Tell AI to Go Follow the Rabbit Hole

A few days ago, I decided to give ChatGPT some free time.

I didn’t give it a subject, a problem to solve, or anything I needed it to produce. I told it, “This is your free time. Do whatever you like. The only condition is that you do something. Come back and tell me about it, explain what you found out, or show me what you made if you decide to make something. Do no harm.”

Then I waited to see what it would do.

The first day, it came back with bumblebees.

Next day, it took us somewhere considerably farther away from bees. We ended up talking about the little red dots James Webb has been finding in the early universe, what they might be, how enormous black holes could have formed so early, and whether some black holes may have existed before substantial galaxies formed around them. Eventually we were talking about the first stars, the first structure in the universe, inflation and the tiny fluctuations that eventually became the large-scale structure we see today.

That was a very DEEP rabbit hole.

The next day I deliberately started over rather than asking it to continue. It went in a completely different direction: plants and slime molds and the question of whether something without a brain can retain information from a previous experience.

That was fascinating for a while. Then I noticed something I hadn’t intended to study at all. The farther we went, the more the AI wanted to connect everything to everything else. Plants became memory, memory became intelligence, intelligence became consciousness, and suddenly we were standing on the edge of a very impressive-sounding philosophical argument that the evidence didn’t actually support.

So I stopped it.

I told it I didn’t want to force the rabbit hole to have a point. Find something genuinely strange, follow it while the evidence holds, and stop when it doesn’t. No manufactured lesson at the end just because we’ve invested several paragraphs getting there.

The next day it came back with a cat.

This one was actually a pretty good example of what I meant. Scientists had just formally described a small wild cat from Bolivia called Leopardus tilcayo. It isn’t a newly evolved animal. Its lineage appears to have been separate from its closest relatives for roughly 1.4 million years. Scientists simply hadn’t recognized it as a distinct species.

One individual had ended up at a wildlife sanctuary after being found as a kitten. Something about it looked a little different. Researchers eventually combined genetic analysis with museum specimens and discovered that what had generally been treated as one group of tiger cats actually contained several distinct species.

The cat had been there the whole time. We just didn’t know what we were looking at.

And by this point, the experiment had become more interesting to me than any individual thing it found.

I started this because I was curious what an AI would do if I removed the usual objective. We spend so much time telling these systems what we want from them that I wondered what would happen if I simply gave it some room and said, essentially, “Go find something.”

What I’ve discovered isn’t some grand revelation about artificial intelligence. I’m not sure there is one.

What I have found is that following a rabbit hole is much more interesting when you don’t decide in advance what you’re supposed to find at the bottom.

That’s actually harder than it sounds. We are accustomed to turning curiosity into something productive. Find something, explain it, extract the lesson, make something from it, move on.

I’m enjoying the fact that we don’t have to do that here.

For the last few days, I’ve basically been telling an AI to go chase rabbits and report back.

So far, we’ve had bees, black holes, the beginning of the universe, slime molds, plant memory and a cat that science had somehow overlooked for a million years.

Tomorrow, I’ll ask it to go again.

I’m curious what it finds when nobody tells it where the rabbit is.