Those who are following the more serious and finally respectable parts of that Blogg and thereby my Hacker side, in terms of the keyboard sound not what MTV did with it...
I have had to change on my Beowulf Cluster's headnode the OS being also down onto again 16GB RAM from 32GB.
Ubuntu Server is running stable with a Minimal Desktop on top. The beauty of Ubuntu and Linux in general is the very different architecture concept and not the price tag in comparison to Microsoft products.
Every developer can deeply embed his creations into the operating system by latterly looking at the source code of the OS, while MS makes everyone learn APIs, which is the difference of understanding a set of locks and keys or having the blue print and building side tools.
The Cyberdeck is a Knowledge System that uses low cost hardware bypassing its limitations using a Beowulf Cluster layer that lets several Computers in a LAN share its CPU and has small size AI models installed on each. This also means to having to have different software run on different computers, eventually. I have a small i7 with two screens attached, another small i7 Laptop next to it, an i5 sitting duck with a broken screen with a headless OS on which top a small Raspi5 with an AI kit and 7inch touch screen sits.
There is even software around to share by software both Keyboard and Mouse over all Computers and screens, but instead I use a Terminal session at this point. I could use OpenOffice Spreadsheets on the main node and use that software to write using OpenOffice Word on the Laptop having the look and feel of one Computer.
Obviously, that's too simple for me.
The Cyberdeck is more than that and part of it is the Obsidian-Deck. I basically downloaded all DeepSeek chats to further process them locally and additionally a Wiki dump of all English Wiki articles, some Open Library Books to combine that data with some Website content.
I can talk about my DeepSeek chats with one AI and have another AI that has wiki and chosen books as its data points.
Then, I had reason to port from Linux Mint to Ubuntu Server LTS by a little nasty USB stick and an issue with booting while having an installable ... to then figure out that Python environments are deeply embedded into the OS with some more side effects than stabilizing the OS called Dependency Conflict. That's red lines of feedback in the Terminal that tell you "Problem! Not happening."
It was time for an upgrade, because the Obsidian Deck worked just fine and there was more capability out there. Porting the scripts as is to Ubuntu Server from Mint just did not happen, because of the different varieties of Linux Operation System flavors.
Having the system architecture design of the next version of the Cyberdeck ready I just in this very moment finished chunking both Wiki/Books and DeepSeek Chats. Chunking makes a heartbeat!
Even so not shown here, I used for RAGing which is creating the data points of an AI or training it, all three capable nodes by using a server software called HAProxy. HAProxy is load balancing software that takes in a batch process now files from a Python Script and distributes the task with file aka the job to one of the listed nodes. The very problem was that both hardware and thereby also Ollama models are a heterogeneous environment in which all three nodes are different. I could not get HAProxy to handle the different response times of three different in power CPUs using three different LLM models no matter chosen to be most efficient on no GPU old hardware.
Having installed an even smaller LLM copy dedicated to RAGing on each node like
all-minilm:latest to create a homogenious LLM chain over the given hardware fixed initial failing load balancing issue.
Being able to have AI batch processes like analyzing text or chunking use several nodes is a major step forward on this upcycling refurbished old hardware.
The RAGing is not perfect and still work in progress, but functional. The Nexus Version of the Cyberdeck is coming to life.
By starting coding from scratch.
Cyberdeck Nexus
AI LLM model load balancing system
[Knowledge System]
├── a)DeepSeek Chat Processor (batch) Obsidian-Deck (several scripts)
├── b)RAG-AI 1 (specific content) Obsidian
├── c)RAG-AI 2 (general content) Wiki OpenLibrary Books Web
├── d)RAG-AI 3 (local content) Chats RAGed abcd
├── e)Multi-source Database Content listings of source files .md, .txt, html, .xml
└── f)[Future: Coding AI, Essay AI, Simulation AI]
[Knowledge System]
├── a)DeepSeek Chat Processor (batch) Obsidian-Deck (several scripts)
├── b)RAG-AI 1 (specific content) Obsidian
├── c)RAG-AI 2 (general content) Wiki OpenLibrary Books Web
├── d)RAG-AI 3 (local content) Chats RAGed abcd
├── e)Multi-source Database Content listings of source files .md, .txt, html, .xml
└── f)[Future: Coding AI, Essay AI, Simulation AI]
In terms of Open Source AI development does that mean that there is a niche for creating dedicated tailored systems much below and out of sight of the Big Data Center large systems AGI DeepSeek and ChatGPT and Pipe Creators using Microsoft based GUI development tools trying to hit the next Hit App for the Apple Store.
I doubt companies having collected millions to create pipe lines into large companies will be able to compete with off the shelf solutions Google and Microsoft will add to their standard software over time being connected with AGI server farms. They might be the "AI Bubble", but AI came to stay no matter Bubbles...
Creating more jobs over time, like every innovation since the Wheel found an axle.
#cyberpunkcoltoure #deggers
