# The Shared Workshop Open tools give both people and AI more power to get things done. The big models can even help build the tools that make smaller ones more useful. - Forfatter: Mikkel Freltoft Krogsholm - Type: Essay - Udgivet: 2026-10-30 - Opdateret: 2026-10-30 - Sprog: en - Emner: ai, open source, værktøjer, skaberkraft - Kanonisk URL: https://mikkelkrogsholm.dk/en/articles/det-faelles-vaerksted/ --- Someone opens an image in PhotoCraft and starts editing. An AI agent can work on the image in the same program. The person clicks the buttons. The agent tells the program what to do. Both use the same tools. [PhotoCraft](https://github.com/storytold/photocraft) is an open-source image editor and part of a whole range of Craft programs for images, film, sound and design. Other projects are about drawing objects and buildings. They have something important in common: they are built for both people and agents. That last point is worth noticing. We talk a lot about which AI is smartest. Which model wins the next test, writes the best code or can solve the hardest problem. Meanwhile, people are building all the things the models can work with. And that can make a big difference. An AI with access to a good drawing program can do something different from the same AI without one. Just as we can do more ourselves when we have proper tools. I think an important part of the coming years' progress will happen here. In the shared workshop growing around the models, which also gives us humans more power to do things. ## Give it something to work with A carpenter needs her saw and her plane to make a table. She also uses methods others have worked out before her. Her skills come to life when she has the tools to work with. The same goes for an AI that needs to get a job done. It needs to find what it needs, do something with it and see whether it worked. If it is fixing a computer program, it helps to be able to open the files, change the code and run a test afterwards. If the test fails, it has something concrete to work on next. In the AI world, this is often called a *harness*. It is everything around the model that helps it get the work done. [Researchers showed in 2024](https://arxiv.org/html/2405.15793v3) how much that can matter. They used the same AI model but changed its tools and the way it received the results. That changed how many programming tasks it could solve. Even the way a code file was displayed made a difference. At Fremtidens Tegnestue, we use [SolidCraft](https://github.com/tangency-labs/solidcraft) and [BIMCraft](https://github.com/IPedrax/BIMCraft). With them, we can make architectural drawings that I find incredibly impressive. The programs give AI a way to work with shapes and buildings that we can change and keep drawing. That is a different kind of help from a nice picture of a house. In BIMCraft, there is a model of the building behind the drawing, with walls, doors and rooms. It can be used to make floor plans, sections and elevations. Imagine a wall needs moving. Then there is something to move, and you can see what it means for the rooms and the drawings. The agent can work on the model, and the person can see the result and carry on making changes. They do not need to operate the program in the same way to work on the same building. For a small architectural practice, that opens up some pretty big possibilities. More ideas can turn into something you can look at and make a decision about. You can take a proposal further instead of leaving it at talk about what you might draw one day. In [*When the Interface Became Optional*](https://mikkelkrogsholm.dk/da/articles/da-graensefladen-blev-valgfri/), I wrote about getting help to use programs without finding all the buttons ourselves. With the Craft tools, we can also keep the buttons and take over ourselves when that makes sense. The person and the agent each have their own way into the same work. ## When the big one builds tools for the little one There is another thought that makes this even more exciting. The big AI models have become good at writing code. So they can also help build tools that smaller models can use afterwards. This has already been tried. In a [research project from 2023](https://arxiv.org/html/2305.17126v2), GPT-4 was asked to write small programs that could solve particular kinds of tasks. The cheaper GPT-3.5 was then given access to the programs and had to use them to find the answers. One task involved keeping track of objects that were swapped around. With ordinary instructions, GPT-3.5 got the answer right in just over 61 per cent of cases. With the tool, it reached almost 100 per cent. The model still had to work out what the task was about, but it could let the program keep track of the swaps. The researchers also tried giving GPT-3.5 the stronger model's explanation of how to think its way to the answer. That did not help in the same way. A useful tool did more good than a longer explanation. The big model had done some work that the cheaper model could benefit from again and again. It did not need to stand beside it and help every time. Imagine that happening with many more kinds of software. People use the strongest models to develop better drawing programs, image editing and calculations. Afterwards, other agents can use the programs, even if they could not build them themselves. The big model's ability then becomes something that reaches beyond its own answers. It helps create a tool that many others can use. Including people who will never work with that particular model. In [*The Revolution Beneath the Surface*](https://mikkelkrogsholm.dk/da/articles/revolutionen-under-overfladen/), I wrote about the many small places where AI may come to make a difference. Here is one way to get there: the biggest models can help make the smaller ones more useful. ## What we pass on That is the whole idea of open source. Someone builds something and shares it so others can use it, change it and build on it. A small team can make something big with code others have written before them. Now agents can benefit from that work too. And when they help build new tools themselves, they can help give something back. A program made by a person and an AI together can become useful to entirely different people and their agents. On 29 October, [computer scientist Fabrizio Montesi wrote in Ingeniøren](https://ing.dk/holdning/nu-kan-ai-skrive-koden-men-mulighederne-raekker-langt-ud-over-hurtigere-programmering) about a related project. CSLib gathers computer science knowledge in a form that computers can work with and check mathematically. Montesi described how that knowledge had helped AI solve a task it had previously had to give up on. A library like that and a drawing program are different things. But both give the next person something to start with. You take over a piece of work that has already been done and can spend your energy moving it forward. And when the tools are open, we can help shape them. We can get a bug fixed, add something we are missing or build in a different direction. That takes work. Someone has to write the code, explain how to use it and look after what others build on. The next user's shortcut is something others have spent time making. For the small architectural practice, it may end up as a drawing we used to struggle to produce. For someone else, it may be a film or a piece of music. In [*Something That Is Ours*](https://mikkelkrogsholm.dk/da/articles/noget-der-er-vores/), I wrote about the joy of being able to create something that only a few people need to be interested in. Better tools give us more things to try. It is worth noticing that the same tools can help both a smaller model and a small team of people. Both can gain access to capabilities they could not have built themselves. A contribution to an open program can therefore have a long life: first it helps an agent solve a task. Then the agent helps a person bring an idea to life. I look forward to seeing what the next big models can do. But also to seeing what they help build and pass on. When a model is replaced, a good tool can still be worth using. The next model may arrive at a better workshop than the one before it. And the door can be open to the rest of us.