▸ BUILDING
production-stabilizing the Localeksi inference stack. PSU-instability-driven CUDA errors finally tamed.
Currently running the FP8 fallback stack at ~23 t/s · 110K ctx while I dial in the
INT4 + TurboQuant + MTP config that hits 37-51 t/s on real ClawdHawk traffic.
Also: thinking about how the Pocket Oracle build codex v1 becomes v2 — what the minimum viable
sovereign inference rig looks like for someone who isn't me.
▸ READING
- papers on speculative decoding and MTP acceptance dynamics
- federal acquisition regulations (the riveting kind)
- Bulkowski's chart pattern encyclopedia (still)
- some philosophy book my wife recommended that I'm pretending to read faster than I am
▸ THINKING ABOUT
How cross-domain insight is the cheat code — depth + breadth = unfair advantage,
but only if you can make the synthesis legible to people who haven't done the work.
Whether the next five years of AI infrastructure look more like the early internet (everyone gets one)
or like power plants (a few people get one, everyone rents). The answer determines a lot about
what Timechain should build.
▸ ON REPEAT
mostly the hum of four GPUs and the soft whir of an Antminer S19J Pro in the garage. very techno.
otherwise: whatever my daughter has commandeered the speaker for.
▸ NOT DOING
- "quick syncs"
- partnership decks with no signal
- raising money I don't need
- missing my kid's school pickup