August 9, 2026
Napster for AI? Cue the panic
Show HN: Lumabri – What if LLMs worked like Napster?
AI Napster drops on HN, and the comments instantly turn into a speed-vs-safety cage match
TLDR: Lumabri wants to let people share AI models across a swarm of everyday computers, borrowing pieces as needed instead of downloading everything at once. Commenters immediately fought over whether it’s brilliant or doomed, with the loudest worries being speed, privacy, safety, and a README one reader called "slop."
A new Show HN post pitched a wild idea in plain English: what if chatting with AI worked more like file-sharing, where one computer hosts the brain and other computers pull only the pieces they need, then keep them cached for later. The creator says it can run on ordinary machines, not just expensive gaming rigs, and even keeps working after the host goes offline once the needed parts are mirrored locally. In other words: the dream is "AI for the people", not just for hardware whales.
But the real fireworks were in the comments, where the community split into camps almost immediately. One side was brutally unimpressed, with one reader snarling that the README was basically unreadable "slop" — the kind of Hacker News drive-by that lands like a slap. Another camp went straight to the nightmare scenarios: what if someone uses your shared AI for something awful, and what if your private prompts get logged by strangers? That turned the thread from cool demo to moral panic in record time.
Then came the performance doomers. "LLMs are already slow on one machine," one commenter groaned, basically asking: you want to make it slower by adding the internet? A more technical critic piled on with the classic three-hit combo — compute, latency, security — arguing the whole thing runs into physics, not vibes. So yes, the project got attention, but the comments read like a reality show reunion: one person yelling about readability, one about murder, and one about speed. Peak internet.
Key Points
- •Lumabri is presented as a peer-to-peer system for running large mixture-of-experts models with the colibri engine in pure C and no dependencies.
- •Client machines fetch only the model bytes touched during inference on first use, cache them locally, and then reuse them from disk for later responses.
- •The system is designed for CPU and SSD first, with GPUs treated as optional accelerators that do not change the output.
- •Users can join a swarm as chat-only participants or donate disk, compute, or both, with disk donations verified against the operator’s signature.
- •A single tracker can index multiple models, and clients can switch live between model architectures such as GLM, OLMoE, and DeepSeek.