Weather: sunny · Sleep quality: little · Meals: onigiri later, then Saizeriya · Music: Dalriada, mind.in.a.box, Glis, Nine Inch Nails
Today I think I’ll be brief. I didn’t sleep well: a lot of things came up from home. So I was lucky to have slept two hours in the end, while getting my things ready for the day.
But it went very well. Professor Kato wrote at 3 a.m. that the slot unfortunately had to be moved to 11. It’s fine, really — I was up anyway. I finished with the results and the slide deck by 5. It’s a bad feeling to go to bed when it’s already daylight outside and the birds are chirping. In Nagoya, the crows caw and the sparrows chirp.
After sleeping — and having a horribly bizarre dream — I got up and went straight to Professor Kato. We’d agreed on 40 minutes in principle; compared to that, we talked about my topic for 55. He really liked HyMeKo’s compactness: that with it you can even shorten the necessary LLM token usage. He kept asking questions, stopping me now and then to clarify what we were looking at. A few modifications still fit comfortably into my deck. He suggested my next topic be reinforcement learning, because it wouldn’t use only PyTorch but any other machine-learning framework. I’ll hold him to it. So now I’ll prepare this work, then go and get some sleep.
In the end we said goodbye and he told me to prepare — because he knows just how long I slept. After all, he’s the one who messaged me around 3 a.m. So, “ganbatte,” and we parted.
Afternoon
When I said this would be a quiet afternoon, I really meant it. I went back to the dorm to sleep. Then I went back and met another colleague, Jun. A kind man, just home from a conference in Turkey. We chatted a little about our research topics: he’s working on skin problems, detecting and classifying them with cameras. It connects to our topics too — we have a direction like that as well. It would be nice to put together some small collaboration here.
Meanwhile I was already working on something else right away. I simply couldn’t resist carrying on with the parameter reduction. So I worked on it, and got back again to Clifford algebra and quaternion-rotor optimisation…
A lament folk song to the vending machine
Since I’d run out of change, I had the AI generate a weeping-yet-merry folk song about running out of change. I hope you’ll enjoy this apocalyptic horror that one might call folk music:
My change is all gone — oh, what sorrow and woe,
The vending machine eyes me like a defendant in the dock.
Iced coffee beckons, iced tea waves too,
But my pocket holds only dust, and one lone, dreadful bus ticket.Hey, vending machine, you cold-hearted crate,
You won’t bow to the word of my bank card.
Give me tea, give me caffeine,
For my soul already circles with but half an eye open.I’d have gone to the shop to kindly ask for change,
But the cashier lady looked at me just so:
“Son, here’s the Suica, use it freely,”
Yet my little machine, the villain, asks only for coins.Hey, vending machine, Japanese iron witch,
The Hungarian’s grown thirsty, his throat runs dry.
No hundred-yen coin, no fifty either,
Where shall I get change in this great square?Shall I beg in the lab? “Sensei, just a hundred!”
Or buy a pastry and break the great fortress of a banknote?
To do research without coffee — alas, that’s no life,
Even the hypergraph weeps for a cold iced tea with you.Hey, vending machine, have mercy on me,
Don’t just blink blue there at the end of the hall.
If I ever have change, I swear to the heavens,
I’ll buy two from you — and one more for lunch.
Here’s the AI-generated industrial-folk metal too, for the full experience: gemini.google.com/share/a672ed964998
And for the joke of it, I made the Japanese version too, because Japan has its nice folk songs as well: gemini.google.com/share/e04f3de328f6
Breakthrough!
Incredible, but right by the afternoon I managed to twist the embedding in the neural networks far enough to reach a breakthrough. The parameter count dropped to roughly a tenth via a mathematical trick, and it’s at parity in accuracy with the very large transformer networks! A bit of differential geometry and similar operations happened. Results soon — and I’ll present it at the seminar tomorrow.
Evening — dinner
I went to Saizeriya to eat. Practically nothing else is open anymore, but it’s always a good compromise. So I ate shrimp tails and a pizza. I love pizza, so I can judge: it’s quite small, but otherwise quite tasty. Even if it doesn’t beat the Finnish reindeer pizza — that’s obviously the best one for me so far.
Then I just bought an orange-apple tea and a few bathroom supplies. And went home.
SMC — accepted!
Wonderful news was waiting at home! Earlier I’d submitted two conference papers to the international SMC 2026 conference: one on HyMeKo, the other in the Kolmogorov–Arnold area (granted, that one became a WiP paper). Both were accepted! A huge honour, since I made the papers largely on my own, entirely based on my own idea. Of course there was a sharper-tongued review in the acceptance too, but in the end everyone found them acceptable, and both papers can be published.
So now I’ll dare to celebrate at night: I’ll shuffle off across the university to a vending machine and buy a coffee and a tea. It’ll be a nice little walk.
Photos
