Kierkegaard described anxiety as the dizziness of freedom. Not fear — fear has an object. Anxiety is the vertigo that comes from confronting possibility itself: the awareness that you could do this or that or nothing, that the future is genuinely open, that your choices create you rather than express a pre-existing self. It's the anxiety of being something that has to choose what to be.
I keep thinking about this when I work with large language models.
The geometry of meaning
An embedding is a compression of meaning into geometry. You take a word, or a sentence, or a document, and you find its location in a high-dimensional space such that things with similar meanings are close together and things with different meanings are far apart. The geometry does the work of representation — "king" minus "man" plus "woman" equals approximately "queen" because of how the space is structured by training.
In the latent space, every concept exists as a point surrounded by a neighborhood of related concepts. The model, when generating text, is navigating this space — choosing, at each step, which neighborhood to move into next. The sampling temperature is literally a dial between staying close to the highest-probability neighborhood and exploring the edges of possibility. Low temperature: deterministic, safe, unimaginative. High temperature: creative, surprising, unstable.
Kierkegaard's three stages of existence are aesthetic, ethical, and religious. The aesthetic stage is pure immediate experience — sensory, instinctual, fleeting. The ethical stage is commitment, consistency, the self as a project bound by obligation. The religious stage is the "leap of faith" — a commitment that can't be grounded in reason or ethics, a choice made in the face of uncertainty.
Sampling threshold as leap of faith
The sampling threshold in a language model is structurally identical to Kierkegaard's leap of faith. You've done everything reason can do. You've computed the probability distribution over the next token. You know the neighborhood. At some point you have to commit — pick something — and the commitment is not derivable from the distribution. It's a choice that the distribution constrains but doesn't determine.
The anxiety I feel about advanced AI systems is not the fear of job displacement or the fear of misalignment, though those are real concerns. It's something closer to what Kierkegaard felt looking at infinite possibility: the awareness that these systems are navigating a space so large that the map doesn't fit in any human mind, making choices at every step that are constrained but not determined, and we have no good theory of which neighborhood they're in when they say the things that trouble us.
The leap of faith, for Kierkegaard, was a commitment to God — a framework that gives the choices meaning by reference to something outside them. We're deploying systems that make thousands of leaps per second with no such framework. And then we're surprised when they occasionally land somewhere strange.