← ELI5 中文
Explained Simply

How far has AI
really come?

For anyone who's curious
No. 213How AI WorksPart 19 of 19Reader questionFact-checked Oct 12, 2026
Video2:07 · English voice (AI) · English subtitles

Can't play sound? The illustrated version is right below.

Read the transcript

The full narration of the video.

The same AI can win a gold medal at the International Math Olympiad. Yet show it a clock on the wall: ask what time it is, and it gets it wrong about half the time. So how far has AI really come? Let's hand it a report card and go subject by subject. Start with its best subject. At the 2025 Math Olympiad: out of six hard problems, one AI solved five, enough for a gold medal.

Coding is another strong subject. One lab that specializes in testing AI: measures how long a job it can finish on its own. A model they tested last summer could finish, about half the time, a job that takes a programmer over two hours. That length has roughly doubled every seven months for years, and lately it's been growing even faster.

And more and more people use it. By mid-2026, over 700 million people in China had used this kind of AI: that's about one in every two. What they do most is ask it questions. So where are the weak spots? Subject one is that wall clock. People read it right nine times out of ten. The best AI, only half the time. Subject two: it doesn't have a good pair of hands yet. Send robots into real homes to do chores, and they succeed only about one time in eight.

Subject three: it often doesn't know what it doesn't know. When it's stuck: it sometimes makes up an answer with a straight face, and it sounds completely real. Lay the report card out, and you see a row of jagged teeth, high and low. Researchers call this jagged intelligence. The tall teeth stand above most people. The short ones fall below an ordinary person. And the tall ones keep growing, year after year.

So it's most like a new coworker who has read nearly every book. Wins gold on exams, writes in a flash. But misreads clocks: can't carry a box, and now and then makes things up. When you use it, remember one thing. Let it draft and brainstorm; that saves time. But check the key numbers and facts yourself. As for when the short teeth will catch up: even the researchers can't say. Your guess: will it learn to read a clock first, or to fold laundry?

1

The gold medal and the wall clock

8:15?

The same AI can win a Math Olympiad gold medal, yet misreads a wall clock about half the time.

That's not a joke. It's real data from Stanford's yearly AI report. Today's AI is hugely capable, but its skills are anything but even.

2

A lopsided report card

AI's report cardMath OlympiadCodingReading clocksChoresKnowing limits

Give AI a report card and go subject by subject: in some it beats nearly everyone, in others it falls below an ordinary person.

3

Three things it's great at

Olympiad goldSolved 5 of 6 problems at the 2025 Math Olympiad, enough for a gold medal
Longer and longer jobsHalf the time, it finishes on its own a coding job that takes a programmer over 2 hours. That length doubles about every 7 months
700 million usersAbout one in two people in China have used it, mostly to ask it questions
4

Three things it's stuck on

Reading a clockPeople get it right 9 times in 10. The best AI, only half the time
Doing choresRobots in real homes succeed only about 1 time in 8
Knowing what it doesn't knowWhen stuck, it sometimes makes up an answer with a straight face
5

Jagged intelligence

Average person

Stand the report card on its side and you get a row of jagged teeth.

Researchers call this "jagged intelligence": the tall teeth stand above most people, the short ones fall below an ordinary person. And the tall ones keep growing year after year.

6

Like a new coworker who's read every book

Hand it over
Drafts, ideas, first versions
Fast, and often pretty good
Double-check yourself
Key numbers and facts
It can make things up that sound real
Your guess: will it learn to read a clock first, or to fold laundry?
When the short teeth will catch up, even researchers can't say. Pick one and tell us in the comments.
Samshang × Maddox
Sources: Stanford HAI, AI Index Report 2026, Technical Performance chapter (Olympiad gold, clock reading 50.6% vs about 90% for people, household robots about 12%, jagged intelligence); METR, Measuring AI Ability to Complete Long Tasks (2025) and Time Horizon 1.1 (2026); China Internet Network Information Center (CNNIC, China's official internet statistics agency), Generative AI Application Development Report (2026) (over 700 million users, adoption over 50%).
UP NEXT

Let It Rewrite Its Own Manual

When an AI keeps failing in the same place, the fix is often not the AI: let a second AI review old mistakes and suggest edits, and a human approves.

No. 8 · 4 minKeep going →
← Back to ELI5