logoalt Hacker News

CSSer • today at 6:10 PM • 1 reply • view on HN

Safety alignment feels like an absolute farce and the completely wrong approach too. These things are guns, or attack dogs. We've never succeeded at reigning in software with this approach in the past, besides maybe DRM, which is itself laughable under many circumstances. This has the unprecedented ability to entertain the base desires of most juvenile threat actors, which is to get a big payout with minimal effort (let alone the pre-existing sophisticated attackers, who are now basically gods). Now we have a turn-key tool that can leverage and chain the most minute vulnerabilities up the ladder to full compromise or discover zero days with unyielding dedication. It feels like it almost makes CVE level classification moot.


Replies

toomuchtodo • today at 6:15 PM

As someone who red teams with open weight models (GLM 5.3 currently) and a domain focused harness, I would agree with this. No regulation will prevent a sufficiently capable open model, nor stop the development of open models. Alignment efforts are “please don’t stop us from selling stock to the unsophisticated” imho. They aren’t stopping anyone competent and/or capable from model development and training.

➕ show 1 reply