https://youtu.be/Xn_YpUtXWT4?t=2252 Really interesting comment from Martin Odersky about why we need proper guardrails, not just some instruction in the prompt if we want to be able to automate stuff with agents. If an agent can do anything, assume that it will do anyhing at some point, including wiping out your mailbox if it can https://www.404media.co/meta-director-of-ai-safety-allows-ai-agent-to-accidentally-delete-her-inbox/