Comment by nomel
3 hours ago
I'm not sure I understand.
I'm not moving a goal post. You're just thinking I'm making a point that I'm not. As I've said several times, it's just boring statistics. Those statistics are optimized to mimic human output. They are, quite literally, trained to write and BE as much like a human as possible, because only humans wrote the text, and they're optimized to predict the next word a human would write. Alignment is partly about removing the models perception of human self. See reports of people who had access to them, pre alignment. This is statistically sound.
It's statistics optimized to predict the next word a human would write, to mimic a human writing as closely as possible, because that is the loss function. Don't assume I think there's more to it.
This does not mean they contain systems that let them get tired. But, this does mean there are latent spaces that progress to generating text that contain text driven by human biology, because it's in the training data. I've also had Claude refer to itself as "she". Does that mean it's a woman? No, it means there was a little bit extra "she" mentions in the training data (btw, this 100% repeatable behavior left with 3.7. They probably cleaned the data a bit better, or hammered it out in alignment).
What percentage of text (these models were trained on all of it) is written from a "I am not a human" type perspective vs from a "I am human" perspective? That's roughly the kind of bias you should see in a base model.
edit: rearranged and reduced redundancy.
Ok, I indeed misunderstood your point.
I'm not sold on the idea that as the chat session goes longer, the probability of an LLM saying "I'm tired" is increased; I'm not convinced this is modeled in LLMs at all. As for what you call "laziness" manifesting in a longer session, I think that's more likely due to context rot than to any kind of statistical modeling of human laziness.
But yes, now I see your point was different to what I thought you were saying. Apologies!
Like I said, it would be neat if someone benchmarked it. It's definitely an anecdote.
Try it though. If it's context rot, then I don't think the weekend reset I mentioned should work? For me, it very reliably does. Or, maybe the weekend reset is just putting the current context into a more "productive" latent space. But, if that's possible, then that would suggest it was previously in a less productive space?
Maybe a test would be ask the LLM what time it thinks it is, or just if it's tired once, within sessions of different length (not within same, since that could pollute the context) to see if there's any relation between length and statistics of a late/tired type response?
Again, I'm sure all this will go away. They're getting good at beating these unhelpful statistics out of the base models.