Internet in a box

Sort by newest Sort by date
JTH

JTH

TSP Legend
Reaction score
1,411
I've been going down numerous rabbit holes, testing Llama AI models, running it in a VM, I'm losing about 10-15% overhead on resources, but we are still in beta and I don't want to alter the HOST system.

It started with the idea I wanted an Internet in a box type of Cyberdeck (fully offline devise) for when the world ends.
You can download all the WIKI, and of course all other sorts of documents you deem relevant to rebuilding the world.
I can then build an extensive library, using purpose built AI models to sift through it when I need quick indexing or answers.

If you have the right data and models, you can build advisors, like a Doctor, Farmer, etc

Anyhow, I rebuilt the Laptop, due to needing one large partition to store models and databases in one space. An 8-core AMD Ryzen 7 4700U with 16GB of ram, and 500GB storage (without a dedicated GPU), can run Small" Llama 3 8B Instruct models if they are heavily quantized. This is about the equivalent of Chat GPT-3.5 but with about half the context memory, so I can run similar (but shorter) task.
 
Last edited:
JTH said:
I've been going down numerous rabbit holes, testing Llama AI models, running it in a VM, I'm losing about 10-15% overhead on resources, but we are still in beta and I don't want to alter the HOST system.

It started with the idea I wanted an Internet in a box type of Cyberdeck (fully offline devise) for when the world ends.
You can download all the WIKI, and of course all other sorts of documents you deem relevant to rebuilding the world.
I can then build an extensive library, using purpose built AI models to sift through it when I need quick indexing or answers.

If you have the right data and models, you can build advisors, like a Doctor, Farmer, etc

Anyhow, I rebuilt the Laptop, due to needing one large partition to store models and databases in one space. An 8-core AMD Ryzen 7 4700U with 16GB of ram, and 500GB storage (without a dedicated GPU), can run Small" Llama 3 8B Instruct models if they are heavily quantized. This is about the equivalent of Chat GPT-3.5 but with about half the context memory, so I can run similar (but shorter) task.
Click to expand...
Interesting. Good luck.
 
  • Like
Reactions: JTH
Cool stuff JTH,

So, which models have stood out to you so far? And what is the largest one you've been able to run on your VM?

When I find time this week, I'm thinking of downloading and playing with Qwen3-4B-Thinking-2507 off of LM studios.
 
TommyIV said:
Cool stuff JTH,

So, which models have stood out to you so far? And what is the largest one you've been able to run on your VM?

When I find time this week, I'm thinking of downloading and playing with Qwen3-4B-Thinking-2507 off of LM studios.
Click to expand...

Nothing stands out just yet, I've only been running basic reasoning test to compare how they answer the questions. Truth be told, it's both better than I expected (but slow), and also disappointing because I know I don't have the specs to run larger models. I'm slamming the CPUs hard which means when running the models I have no room for real-world task.

We both know what's going to happen here, I'm going to have build a machine...

Screenshot_2026-03-23_19-57-41.png
 
Progress report.

I've moved the AI software unto the Host. Also managed to get my integrated graphics card to kick in some of the work on the smaller models. The Llama Server was easier, the Open Web UI has harder but it has a lot of options.

There is a ton of features under the hood here, everything you wished GPT can do, this already has it, you can control many parameters, like logic/creative temp, context size and so much more. I have the Voice Module loaded, but need to configure this as it's own server service. Next I'll work on the project folders, so I can store specialized instruction sets. Then at some point expand out and feed it news (it's not connected to the internet) so some of these models are unaware of present circumstances.

Screenshot_2026-03-27_18-24-45.png
 
For the internet in a box topic, I've developed a fascination with 12-Volt low-powered processors like the Intel N200, something that sips little power and can be re-charged easily. Something small that can be carried, and can be protected with a faraday pouch. Not sure if you've noticed (it's been 10 years since I've built a PC) it now seems cheaper to buy vs. build. And frankly all of it is overpriced.

As for the AI project, the laptop can run up to 9B-Q5-KM models with no IGPU at about 4 words per sec (Super slow) I just don't have enough juice to run these at a fast pace. Smaller models run faster, but they are dumber like a 3rd grader. I did get internet searches to work but seem to be having issues with them parsing internet pages, they love to give you links, but they don't want to read the data. I broke all of it today trying to add Ollama to my Llama for the Open WebUI. So I'm going to remove it all and rebuild from scratch (again).

Also I put Linux on my Arm-Based ChromeOS tablet, cool right?
It is until you find out your USB-C Video output to external monitor no longer works... nor does your Netflix because they haven't developed the DRM (Digital Rights Mgmt). So yes, I'm the proud owner of a linux ARM tablet, but I feel like I only have about 60% of the linux features my X86-64 MX Linux has.
 
