AI@SGMK

We want learn about how AI works, how we can use it and apply it to our work. As a starting point, we'll organize a workshop to kick things off. We will organize further sessions if folks are interested or want to contribute their own topics, demos and presentations.

Tiny chips, great power —
Neural networks in theory,
Local AI takes flight.
— Qwen 3B

Sunday, Date to be defined, 13:00–16:00, at Bitwäscherei in Zurich

We have not yet decided on a new date and there was a suggestion for an alternative location, if that would give us access to other setups. Otherwise we'll fall back to the Bitwäscherei. Watch this space.

Sunday, 10 May 2026, 13:00–16:00, at Bitwäscherei in Zurich

Having gotten a solid introduction into the theory behind AI, the hardware thats currently out there and some practical local AI LLM demos, the attendees asked get to do some more hands on work with it.

AI Install Party

The idea is to bring your AI capable devices to the workshop (don't forget to also bring power cables) and we can help each other getting some AI installed on it. The Bitwäscherei offers fast internet, so slurping in those models from Hugging Face and Co. should take just a blink, but of course we will still be limited by the memory on your device.

Projects Tackled

• Ollama with Apertus and Gemma models got installed on a Macbook (using homebrew) and on a Raspberry Pi with AI-hat
• We explored ComfyUI, as Zoe could show us her setup. Markus had a local install of it and you can read his short intro here: File:ComfyUI VeryShortIntro.pdf
• On the Radxa SBC presented at the last event we did manage to install Open WebUI and connected it to the RKllama service on the device, but failed selecting the installed models or to install Apertus - This setup will be worked on between sessions and hopefully we could set this up as a service at the Bitwäscherei for folks to toy with
• We did experiment with text and images uploaded to Open WebUI (on an earlier installation) and ollama on the Macbook - Text works very well, Open WebUI lets you provide websites as context and some models like gemma on the Mac did provide decent image-to-text output
• On the Raspberry Pi AI hat, image recognition got setup (editors note: please add some more details on this)
• We didn't yet look into the ESP32-P4X-EYE camera & microcontroller

Sunday, 26 April 2026, 13:00–15:00, at Bitwäscherei in Zurich

Overview

• AI Hardware - From MCUs to server grade NPUs
• Basic Theory - Neural Networks
• Running AI locally on single board computers
• What's next? Which topics do we want to dive deeper into? When do we plan to do a next session?

AI Hardware

Chrismicro will introduce this workshop series, and will give us an overview of what AI-related hardware is out there, from tiny micro controllers (MCUs), consumer graphics cards (GPU), up to server grade neural processing units (NPU).

What to expect:

• What do artificial neural networks really have in common with the human brain?
• How much AI can fit on a microcontroller?
• Why are memory prices rising right now—and what does that mean for AI?
• Can AI run locally, without cloud services and data sharing?
• What hardware is worth investing in today—and what’s coming next?

Local AI – take control of your data:

More and more people don’t want to share their data with the cloud. Local AI opens up powerful alternatives.

• What exactly is an AI model?
• What hardware is required?

Theoretical Basics

Markus will provide us with the theoretical tools to grasp how neural network operate, so we can start to make sense of all the new acronyms and terms that can make your head spin when dealing with AI.

Presentation 26.4.2026: File:AI SomeTheoreticalBasics.pdf

Running Local AI Inference

Simon / El RIDO will show some practical applications of small LLMs, that we can run locally on single board computers (SBC) with a small NPU. For practical trials, a Radxa Rock 5b, an SBC based on the Rockchip RK3588 chipset with 16 GB of RAM and a NPU with 6 TOPS will be available on site, letting us run local, offline inference using the RKNN-LLM framework in a terminal based chat. For some more bells and whistles, such an SBC or a gaming GPU (300 TOPS and more) can be combined with the Open WebUI dashboard through the ollama-API. This lets us come close to the experiance of cloudbased commercial AI chat solutions.

Presentation: File:Running local AI.zip

Planning Future Sessions

We'll collect what topics are of interest to the participants and will try to organize further sessions to dive deeper into the topics or work on concrete applications with the tools at hand. Let's have some fun!

If you have any suggestions, please add them here below:

• Ask Dusjagr to give us a Demo of his Vibe Coding setup for the Attiny Microcontrollers