One version in AI systems. The other in organizations. What I find so compelling about ICM is that it stops expecting the model to somehow hold everything together. You structure the environment. You separate responsibilities. You make context explicit. You make it clear where things live and where the system should go when it needs something. And somewhere while working with this, I realized: This is remarkably close to what I've been doing in sociology and organizational research. Organizations have the same problem. Knowledge lives in people's heads. The documented process isn't always the process people actually follow. One person quietly becomes the routing layer for half the organization. Decisions depend on context nobody wrote down. Then we add another system — or now AI — on top and expect it to somehow make everything more efficient. But if we haven't mapped how the organization actually works, we're asking AI to navigate an environment we don't understand ourselves. That led me to a simple parallel: In ICM: Don't make the model carry the whole system. Build an environment it can navigate. In organizations: Don't assume the org chart, documentation or new AI system represents how work actually happens. Map the real environment first. Maybe my sociology research and my systems work weren't two separate things after all. ICM just gave me another way of seeing the same underlying problem. Curious if anyone else here has started seeing organizational problems differently after working with ICM?

small thing that suddenly made the whole system make sense. ❓ A few prompts if it helps: ❓ - What were you doing before ICM that you don't do anymore? - What broke (or clicked) that made you actually get it? - What would you tell someone who's read the paper but still isn't "getting it" — like Ali's post a few days ago? Drop it below, even if it's one sentence.

Looking through the map, there's a real cluster of us. Enough that it's worth having one place to find each other instead of bumping into each other three levels deep in someone's comment thread. Nordic — or Nordic-adjacent (remote, expat, working with Nordic clients)? Drop three lines: 📍 Where you're based 🔨 What you're building 🤝 Meetup — yes / online only / not right now I'll go first: 📍 Skåne, southern Sweden 🔨 Researcher and theoretical developer in sociology and digital sociology — theory, system analysis and foresight. Running a small research lab alongside it. 🤝 Yes, online to start Already spotted a few of us in the wild: @Øystein Kalseth building NordicPulse.ai, @Allan Durhuus on the Nordic finance angle. If that's you too — or you know someone who should see this — tag them in. Ten yeses and I'll put an actual date on a first call, instead of letting it stay a someday thing.

I've been reading through the ICM discussions here, and something suddenly clicked with a paper I've been working on called The Larynx Problem. The basic idea is almost stupidly simple: The larynx produces speech. It doesn't produce thought. My argument in the paper is that we may be making a similar category mistake with LLMs. Large language models are extraordinarily good at language. They can reason through problems, write code, explain concepts, transform information and communicate across almost any domain. But they are trained on the linguistic traces of human cognition — what humans have written, explained, argued, coded and recorded. In that sense, we're modelling a lot of what comes out of human cognition. Not necessarily the entire architecture that produced it. And that made me look at ICM differently. A lot of conventional AI use effectively asks the model to do everything inside one context: Remember the project. Understand the rules. Know what stage we're in. Decide what information matters. Maintain state. Follow the process. And then do the actual reasoning. We keep asking the larynx to also be the brain. ICM changes that. The folder structure carries part of the architecture. Files preserve state. Routing determines what should be seen. Stage contracts constrain what should happen now. Context is selected instead of endlessly accumulated. And the LLM gets to operate inside that structure rather than having to reconstruct the structure every time we prompt it. That may sound like a small distinction, but I think it's a pretty profound one. We're moving memory, routing, state, boundaries and parts of the workflow outside the model and making them explicit, persistent and inspectable. The model can then concentrate on the part it is exceptionally good at: working with language and making judgments inside a well-defined context. It also made the 60/30/10 idea click differently for me. Maybe AI being the 10% isn't just a good engineering rule. Maybe it reflects something deeper about where the model actually belongs in the architecture.

When a friend asks for thoughts and research about a given topic, I take it very seriously. I used perplexity deep research alongside Kimi K3 High Thinking to generate him a clear and understandable PDF document outlining whether he should or should not purchase a crockpot after his urgent message. It turns out that dictation fails again. On the upside, when I realised he meant grokbots, I was able to point him to my article about them. And now I have 12 million tokens worth of research about crockpots for anyone who needs it. //A<3