I asked grok to dumb down JTH's last post for me: 😁

Hey kiddo! Imagine a grown-up guy who's super into building his own cool computer toys. He's writing a note to a friend about two fun (but kinda frustrating) projects he's messing with. I'll explain it like a bedtime story, super simple:

The "Internet in a Box" Idea

He's dreaming of making a magic little box that has the whole internet inside it — like a backpack-sized computer you can grab and take anywhere (even if the power goes out or something bad happens). He’s really excited about tiny, battery-friendly computer brains (like one called the Intel N200). These brains are awesome because they sip power like a tiny sip of juice — not a big gulp. You can recharge them easy-peasy with a car battery or solar charger. The whole thing is small enough to carry in your pocket almost, and he wants to tuck it into a special shiny bag called a "Faraday pouch." That bag is like an invisible superhero shield that blocks zappy radio waves so no one can spy on it or fry it with lightning or weird storms.

He also grumbles that back when he built big computers 10 years ago, it was cheaper to make your own. Now? It’s actually cheaper (and easier) to just buy one ready-made… but he thinks everything is way too expensive anyway. “Boo, prices!”
😤


His AI Robot Brain Project

Next, he’s trying to build his own smart robot friend (called AI) right on his laptop.
  • He can make the AI “think” using medium-big robot brains (the 9-billion-parameter kind). But it’s slooooow — only spits out about 4 words every second! That’s like waiting forever for your friend to finish a sentence.
  • His laptop doesn’t have enough “juice” (power) to make it run fast.
  • Tiny robot brains run quicker… but they’re super dumb, like a silly third-grader who doesn’t know much.
  • He got the AI to search the real internet, which is cool… but it’s being lazy. It just hands you links (“go read this yourself!”) instead of actually reading the page and telling you the good stuff.
Then today he tried adding some extra magic software (stuff called Ollama, Llama, and Open WebUI — basically ways to talk to the AI easier). Oops! He broke EVERYTHING. So now he has to delete it all and start over from scratch… again. Classic “I touched it and it exploded” moment.

The Cool Tablet Adventure

Last part: He took his tablet (the kind that normally runs Google’s ChromeOS and has an ARM brain — that’s just a different family of computer parts) and put real Linux on it instead. Linux is a free, super-powerful computer operating system.

He says, “Cool, right?!” …until he found out two things broke:
  • He can’t plug it into a big TV or monitor anymore with the USB-C cord (video stopped working).
  • Netflix won’t play because it needs special “movie locks” (called DRM) that the new setup doesn’t have yet.

So now he’s the proud owner of a Linux tablet… but it only feels like 60% as good as his normal powerful Linux computer (the one with the x86-64 brain). It’s like getting a shiny new bike but two of the gears don’t work.

That’s the whole message! He’s having a blast tinkering but also hitting a bunch of “why won’t you work?!” moments. Grown-ups who love computers call this “fun frustration.”
😄
Click to expand...
 
Since doing fresh installs on the Tablet and Laptop, I finally got around to using a note taking app to keep track of the steps used to configure the systems. I settled on Joplin, I like the MS-Outlook type gui and it's a small footprint. I'm using Syncthing to keep the same database between the Tablet and Laptop, thus it's like bypassing my usual Google Docs route.

I really have no idea what I'm doing, I am literally the dumbest nerd you will ever meet.
 
I've been using Virtualbox since the beginning of time. Like an old car, when it breaks down you know how to fix it.

But, I just recently switched over to QEMU/KVM, on this system, its faster and lighter on every metric.
8% faster CPU performance
11% faster Browser benchmarks.
4-CPU/4GB Ram Bootup in 21 seconds, vs. Virtualbox's 28 seconds.
A fresh build full linux from scratch took <5minutes
Only downside I see is it consumes 380MB more ram from the host at boot, but the performance difference is worth the tradeoff.
 
Internet in a Box Version 2.0

Picked up a used 1-Liter business PC, post-lease Class A (minimal wear) from Jul-2020.
Lenovo ThinkCentre M75q Gen 2 Tiny Desktop - Ryzen 5 Pro 4560GE
  • Will upgrade from 8GB to 64GB Ram, and add more storage,
  • Also upgrade from 65-Watt to 90-Watt Pwr Supply and convert over to USB-C
  • Other plans, perhaps additional cooling, RealTek Software Defined Radio (HF/VHF/UHF)
Overall it's a good low-watt desktop, I'm expecting a 10-30% performance boost over the laptop.

1PC.png2PC.png
 
The upgraded 200 Watt USB power supply works great on the PC. While it did create more headroom to draw more power, it did not yield much of an increase in performance due to the CPU's 35-Watt (Energy saving) cap. But now I can power all my devices with 1 power supply, so their's less stuff to drag around.

