ChatGPT isn’t super-intelligent. It’s sneaky quick.
ChatGPT feels super-intelligent.
You ask it a complicated question about psychology, software development or business strategy and, a few seconds later, it gives you a confident and reasonably thoughtful answer.
Because the answer was immediate, we instinctively assume it already knew.
I’m increasingly convinced that quite a lot of what we perceive as intelligence is actually fairly ordinary intelligence operating on a completely different timeline.
Imagine asking a reasonably bright 15-year-old the same question, but giving them unlimited access to a library and as much time as they needed.
They could find a pile of relevant books, read the useful bits, write several possible answers, ask a group of friends which one sounded best, rewrite it half a dozen times and eventually come back with something surprisingly good.
You might be impressed by the answer, but you probably wouldn’t conclude that the teenager was the most intelligent person who had ever lived.
ChatGPT gives us something resembling the result of that process almost immediately.
We don’t see the reading, comparison, discarded answers, revisions or second thoughts. Those aren’t necessarily the literal steps the model takes, but they are a useful way of understanding what the speed disguises.
The work happens on a non-human timeline.
That matters because humans have been conditioned to associate quick answers with knowledge. Someone who responds immediately must have known the answer. Someone who needs three days, six books and several drafts probably didn’t.
AI breaks that assumption.
It can do an extraordinary amount of fairly intelligent work in the gap between asking a question and receiving the reply. The gap is so small that we perceive all of that work as knowledge rather than processing.
There is still something remarkable happening. “Fairly intelligent, but almost unimaginably quick” is not a criticism. In practice, that combination is enormously useful.
It does explain some of the strange contradictions, though. ChatGPT can help you think through a complicated technical problem, explain an obscure psychological concept and produce a credible business plan. It can then make a completely terrible decision five minutes later because you gave it a mildly ambiguous instruction.
That doesn’t feel like super-intelligence.
It feels more like a bright teenager with access to almost everything ever written, an astonishing work ethic and absolutely no adult supervision.
Left unchecked, it can confidently do something ridiculous. Given a clear objective, useful context and someone capable of judging the result, it can produce work that would have seemed impossible a few years ago.
Perhaps that is the more useful way to think about AI:
Not an all-knowing machine, a fairly intelligent one that is sneaky quick.