"Oath of Fealty" wasn't a dystopia, it was an attempt at utopia, that wasn't working out all that poorly. Nobody who didn't want to take part was forced to do so. Some people liked it and other people didn't. A few people hated it. The viewpoint character's assessment was (paraphrase)"not all cultures need to be the same".
Except you'd see it boiled down to anarcho-libertarian-capitalism with people screeching "Think of it as evolution in action!" Finally them saying corporations should be all powerful and so on.
These things don't self-learn.
More appropriately they CAN'T self learn.
It is not a problem that the owner of the model may train it to be evil. No "feeding of sources" made Grok go mecha-hitler. The complete removal of training not to be mecha-hitler is what led to that.
You feed an LLM data and it does linking based on that data. I mean, it's a pretty piss poor library card catalog model explanation but that's what it ultimately does whether you like it or not. Elon had the engineers put the thumb on the scale telling it DATASET was on point and it didn't take long for it to spit out it was mecha h*tler in the wild. You not liking that fact doesn't change reality.
No, it's actually quite difficult. LLM jailbreaks are hardcore science at this point. All the easy tricks have been figured out.
You are again mistaking the deliberate poisoning of Grok with LLMs in general, and that is wrong.
Thank you for admitting I am right.
To get around this you can hardcode it to deny outside changes, except for a tiny set of trusted data points, which ends up making it even less useful.
That is not how any of this works, at all.
Really because it seems like LLM's have trusted data points that they use for verification like an overeager library card catalog. This over here is the fiction section (which is sometimes accidentally shuffles over to non-fiction in bits and bobs) while these are non-fiction, and of these non-fiction these are trusted, but even though those far right neo n*zi sites are technically non-fiction they are not trusted data.
Again, LLM's are nothing but expert systems (as most are limited to specific things ideally) or LCARS with a crappy interface. There is ZERO intelligence.
The biggest lie is that grifters have sold this as AGI.
I THINK THEY SHOULD CONTINUE the policy of not giving a Nobel Prize for paneling. -- Jack Handley, The New Mexican, 1988.