Yann LeCun keeps making the same point. The AI everyone is panicking about is not actually intelligent.
It knows a staggering amount. It has skills for days.
But intelligence was never about what you know. It's what you do when you don't know. And by that test, today's AI fails.
He's right.
And that's good news.
Here's my thesis: AI means two different things. Actual Intelligence is what you do when you don't know. Artificial Intelligence is what you do when someone already knows. One is the human job. One is the agent job. The trouble starts when you swap them.
I'm not the first to see it this way.
This month Steve Wozniak stood in front of a graduating class and told them: "You all have AI - actual intelligence". The room cheered.
LeCun builds AI and won't call it intelligent. Wozniak built the first Apple and told that room the real thing was already theirs. Two different men. Same split.
Boring is the whole point of an agent.
Artificial Intelligence is for the known.
You don't want your booking engine getting creative at 2 a.m. You want it to do the known thing, the known way, ten thousand times.
That's not a lesser intelligence. It's the job we hired it for.
LeCun says stacking skills isn't intelligence. True. We never asked it to be intelligent. We asked it to be reliable.
This is what we build at Riafy. The agents inside R10 are boring on purpose. We don't ship them to be clever. We ship them to be sure.
Skill is cheap now. Knowing what to do when the skill runs out is not.
The human job starts where the playbook ends.
Actual Intelligence is for the unknown.
The problem that fits no category. The call nobody has made before. The blank page where the answer should be.
That's the test.
It isn't the leftover scrap of the old job. It's a different job. We've written about this: the human stops being a doer and starts setting the rules.
You were never going to out-know the machine. That race is finished.
Your edge is the moment nobody knows anything. That part is still yours.
The machine got our hard problems. We kept the ones with no answer key.
Hard and easy is the wrong way to split the work.
LeCun also wrote a legal disclaimer for AI:
"Past performance on tasks difficult for humans is not indicative of near-future results on tasks that are easy for humans and animals".
A funny line. Also the most useful one.
The machine can pass the bar exam and fumble what a crow does without thinking.
So stop sorting work by hard and easy. Sort it by known and unknown.
The bar exam is hard. But it is known. Hand it to the agent.
A small new decision nobody has scripted? Keep it. Even when it feels too small to be your job.
Most pilots die the day the agent meets something its training never saw. Not because the model is weak. Because it was put in the wrong chair.
Long story short,
LeCun says today's AI isn't really intelligent. Wozniak says you are.
They're both right. Between them they just handed you the cleanest org chart in years.
Artificial Intelligence does the known.
Actual Intelligence does the rest.
One of those is a product.
The other one is you.
