§04 · WRITING · TRANSMISSIONS RECEIVED ON DESCENT
Notes from running big models on small hardware.
Measured results, including the things that did not work. Mostly local LLM infrastructure, multi-agent systems, and automation.
RX // SIGNAL CLEAR · 2 TRANSMISSIONS ON RECORD
LOG
Transmissions
TX 04.2
2026-08-20
2026-08-20
2026-08-20
I trained a language model from scratch on my gaming PC
29.9M parameters, my own architecture code and my own tokenizer, random weights to coherent prose in sixteen minutes. Nine hours on one RTX 3070 Ti — including the two silent bugs that made the first attempt eighty-six times too slow.
TX 04.12026-08-20
What actually makes a 27B model faster on an 8 GB GPU
Six experiments on a single RTX 3070 Ti. Multi-token prediction doubled throughput for free, speculative decoding made things worse, and the biggest win turned out to be a Windows setting that was silently moving VRAM into system RAM.