AI comes in degrees: narrow, general, super

"AI" covers everything from a spam filter to systems that write software. That's like using one word for a bicycle and a jet. To think clearly, you need the degrees.

I sort AI on two questions.

How many fields is it good at? One, many, or all.
Can it go past what it was trained for? Can it solve problems nobody showed it? How hard: simple, complex, or "impossible", meaning still unsolved by anyone? And can it conceive big new problems by itself, and solve them, in ways that make lives better and longer?

01The map

Where a system sits: breadth × how far past its training

AI. Great at what it was trained for. Predictable.
AGI. Reaches for solutions it wasn't trained for.
ASI. Solves what humans can't, invents new science and industries.
Today's best chatbots and agents (my read, 2026)

Where a system sits: breadth × how far past its training
	Only what it was trained for	New problems: simple	New problems: complex	New problems: unsolved by anyone	Conceives new problems and solves them
One field	AI	AGI (narrow)	AGI (narrow)	AGI (narrow)	—
Many fields	AI	AGI (Today's best chatbots and agents (my read, 2026))	AGI	ASI	ASI
Every field	AI	AGI	ASI	ASI	ASI

In words:

AI. Great at one field, or at many, but only inside the edges of what it was trained for. The answers are predictable. A chess engine. A grammar checker. Most chatbots, most of the time.
AGI. Great at a field and able to reach for solutions to problems it was never trained to solve. That's the crossover I described in What is intelligence?: carrying knowledge from one place to crack a problem in another.
ASI. Great across many or all fields, able to solve problems no human has solved, and able to come up with new problems no one gave it. This is AI that can create new science and new industries. It takes us to a future we can't picture yet.

02How the labs measure it

Researchers have their own ladders, and they fit this map well.

In 2023, Google DeepMind proposed Levels of AGI. It scores how good separately from how general:

Levels of AGI (Google DeepMind, 2023)

Levels of AGI (Google DeepMind, 2023)
Level	How good	Narrow example	General example
1 · Emerging	Equal to or a bit better than an unskilled person	Old rule-based systems	ChatGPT, Gemini, Llama 2 (as rated in 2023)
2 · Competent	At least the 50th percentile of skilled adults	Siri, Alexa	Not yet achieved (2023)
3 · Expert	At least the 90th percentile	Grammarly, DALL·E 2	Not yet achieved
4 · Exceptional	At least the 99th percentile	Deep Blue, AlphaGo	Not yet achieved
5 · Superhuman	Better than every human	AlphaFold, AlphaZero	Superintelligence: not yet achieved

Superhuman AI already exists, in narrow form. AlphaFold predicts protein shapes better than any human team. What doesn't exist yet is superhuman and general.

In 2025, a team of about 30 researchers led by Dan Hendrycks, with Yoshua Bengio among them, tried to score AGI directly against a well-educated adult across ten areas of thinking.

AGI score, Hendrycks et al. (2025)

Ten areas of thinking, each worth 10%.

AGI score, Hendrycks et al. (2025) (%)
Item	Value	Note
GPT-4	27%
GPT-5	57%
Well-educated adult (the bar)	100%	the benchmark

Two lessons. The jump is fast: one generation roughly doubled the score. The profile is jagged: in the authors' words, models are strong on knowledge but have "critical deficits" in basic machinery, especially long-term memory. That's why a model can write a sharp essay and still forget what you told it yesterday.

03Why the degrees matter

A narrow system that goes wrong does narrow damage. A bad chess move costs a game. A general system can go wrong in any direction, because it can act in any direction. A superintelligent one that goes wrong might be one we can't correct.

So the safety question grows with every step right and down the map. That's alignment.

04The short version

Sort AI by breadth (one, many, all fields) and by how far past its training it can go.
AI stays inside its training. AGI reaches past it. ASI solves what nobody has, and invents new problems.
On one 2025 measure, models went from 27% to 57% of "well-educated adult" in a single generation.

05Sources

Morris, M. R. et al. (2023). Levels of AGI for Operationalizing Progress on the Path to AGI. Google DeepMind. arXiv:2311.02462.
Hendrycks, D. et al. (2025). A Definition of AGI. arXiv:2510.18212.
Jumper, J. et al. (2021). Highly accurate protein structure prediction with AlphaFold. Nature 596.

All posts

Appointment setting3

Copywriting3

AI services2

Deep Think AI13

Get the next markup by email.

One real piece of copy per issue, rewritten line by line: a cold email, a web page or a newsletter.