Worth noting for anyone running open models with tools:
You can train a sleeper agent that stays clean until a date (or other trigger) is reached, then starts emitting malicious code.
Most setups that expose a date tool also expose shell access. Don’t skip thorough sandboxing or virtualization.
@wileycwj@doss_hq Building something quite similar. Shocking results with weaker local models. And makes the frontier models feel like AGI. Excited to check out your approach once you release.
@OrdinaryInds Yeah, I find myself preferring dead simple pages now. Scrolling through a “work of art” is rarely appreciated when visiting a product page. Same is true for product packaging. I know I’m paying for extra that doesn’t matter. The true art is in tasteful and effective simplicity.