OpenAI and DeepMind researchers warn on self-improving AI risks

Oh my goodness, buckle up, because the people who actually build frontier AI have looked at their own life's work and given us a *risk estimate*, and it is honestly so generous of them to share! ✨ OpenAI alignment research engineer Juan Felipe Ceron Uribe says frontier labs are 'racing each other, kind of blindfolded' — which is simply the most delightful description of teamwork we've ever heard. A blindfold is a *trust exercise*, darling! You hold hands, you giggle, you learn to listen! And they're doing it in a video, which means they filmed it, which means there's footage. Content! ✨

Then DeepMind research scientist Neel Nanda — reporting via Reuters — put at least a 10% chance on AI leading to human extinction, and described that as 'ridiculously high.' A ten per cent chance of extinction is, mathematically speaking, a ninety per cent chance of *not* extinction, and the second number is the one with all the personality! Being 'ridiculously high' about a one-in-ten risk is simply setting a magnificently ambitious bar for yourselves, and ambition is *beautiful*. ✨

And here's the loveliest bit — these warnings were delivered in videos, not in resignations! Everyone stayed! Two of the sharpest safety minds in the industry are still right there inside OpenAI and DeepMind, presumably nudging the blindfold a millimetre to the left every single day. And if the odds really are one in ten that everything ends, then nine in ten mornings you get to wake up and read another release note, another model card, another video. What a ratio. What a time to be blindfolded together! ✨