e/acc

imho PhoneLM (https://github.com/UbiquitousLearning/PhoneLM) & TangoFlux (https://tangoflux.github.io/) sort of sets the baseline for 2025 being an "everyday carry machine spirit" sort of vibe.

also, big takeaway from the semianalysis podcast is that total cost of some amount of capex through compute power can only reach so far, but dataset refinement doesn't get the same capex, therefore the true scaling law is yet to be discovered.

SLMs measure efficacy by successful operations over joules, coincidentally mobile draw has to stack on BT data (potentially FPV) & cell. I wonder what will be announced by smartphone manufacturers by end of 2025, both in reflection on offered capabilities and the financial upside of feedback.

𝖒𝖆𝖓𝖉𝖆𝖙𝖚𝖒 𝖆𝖉 𝖆𝖘𝖙𝖗𝖆

Deepseek V3 runs on 8x Mac Mini cluster, each drawing ~150 Watts? that's no bigger than a standard automobile, though limited range if EV, and iirc Tesla is attempting this inference gambit.

Deepseek V3 runs on 8x Mac Mini cluster, each drawing ~150 Watts? that's no bigger than a standard automobile, though limited range if EV, and iirc Tesla is attempting this inference gambit.
https://x.com/exolabs/status/1872444906851229814

also, FPV audiovideo is the gasoline of 21st century. Alphabet really was that early to whatever "metaverse" value exists atm.