Spent the better part of a week fighting QEMU-Resizing/Docker/Nextcloud. It all works, but I did not foresee that my simcard-router will not port forward. There are alternatives, but I don't like them, so I'm going back to simple Synthing of shared folders.

00A.jpeg
 
JTH said:
The upgraded 200 Watt USB power supply works great on the PC. While it did create more headroom to draw more power, it did not yield much of an increase in performance due to the CPU's 35-Watt (Energy saving) cap. But now I can power all my devices with 1 power supply, so their's less stuff to drag around.
Click to expand...

A small correction, yes the power supply works great, but it is not a good fit for a desktop machine without it's own battery like a laptop has.

Reason: When you plug a 2nd device into the power supply (say a phone), the GaN charger briefly cuts power because it reallocates wattage between ports and renegotiates USB-C power delivery. It's not an issue for a Laptop/Phone but the desktop loses power instantly because it has no battery backup.

So this power supply will just have to be allocated to the laptop/tablet.Screenshot_2026-05-17_11-08-56.png
 
Consider, if you will: a future where the world remains connected by wires and screens, yet every word, image, and truth is quietly filtered by machines that decide what each nation and each person is allowed to know.
  • The internet no longer spans the globe; it ends at the border, reshaped by every government that controls it.
  • Artificial intelligence becomes the unseen editor, rewriting reality one harmless-looking summary at a time.
  • Censorship arrives not with force, but with comfort, convenience, and the promise of safety.
  • Most never notice the walls closing in, because the cage is built slowly, and tailored to feel like freedom.
  • In the static of old shortwave radios, voices from beyond the border return, crackling reminders that another version of the world still exists.

I've finished building my Speech-to-Text Natural Language Pipeline. This system automatically turns live foreign audio broadcasts into readable English summaries, it's a Local-AI-powered monitoring station. The process is simple, Record-Transcribe-Translate-Summarize.

  • Simple pipeline: how it works​

    • Step 1 — Record audio
      • Script continuously captures live audio (radio, stream, system sound).
      • It automatically splits it into small files (like every 2, 5, or 10 minutes).
    • Step 2 — Transcribe speech to text
      • Each audio file is sent to a speech-recognition Whisper-based server.
        The system processes the audio and converts spoken language into written text (a transcript).
    • Step 3 — Translate to English
      • If the source language is Polish, Russian, etc., a Local-AI model translates it into English.
      • This makes foreign broadcasts readable in one common language.
    • Step 4 — Summarize
      • A Local-AI model reads the translated text.
      • It creates short bullet-point notes (“cliff notes”) of the important points.
So the proof of concept works, I've tested on Polish/Russian/Mandarin Chinese, an hour of audio takes about 5-10 minutes to fully transcribe/translate/summarize.

The next step is to purchase a magnetic loop antenna and a software defined radio, then configure this on the PC. The GROK image below is not accurate, but close enough to where you undersatnd the setup.
 

Attachments

  • Untitled.jpeg
    Untitled.jpeg
    111 KB · Views: 8
JTH said:
Consider, if you will: a future where the world remains connected by wires and screens, yet every word, image, and truth is quietly filtered by machines that decide what each nation and each person is allowed to know.
Click to expand...

More than 340 newspapers block the Wayback Machine. What that means for the future of the internet​

But by this month, the number of U.S. news websites throttling access to the Wayback Machine had grown to more than 340. Most are local outlets controlled by just five conglomerates: USA Today Co., McClatchy, Advance Local, MediaNews Group and Tribune Publishing.

Eighty percent are local newspaper websites owned by USA Today.
 
JTH said:
I've finished building my Speech-to-Text Natural Language Pipeline. This system automatically turns live foreign audio broadcasts into readable English summaries, it's a Local-AI-powered monitoring station. The process is simple, Record-Transcribe-Translate-Summarize.
Click to expand...

The above Local-AI-powered monitoring station is more for static scheduled monitoring, but when looking for new sources, I need a live version, so I've added a Live Feature, with Auto-Word-Tags.

We have sets of different Word-Tag list based on what we want to ID, key words like "Chemical Spill" or "Explosion"

In the example below, the word "Stocks" was tagged on a Chinese broadcast, which triggers a pop-up sound notification.

Upper left screen: Live Native Language Transcription.
Upper right screen: Live English Language Translation.
Lower left screen: Live English-Word Tagging.
Lower right screen: Live Video/Audio Feed + the Pop-Up notification.

Screenshot_2026-05-24_12-04-21.png

This guest machine is running the live setup, consuming 15% of 3-Threads throttled to a 3.6GHz limit, and 38% memory from 6GB.
The Host is doing most the the heavy lifting, running the AI/Whisper-Live Dockers, but the Dockers are also throttled to 2-Threads and 8GB memory.
 
Last edited:
You must log in or register to reply here.
Back
Top