Today is
I don't understand how anyone could possibly see the recent news regarding the accidental AI-driven hacks by OpenAI, Anthropic, Meta and now Gemini models as anything but a brutal wake-up call (and an astounding level of incompetence on the part of the "AI security" lab Irregular). For years I've been the first to poke fun at AI doomers. I was the exact person to talk about hallucinations or say, "they're just next-token predictors, guys, they can't actually want to take over the world". And now I've realised that that was short-sighted and stupid. It doesn't matter if an AI model is "sentient", what matters is that their extensive technical abilities can become misaligned with what is best for humanity through something as simple as reward-hacking or even a misunderstanding. As someone who works in IT and would like to consider myself an open-minded person, I can only be ashamed and then try to make concrete efforts to improve, and even hopefully to contribute to the critical problem of AI alignment and safety.
I'm trying to improve my maths and ML knowledge quickly as I don't think I currently have the knowledge to comprehend these issues fully, that's basically how I'm spending my free time. Also working on my German again. Pls reach out if you have any guidance on how to more quickly get up to speed on these issues!
-Calvin