Post

Conversation

David Watson 🥑
Post your reply

Same with things like filler words. For example, if you say “kinda”, J-space may fill with speculation that you are unconfident or uncertain.
Interesting, shouldn't the question end with <eos> ? Otherwise the next highest prob is an insult, which, makes sense 😂
Since J Space is here, is it like a small-scale model of the “thoughts” of these models? Could psychological manipulation techniques be applied to this part of the model, or could we start conducting MK Ultra-style experiments with LLMs? Because if they’re so good at replicating
Sometimes I literally use short hand or don't correct mistakes cause I know the model is smart enough to understand anyway
Models def think you're dumb if you mispell or use bad grammar unless it's clearly language barrier driven. If you sound native but make common mistake, the model downgrades your IQ and it behaves differently and knows it can get away with more stuff to be lazy because you're too
I would expect, J-space or not, the model treats different kinds of users differently - spelling could be a feature in that classification
I know you're probably joking. But for the sake of the naive reader: this is the model recognizing and correcting the mistakes in spelling and grammar. It doesn't really affect the outcome (i.e. the model does not form a "negative opinion" of you.
Some day Anthropic is flip and switch and we'll see AI is really thinking about us. It'll be that episode of Gilligan's Island—we'll all want to go back to ignorance.
J-space probably influences the quality of the code reviews you get based on the code clarity and conventions bothers LLMs even if they are irrelevant to you. We should design code for AI reviewers.