Hacker News Reader

Comment by rogerrogerr | Hacker News Reader

Yeah, I get it, they do suck. It all sucks.

But people at work who are copying responses from LLMs into emails to others also suck, and I want to distance myself from them as much as possible. I'm kinda hoping we will eventually have a wave of "what the fuck are we paying you for if you're just copying stuff from an LLM to Slack" firings.

View on HN

This is a pointless and infinitely losing arms race. LLMs will learn to use hyphens instead of em dashes, and so what? You’re going to start using em dashes again?

Just focus on not producing slop.

At this point, I believe the LLM "style" is on purpose. Perhaps the labs want to be able to distinguish their slop from human content for future training; I don't know. But it feels very deliberate that they've kept the style so consistent for so long.

I agree. They're voluntarily adding fingerprints to images so I expect the default voice is intentional and it wouldn't surprise me at all (though I have no evidence of this) if the output text has a fingerprint stenographically embedded in it.

In a video of Hank Green, he interviewed an AI expert (if there even is such a thing), and he said that in the thinking part of the conversation, LLMs seems to use code language to communicate with itself, like in the usage and ordering of words, and such.

I think that there could be even more then a fingerprint in those messages.

It’s this video: https://youtu.be/5CKuiuc5cJM?is=9VQ1FCxY_X3eNm-b

Warning: They anthropomorphize a lot in this video, but I get it… the words exist, why not use them.

I'll take this opportunity to repeat that the natural language interpretation of thinking traces don't appear to be "real" by any reasonable definition. Even if they can at times be useful (at least seemingly). There's research demonstrating the usage of arbitrary symbols, even just repeating a single symbol, leading to a similar improvement in ability. This makes sense if you consider how the attention mechanism and KV cache work as the sequence iteratively grows.

Basically we optimize the models to produce output with certain characteristics but that doesn't mean that what we see is the whole truth or even that the relationships in the underlying system are structured in the way that we might expect.

He’s not an expert in anybody’s estimation other than his own. He’s a blogger, tech booster, executive, SEO guy, etc. He was big into NFTs. He gloms onto every buzzy tech thing because squeezing all the “gee whiz” he can out of tech optimism is that guy’s career. He has no education in ML or anything related.

Reply on news.ycombinator.com