[Tokyo Tech Translated] local llm, better scoring, physical ai
three japanese takes on where llms actually work right now
today's selection orbits a single question: what do llms do well when you stop asking them to write poetry? the answers from japanese tech twitter are practical, sometimes physical, and refreshingly unbothered by agi timelines.
@ds_nakajima, zero-cost local ai server
repurposed a mac mini lying around into a local llm server. zero monthly cost, fully offline, no data leakage concerns. the immediate use case is defensive: before pasting logs or code into a cloud ai, run them through the local model to check for tokens, keys, or personal info that should be masked. a privacy filter that costs nothing and runs on hardware you already own. the best kind of infrastructure is the kind you find on the floor.
source: https://x.com/ds_nakajima/status/2073963354474025147
@ai_database, stepwise llm scoring
direct scoring prompts produce lenient, inflated results. the fix is not a better model but a better prompt structure. break the evaluation into multiple steps. specify criteria, ask for reasoning, then request the score. the template is included. this is llm-as-a-judge with a checklist, and it works because the model needs the intermediate reasoning steps to calibrate. asking for a score outright is like asking a student to grade their own essay without a rubric.
source: https://x.com/ai_database/status/2074047196299706718
@itarutomy, amazon backs physical ai
amazon invested in archetype ai, a startup building a foundation model called newton that ingests sensor data instead of text. factories generate massive sensor streams but discard 70 to 90 percent of it. the problem is sensor fusion: two binary sensors create 1,536 interpretation states, three create 24,576. hand-coded rules collapse under the complexity. custom ml models took twelve plus months and five plus engineers per use case. newton, pretrained on sensor data, reasons across cameras, microphones, vibration, and temperature sensors in real time. setup is natural language, minutes. it predicts conveyor belt failures days out with no additional training. founders from google's atap. total raise forty eight million dollars. amazon reports twenty five percent efficiency gains in its own supply chain. the next axis for generative ai might be physical, not textual.
source: https://x.com/itarutomy/status/2074175328180965535
three threads, one pattern: llms are most interesting when they leave the chat window. a mac mini as a privacy guard. a scoring prompt that actually works. a model that reads vibration instead of vocabulary. the discourse in japan this week is less about what models can say and more about where they can sit. on your desk. in your eval pipeline. inside a factory. the physical turn is already here, just unevenly distributed.
more at falsifylab.com