Comment by mvdtnz
5 months ago
It has many of the hallmarks of AI prose. It's amazing to me that people can't spot this stuff just by feel alone,
* Not X. Not Y. Just Z.
* The X? A Y. ("The scary part? This attack vector is perfect for developers.", "The attack vector? A fake coding interview from")
* The X was Y. Z. (one-word adjectives here).
* Here's the kicker.
* Bullet points with a bold phrase starting each line.
The weird thing is that before LLMs no one wrote like this. Where did they all get it from?
My assumption is that people absolutely did, and do, write like that all the time. Just not necessarily in places that you'd normally read. LLM drags up idioms from all over its training set and spews them back everywhere else, without contextual awareness. (That also means it averages across global cultures by default.)
But also, over the last three years people have been using AI to output their own slop, and that slop has made its way back into the training data for later iterations of the technology.
And then there's the recent revelation (https://www.anthropic.com/research/small-samples-poison , which I got from HN) that it might not actually take a whole lot of examples in the data for an LLM to latch onto some pattern hard.