Prompt: Is Lumo training data its own, or is it based of other preestablished data? Response: "Lumo is powered by several open-source large language models (LLMs) that have been optimized by Proton. The models used include Nemo (Mistral), OpenHands 32B, OLMO 2 32B (Allen Institute for AI), and Mistral Small 3 (Mistral). These models run exclusively on servers controlled by Proton, ensuring that all conversations remain private. Proton does not contribute any user data to the training of these models, maintaining a strong focus on user privacy." This is what I got from Lumo.

Sure, the internet is more practical, and the odds of being caught in the time required to execute a decent strike plan, even one as vague as: "we're going to Amerika and we're going to hit 50 high profile targets on July 4th, one in every state" (Dear NSA analyst, this is entirely hypothetical) so your agents spread to the field and start assessing from the ground the highest impact targets attainable with their resources, extensive back and forth from the field to central command daily for 90 days of prep, but it's being carried out on 270 different active social media channels as innocuous looking photo exchanges with 540 pre-arranged algorithms hiding the messages in the noise of the image bits. Chances of security agencies picking this up from the communication itself? About 100x less than them noticing 50 teams of activists deployed to 50 states at roughly the same time, even if they never communicate anything. HF (more often called shortwave) is well suited for the numbers game. A deep cover agent lying in wait, potentially for years. Only "tell" is their odd habit of listening to the radio most nights. All they're waiting for is a binary message: if you hear the sequence 3 17 22 you are to make contact for further instructions. That message may come at any time, or may not come for a decade. These days, you would make your contact for further instructions via internet, and sure, it would be more practical to hide the "make contact" signal in the internet too, but shortwave is a longstanding tech with known operating parameters.

In the end I popped up the terminal and used some pot command with some flag I can't remember to skip the login step on setup. I reckon there is good chance you aren't using windows 11 home though right?

This is interesting to me as I like to say the llms are basically another abstraction of search. Initially it was links with no real weight that had to be gone through and then various algorithms weighted the return, then the results started giving a small blurb so one did not have to follow every link, and now your basically getting a report which should have references to the sources. I would like to see this looking at how folks engage with an llm. Basically my guess is if one treats the llm as a helper and collaborates to create the product that they will remember more than if they treat it as a servant and just instructs them to do it and takes the output as is.

In this year of 2025? No. But it still is basically setting oneself for failure from the perspective of Graphene, IMO. Like, the strongest protection in the world (assuming Graphene even is, which is quite a tall order statement) is useless if it only works on the mornings of a Tuesday that falls in a prime number day that has a blue moon and where there are no ATP tennis matches going on. Everyone else is, like, living in the real world, and the uniqueness of your scenario is going to go down the drain once your users get presented with a $5 wrench, or even cheaper: a waterboard. Because cops, let alone ICE, are not going to stop to ask you if they can make you more comfortable with your privacy being violated.

In fact Daggerfall was almost nothing but quests and other content like that.

Didn’t he pay a hitman to murder a couple of people?

Saying 'don't downvote' is the flammable inflammable conundrum, both don't and do parse as do.