LLMs Believe
ArsTechnica has become my latest favorite punching bag recently. They published this slop with the title "LLMs believe false statements even after explicit warnings that they’re false".First of all, LLMs don't "believe" anything. Stop anthropomorphizing them. It doesn't help understand what is going on whatsoever.
Second, it's well fucking understood that fine-tuning doesn't alter the underlying training data, it only shifts it a bit. So if an LLM has modeled that something matches, it isn't going to suddenly do the opposite of that after fine tuning.
This isn't new news. It's been known since the beginning of fine tuning. This is bad reporting because it's like "did you know water is wet". Yes, no shit genius.
Plus, TFA doesn't even explain what I just said.
Also it goes further into repeating additional well repeated bullshit: "LLMs frequently hallucinate". Uh. Hullucination isn't made up nonsense. It's just probabalistic matching with training data. We may think of it as if it is hullucination ourselves as humans, but that just distorts human understand of what is going on.
There is no fucking "implications on... how data should be structured" as Kyle is fucking saying. That's idiotic. LLMs "hallucinate" because that's how LLMs work.
Further Kyle doesn't seem to understand what Lying is. Lying is saying something with the intent to deceive. LLMs have no fucking intent only their model, so they can't lie. What they say may be false, but using the word 'lie' or 'lying' with LLMs makes no damned sense.
TFA goes on to talk about LLM reasoning. Do I even need to dig into that? LLMs can't reason. Period. They can iteratively prompt again ( the so called "chain of thought" ), which is effectively what reasoning is. It has almost nothing to do with what we humans consider to be "reasoning."
Essentially, you aren't going to learn anything from that article and you should take care not to think of LLMs as human or working in anything approaching a human way. The adoption of terminology that is used with human thinking to talk about LLMs is horrible and just confuses everything pointlessly.
Go check out my article on AI Psychosis as it deeply relates to this whole behavior of believing LLMs are doing things LLMs are not.
AI as usual isn't fond of what I've said here. Here is the AI sanded down version:
The biggest problem in AI reporting is not that journalists get facts wrong. It's that they consistently describe LLMs using language intended for human minds. They talk about belief, reasoning, lying, understanding, memory, knowledge, confidence, and intention. Then they write articles about surprising behaviors that arise from those concepts. The surprise exists only because the terminology was wrong from the start.
LLMs are pattern-matching systems trained on enormous datasets. They do not possess beliefs, intentions, or understanding in the human sense. Once you stop pretending they do, most of the "shocking discoveries" reported in AI journalism become entirely predictable.