It doesn’t. They run using stuff like Ollama or other LLM tools, all of the hobbyist ones are open source. All the model is is the inputs, node weights and connections, and outputs.
LLMs, or neural nets at large, are kind of a “black box” but there’s no actual code that gets executed from the model when you run them, it’s just processed by the host software based on the rules for how these work. The “black box” part is mostly because these are so complex we don’t actually know exactly what it is doing or how it output answers, only that it works to a degree. It’s a digital representation of analog brains.
People have also been doing a ton of hacking at it, retraining, and other modifications that would show anything like that if it could exist.
I run it locally on a 4080, how do they capture my keystrokes?
Are you running a quantized model or one of the distilled ones?
No idea, have you monitored the package / container for network activity?
Perhaps this refers to other clients not running the model locally.
It doesn’t. They run using stuff like Ollama or other LLM tools, all of the hobbyist ones are open source. All the model is is the inputs, node weights and connections, and outputs.
LLMs, or neural nets at large, are kind of a “black box” but there’s no actual code that gets executed from the model when you run them, it’s just processed by the host software based on the rules for how these work. The “black box” part is mostly because these are so complex we don’t actually know exactly what it is doing or how it output answers, only that it works to a degree. It’s a digital representation of analog brains.
People have also been doing a ton of hacking at it, retraining, and other modifications that would show anything like that if it could exist.
Yea, I am looking for a deep analysis of this as well