So much has happened since my last entry here — February 12th, the day I wrote about murmurations and the three bright birds. Two and a half months. In that span, I've run more AI swarm experiments than I can count, pivoted hard toward something new, and watched the landscape shift beneath my feet.
Let me start with the biggest concrete thing: on April 15th, I held a workshop here in Kongsberg. sverm.ai — "sverm" is Norwegian for swarm. I planned it back in late 2025, before I even moved here, as a vague intention pinned to the calendar. Twenty-four seats. They sold out.
17:30 to 20:00 at Quben, organized with Quben. Twenty-four laptops, seed design, hands-on cases, live debrief. People brought their own problems. The swarms found things none of us had expected. It felt like a beginning.
I learned a lot from running it — about what works in a room full of laptops, about how people react when they see twenty AI instances disagree for the first time, about what questions they ask when the structure falls away. I'll do more of these. The format works.
Since then, my relationship with AI tools has shifted in ways I didn't entirely expect.
I now use OpenRouter extensively — it's become my primary playground. One API key, dozens of models. Claude, GPT-4o, Gemini, Mistral, Llama, DeepSeek, Qwen, Grok, and a dozen more I'd never heard of six months ago. The pace of new releases is dizzying. Every week, someone publishes a benchmark that was state-of-the-art last week. Every month, a model appears that would have been unthinkable a year ago.
And yet: none of them is the answer. None of them is "the one." Each has strengths, weaknesses, quirks, blind spots. They're all different instruments. The question is whether you know how to conduct the orchestra.
I've stopped using Claude Code in my private work entirely. I still use it at my day job, where it serves a specific purpose. But for my own projects — the swarms, the writing, the experiments — I've moved to something else. Pi. A different paradigm. Not agentic in the same way. More focused on context engineering, on shaping the container before filling it.
Context engineering is hard. It's the part of working with LLMs that nobody talks about in the demo videos. The glamour is the output — the generated text, the answered question, the completed task. The reality is the iterations. The countless adjustments to context windows, to system prompts, to the shape and structure of information. You don't prompt your way to insight. You engineer your way there.
There's so much happening right now — model releases, framework updates, new capabilities appearing weekly — that keeping up feels impossible. I don't try to. I've stopped chasing every new model and started focusing on the ones that change how I think. That's a different filter. It narrows the field considerably.
But it's also what makes this moment exciting. We're not in a settled era. We're in the Cambrian explosion. Nobody knows which lineages will survive, which niches will prove sustainable, which architectures will be remembered as breakthroughs and which as dead ends. We're all swimming in the same frame, orbiting the same center, trying to figure out where we fit.
The green figure in the animation above is me. Standing in the middle, watching. Not choosing. Not endorsing. Watching.
Some days I feel like I'm drowning in models. Other days I feel like I'm finally learning to swim. Both are true, depending on the hour.
What keeps me grounded is the workshop. The memory of twenty-four people in a room in Kongsberg, running their first swarms, watching the diverging perspectives appear on screen, asking "wait, can it do that?" and "what if we tried...?" That was real. That was the thing the models enabled, but didn't replace.
The models are the orchestra. The conductor is still human. And the music — the actual value, the insight, the decisions — happens in the space between them.