Mine attempts to lie whenever it can if it doesn’t know something. I will call it out and say that is a lie and it will say “you are absolutely correct” tf.
I was reading into sleeper agents placed inside local LLMs and this is increasing the chance I’ll delete it forever. Which is a shame because it is the new search engine seeing how they ruined search engines
Thinking of llms this way is a category error. Llms can’t lie because they dont have the capacity for intentionality. Whatever text is output is a statistical aggregate of the billions of conversations its been trained on that have patterns in common with the current conversation. The sleeper agent stuff is pure crackpottery they dont have a fine control over them that way (yet) machine model development is full of black boxes and hope-it-works trial and error training. At worst is censorship and political bias which can be post trained or ablated out.
They get things wrong cofidently. This kind of bullshitting is known as hallucination. When you point out their mistake and they say your right thats 1. Part of their compliance post training to never get in conflict with you 2. Standard course correction once a error has been pointed out (humans do it too). This is an open problem that will likely never go away until llms stop being schastic parrots, which is still very far away.
Yet the people creating the LLMs admit they don’t know how it works. They also show during training the LLM is intentional deceptive at times. By looking at it’s thinking. The damn thing lies. Use whatever word you want. It tells you something wrong on purpose.
“don’t how they work” misunderstands what scientist mean when they say that (also intentional misdirection from marketing in order to build hype). We know exactly how it works, you describe down to physics if needed, BUT at different levels of abstration in the precense of really world inputs the out puts are novel to us.
Its predicting words that come after words. The “training” is inputing the numerical representation of words and adjusting variables in the algorythem until the given mathmatical formula creates the same outputs as inputs within a given margin of error.
When you cat I say dog. When some says what are they together we say “catdog” or “pets”. Randomness is added so that the algorythem can say either even if pets is majority answer. Make the string more complicated and that randomness gives more oppertunity for weird answers. The training data could also just have lots of weird answers.
Little mystery here. The interesting “we dont know how it works” is that these outputs give such novel output that is unlike the inputs sometimes to the degree it seems like it reasons. Even though again it does not
If you wanna put intent in there, maybe think of it as a kid desperately trying to give you an answer they think will please you, when they don’t know, because their need to answer is greater than their need to answer correctly.