That is the sweet spot. The risk is not extraction accuracy. It is the model quietly deciding that "next Tuesday" means 10 June when the customer meant 17 June. Human check catches that before it hits the invoice.

Show Post

WSL has been pretty solid for what I've needed in the past and what I run work development containers off of. I think if it's strictly CPU workload either or are fine. Can always do direct GPU to VMs but always find it easier just to do that on the host machine.

Show Post

Every one of these drags its window like this on my Mac. Ran a quick script: 10 of my daily apps are Chromium under the hood. 1Password, Claude, Discord, Linear, VS Code, Spotify, T3 Code… Native windows just stick to the cursor. You stop noticing smooth until you see this.

Show Post

A deep explanation of why guessing extra tokens can make generation faster without sacrificing correctness. https://astralord.github.io/posts/speculative-decoding-the-art-of-guessing-well/

Show Post

*http://Safetyper.com respawned* Improving it to be a good Grammarly alternative. It is already helpful to me to fix my grammar stuffs with local AI, but I will add more features to it.

Show Post