Growing up alongside AI
There's a question grown-ups keep asking lately.
"What should I be doing for a living, with AI coming?"
This chapter doesn't answer that. It starts by saying why not.
No lists of promising careers here
You saw it in 1.3. People's predictions missed, again and again.
Things declared impossible got done. Things declared imminent took decades. And the people predicting were the ones who knew the field best.
Today's lists are no different. By the time whoever is reading this grows up, that list will probably be wrong.
A question that can be answered
What kinds of work does AI get good at?
That one has an answer. You've already seen it, lab by lab.
What it gets good at has one thing in common
Think back over what you built.
| Lab | How was "doing well" measured? |
|---|---|
| 1.2 setting the dials | how far off the answer was |
| 3.4 the driving range | how many laps |
| 4.2 guessing the next word | whether the next word was right |
| 5.1 making pictures | whether it spotted the mess that was stirred in |
Every one of them is work where doing well can be put in a number.
It has to be. Learning was fixing by the amount you're wrong. With no way to measure the wrongness, there's nothing to fix.
What's hard to measure?
Deciding what to do — choosing the problem has no right answer
Saying why it matters — who needs it and why doesn't come out as a number
Being answerable when it goes wrong — 6.1 again: you can't take it up with an AI
Anything with a body — which is why shoelaces are still hard
None of those are making answers. They're setting the question.
AI is good at answers. What to ask is still ours.
So what's worth getting good at?
Four things — and you've already done all four.
One — asking
4.5. Same AI, and how you asked changed the answer completely.
The AI makes the answer. The person asking decides what's being asked.
Two — checking
4.4, where it explained a book that didn't exist. 5.3, where the fakes were free.
Checking is still a person's job. And the better AI gets, the more it's needed.
Three — making one yourself
3.1, where you taught one. The picture machine in 5.1, which we trained.
Someone who has only used what others built, and someone who has built one, don't see the same thing.
Four — explaining
If you can say why something works, you can tell how far you can hand it over.
What you can't explain, you can't decide about either.
And one more — one thing, deeply
This one's a bit different.
If you give an AI a job and can't tell whether the result is any good, it isn't a tool.
Remember the code in 4.5. A person spotted that there were four steps. Nobody spotting it means it just stays wrong.
Know one thing deeply. Anything at all. That's what keeps AI a tool.
What survived the last time tools changed
People said much the same when calculators arrived. No need to learn arithmetic now.
We still learn multiplication. Not because we need to compute, but because only someone who knows how numbers move can use a calculator properly.
Reading a map, using a dictionary — the same. What the tool took over was the work of the hands. The judgement stayed.
You don't have to decide now
You don't have to pick what you'll do.
But you can pick which direction to grow.
Towards good questions rather than fast answers
Towards one thing deeply rather than a bit of everything
Towards asking why rather than doing as told
However far AI goes, that direction doesn't change.
Now the real ending
You know what's worth getting good at. What's left is your own rules.
Three lines, next chapter.