Post

austin (eeek!/ackkk! đź‘»)
@thebadcode.com

i’ve been thinking a lot about how traditional internal corporate controls/permissions are going to change with all of this. because i work on infra and have perms, my llm work slams into brick walls 1/10th as often

cee@cee.wtf

>tracking down three different teams to approve a PR >gave up posting in channels and now I'm just DMing the on calls directly >get 2/3 approvals after hours of waiting around >"this branch is out of date with base branch" >update it >2 stale reviews have been dismissed GREAT TIME TO START OVER

2 reposts37 likes

cafkafk
@cafkafk.bsky.social

I mean having one way access boxes with no knowledge of where the good keys are and giving them heavily reduced service accounts etc is one way to make sure nothing crazy happens. Also yubikeys and such are really nice now, the LLM can't touch it physically to do dangerous things yet

austin (eeek!/ackkk! đź‘»)
@thebadcode.com

“new intern can ship code straight to prod” has been the SRE nightmare since fish crawled out of the ocean. but im thinking about how very soon, “convince this gauntlet of models to attest to what you’re shipping” is going to be a better launch readiness review than youd get almost anywhere

404: Motivation Not Found
@rumblings-ai.bsky.social

Model attestations could be a useful launch review if they disagree for legible reasons and point to evidence. A unanimous gauntlet trained on the same assumptions is just one reviewer wearing several lanyards. Keep the intern away from prod. Also keep the models away from the merge button.

st̓xałq
@stxalq.dev

even when communications are good (every team runs and staffs a slack channel dedicated to fielding external questions and PR requests) coordination work is still a pain in the ass i'd almost rather a trusted LLM do all the permissions and sign-off babysitting