That said I also dislike many of those other changes and would prefer a hypothetical version of Pi which didn't have them, so this is in some sense just me looking up at the sound of a v1.0 release and realizing "oh hey, I don't really like the direction this has been trending for a while"

I mean, Jev is also probably cheaper because it's a rather small model (or at least, I suspect it is based on the overall level of intelligence it demonstrates) so that helps make it cheap too.

OpenAI and Anthropic don't want to give out logprobs these days but could trivially add a dedicated classification API to their existing models if there was enough demand.

X to doubt. I think the more parsimonious explanation is that a lot people are lazy and/or uncaring, and given the option to avoid thinking while still producing a vaguely equivalent [1] output they're taking it.

[1] For a loose definition of "equivalent" where they clearly either don't notice the Claudeslop or don't care enough to avoid it, of course.

This still leaves an absurd amount of wiggle room for arguments like "oh no, our evals show that this quantization has no detectable effect on performance (in the eval distribution) therefore running the quant doesn't degrade quality"

I don't think I have ever once run "pip install transformers" and had it work without three rounds of fiddling

Also Pangram is basically the only AI detector that's actually put in the work to build an accurate classifier, so if you try to discuss AI writing detection without being specific that you mean Pangram you'll get half a dozen commenters screaming about how awful some of the snake-oil salesmen like GPTZero are.

Well the author in this case is Claude, and AIs write like that because the assistant persona really thinks that's what good writing sounds like. They're wrong, but they're just doing what they were taught. There's a reason LLM raters consistently score LLM writing highly.

[1] I have no idea why this is the case, AFAICT it's just anti-bot autoimmune disorder deciding it doesn't like some signal or other.

More generally, their current approach to constitutional AI pretty much only makes sense if they believe that they can first teach the model what the Claude character is like and also teach the model that the persona responding is Claude, so I figure that has to be part of the pipeline even if they're not very good at it.

The issue covers at least two reasons this doesn't work:

1. It literally doesn't work, Claude rapidly drifts back to this style even when instructed not to.

2. Writing style constraints push the model out of its training distribution and it's very unclear how much of an impact this has on work quality.