The Eighteenth
Message

Friluftsfrämjandet have spent more than a century getting people outdoors, and they know that lessons outside are good for learning, for health, and for how a class holds together.

But it is hard to know exactly how to take a maths lesson outside. How do you teach fractions in the woods? How does it connect to the curriculum? A teacher with a full week and no training in outdoor pedagogy is not going to work that out on a whim. The knowledge exists, at Friluftsfrämjandet. It just is not where the teacher is at nine on a Sunday evening.

Asking GPT or Claude has become the natural first thought. For Friluftsfrämjandet it came with a second question, one being asked far beyond them: what does AI cost the planet? The electricity behind these models, and the water that cools the data centres running them, are now part of any serious conversation about using them. An organisation whose purpose is nature could not help teachers with one hand and add to that problem with the other. So before building anything, we looked at what happens when a teacher plans a lesson in a chat.

What happens in a chat

You send text to a model, it processes it, and text comes back. The text is counted in tokens, roughly fragments of words. Every token the model reads or writes takes computing, and computing takes electricity and cooling water. The token count is the closest thing to a meter on a chat.

What is easier to miss is that the model keeps nothing between messages. Each request starts from zero. A chat feels like a conversation because the app sends the whole transcript along every time: your second message goes in with your first and its answer, your third with all of that. It is a bit like a colleague with no memory, where every question begins with a printout of everything the two of you have said so far. The printout grows.

So a teacher planning an outdoor lesson in a chat might describe the class and the subject, paste in the curriculum at message three, run a web search for activities at message five and another at message nine, and refine over eighteen messages. Everything pasted or found stays in the transcript. The figure shows what the model reads at each step.

The teacher sees one lesson plan. The model has read about half a million tokens to produce it, and the curriculum sixteen times over. None of this shows up anywhere. There is no counter in the corner of the window. It is not a flaw in any particular product, just how the technology works, and it makes a long conversation easy to underestimate.

State-of-the-art (SOTA) models are highly capable and “know” a great deal, but they know it in fragments, from every country’s curriculum and every tradition of teaching at once. Asked for a Swedish outdoor lesson for year five, they are bound to improvise a little too much, or go to the web and cook something from whatever they find.

Deciding once what the model should know

Instead of every teacher assembling the background from scratch, Outomaten has it assembled once, by domain experts. Friluftsfrämjandet’s outdoor pedagogues wrote and reviewed the material: how to run a class outside, what works on a wet morning in November, how to keep everyone safe. We added the Swedish curriculum excerpts and the National Agency for Education’s commentary for each subject and stage. Thirty-seven files in all, curated and fact-checked. This gives the model a much better starting point and raises the quality of what comes out by a lot, even with SOTA models.

The teacher answers a few questions. Which year group, which subject, where you will be, anything particular about the class. That is the only part that changes. The model reads the fixed material with those answers, once, and writes the lesson in a single call. About sixty seconds. Free, with no subscription.

The numbers

Building the material also cost tokens. We put it at around six million, deliberately generous. That cost was paid once and is shared by every lesson the service will ever produce. It is recovered after about twelve lessons.

What this has to do with nature

Per token, the electricity and the cooling water are tiny amounts. Multiply by half a million tokens per lesson, and then by every teacher in the country planning a few lessons a term, and they stop being tiny.

The whole idea

There is no special model inside Outomaten. It is the same kind of model you can chat with anywhere. The difference is that someone decided once what it should know, so that no teacher has to explain it eighteen times. Experts did it once. Everyone after them builds on that work. The quality can be checked before anyone relies on it. Teachers now start from a finished lesson, with a worked-out structure built from curated knowledge, instead of from a blank page. It is better, it is lighter on the planet, and it happens to be cheaper too, all for the same reason.

This has a name: context engineering. Deciding what the model should know, and from whom, before anyone asks it anything. Most organisations have a few things they do the same way many times over, and for those the question is not whether to use AI. It is whether to let everyone assemble the background on their own each time, or to do it once, with the people who know.

If you would like to talk about how your organisation could think about AI more responsibly than the token-maxxing trend we are sadly seeing now, give me a poke.