LAIT Night: Linux and AI — April 2026
Tuesday, 21st April 2026
Linux & AI working group
Our April LAIT Night (Linux and AI Tuesday) was an open-floor discussion that covered a lot of ground. Here's what came up.
Multi-modal models and agents
The group touched on the expanding landscape beyond text: Vision-Language-Action (VLA) models that connect perception to physical action, Large Action Models (LAM), and the growing ecosystem around tiny/smol models for constrained hardware. On the agentic side, Agent-to-Agent (A2A) communication protocols and MCP (Model Context Protocol) tool calling are picking up traction as practical ways to wire models into real systems.
Robotics and simulation
Deepbots came up as a Python wrapper for reinforcement learning inside the Webots robot simulator — a practical entry point for experimenting with RL without physical hardware. Physics simulation more broadly is important for training robots before deploying them. One notable example: GTA was used as a training environment for an AI model, which says something about how far synthetic environments have come.
Hardware for local AI
There was plenty of hardware talk. Options on the table ranged from Macs and gaming PCs to the Framework Desktop AI Max+ 395 (128 GB RAM, though memory bandwidth trails Apple Silicon) and multi-GPU SXM2/SXM4 boards (A100/V100) available secondhand on eBay. The RTX Pro A6000 Blackwell with 96 GB VRAM is at the high end for a single workstation card.
The CPU-only reality check: a 52-core HP Z6G4 with 300 GB DDR4 gets around 1 token/second running MiniMax 2.5 at 5-bit quantisation with a 200k context window. Usable for some workloads, not for interactive chat.
One pragmatic take from the floor: wait for some AI companies to fail — their hardware will end up on eBay.
Homelabbing
Running AI locally means running infrastructure. The conversation covered Tailscale for zero-config VPN access to home lab machines, Unraid as a flexible home server OS, and running agents inside VMs. NixAI was mentioned for NixOS users wanting an AI-assisted configuration companion. SuperHouse is worth a look for home automation hardware projects.
Home Assistant and local AI
Integrating local AI with home automation is a natural fit. Home Assistant with Zigbee is the recommended stack. On the IoT device side: Tuya uses a proprietary protocol and is best avoided. Older Tuya devices can be reflashed to Tasmota if the firmware version is old enough. Tapo is reasonably straightforward to work with. Ewelink (used for some sensors) has been more problematic.
Linux distros and AI
Fedora is well-positioned for AI workloads. COSMIC (the new desktop from System76 for Pop!_OS) is worth watching for its AI integration direction.
Coding with AI
Harnesses and coding agents were discussed — the tooling around directing AI to write and run code in structured ways. A practical area where the tooling is evolving quickly.
LAIT Night runs monthly. See upcoming events for the next session.