No they hallucinate because they lose context but are tasked to answer anyways.
Nothing to do with context. Context windows are an LLM specific thing, but all neural networks no matter the kind or size "hallucinate" because they are a piece of linear algebra where you put numbers in on one side and get numbers out the other.
If you can’t see the difference between a compiler and a large language model
If you can't see the difference between a given example and the underlying logic...
And if the entirety of your argument is a mystic “and all that stuff”
What, you want me to list the entirety of sciences downstream from physics that are involved in generating and predicting movement in mammals? Because that could, like, take a while...
That's... the joke. The original proposition was that AI code bans can't be enforced, a thought which I was leading ad absurdum.
I hate this attitude that if a rule or standard can’t be enforced perfectly, we shouldn’t strive for it.
"Murder bans may be impossible to enforce in the first place"
What do you mean? That's exactly how LLMs work. That's why they "hallucinate".
I’m just a movement generator.
Which is way more complex than just predicting text, because you have to predict physics and all that stuff. There is a reason most robots controlled by neural nets have failed hilariously so far.
then you’ll have to acknowledge that some form of reasoning is going on, no?
By that logic you could say a compiler is reasoning. But the reasoning displayed was happening when the compiler (or LLM training material respectively) was written.
Might well be behind here, but to my knowledge we don't have a single robot outpacing a single human in most tasks. They don't do one-shot learning from their mistakes, they don't learn new movements randomly. Because most basically just use static weights when operating, because you don't really want an industrial robot to get ideas. But even test systems that can learn usually just have an equivalent to the movement model of a brain, maybe a vision model, just like an LLM is just a language model. You would need a system that has all of those, plus the other brain areas, especially a prefrontal cortex like model for integration.
I'm sure people are working on it, but I haven't heard of anything successful yet. I imagine there might be a being like that in secret which is currently tortured in some billionaires tech dungeon. Poor thing.
Yeah but that's coincidental. It's the model weights, prompt, and the RNG aligning. They can mock reasoning, because they do it by what they always do, predict more text, but it's not like this has any effect on themselves. They aren't really understanding a mistake when you point it out and growing neurons and synapses, i.e. they won't have different model weights the next time you ask the same question. They can only really change when the powers that be release an update, which includes new training data, and hence model weights.