This one began as heavy breathing and half sentences, recorded on my watch during a Friday-afternoon run through Värsnäs, the coastal reserve just north of Kalmar. It is deliberately dated — July 17, 2026 — and it will age badly, on purpose. That’s the point.
It is the seventeenth of July, 2026. A Friday, late afternoon, twenty-six degrees — which in Sweden counts as a heat wave. I am running the trail out through Värsnäs: oak woods, open meadows, the sea flashing between the trees. No phone. Just the watch on my wrist and headphones in my ears, and the watch is recording, because this is how I think. Out loud. One footstep at a time.
Everything in this episode — and this post — began as that voice memo. A machine transcribed it afterwards, imperfectly: at one point it turned Lars, a researcher from Linköping, into a fisherman from Linköping. Somewhere between that tape and this sentence, something had to know what I meant. Hold that thought.
The joke in Glasgow
May 2019. A small conference on ship energy systems — MOSES — seventy marine engineers from eighteen countries, three days of turbochargers and fuel models in an auditorium on Cathedral Street. On the Thursday night there is a conference dinner at the Corinthian on Ingram Street, all chandeliers and Victorian plasterwork. I have a glass of wine in me and a talk to give the next morning, second-to-last slot of the entire conference.
Instead of rehearsing, I sit up that night and feed every abstract of the conference into a language model, and make it write a new one. A fake abstract — plausible nonsense in perfect conference English. I drop it at the end of my talk as a joke, and the room laughs, because in May of 2019, what else would you possibly do with that?
Then a hand goes up. A researcher from Linköping, named Lars: “Did you use GPT-2?”
I had never heard of it. In my defense, that is less strange than it sounds — I built machine-learning models for a living, but on energy data. Sensors, fuel flows, time series. Language was a different continent of the same country. But still: GPT-2 was three months old that week, the model the newspapers were writing about, and it had passed me by completely.
And I thought I was keeping up.
A runner cannot feel pace
That is the whole problem with exponentials, and I can explain it with the watch on my wrist. On a good day, eleven kilometers an hour feels like nine. On a heavy day it feels like fourteen. The legs lie, the breath lies. That is why the watch keeps splits — not because a number feels like anything, but because a split is a photograph of a moment, and photographs can be compared. Feelings cannot.
You cannot feel an exponential either. Nobody can. From inside, every single day feels locally linear. The doubling never announces itself; there is no bell. The only instrument that works is the split: mark the trail, write down the time, let the future measure the distance.
I know my own instruments are broken, because of Glasgow. On the tape from this run I tell that story as 2018 — my last PhD year, I would have sworn it. We checked. The conference program still exists, a twelve-page PDF with my name on page seven. It was May 2019. Right city, right dinner, right Lars — and the year off by exactly one. Memory, it turns out, is an insider too.
So this episode is a split. Cut deliberately, dated the way milk is dated. In two years it will sound like a man carefully describing his first telephone. Good. That quaintness is not the failure — it is the measurement.
The splits
2015–2018. On paper I was the wrong man for machine learning: marine engineer, officer’s training, no computer science at all. scikit-learn, Matlab then Python, neural networks training overnight while I slept — me checking them in the morning the way you check nets left in the water. I said the same thing to anyone who asked: these models are my hammer. My wrench. A tool. You pick it up, you use it, you put it down. And the tool stayed where you left it. Remember that part.
2019. Because of Lars I looked up GPT-2. The lab behind it had declared it too dangerous to release in full; one British paper ran, roughly, an AI so powerful it must be kept locked up for the good of humanity. The famous demo? Unicorns. Researchers gave it two sentences about English-speaking unicorns in an unexplored Andean valley, and it wrote the rest of the news story straight-faced. That was the machine locked up for the good of humanity: a unicorn journalist. The full model came out in November the same year. The danger had a shelf life of nine months.
2020–2022. A confession: GPT-3 came out in the summer of 2020 — the model of its era, the biggest thing in my own field. I touched it for the first time in 2022. Two years late. And I thought I was keeping up.
