Latest
Four hundred apps, one at a time
NucleoOS-P4 has a Store with over 400 apps and an assistant, on a microcontroller. I chose the features. Memory decided the shape.
5 min read
- esp32
- operating systems
- memory
Writing
Writing
Notes on how the things above are built.
-
3 min read
A computer sold as a component
The most powerful chip in the ESP32 family has no wifi: a second chip sits next to it. The family born as a radio module took the radio out.
- esp32
- embedded
- hardware
-
3 min read
A DJ who does not exist
At the top of the file that decides what the Radio Index DJ does it says: no ML, pure deterministic logic. And that is the part nobody should notice.
- radio
- audio
- ai
-
3 min read
The smallest useful amount of intelligence
On the Cardputer, once it has booted, about 18 KB of memory are left. A language model does not fit. That is not why the assistant isn't one.
- on-device ai
- assistants
- constraints
-
4 min read
Leaving memory for a machine
My game's instruction file had a rule against copying the project's numbers. Thirty lines above it, the numbers were copied, and wrong.
- craft
- documentation
- ai
-
4 min read
Trusting code you didn't write
I changed my game's AI and every metric improved. Then I put the new version against the old one, and it lost eight to sixteen.
- craft
- testing
- ai
-
3 min read
A constraint is a specification
The cover of 512 KILOBYTE shows 512 memory cells, eighteen of them still free. It is not a metaphor. It is a readout.
- book
- operating systems
- memory
-
4 min read
OpenClaw is not a coding agent
I installed it expecting a rival to Claude Code. It is another thing. Once you know what it is, it becomes useful: the local model, reachable from Telegram.
- openclaw
- agents
- local inference
-
4 min read
On eight gigabytes of VRAM, size is not what matters
A dense 27B runs at 2.5 tokens per second. A 35-billion MoE, on the same card, does thirty. The difference is not size.
- local inference
- llama.cpp
- gpu