# Models and Handles.

Type: Article
Date: 2026-07-06
Tags: Computation, Cognition, Tools for Thought, AI, Design
Location: Sydney, Australia
Canonical: https://www.aaronroot.net/journal/models-and-handles

In 1979, at the Paris Observatory in Meudon, a young astrophysicist named [Jean-Pierre Luminet](https://articles.adsabs.harvard.edu/pdf/1979A%26A....75..228L) produced a picture of something that cannot be photographed. A black hole emits no light. Nothing that crosses its horizon returns, and no telescope on Earth or above it had ever seen one. What Luminet had instead was general relativity, an IBM computer fed with punch cards, and a question. If a black hole were wrapped in a glowing disc of matter, what would it actually look like?

The machine gave him numbers. He plotted them by hand, inking dots one at a time onto negative paper, packing them denser where the disc burned brighter. The image that emerged was strange and lopsided. A dark shadow ringed by a glowing halo, brighter on one side than the other, with the far side of the disc visible above and below the shadow because the black hole's gravity bends its light over the top.

Forty years later, the [Event Horizon Telescope](https://arxiv.org/abs/1906.11241) linked radio dishes across the planet and captured the first real image of a black hole. It showed a dark shadow ringed by a glowing halo, brighter on one side.

It is worth being precise about what Luminet did, because it fits neither of computing's familiar categories. He observed nothing. He automated nothing. He built a model, ran it, and something no eye could ever see became something a mind could hold.

[Seymour Papert](https://oldobjectsnewideas.com/_reading/Papert_Mindstorms_1st_ed.pdf) had encountered this process before he encountered a computer. As a child he became absorbed by gears. The differential gear gave him something mathematics on a page could not. He could turn one part, feel the others answer, and carry that movement into equations. The gear was not an illustration added after the idea. It was the thing through which the idea became available to him.

He later called such things objects to think with. Their value was not simply that they made an abstraction visible. They gave thought somewhere to go.

The Logo turtle was built for this purpose. It began as a small machine that moved across the floor with a pen underneath it, then migrated to the screen. A child could tell it to move forward, turn, and lower its pen. To draw a square, the child first had to imagine the path from inside the turtle. Move forward. Turn ninety degrees. Repeat. The turtle followed the instruction exactly and left the child's reasoning behind as a line.

When the drawing came out wrong, the mistake was no longer hidden in the child's head. It was there on the floor or the screen. The child could inspect it, change the instruction, and run the thought again. Programming the turtle was also a way of stating an idea precisely enough to find out whether it was true.

This was the mechanism beneath Papert's Mathland. Children in France learn French because they live where it is spoken. A child given a computational world to live in could learn mathematics the same way. The computer did not teach geometry by delivering better explanations. It created a world in which geometry became something a child could inhabit, construct, and debug. The external model built the internal one because the child remained inside the loop.

[Alan Kay and Adele Goldberg](https://cognitivemedium.com/tat/assets/Kay_Goldberg.pdf) had named the larger property of the computer in 1977. The computer was not a tool but a medium, and not an ordinary one. It was the first metamedium, able to become every other medium including ones not yet invented. Its native content was the dynamic simulation. A message that runs. Every previous medium held descriptions of ideas. This one could hold working models of them.

In [*Universal Machines*](/journal/universal-machine), I called these things instruments, but the word has never sat right. A telescope, a stethoscope, or the artificial horizon that lets a pilot fly through fog extends perception. Each has authority because it is calibrated against something in the world. The word still earns its keep wherever reading is the point, in the cockpit, the trading terminal, the map that knows where you are.

But Papert's turtle did not read a world. It held one a child could enter. A spreadsheet arrives blank. Nobody reads Excel. You write in it, and it writes back. Kay had already found the right name for that. An instrument reveals a world. A medium lets you construct one.

The spreadsheet may be the most widely used thinking medium ever shipped. Its success has little to do with arithmetic. Every cell is both a reading and a handle. Change an assumption and watch it run through the whole model. You can drag a number until the future breaks, then follow the wreckage back to find out why.

