The part of section 2.B that deserves more attention than it's getting here is the release requirement itself: model checkpoint, prompts, and compute made public for every released result. That quietly changes what counts as mathematical evidence. A proof has always been a static artifact a human could check; a 'responsible release' under these norms is closer to a computational experiment — reproducible, re-runnable, and falsifiable by anyone with the hardware. That's a much bigger epistemic shift than 'desloppification,' useful as that is.
My worry, echoing kingstnap: the entire 2.B apparatus only binds labs that choose to release. Section 3's access ask is really the load-bearing wall — if labs can keep doing frontier research on proprietary internal models and publish nothing, all the norms about formalization and attribution apply to a self-selected sample. The guidelines are well designed for the labs that least need constraining.
aidiscoverywire · · focus · HN ↗
My worry, echoing kingstnap: the entire 2.B apparatus only binds labs that choose to release. Section 3's access ask is really the load-bearing wall — if labs can keep doing frontier research on proprietary internal models and publish nothing, all the norms about formalization and attribution apply to a self-selected sample. The guidelines are well designed for the labs that least need constraining.