[1]: https://github.com/aivis-project/AivisSpeech-Engine

    Model                         pp (t/s)    tg (t/s)
    Qwen 3.6 27B            900           29
    Qwen 3.6 35B-A3B   2100          85
    Gemma 4 31B            750           28
    Gemma 4 26B-A4B   2500         90

- Layer splitting used. Tensor splitting is ~1.2x faster in TG, but power spikes from 150W to 380W.

> Von Neumann computer: powered by programming vs. Personal robot: powered by family happiness. https://i.gzn.jp/img/2014/06/05/softbank-conference/snap3014...

His sense of humor has only reached new heights since then.

IMO, the tight-knit division of labor between Toyota and its subcontractors is a slightly different story from the broad diversification within a single corporation. While the latter was historically bolstered by strong industry-academia ties (often driven by university cliques), we rarely see this kind of broad diversification happening in recent years. That said, Japan's traditional "membership-based" employment system, combined with a cultural reluctance to shut down unprofitable business units, is likely what has allowed this diversification to persist for so long.

In any case, Japanese companies are currently struggling with the friction between their traditional corporate culture and the superficial adoption of Western concepts like DX, Agile, meritocracy, job-based employment, and a startup-centric mindset. I suspect Korea might be facing similar structural clashes, though perhaps you are adapting at a much faster pace.

By the way, even while a custom domain is still pending verification, the GitHub Pages LB will route the request based on the Host header, allowing for the following:

    dig +short github.io | head -1
    185.199.108.153

    curl -H "Host: 42.news.ycombinator.com" 185.199.109.153
    hello

Comments URL: https://news.ycombinator.com/item?id=48168982

Points: 2

# Comments: 0

I admit I sometimes get caught up in the tooling for its own sake, but I find local models useful for specific tasks like migrating configuration schemas, writing homelab scripts, or exploring financial data.

It might sound a bit paranoid, but privacy is another major driver for me. Keeping credentials and private information off cloud services is worth the extra friction.

The system prompt and tools have very little overhead (<2k tokens), making the prefill latency feel noticeably snappier compared to Opencode.

[0] https://www.npmjs.com/package/@mariozechner/pi-coding-agent#...

    ```
    # platformio.ini
    platform = https://github.com/pioarduino/platform-espressif32.git#55.03.37
    framework = arduino
    ```

This means a 100k token request counts the same as a 100-token one. I’ve made about 8000 requests in the last two weeks, averaging around 80k tokens per request. It feels like they’re subsidizing this just to gather data on agentic workflows.

On the downside, the speed is mediocre (15–30 tg/s for GLM-5), and I’ve seen the model glitch or produce broken output about 10 times out of those 8k requests.

Needless to say, benchmarks are limited and impressions vary widely by problem domain, harness, written language, and personal preference (simplicity vs detail, tone, etc.). If personal experience is the only true measure, as with wine, solving this discovery gap is an interesting challenge (LLM sommelier!), even if model evolution eventually makes the choice trivial. (I prefer Gemini 3 for its wide knowledge, Sonnet 4.6 for balance, and GLM-5 for simplicity.)

Most of these are only for formal settings. Honestly, I haven't even heard of some of them. Aside from Tatebashi (sticking chopsticks in rice), they’re mostly avoided for hygiene reasons. As for Nigiribashi (clutching them in a fist), it just looks a bit strange for an adult to do.

Regarding lightweight laptops, the Fujitsu FMV Note U series (14-inch) weighs only 634g-917g with Arrow Lake 255H and a replaceable battery.