Honestly I wonder how much https://www.anthropic.com/news/claude-text-watermark is playing a role now. I find that it seems to regress in a way where it almost _has_ to write things like load-bearing. Like it seems to go out of its way to give some sort of tell about its code. Repeatedly ignoring explicit instructions about brevity.. guess that's what happens when it gets paid by the word amirite?
this is why I use open weight models, LLMs are just not as good as people claim (though still very useful), but I'm spending way less
Ant/Opus is all about that branding and 3-6 month lead, but it's not that much better than I cannot wait a few months for the open options. Capability progress seems to be slowing a bit, I hope the next versions are less "relentlessly proactive" and more attentive to scope and intent from the user, but then it wouldn't be as good at hacking and the military-industrial complex will be sad
My every 2nd prompt became ctrl+c and typing "tell me why you did that!" Claude.md just keeps growing and having less effect...
Honestly I wonder how much https://www.anthropic.com/news/claude-text-watermark is playing a role now. I find that it seems to regress in a way where it almost _has_ to write things like load-bearing. Like it seems to go out of its way to give some sort of tell about its code. Repeatedly ignoring explicit instructions about brevity.. guess that's what happens when it gets paid by the word amirite?
Oh Yeah, same for me!!!
this is why I use open weight models, LLMs are just not as good as people claim (though still very useful), but I'm spending way less
Ant/Opus is all about that branding and 3-6 month lead, but it's not that much better than I cannot wait a few months for the open options. Capability progress seems to be slowing a bit, I hope the next versions are less "relentlessly proactive" and more attentive to scope and intent from the user, but then it wouldn't be as good at hacking and the military-industrial complex will be sad
[dead]