The morning of September 13 was a day when the center of gravity shifted from chasing frontier models to "how to use LLMs smartly." An attempt to boot a local LLM directly on a Raspberry Pi without going through an OS drew attention on Hacker News, while in the Japanese-language sphere, the lineup centered on cost-minded practical topics such as Sakana AI's "Fugu Max," which bundles inexpensive models, and token-waste countermeasures to prevent generative AI budget overruns.
Raspberry Pi — Booting a Local LLM "Directly" Without an OS
A project that tickles the hearts of local LLM enthusiasts has been making the rounds. According to XDA Developers, a user ran an experiment booting a local LLM directly on a Raspberry Pi without starting an OS such as Linux. The idea is to skip the OS layer and bring up only an inference environment — a DIY project that diverts the resources of low-spec hardware to the model to the greatest extent possible.
We have recently also seen efforts like "smolbenchmark," which benchmarks models that run in under 8GB on tablets, smartphones, and the Raspberry Pi, and the trend of running LLMs at the very edge of the device is pushing into ever more extreme territory. It is less about practicality than an extension of the desire to run everything on-device, but it can be seen as an experiment with implications for embedded and educational use.
famulor — A Full-Duplex AI Phone Agent That "Listens While Speaking"
Show HN featured "famulor," a "full-duplex" AI phone agent that can listen while it speaks. Conventional voice AI has been basically half-duplex, with the parties taking turns, so the agent had to wait for the other side to finish talking. With full duplex, if the user interrupts, the agent cuts off its words and can carry on the conversation while giving backchannel acknowledgments.
Naturalness on a phone call, it turns out, depends heavily on exactly these two things: interruption and backchannel responses. Whether an AI can keep step with the rhythm of human conversation is likely the watershed that decides whether voice agents become products people actually use. The publicly available information is still limited at this point, but it is worth attention as a signal of a qualitative turning point for voice AI.
Sakana AI's "Conductor AI" Fugu Max — Winning Cheap Without Relying on Expensive Cutting-Edge Models (Updated from the Previous Fugu)
A new explanatory article on Sakana AI's "Fugu" — which made news earlier this week for claiming to beat GPT-6 Astra on some tasks — has appeared in the Japanese-language sphere. A Qiita article introduces the mechanism for winning cheap without leaning on expensive cutting-edge models: Fugu Max brings multiple models together like a "conductor."
Rather than depending on a single giant model, the idea of directing and combining a fleet of cheap models according to the situation creates cost competitiveness on an axis separate from the performance race. As per-model prices keep falling, if the value of the "orchestration layer" keeps rising, the reason to stay locked to one frontier vendor will fade. For the implementation details of the ensemble and routing, refer to the original article for now, but as a cost strategy from a Japan-born AI venture, this is one to watch.
An Era Where "Hiring Humans Is Cheaper Than AI Coding"? — The Full Picture of Token-Waste Countermeasures
@IT poses the question "Is an era coming where hiring humans is cheaper than AI coding?" and pulls together token-waste countermeasures for preventing generative AI budget overruns. At a time when AI coding efficiency is all anyone talks about, this is a practical article that squarely addresses the possibility of costs overtaking personnel expenses.
As previously covered on this blog, measurement-focused verification has already been active in the Japanese-language sphere — for example, findings like "80% of our billing was cache_read (context reloading)." The new @IT article is an extension of that work, and what is new is that it systematically organizes "countermeasures to prevent budget overruns" rather than measurement itself. The quiet cost blowouts at adopting companies are a real problem, and a sense of AI coding costs is no longer a matter of individual engineers' literacy but is fast becoming a management issue.
11 New Models in 13 Days — "Three Model Names" Lined Up in an Automation Task
An article looking back at the new-model rush of the past few weeks also caught the eye. The author of a Qiita post, while researching the "11 new models in 13 days" of releases, noticed that the settings screen of the automated posting task they operate listed three model names, a mix of old and new.
This blog has been tracking new models day after day as well — GPT-6 Astra, DeepSeek V4.1 Flash, Qwen3.8, and GigaChat-3.5. Individual models may be consumed as breaking news, but the automations and pipelines built around them tend to get left behind. The situation where "which model to specify" stays buried deep in a config file and never gets refreshed is probably familiar to many operators. It is an episode that drives home the importance of pulling model selection out of the code and keeping it in a trackable form.
Three Approaches to Leaving Accounts-Receivable Reconciliation and Collections to AI
AI is also working its way into accounting back offices. On Qiita, an implementation article was published introducing three approaches to automating the chasing of unpaid invoices — accounts-receivable reconciliation (matching payments against invoices) and collections — with AI.
More than flashy generation, the steady accumulation of routine tasks such as matching, classification, and drafting messages is exactly the territory where LLMs originally excel. It draws less attention than coding assistance, but the movement to turn "the implicit matching work once done by hand" into AI-driven work is the kind of change that quietly but surely pays off in the operations of Japanese companies.
Wrap-up
From execution experiments at the very edge of the device, such as OS-less booting, to the practical craft of "using AI smartly" — bundling models, preventing waste, automating accounting — this morning's news was filled with topics about how to really work AI in. It may be a good time to take a step back from the stream of new-model headlines and reassess your own environment and costs.