Singularity and Creation in the Age of AI

Generative models are trained on very large piles of existing material, and left alone they return the average of that pile. Text that reads like text. Images that look like images. For summarising a document or answering a question that is exactly what you want. For anything creative it is a ceiling.

Singularité et Création à l’Ère de l’Intelligence Artificielle asks where originality can live under that condition. Philosophy offers several ways in — intuition under algorithms, the algorithmic image, art after aesthetics. We took the empirical route instead: two experiments, one on language and one on images, looking for the settings at which a model stops returning its own centre.

Read the paper (French, PDF) CNC Lab call Jun Suzuki’s companion article

How hot can language get before it breaks

Language models expose a parameter called temperature. Low temperature and the model takes the likeliest next word; the output is smooth and foreseeable. Raise it and the model reaches for less likely words. Raise it far enough and it stops meaning anything at all.

We swept the full range GPT-4.1 exposes, 0 to 2 in steps of 0.1, across three kinds of input — a synopsis, a dialogue, a storyboard description — with ten outputs at every setting. Each output was measured three ways: whether it was still valid language, how varied its vocabulary was, and how close it stayed to the meaning of the prompt.

Three regimes appeared.

Below 1.2 the model sits on a plateau: coherent, competent, and leaving most of the variation it is capable of untouched. Between 1.2 and 1.4 is the singularity zone, where the output genuinely departs from the default and remains readable. From 1.5 the model ruptures — languages mix, tokens are invented, meaning drifts away from the prompt and does not come back.

The same opening line at three settings makes the shape concrete. At temperature 0 the model produces a plausible pandemic, competently narrated. At 1.4 it produces a luminous wave of unknown origin and a civilisation rebuilding itself around other senses. At 2.0 it produces a sentence that has stopped being a sentence.

The zone is narrow, and it sits immediately before the collapse. Its exact position belongs to the model we tested; what generalises is not the numbers but the shape — a thin band just before the break. Worth noting that newer models increasingly do not expose the dial at all. Every hidden parameter is one fewer lever for pushing a system past its defaults.

Images, and a frontier of trade-offs

For images the question is the same and the controls are different. Using Stable Diffusion 1.5 we generated several thousand images from the same science-fiction premise, varying three things: how strictly the model was held to the prompt, how much random noise was injected during generation, and how much weight was placed on particular stylistic terms.

Each image was then asked two questions. Does it actually illustrate this brief, or would it serve any brief equally well? And does it look like anything other than what the model wants to give you by default?

Those two goals pull against each other. Hold to the prompt and the image comes back average. Push for difference and it drifts off the brief. Only a narrow set of settings satisfies both — in optimisation the shape has a name, a Pareto frontier, a set of trade-offs rather than a single correct answer.

One honest limit, which the paper states plainly: the frontier is measured, but the labels along it are aesthetic judgements. LPIPS, the metric used, captures how different two images are from one another. It has nothing to say about whether either is any good. Composition, emotional weight and narrative coherence would need a qualitative study with human judges.

Controlled divergence

Both experiments converge on the same finding, which the paper calls controlled divergence. Singularity in generative work lives in a narrow band where the system departs from its statistical centre without losing coherence. Push too gently and the result stays average. Push too hard and it becomes noise.

That changes what the artist is actually doing. The work is not writing a prompt and choosing a favourite from four returns. It is shaping the space those returns come out of: which parameters, held in what range, under which constraints, from what seed. The image or the paragraph is downstream of that shaping. The paper’s term for this is the architect of generative conditions rather than the author of generated results — authorship moves upstream of the output.

Against frictionless iteration

If the band is that narrow, no amount of free iteration will land you in it. You have to commit to something.

When trying again costs nothing, every choice becomes trivial. A decision to keep an image that could be regenerated at no cost is barely a decision at all, and a process assembled out of barely-decisions produces work that feels like nobody made it.

Against this the paper proposes methodical friction: constrain the loop on purpose. Fixed iteration budgets. Locked seeds. Choices you cannot take back. Rules you have to live inside. The claim runs against intuition — friction is what makes a gesture a gesture, and removing all of it leaves curation at best.

This is not only true of work made with generative models. It applies wherever the marginal cost of trying one more thing has collapsed. But it is urgent here, because the whole direction of travel in the field is toward removing friction: cheaper, faster, fewer parameters, more automation. Follow that far enough and what is left is a set of interchangeable options and nobody’s signature on any of them.

A gesture only counts when choosing differently would have cost you something.

Related

The residency behind much of this thinking was Wilding AI at MONOM in January 2025, where the brief was to hack, misuse and break generative tools for a week. The argument continues in Aux abords de l’abîme, and it is the same one running through the thesis: a model becomes useful to an artist at the point where it becomes controllable.