Hilbert is important to the Lab because it reframes local AI around capacity.
The system is based on an AMD Ryzen AI Max+ 395 class machine with 128GB unified memory and ROCm-oriented local inference. That makes it different from a conventional desktop GPU box: it can attempt models and contexts that are awkward or impossible on smaller discrete cards, even when those cards are faster.
01 – The Workstation Stack
The v2 archive recovers a July 2026 layer of the open-source workstation stack:
- local inference and chat;
- coding agents;
- generative media;
- voice.
The point was not to install one impressive demo. It was to assemble a workstation where local models, coding tools, ComfyUI-style media workflows and voice systems could coexist.
02 – Hilbert Unofficial Suite
The Hilbert Unofficial Suite combined several layers:
- Hilbert Studio, a ComfyUI workflow platform;
- Genesis Runtime, an assistant/agent runtime;
- an AI code-build pipeline;
- ai-switch, a switcher intended to keep GPU-heavy stacks from trampling each other.
The suite's workflow layer was manifest-driven: declared inputs and outputs, model requirements, hardware requirements, tags, presets, registry discovery, schema validation, batch processing and quality-control scoring.
At one point the public registry was recorded as 12 categories and 89 workflows. Adult and community-model checkpoint categories were intentionally excluded from the public registry.
03 – ROCm As A Constraint
The ROCm focus is not incidental.
Most generative-AI tooling assumes NVIDIA/CUDA by default. Hilbert made ROCm and unified memory first-class constraints, which is exactly the kind of constraint the Lab should document: the engineering work required to make AI practical on the hardware actually available.
04 – Failures Count
The Hilbert archive includes operational failures:
- ROCm/unified-memory messages involving Qwen3 Coder Next and SVM allocation;
- sleep/resume instability after moving to Pop!_OS;
- a disk-capacity incident where Code state and the AI suite consumed enormous storage.
Those are not footnotes. They are part of the research output.
05 – Relationship To The GPU Farm
Hilbert and the older GPU farm solve different problems:
- Hilbert provides capacity through unified memory;
- discrete GPUs provide speed and throughput for narrower specialised tasks.
That distinction feeds directly into Omega's cognitive-cluster direction.
Source note: assembled from v2 sections on Hilbert, the recovered July 2026 workstation layer and the Hilbert Unofficial Suite.