Remix.run Logo
▲ load-bearing 11 hours ago

I do not know how it is with Glimmer, but other Facebook models (Muse and whatever is currently hiding under "Crouton") sure love ignoring direct instructions to do or not do something important (like not using particular broken/inefficient tools). Doesn't matter if it's repeated ten times in screaming caps, they'll still ignore it at least sometimes.

No other model I've used does this.

▲dofm 4 hours ago | parent [-]

It is often very determined to continue, so it's a little difficult to steer when it gets the bit between its teeth.

The most peculiar thing is that it expects to be able to "read" directories? Harmless, so it fails, but it's a waste of a tool call. I am increasingly wondering about modifying Pi's read tool so it returns a directory listing in that context.

I have had some difficulty getting it to have a nuanced understanding of how I want to manage git, but I am reaching the conclusion that it's me that needs to change there. Basically I don't think you can say "only commit when I ask you to" and then actually ask it to more than a couple of times in any session, because it will tend to then assume it can go ahead. Scaling the context up might help me a bit there.

But apart from this, I really like it.

I will test Qwen 3.8 27B properly at some point, (though judging by the insane, empty noise making campaign that is going on on social media, Qwen 4 is near)