November 2022. ChatGPT. A hundred million people in two months. Then it started for real — and even from the inside, seven years inside, it felt sudden.
December 2025. Claude Code, Opus 4.5, agents orchestrating agents — the month the hands came off the wheel. You hand over a task in the evening and check the nets in the morning. Ten years later, the same ritual — except the nets got smarter than the fisherman. And since then the harness hasn’t changed; only the mind inside it keeps upgrading. Prompt engineering was a career in 2024. It quietly stopped mattering. Every clever workflow rots faster than it can be evaluated. The half-life of know-how is now shorter than the release cycle.
The photograph: the week of July 17, 2026
Dear listener in 2027 — or whenever you are. You already know how these stories end. I don’t. That is the entire reason this exists.
Fable 5. On June 9 — a Tuesday — Anthropic released the most capable model ever made public. By Friday it was gone: a US export directive, after researchers showed it could find security holes and write working exploits. A language model, briefly handled like a weapons shipment. Back July 1. That was five weeks ago, and I used it today the way I use tap water. The scandal of June is the plumbing of July.
The usage war. GPT-5.6 shipped eight days ago; its top tier is called Sol, so the frontier duel of this summer is literally named Fable versus Sol — a fairy tale against the sun. Both labs keep resetting the usage meters in a war for users, which lands us somewhere historically strange: the two most capable models ever built are, this week, effectively unlimited, flat rate.
The tests. On the Mensa-style IQ tests people run on these systems, frontier models now score around 130 — on the offline versions, written fresh so they cannot exist in any training data. Mensa membership starts at 131.
And this morning, while I laced these shoes: Beijing announced Kimi K3 — 2.8 trillion parameters, the largest open model there has ever been, frontier class, free weights in ten days. In 2023 that would have been the story of the year. Today it was barely a headline by lunchtime.
Where was I
Mid-run, mid-sentence — right as I was reaching for what all of this means — a call came in, straight into the headphones, and killed the recording. Mid-word, in fact: the tape dies on the Swedish word for but. In the episode you can hear it happen, and then you hear the next recording start with the only sentence available to a human being in that situation: Var var jag någonstans? Where was I.
I never found that thread again. And later in the same run I lost it a second time, no call involved. Nobody can hold an exponential in their head — not even for the length of a run, not even the person making the episode about it.
Living inside one is a cycle: overwhelmed, then numb, then you catch up a little, then a release note lands and you are overwhelmed again. The numbness is not laziness or denial — it is a nervous system doing exactly its job, protecting you from a signal that will not stop firing. A few weeks ago I read a post on X about someone who had checked themselves into care — burned out, not from using these tools, but from trying to keep up with them. One person, one post, no statistics. But I recognized the shape of it instantly. That is what keeping up costs when the pace you are keeping is not your own.
So what survives, when know-how rots in months? Taste. Judgment. Knowing what you want and recognizing it when you see it. Every model in the pipeline that made this episode was superhuman at something — and the one thing the whole chain still needed was somebody who knew that Lars was not a fisherman.
Somewhere in these ten years the hammer woke up. It works through the night now; it does not get tired; this week it scores Mensa numbers and builds other hammers. What is left in your hand, when the hammer no longer needs the hand? The choosing. That is all, and that is everything: what to build, what to point it at, what any of it is for.
The runner’s now
Kilometer ten, my street, walking pace. On the tape, right about here, I say the thing this whole run was actually for — and you can hear the stumble and the breathing, and I kept both, because that is what conviction sounds like when nobody is performing it:
Jag är inte rädd. Jag är inte orolig. Jag tycker det är fantastiskt kul. I am not afraid. I am not worried. I think this is tremendous fun.
I cannot control the curve. What I can control is what I point myself at — what I love, what carries meaning, and the pace I run while doing it. A run happens one stride at a time, and so does a technological revolution, if you are one of the people living inside it.
I have stopped trying to keep up. I am running my own splits now.
This was July 17, 2026. If you are listening from later — and you are, everyone always is — I know how this sounds. We could not see it either. Nobody can, from inside. That is why I left you this photograph.
This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit frahlg.substack.com
