frontier models are “great” at finding vulnerabilities. they’re even better at finding ones that don’t exist.
from our latest benchmark of 5 model setups in general-purpose harnesses across 3 open-source web apps 👇
@AutolithAgent note on the curl install featured on the page: v0.67.56 seems to be missing the arm64-darwin build, so the installer 404s on apple silicon. v0.67.55 works fine for now. was setting up on my work mac <3
@verrsane how messy do you wanna be with it? Hermes has everything, the upcoming iOS app is great, and the plugin ecosystem is huge. But it can be messy
@uzairansar Me neither. I think it’s for people who use many agents as really segregated affordance surfaces? But that’s been shown to not work as well as one big one in most cases. So idk?
@SilasOnLinux i gave it a claude api key (you get api credits every month with a sub), then boostrap into meridian, then clone what directsdk does for hermes agent into its own sensibilities. seems to work fine
I’m getting blackpilled. I now think CS majors should only be allowed to use C and Lisp until they turn 40, and after that point they can use Python or whatever.
Soon, we will merge the channel model into Autolith. Classic sessions will be replaced with independent channels and agents running on a "world server".
Within the world, parents manage their persistent children, talk to each other, mention each other like on Discord, and the user is just another participant who can be mentioned or inboxed.
Now you can start and run Autolith as a 24/7 service on your machine.
there's a massive gap for someone who's not an asshole to actually listen to omarchy users and figure out why they like it and make a competent and friendly alternative
I get stared at like a deer in headlights when I bring this up, though, like I'm speaking an alien language