Comment

Comment on Lumo: the least open 'open' AI assistant

UnfortunateShort@lemmy.world ⁨4⁩ ⁨days⁩ ago

TLDR: Apps are open, models are open weight but you don’t know which one you get, and they have not yet disclosed any tweaks they may have made (although for me, it never sounded like they actually tweak the models all too much)

source

Sort:hotnew top

XLE@piefed.social ⁨4⁩ ⁨days⁩ ago
Your tl;dr appears to be missing some important data. You can have an opinion but please don't represent it as an accurate summary.

Things you crucially missed:

Less open than every other service available

Bills itself as the most open

Server side source code is MIA

No model card available. Evaluations, risks, biases, guardrails and safety measures unclear.
source
- fmstrat@lemmy.nowsci.com ⁨4⁩ ⁨days⁩ ago
  It get worse, and the model weights is a bit inaccurate with the Sept update:
  
  The only open source code we have found is for the Lumo mobile and web apps. Proton calling the Lumo AI assistant open source based on that is a bit like Microsoft calling Windows open source just because there’s a github repository for Windows Terminal.
  
  The models listed on Lumo’s privacy policy page are “Nemo, OpenHands 32B, OLMO 2 32B, and Mistral Small 3”. OpenHands is a QWEN fine-tune, and Nemo and Mistral Small are both Mistral models. Since Proton has open-sourced neither the Lumo system prompt nor the mysterious routing methods that decide which model will handle your query, you never know what you are going to get.
  
  So if the server isn’t open source, and the server does all the work, this system is simply not Open Source.
  
  source