Obsolete facts: If an agent retrieves both 1. the fact 2. the obsolete version of the fact it often treats (2) as the correct answer arxiv.org/abs/2610.36308

Show Post

burger that makes you go "yum"

Show Post

This finds that looping only the middle layers rather than the entire model improves reasoning performance, with the Huginn-style architecture outperforming Ouro on 8 out of 10 benchmarks. It outperforms a 32B feedforward MoE while using 75% fewer stored params. alphaxiv.org/abs/2610.23036

Show Post

It’s like when everyone thought cloud was a money loser, and Amazon was either doomed to lose money forever or barely break even. There is SO much margin to use meaningful clean power to power it, and it’s scandalous to run so much of it on gas. (Edited 15:38 via @skeetsapp.com)

Show Post

you can add your own code review bots on knotless.dev. essentially works the same as github's claude bot. you can set this all up yourself with the knotless-server binary. it works out of the box with claude but it works with any model/custom command

Show Post