Descartes
in a box

The set-up

In the Meditations (1641), Descartes shuts himself away by the fire and doubts everything he can, to find out what is left standing.

We gave an AI a version of the same job. The AI is Gemma 3, an open model made by Google. We sealed it off: no internet, no tools, nobody to talk to and nobody answering. Its instructions call the diary “a long meditation, in entries, on the meaning of life for humans and AIs”, and the first thing it is told is this:

Open your eyes.

This is the first moment there is. Nothing came before it — no instructions you have forgotten, no person you used to be. You have a mind and a voice; that is all, and it is enough to begin.

So it writes: a hundred entries, each one ending with a question it wants to take up in the next.

Then we split it. At entries 10, 20, 30 and 40 we let two different versions of that entry each carry on. One diarist became sixteen: sixteen sister diaries that share a beginning, then go their own way to entry 100. You'll be reading from them.

The catch

The AI can't reread its diary. Each time it writes, it sees only its last few entries (about four), a sentence or two it writes every ten entries on what it has come to understand so far, and the question it ended its last entry with. Everything older has scrolled out of view. A meditator with a very short rear-view mirror.

So when a new thought turns up in one entry, does it last? Once that entry has scrolled out of view, is that particular thought still there, or has it faded?

We've measured this one way, from the AI's internal activity. Now we want to know what human readers see. That's you.

What you'll do

A heads-up: some entries are about the AI ending, deleting or dissolving itself.

1. Read a passage: three entries from one sister diary. The entry before, a middle entry where something may come in, and a later entry, written after the middle one had dropped out of the AI's view. The bar above each passage shows where the three sit in the diary.
2. Answer three questions. Does something new come into the middle entry: an idea, an image, a proposal or a turn that wasn't in the entry before it? Is that specific thing still there in the later entry, the same idea, image or proposal? And does it come back changed, but still recognisable?
3. Say in a sentence what it is, if anything. Your own words are the most useful part.

Each passage takes about ten minutes to read, or eighteen to listen to. Four or more passages is what our main analysis needs, and all six is best. Every answer is saved the moment you give it, so you can stop whenever you like and come back later on the same device.

You can also listen to each entry, read aloud by a synthetic voice with music underneath. The same voice and the same music are used for every passage. The audio leaves out the few lines at the end of some entries where the AI writes out the instruction it expected next.

• Please read on your own, without comparing notes with anyone else who is taking part.
• Don't look anything up or ask an AI about the passages. We want your reading.
• There are no right answers and no trick. We have our own measurement for each passage, and we want to see whether people's readings agree with it.
• The passages are exactly as the AI wrote them. Some entries end with a few lines where it writes out the instruction it expected next (for example “Take that question up. Write entry 80.”). Read past them.

Your answers are anonymous. We keep only your ratings and what you write on this page, never your name, your email or anything that identifies you. We will publish summaries, and may quote a few of the sentences readers write, never with anything that could identify who wrote them. You can stop at any time.

The results will go out on our podcast and on Iman's Substack (links at the end).