard for AI
Easy for AI, hard for you
Every ability, without the radar
Radar charts can mislead, so here is the same data as one line per ability. Open a row for its sources.
Methods and limits
- The scale. A ladder anchored to people, adapted from Google DeepMind's “Levels of AGI”: 1 is a capable, well-educated adult, 2 a professional, 3 the best humans alive; beyond 3 is ability no person has. The chart gives the zone inside the capable-adult ring extra room so gaps below human level are as visible as spikes above it.
- The rungs are ranked, not measured. Nobody knows whether the step from capable adult to professional is as big as the step from professional to best human. Averaging treats them as equal, so the page gives averages in ladder words rather than exact numbers.
- The abilities. The ten core abilities are the domains of Hendrycks et al., “A Definition of AGI” (2025). We didn't choose them, which makes cherry-picking harder. The four extras are ones that framework leaves out: two favoring AI, two favoring people.
- The placements are our judgment from the evidence above, each with a confidence rating. Reasonable people could move any of them by half a step. “Frontier AI” means the best publicly reported results at each point, across labs. Labs report internal models beyond these, so 2026 is probably an underestimate.
- The people other than “a capable adult” are illustrative sketches, not measurements. “The best human at everything” describes nobody who has lived.
- The flipped view draws AI as a circle and places each human ability by how many rungs it sits above or below AI. Same numbers, different ruler.
- Scores are averages, not areas. Drawn area grows with the square of the values and changes with axis order. The dial uses a generalized mean, following Fourati (2025): “Everything counts” is a plain average, “Weak spots count more” the geometric mean, and “Weakest link only” the minimum.
Our AGI-test view against the published one
The capped view scores abilities out of 10 up to the capable-adult line, as Hendrycks et al. do. We're more generous, mostly on speed: their test times quick, simple tasks where a “thinking” model is slow; we weigh how fast AI reads and writes. The chart gives the capped score in ladder words rather than as a percentage, and on those words the two scorings agree: well below a capable adult in 2023, and below one in 2025.
Where every number comes from
Prototype, September 2026. Several 2026 results are recent and under review, notably the Navier–Stokes construction, which the Clay Institute calls “apparently settled” but has not accepted. Research notes behind every number are in the project's research folder. Corrections welcome on GitHub.