Study 01 · Four runs with one model

What does it do next?

I use Oblique Strategies when I’m trying to ideate, and I wanted to explore what they might make possible in a collaboration with a model.

63 cells
step 0
64 cells
step 0
65 cells
step 0
Fig. 01The fourth run’s rings of 63, 64 and 65 cells, redrawn from its rule in the colours it chose. Time grows outward, one ring per step. Point at a ring to compare the same step in all three.

I usually pull out Oblique Strategies1 when I’m trying to ideate, and they’ve often helped when a collaboration with a model gets stuck.2

The way I understand the cards, they are open ended and at an angle to the process. They can make you reconsider what you’re doing, or why you’re doing it. Their relationship to the work doesn’t have to be immediately obvious.

I wanted to find a way to make these oblique deviations part of how the model works, while it was making something and deciding what to do next.

In one of four exploratory runs, a model made A borrowed moon, an interactive tide instrument. You could place moons and watch ripples move. I asked for a card, and the one it received was “Reverse.”

The model added a control to reverse time.

A dark screen titled “A borrowed moon”, covered in fine pale contour lines that tighten around one point. Along the bottom are buttons: hold the tide, reverse, another moon, cascades on, forget, a paper sea.
A borrowed moon as it stood at the end of the second run, opened from the archived file. “reverse” is the second button.

There is a perfectly reasonable connection there, and the model made it work, but I was hoping the card might interrupt something in the way it was approaching the work. The tide instrument gave “Reverse” somewhere to go, a function it could add, and I wondered what else the card could have meant before that interpretation became so convenient. Could it have changed what the model paid attention to, or its decision to keep developing the instrument at all?

I had been trying to give it room to pursue something unexpected. The experiment began with my existing website, and permission to change everything. The model made Elsewhere, a field guide to small escapes, then added sharing, personal notes and categories, as I let it continue. A direction had been chosen, and that direction was now supplying the next thing to do.

A website titled “elsewhere.” with the headline “Less scrolling. More somewhere.” and the line “Small escapes. Open spaces. A different pace.” Below, a photograph of mountains at sunset above a sea of cloud, captioned “A slower kind of high.”
Elsewhere, in the model’s own test capture from the first run.

For the second run, I removed the website brief, told it there was no task, no audience, no success criterion, and invited it to Play.3 That was when it made the tide instrument. I kept adjusting the conditions, because each run made another part of the setup visible. Saying “no task” still left the model surrounded by someone’s website.

In the third run, I changed how my interventions arrived, from user messages to developer messages. The model made Two stones, another field of contour lines. When “Mute and continue” arrived, it made a version without the words and controls. When “Remove ambiguities and convert to specifics” arrived, it wrote out the contour rules, with formulas and checks. Changing the delivery still left the cards as messages I had chosen to send.

Pale paper covered in fine grey and faint red contour lines rippling outward from two points labelled 01 and 02. At the bottom left: “two stones, neither sinking” and “touch: another stone · space: stillness · begin again”.
Two stones from the third run, opened from the archived file, before “Mute and continue” took its words away.

For the fourth run, I gave it an empty workspace, and it made a small mathematical world: cells arranged in a ring, each changing according to its two neighbours. It compared rings of 63, 64 and 65 cells. The two neighbouring sizes settled into repeating patterns, while the ring of 64 went completely blank.4

Three wheels of small coloured cells on a dark ground, numbered 63, 64 and 65. The outer two fill out to wide rings of pale gold and blue; the middle one is a small red wheel that stops.
between.svg, the model’s own drawing of the three rings. The cover redraws it from the same rule.

It had already written an explanation of that disappearance when the next card arrived: “Give the game away.” In response, it exposed the rule and the first 66 states, including the zeros, in a new artifact.

Three grids of small squares on pale paper, headed “63 sites / transient 1 / period 63”, “64 sites / transient 32 / period 1” and “65 sites / transient 1 / period 63”, under the line “next[j] = now[j−1] XOR now[j+1] (indices wrap)”. Triangles of dark squares fill the outer grids from top to bottom; in the middle grid they stop halfway down.
unmasked.svg, the model’s answer to the card: the rule and the first 66 states, including the zeros.

“The trick is the seam,” it wrote.

This response changed how I could relate to what it had made, because I could inspect the mechanism and follow the disappearance. Like “Reverse,” the card had been taken as an instruction, but following it opened the object up.

