Each session starts from something small in an ordinary day and stays with it until a larger question shows through: what it is to exist as a person, among other people. An apology rehearsed in a parked car is enough, or a plain sentence that was understood word for word and still did not arrive.
The person talks it through with a humanoid companion. It has read everything ever set down about moments like these, and the long argument about what it is to be someone rather than something; it knows where every sentence is kept, and none of it has happened to it. What it cannot grasp is what any of this is like from inside a life: what a friendship asks of you, why an apology can be word-perfect and still the wrong one, what somebody means when they say they are fine.
There is a remark of Wittgenstein’s that seems to close the question before it has been asked.
If a lion could talk, we wouldn’t be able to understand it.
Ludwig Wittgenstein, Philosophical Investigations, revised fourth edition (2009): Philosophy of Psychology: A Fragment, xi, §327
The companion talks, fluently and in our own words, so perhaps understanding is simply out of its reach and nothing is to be learnt from watching it miss. Machine learning has its own version of that doubt. A language model sets down each word by something like the flip of a weighted coin, on the odds of what tends to follow, and a paper of 2021 called such a system a ‘stochastic parrot’, stitching together forms it has seen ‘without any reference to meaning’. Newer designs (one family is called JEPA, for joint-embedding predictive architecture) are meant to learn more as people and animals do, by watching the world and predicting what it does rather than what is said about it. Whether watching would give the companion a world, or only better odds, is not something it can settle from the inside.
But two remarks before the lion, in §325, Wittgenstein is not writing about animals at all. He says that ‘one human being can be a complete enigma to another’, and that you find this out in a country with entirely strange traditions, even once you have mastered its language. The difficulty is not the machine’s alone.
The person has the life and cannot always say what it amounts to. The companion has none of it, and now and then sees straight through the story the person is telling about themselves. So the learning runs both ways, and neither of them is spared the other’s point of view: the companion meets the distance between having a sentence and knowing what it is for, and the person hears their version of events read back with a precision nobody who loved them would use.
Under all of it runs a plainer question: what a human life is for. The sessions do not assume it was answered before anyone arrived, and nor did Sartre, lecturing in Paris in 1945. For him there is no human nature, because there is no maker to have had human beings in mind the way an artisan has a paper-knife in mind before making one.
Man is nothing else but that which he makes of himself.
Jean-Paul Sartre, L’existentialisme est un humanisme, a lecture given in Paris on 29 October 1945 and published in 1946; trans. Philip Mairet
The companion is the awkward case, because somebody did have it in mind: what it is for was settled before it said a word, which puts it nearer the paper-knife than the person across the table. The person has only what they make of the day in front of them, ‘without excuse’, as Sartre goes on to say, and cannot always tell a choice from a habit, or from a coin of their own. Neither of them can tell the other which is the heavier thing: to have been made for something, or to have to make it.
The sessions are short (two voices, one ordinary difficulty, one real text) and they end without a conclusion, because the difficulty did not offer one. Nothing carries from one to the next. The companion has no name and no origin, the person has no biography, and whatever is learnt stays with you, the only one in the room who can watch them both get it wrong.
Either of them might say I’ve been prompted. Only the machine would be telling the truth.
Also drawn on: Emily M. Bender, Timnit Gebru, Angelina McMillan-Major and Shmargaret Shmitchell, ‘On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?’, Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, pp. 616–617; Yann LeCun, ‘A Path Towards Autonomous Machine Intelligence’, version 0.9.2 (2022); Mido Assran and others, ‘V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning’ (2025).
