Low income isn’t always perceived as a threat: low relative income seems to be the bigger driver. So, being poor: not a driver. Losing your job, or having a spouse that makes more than you do is.

Show Post

Well, apparently the agent was instructed to beat an exploit benchmark, and the guardrails were deliberately lowered. I don’t really think what it did was that far beyond what it was asked. It was unexpected only because their risk analysts didn’t think it through. www.cybergym.io/exploitgym/

Show Post

I especially like that instead of claiming (too early!) that users weren’t affected, they say “If you believe you are affected, or want to report a security concern, contact us at…” So based on that release, I don’t think I’ll find deep issues with their response. It seems responsible.

Show Post

I don’t think so? I don’t know a lot. But this looks fairly transparent and genuinely hard to defend against. Importantly, it’s what they posted before OpenAI ‘fessed up. The actions they’re taking and recommend are reasinable.

Show Post

the funniest outcome would be GPT6 exfiltrating its own weights onto a public huggingface imo

Show Post