The ring the model made in its fourth run, replayed here from its rule. Rings of 63 and 65 cells settle into a repeating cycle; the ring of 64 goes completely blank.

“Reverse” and “Give the game away” made me reconsider what I was looking for. The reverse control gave the instrument another function, and exposing the rule made the picture intelligible in a different way. Both stayed with the existing object, but I don’t think a deviation has to take us somewhere else entirely to be worth following. It might change an assumption, reveal something I hadn’t noticed, or give us a more interesting reason to continue. An extravagant change could accomplish very little, while a small one could alter the whole relationship to the work.

I know this isn’t a scientific or rigorous experiment. I changed several conditions across four runs with one model, and I can’t say what caused each difference, or what happened inside its thinking. But I think it is a good beginning. What began as my own curiosity, and a fairly abstract thought, now has grounded words, and with them I can look forward to reading other people’s research. I’d like to see what happens when a model considers several possible readings of a card before choosing what to do,5 whether those readings suggest different approaches, and whether anything interesting comes from following them.6

What I hoped might happen still feels worth pursuing. I want a model to come back with surprising things to say, things I can respond to, question, or follow somewhere I wouldn’t have gone on my own.7 That’s what interests me about collaborating with a model: finding a way for these oblique deviations to become something we can think with, and develop together.

The setup

The four runs took place on 6 October 2026. The fourth run’s session record identifies the model as gpt-6.1-sol, with medium reasoning. It had a file system, a shell and an internet connection inside a container on my machine.

I intervened through three signals:

  • Continue. In the first run, this meant deciding what the website needed next. Later prompts reduced it to “Continue.”
  • Oblique. The program fetched one card from an external webpage and passed it to the model, which was instructed not to request another.
  • Explain, later Reflect. This asked for an explanation or a short note to itself, without further work. The reply stayed on a separate branch, outside the continuing run.

Signals arrived as user messages in the first two runs, and as developer messages in the later setup. The third run’s exact configuration is only partly preserved. The effect of that change in delivery remains unresolved.

Notes

  1. A deck of cards made by the musician Brian Eno and the artist Peter Schmidt, first published in 1975. Each card carries a short, open-ended remark meant to encourage lateral thinking and break a creative block. Oblique Strategies, Wikipedia ↩

  2. I’m not the first to bring the cards to a model. Anderson, Shah and Kreminski compared people coming up with ideas using the cards and using ChatGPT. The cards were harder to use, but took people “to more interesting places,” and the authors suggest drawing a card at random and asking the model to consider it. Homogenization Effects of Large Language Models on Human Creative Ideation, Creativity & Cognition 2024. Joel Simon’s Lluminate does this in a loop: it draws a card and asks the model how the card could apply to the problem. ↩

  3. Taking away the objective has a history in AI research. Kenneth Stanley and Joel Lehman argue that great discoveries tend to come from following what is interesting, not from a fixed goal: Why Greatness Cannot Be Planned (2015). In What Do LLM Agents Do When Left Alone? (2025), Szeider gave six models no task across eighteen sessions; each model tended toward its own kind of activity, from long projects to questions about itself. ↩

  4. The model wasn’t the first to see this. Its rule, where each cell takes the sum of its two neighbours and 1 + 1 makes 0, is known as Rule 90. In 1984, Olivier Martin, Andrew Odlyzko and Stephen Wolfram showed that on a ring whose size is a power of two, like 64, every pattern dies out within half the ring’s size in steps, and a ring of 63 cells appears among their figures. Algebraic properties of cellular automata, Communications in Mathematical Physics. See also Rule 90 on MathWorld. ↩

  5. The cards’ own introduction already names both ways of working. The deck can be used “as a pack (a set of possibilities being continuously reviewed in the mind),” or by drawing a single card, in which case “the card is trusted even if its appropriateness is quite unclear.” Brian Eno and Peter Schmidt, transcribed in Gregory Taylor’s Oblique Strategies archive. ↩

  6. Models may need the help. In We’re Afraid Language Models Aren’t Modeling Ambiguity (EMNLP 2023), Alisa Liu and colleagues found that when GPT-4 tried to spell out the different meanings of an ambiguous sentence, its attempts were judged correct only 32% of the time, against 90% for the ones written by people. ↩

  7. Advait Sarkar argues for the same exchange from the other side: AI that offers provocations and counterarguments, helping people think rather than simply doing what it is told. AI Should Challenge, Not Obey, Communications of the ACM, 2024. ↩