Eva Behrens

Or, y'know, we could have a global non-proliferation treaty, starting with a bilateral treaty between the US and the CCP.

Here are 5 policy recommendations for the upcoming AI Safety Summit in Seoul, from me and my colleagues at ICFG. In Bletchley, world leaders discussed major risks of frontier AI development. In Seoul, they should agree on concrete next steps to address them.

Here are 5 policy recommendations for the upcoming AI Safety Summit in Seoul, from me and my colleagues at ICFG. In Bletchley, world leaders discussed major risks of frontier AI development. In Seoul, they should agree on concrete next steps to address them.

A case study in parrot misalignment: When we first got her, sometimes she'd get scared of new toys or other unfamiliar objects and we would reassure her with "it's ok". So now whenever she's scared of something, she says "it's ok" over and over while rapidly backing away from it.

1) What CAPTCHA means Completely Automated Public Turing Test to Tell Computers and Humans Apart???

I think 🇫🇷🇩🇪🇮🇹 leaders are scoring an own goal without actually realizing it. “Let’s catch up with 🇺🇸 on AI by not regulating the most powerful AI” doesn't sound crazy, unless you realize that the LLM risk assessment and red-teaming requirements wouldn’t apply to any EU companies: they’d be allowed to do whatever they wanted as long as they share basic info. Here’s the Spanish compromise proposal they rejected, for a 10^24 FLOP limit:

One specific lesson from that history: It's better if liability rests with whoever can address the *root causes* of risks, and if it rests with larger organizations that have the resources to invest in safety engineering, as opposed to with small businesses or individuals.