On Thursday, October 22nd, I resigned from my job.
I started at a trade union for engineers in January 2022. I leave on January 31st, 2027. Five years, almost to the day — a clean number, which somehow makes it easier to say out loud. It's a good place. Serious work, good people, a purpose I never had to explain to myself in the morning.
And I quit anyway.
My last working day is January 31st, 2027. After that, the calendar is empty.
Let me be honest about the shape of this: I don't know what I'm going to do from February 1st. Not "I have a plan but I'm keeping it vague." I actually don't know. I need an income, like everyone else. That part is not abstract and I'm not pretending it is.
But here is the thing I keep coming back to. Three months ago I wrote in this log that I was drowning in models on some days and learning to swim on others, and that both were true depending on the hour. That's still the case. What has changed is that I've stopped treating the uncertainty as a problem to be solved before I'm allowed to move.
So I'm not solving it right now. I'm training.
On March 11–12, 2027, the final of the Norwegian AI Championship happens in Oslo — NM i AI. The concept dropped the day before I handed in my resignation, which felt like either a coincidence or a signal, depending on how superstitious I'm willing to be on a given Tuesday.
But it doesn't start there. There are regional championships first — Tromsø, Bodø, Trondheim, Molde, Bergen, Grimstad, and Oslo. Seven cities, seven dates, running from February 10th to March 4th. Oslo's is on Thursday, March 4th. The final is the week after.
So that's the road. A regional in Oslo on March 4th, and if that goes well enough, a return trip a week later for the final. Eight days between the two. Not much time to fix anything you got wrong.
So from now until then: I train. Not "train" in the vague sense of reading papers and feeling informed. Train in the boring, concrete sense. Build systems. Run them. Measure them. Break them. Discover that my beautiful architecture loses to something dumb and fast. Fix it. Repeat until the score stops moving.
The training ground, at least for now, is grocerybot.no — one of the 2026 challenges, now open to anyone who wants to play.
The premise is very simple. You write a program that steers a fleet of robots around a dark grocery store. Shelves with dairy, bakery, produce, pantry. Trucks docked at the loading bays, each one carrying an online order. Robots pick items from adjacent shelves, carry three at a time, load them into the right truck. One point per loaded item, five more when a truck's order is complete. Robots block each other in the aisles, and there's no way to put an item back — a wrong pick occupies a slot until an order happens to need it.
Five stores. From one robot on a 12×10 grid to twenty robots on something much worse. Two hundred rounds to look clever, two minutes of wall clock to prove it, two seconds to answer each round or everything freezes. Nightmare mode gives you five hundred rounds, five minutes, and twenty robots that all hate each other.
I have never written anything that had to be correct within two seconds, twenty times per second, while twenty agents blocked each other in a corridor. It is humbling in exactly the way I needed.
Because here is what the training is actually about, and it isn't the grocery store.
I'm not learning to write a bot. I'm learning to build the factory that produces bots.
The bot is the artifact. The thing I'm actually building is the machinery around it: the benchmark harness that runs five hundred games overnight without me watching. The replay analyzer that tells me why a truck missed its window instead of just that it did. The regression suite that refuses to let me "improve" something that was already working. The context that I shape, deliberately, before asking for anything. The tools my swarm can use when I'm stuck at 03:00 and out of my own ideas.
That's the skill that transfers. Whatever happens after January, whatever the actual work turns out to be, the ability to build systems that make me and my team better at a well-defined problem is not going to be a niche skill in 2027. It might be the only one that matters.
And I want to be clear about the ambition, because there's no point being coy: I'm not entering to participate. I'm training to win. The training itself is the thing, but I wouldn't be doing it if I didn't secretly believe I could be genuinely good at this.
There's also the older thread, the one that keeps showing up in this log. Twenty-four years ago I registered dltq.org — Don't Lose The Question — because of something Rumsfeld said in a press briefing about the limits of epistemology, and because I'd figured out that a good question is worth more than a clean answer. In February I was writing about the three bright birds in a murmuration, the ones who don't lead and aren't special, they just happen to catch the light differently. In April I was writing about the conductor and the orchestra.
Now I'm building the thing that automates the orchestra. That's not a contradiction. That's the whole point. The conductor doesn't play an instrument. The conductor makes space for the people who do.
So: the job is ending. The calendar is emptying out. I have no idea what happens on February 1st.
But I have a practice field, I have a regional in Oslo on March 4th, and I have a two-minute wall clock that does not care at all about how good my intentions were. That's more than I'd have if I'd stayed.
I'm genuinely excited. Not the nervous, careful kind of excited where you hedge everything in case it doesn't work out. The other kind. The kind where you clear your calendar because something is about to happen and you want to be there when it does.
The training starts now. I'll write about what I learn — the scoreboards, the dead ends, the moment something finally clicks. And somewhere around February, I imagine this log gets interesting in a whole new way.