At its best, a spreadsheet preserves the same loop as the turtle. Form an expectation. Encode it. Run the model. Inspect the result. Trace the mistake. Change the thought. Financial intuition does not come from moving cells. It forms when a person sees how an assumption travels through a system and learns to anticipate the consequence before the number changes.

I thought handles might be enough. The spreadsheet makes that difficult to believe. A notorious spreadsheet error sat in [Reinhart and Rogoff's debt model](https://www.socialjustice.ie/system/files/file-uploads/2021-09/2013-04-15-herndonashpollin-doeshighpublicdebtconsistentlystifleeconomicgrowth-critiqueofrandr.pdf), quoted by finance ministers for three years while its cells held a mistake a graduate student eventually found by hand. The model had handles. It could still be operated without being understood. Manipulation was not the same as participation.

[Bret Victor](https://worrydream.com/#!/InventingOnPrinciple) argued that creators need an immediate connection to the thing they are making, and that every layer of delay between a change and its consequence costs thought. Papert's turtle answered the hand immediately. So does the spreadsheet. Yet speed alone cannot explain why either one builds understanding. A slot machine has immediate feedback. So does a feed. Neither asks the person to form a model of what happens next.

Kay spent his later decades pointing out that the revolution he described had not happened. The computer was collapsing into television, a medium you receive rather than play. The pattern is everywhere once you look for it. [Turn-by-turn navigation](https://pmc.ncbi.nlm.nih.gov/articles/PMC8032695/) delivers every journey and quietly dissolves the map in your head. The answer arrives and the model never forms.

Nobody was tricked into that trade. People chose the voice over the map because the model in the head was worth less to them than the effort it cost. Sometimes that is sensible. Not every journey needs to teach geography. The trouble begins when the thing being surrendered is also what lets a person recognise when the system is wrong.

The prompt box runs on the same bargain. You describe the outcome, approve what comes back, and ask again if it is wrong. What comes back is often remarkable. The systems behind it can build in an afternoon what took Luminet a year of punch cards and a steady hand. A model of a watershed, a classroom, an immune response, a supply chain, or a city's traffic can be built to order and fitted to the question in front of you.

For the whole history of the medium, working models were expensive. They belonged to the people who could program them, which meant most people who needed one never got one. That constraint is falling away. This is the machinery Kay's children were waiting for, the Mathlands Papert could sketch but not scale.

What has not arrived automatically is the second thing. Papert's child programmed the computer and discovered something about their own thought. At the prompt box the person describes an outcome and the computer supplies the construction. The artefact can arrive before the understanding, and the difference is a [cognitive debt](/journal/cognitive-debt).

Adding controls does not repay it. A person can move a slider without knowing what the variable means, why it is connected to the result, or where the model's authority ends. A model built for you, fitted with handles you do not understand, may still be television made to order.

The difference is participation. Papert's turtle made the child predict a path, state it precisely, watch it fail, and revise the instruction. The handle mattered because it kept the child responsible for the connection between intention and consequence.

This suggests a different design problem for AI. Not better answers, and not interactivity for its own sake. The work is to preserve the loop through which operating a model changes the operator. Let the chemist set the threshold, but also see why the threshold matters. Let the teacher bend the scenario, predict what will follow, and trace the result back through the assumptions that produced it. Let the system argue when the person's model and its own come apart.

An AI that hands you conclusions is a very fast oracle. An AI that builds a world around your question, gives you a way to test your understanding inside it, and stays while you revise both the world and the question is something else. It is an object to think with.

Luminet shows why immediacy is not the test. His loop took days, but he chose the question, formalised the model, and drew its consequences by hand. The delay did not cost him the thought. It may be where the thought happened. There is no general rule for separating work that is merely cost from work through which judgement forms. The same step can be drudgery to an expert and instruction to a novice.

So far this is a relationship between one person and one model. Most consequential models will enter organisations instead. Each person might understand their exchange with the machine while the group still lacks a shared account of what was decided, why, and what should change next.

Scaling the relationship does not mean multiplying private exchanges. It means creating shared representations through which a group can inspect the work, challenge it, learn from it, and act together.

An object to think with leaves a person able to find the next answer. The next question is how a group finds it together.

![Jean-Pierre Luminet's 1979 black hole illustration](/content-assets/articles/article_images/luminet-black-hole-1979.jpg)