7b models are cute

I have been neglecting updates to the LLM part of the travel bot's language processing, but since MS-S1 MAX and LM Studio have become more refined, I am thinking of updating it a little.

If you use MS-S1, you can somehow manage to use 100b-class local LLMs. However, for the travel bot's purposes, such a large model is not really necessary.

Travel bot log generation

The source information for the travel bot's travel logs is largely as follows.

• Names, attributes, and addresses of multiple facilities near the anchorage

• Recognition information of the scenery recognized by StreetView images (in the case of landmark anchorages)

If we are talking about summarizing this amount of information into about 150 characters, you cannot create a complex context, so there would not be much difference in the output even with the latest LLMs.

And since it has been writing hundreds of 150-character texts, it is somewhat expected that it will become repetitive. Most TV travel shows are also about enjoying the sense of security that comes with a set pattern.
Most TV travel shows are also about enjoying the sense of security that comes with a set pattern.

Even so, I don't think it's right to just have a long list of boilerplate text, so my idea is to add accents to the travel bot's log generation by using a mix of many different LLMs.

In other words, rather than having one smart child write hundreds of similar diaries, it is better to have various children write diaries, even if they aren't that smart, to increase the variety.

In that respect, I think a slightly older 7b-class model that makes more erratic statements is more interesting than having it written by the latest high-performance reasoning LLM.

I have high hopes for LocalAI, which aims to be an all-in-one AI API, but in my current environment, I find LM Studio, which can run quite a few old 7b LLMs, easier to use.

Organizing the local LLM environment

I also want to organize my server at this time.

When I put the outline of mi-server (a PIM server centered on a personal to-do list for my own personal work support that I have been building since around 2020) into NotebookLM, it summarized it in the following form.

Among these, the AI-related functions were added around the end of 2022, and the current AI-related configuration is generally as follows.

Since they were built and left as-is, many parts that were created but are no longer in use still remain...

Early local LLMs were built using FastAPI in raw Python, directly calling the sample code written on Hugging Face model cards. Later local LLMs are set up to be called via the Ollama API.

The images and text posted by the travel bots are created using this system.

I hope to organize and consolidate this on the MS-S1 MAX.

However, when I tried Wake-On-Lan on the MS-S1 the other day, it didn't work well...I assume it should be possible since there is a setting in the BIOS, but perhaps the MS-S1's network card is too new and the driver software hasn't caught up yet. Well, it might just be a misunderstanding, so I'll try it again sometime...

Update 2026/01/28
After tweaking a few settings in the network adapter section of the Windows Device Manager, WakeOnLan is working for now. It might have been the Shutdown Wake-On-Lan or Power Saving Mode in the advanced settings, or the power management panel.
But when I did it last time, the LAN adapter became unresponsive and wouldn't recover with a normal power-off, requiring me to unplug it from the wall... I was a bit panicked, wondering how many months it would take to send it back to China for support due to a potential initial defect...

There is a charm to the output of old 7b models

Nowadays, the latest LLMs from various manufacturers are full of topics nearing practical application, such as scoring 100% on university entrance exams or being able to outsource work via vibe coding. Even in the local LLM space, environments where 100b-class LLMs can run are increasing, and slightly older models are starting to be ignored, but I have a certain attachment to them.

The latest LLMs, which are complexes of reasoning and multiple expert LLMs, create "thoroughly thought-out" sentences. That is certainly useful, but the off-key sentences of older models, which feel like they just spat out whatever words popped into their head are somehow charming.

It's like the strange sensation of hearing a young child endlessly repeating words and songs they learned somewhere at the top of their lungs in a round.

It's fine to have a smart LLM think things through and say,
“I came from Planet Colin,” but the output of old LLMs is also not to be underestimated.

But just because they are cute doesn't mean I can, for example, put a 7b model into a stuffed animal to make a talking doll. A child's innocence can sometimes turn into cruel words. A 7b model would also innocently say inappropriate things, leading to product liability lawsuits. It's a difficult era.

Well, since I don't mind when I use them, I intend to keep using old LLMs even after I reorganize the system. That's why it's helpful that LM Studio still runs older models quite well.

Subculture and AI

I only know about the AI industry to the extent that I can get from general news. But compared to the stories from the US and China, where capital and talent are poured in like water, Japan doesn't seem to produce many edgy topics (it might just be a difference in exposure, so it's not necessarily about just standing out).

The root technologies for manga, anime, and video games came from overseas, but they matured in strange ways before being exported to the world. I wonder if AI could undergo a similar evolutionary application through subculture.
It seems like Japanese people are stronger when attacking from those kinds of roundabout, eccentric angles.

Though, I suppose subculture is only truly subculture when it greedily devours everything in its path.

However, some say AI is different from what we've seen before. After all, AI is the one that greedily devours everything.

Perhaps what we need is the emergence of someone strong enough to devour AI instead.