The revolt of the reader
96 points by prattmic
96 points by prattmic
Marketing plugs aside it's nice to see a leader taking human communication seriously. I've had to have this discussion constantly.
This seems like more like an ad for their service. There's no public weights for this. It's a very interesting read though.
A confession: with particularly egregious pieces, I have fantasized about sentencing the author to read them aloud, certain that they themselves will be unable to endure the slop that they are foisting upon the rest of us.)
For years now, when I’ve seen a particularly rage bait take by someone who should know better I’ve imagined sentencing them to deliver a 15 minute talk defending that take to people they intellectually respect.
It works on yourself as well. Sometimes I delete something after thinking “I wouldn’t feel good trying to defend this to X”.
The war on spam started to turn when we could identify it at scale; could something similar happen to LLM-authored writing?
I imagine not. There have already been cases of people being falsely flagged by algorithms when they could demonstrate the content was their own. There have been cases right here on lobst.ers where humans who can "definitely tell" mis diagnosed a piece as LLM generated.
It's one thing to have your well intended email to a peer end up in their junk box, I think having a long form communication be classified as LLM (either by user gut feel or algorithm) feels more off putting and would likely have a negative effect on production.
As "the reader", I prefer having a successful heuristic that might misfire and drawing my own conclusions rather than not having it at all.
EU regulation requires LLM output to be marked, and Anthropic is enabling their fingerprinting for Claude worldwide.
Of course that’s not gonna stop people who run local models but that already raises the required investment quite a bit.
There have been cases right here on lobst.ers where humans who can "definitely tell" mis diagnosed a piece as LLM generated.
One thing I'm afraid of is that reading too much robot writing could make me sound like a robot. It's one of the reasons why I avoid LLM generated content even in cases it might sound interesting, and part of why I'm increasingly afraid of asking an LLM for anything.
To add to this, LLMs are designed to mimic the human language distribution they were trained on. The only reliable perturbations to that distribution are going to arise from RLHF and other fine-tuning processes after the baseline training, and I don't think those are going to introduce biases that reliably indicate LLM generation. Spam, on the other hand, can be reliably identified based on its content - that indicates there is something in the content that is characteristically spammy. LLMs don't work that way.
How is this related to coding at all?