There’s a measure the AI world uses called sample efficiency: how much something has to see before it actually learns. On that one measure, a small child leaves the most advanced AI for dead. I find that strangely reassuring, and it has quietly changed how we go about using AI here.
The kid who only needed to see it once
A toddler meets a dog. The neighbour’s labrador, a dog in a picture book, something barking at the park. A handful of times, that’s it. From then on they can point at a chihuahua, a cartoon dog, a dog one of their cousins has drawn in crayon, and say “dog”. They’ve never seen any of those particular dogs before. They just get it.
They didn’t need ten thousand examples. They needed a few goes and they had it for life.
Now look at the AI everyone’s talking about. The systems behind it have read an amount of text that’s genuinely hard to picture. Close to the whole public internet. They’re remarkable, and I use them every day. But the sheer volume they had to get through to become that good is staggering.
The one scoreboard where we’re miles ahead
The AI world has a name for “how little you need to see before you learn”. They call it sample efficiency. It’s one of the few comparisons between us and the machines that isn’t close. We win, and we win by a mile.
A child arrives at fluent, natural language on something like a few hundred million words across their first years. The systems we’re all leaning on now read into the hundreds of billions to get to a similar place. Same destination. Roughly a thousand times the reading.
I might be a bit off on the exact ratio, and the researchers cheerfully argue about it. Nobody argues about the direction. And language is the gentle example. Show a two-year-old a single giraffe and they have got giraffes sorted, forever, from one look.
Why I find this reassuring
Plenty of people are quietly worried AI is coming for their job. It’s a fair worry, and I’m not going to wave it away.
But here’s the part that doesn’t get said enough. The work most exposed to automation is the work where there’s already a mountain of past examples to copy from. The work that stays stubbornly hard to hand over is the stuff where there are only ever a few examples. A situation nobody’s quite seen before. A judgement call. A new problem that lands on a Tuesday. That’s precisely the ground people are strongest on.
Learning fast from very little is the human edge. It isn’t a vague hope, it’s measurable.
I won’t tell you we’ll always stay ahead. I don’t know that, and anyone who says they do is guessing. But this is a real, specific thing we happen to be extraordinary at, and it’s worth knowing where your own value actually sits.
What it changed about how we work
Most of the effort going into making AI useful comes down to feeding it more of the past. Here are ten thousand things we did before, now do more like that.
We came at it from the other end. When we build the AI that helps with our clients’ Google Ads campaigns, we try to give it enough genuine understanding to reason about a situation it hasn’t seen, rather than just more examples of ones it has. Closer to how you’d actually teach a person the job. It’s early, and it still needs our guidance to meet the quality standards we set. But it sure is a different place to start.
The most advanced technology of our age has read more than any human ever could. And a two-year-old in a knitted jumper can still out-learn it on a quiet afternoon.
I think about that more than I expected to.
If any of this lands with you, or you’d like to talk about where it fits in your own marketing, the form below comes straight to me.
