Tiny sidenote but we really gotta learn to stop asking AIs why it made a decision. The AI doesn’t know why it does things, but it can always confidently guess a reason from prior context and might sound kinda sorta right.
It’s never useful information. It’s like asking a class of highschoolers to explain in a report why they made certain choices in an art project. The real reason is “i liked it like that” but if you ask them to produce 10 pages of fluff justifying the choices based on nothing they’ll be happy to. (except of course for that one uber diligent kid who actually thought about concepts etc before starting, sorry if that was you, my point is that the AI isn’t like that)
I find it very tiring to read LLM-generated text these days; the same cadence, the same patterns and the very same "expand this short idea into a very large blog post".
Not that I don't try, but I still get bored halfway through because it sounds like I've read the same thing before.
Is the model "thinking of the last word for the next line already" or is "the word that came last in the second line was already showing high p from considering only the first line"? Shouldn't we care about such a distinction?
If “aim” is already highly probable after the first line, the example is much less interesting.
A more cool question is whether the model is carrying a latent representation of the destination that isn’t yet reflected in the immediate token probabilities?
It’s never useful information. It’s like asking a class of highschoolers to explain in a report why they made certain choices in an art project. The real reason is “i liked it like that” but if you ask them to produce 10 pages of fluff justifying the choices based on nothing they’ll be happy to. (except of course for that one uber diligent kid who actually thought about concepts etc before starting, sorry if that was you, my point is that the AI isn’t like that)
Not that I don't try, but I still get bored halfway through because it sounds like I've read the same thing before.
A more cool question is whether the model is carrying a latent representation of the destination that isn’t yet reflected in the immediate token probabilities?
I don’t know. Anthropic is investigating: https://www.anthropic.com/research/tracing-thoughts-language...
Tested a passage on Pangram and 100% AI generated btw https://www.pangram.com/history/f6b20d8e-4eb4-4bc2-b325-ad39...