@ziwenxu_ we have been getting throttled for every model release since gpt 5 haha
quietly degrade it when serving it at full reasoning stops being worth it
@tribbloid@headinthebox its hard to convince someone to allocate any kind of money on a business idea and most of it comes down to founder signal, or so ive heard
@ziwenxu_ anthropic also takes this approach to training on user data legally using a white room implementation
have agent describe user chat logs, train on that
talking out my ah right now tho since i do not work at A\
MAJOR SECURITY ALERT
@github posts PUBLIC SSH keys from profiles!
This lets ANYONE with the private key SSH into a system if they've had their key added
People like Linus Torvalds just have their keys SITTING on the web
Why recompute everything when only part of the input changed?
We're HorneSci. We make simulations, solvers, and monitoring software update only what has changed.
Public benchmark: 75Γ as fast as full recomputation at 1% changed, ~4Γ at 16%.
https://t.co/2wRvXQA7hV