Podcast / Algorütm, episode 347
Estonia's weekly developer podcast, recorded 4 August 2026 and published 27 August. One Steam Machine instead of a basement GPU rig, and a live demo where the assistant said it had no data.
Algorütm is Estonia's weekly podcast for developers. Priit Liivak, Erik Jõgi and Martin Kapp have covered self-hosted AI on the show before, but always the multi-GPU basement rig. I brought the opposite: a Steam Machine that was bought to play games and does its real work at night.
8 GB of VRAM fits exactly one model at a time, and that single fact decides the whole design. A tool loads the model it needs and evicts whatever was there. Idle models unload after two minutes. You stop interleaving work and start batching it by what is loaded. One model at a time turns out to be enough for a personal assistant.
The episode is in Estonian. It was recorded on and published on . Chapters below, then the numbers, then the full machine transcript on its own page.
Episode 347, hosts Priit Liivak, Erik Jõgi and Martin Kapp, with me as the only guest. Priit draws the contrast himself: the previous self-hosted AI guest brought a basement GPU cluster with 25G networking. This one runs on a games console.
Arch Linux underneath SteamOS, 16 GB of RAM, 8 GB of VRAM, a semi-custom AMD part, low power draw, and bought to play games. The framing I keep coming back to: right tool per need, and for personal needs the small quiet box is enough.
A notebook with more capability. It reads the calendar and the mail through custom MCP servers, including one for the kindergarten's notices. The thinking stays with people: it collects and drafts, I decide. The honest limit named out loud is context, roughly 25 pages on the box against a million in the cloud. Then the architecture, a cron job starting a Claude Code session that reaches the box's MCP server over Tailscale. Martin's summary is the best line of the segment: usually the MCP server is out on the internet and your tool reaches out to it, here the direction is inverted.
Tailscale, SteamOS on top of Arch, and models kept resident on disk rather than pulled across the network on every switch, which is the real contrast with a fast basement rig. Unload plus RAM cleanup on demand.
The cloud is the brain, the box is the senses. The night researcher gets the network, the night coder does not and works inside Docker, and I have seen the plan before anything runs. I take the sandbox-escape risk seriously on air rather than waving it away.
Local models on ordinary devices, Estonian open weights on a phone, and the honest gap: Estonian speech synthesis has good diction while the sentence structure coming out of a cloud model still reads clumsy in Estonian.
I ask the assistant on air which Algorütm episode is the best one. It answers that it has no data on that. I will take that over a confident guess every time. Then the embedding model and the podcast corpus behind the question.
Including the run that generated 80 music files overnight while I slept, to find one speaking voice.
Instructions hidden in audio, and the laser-through-a-window attack on smart speakers.
Priit encourages open sourcing it, unprompted, and names exactly the valuable part: the model load and unload and the memory management that make weak hardware usable.
Not quoted from anyone's documentation. A machine's own docs are a claim, not a measurement, which is a lesson this project learned the hard way.
The full transcript
YouTube's automatic captions for the episode, unedited and in Estonian, are on their own page: the machine transcript. It is ASR, so it mangles names and terms and should be used to find a moment, never to quote one.