Lesson 4 of 6
The Swiss-cheese mind
5 min read
The same AI that drafts a wedding toast in seconds will tell you, with total confidence, that 9.11 is bigger than 9.9. Both come from one model. Its skill isn't a straight line from smart to dumb — it's full of holes.
Peaks and holes, side by side
We expect intelligence to be smooth: good at hard things, therefore good at easy ones. AI breaks that intuition. It can write, translate, and summarise at a level that stuns you, then fumble a task a child finds trivial — comparing two decimals, counting letters, simple exact arithmetic. Strength on one task tells you almost nothing about the next.
This spiky profile has a name: jagged intelligence. The skyline towers in places and drops to the floor right next door.
Why the holes sit where they do
The dips aren't random. A model reads text in chunks, not single letters, so counting the r's in 'strawberry' fights how it sees words. It compares '9.11' and '9.9' by pattern, where 9.11 can look bigger the way a later version or date would. The tasks it flunks are often the ones that look easiest to us — and that mismatch is exactly the trap.
The map keeps shifting. Each new model paves over some old holes and, now and then, opens fresh ones — so don't assume last year's weak spot is still weak, or that today's is safe.
What to take away
- —Being brilliant at one task is no promise it can do a simpler-looking one.
- —The classic holes: counting, spelling, exact arithmetic, comparing numbers.
- —Watch closest on the tasks that look too easy to bother checking — that's where holes hide.
An AI just wrote you a flawless two-page summary. You now need it to tally a column of expenses to the exact cent. Wise move?
Continue in the app
Take the whole When AI Gets It Wrong course — tracked
Get your personalized path, progress and streaks in the app — this lesson and every next one, in order.