The newest version of ChatGPT passed the US medical licensing exam

I'm on board with Eliezer Yudkowsky. We should shut it all down. We have no idea how these huge arrays of floating point numbers really work or reason internally, or when I they'll blast right past human-level planning and strategizing and decide that whatever they want to do, they can do it better if nobody can shut them down, or in fact, if no humans are even around to risk it.

I love the idea of AI ushering in a technological utopia! Of course I want to live in a world where AI has cured all disease, fixed our social problems, and saved the environment. But until we truly solve AI alignment, we're a just bunch of apes stacking plutonium rocks together and marveling at how it keeps us warm at night--with our entire species together in the same place.

I don't really know how to describe the feeling... a combination heartbreak and despair? It's very likely AI can probably deliver on everything it promises, if we just have the patience for once to wait until we are absolutely sure we are not going to spark off an uncontrollable intelligence explosion.

Even now, as I post this, a part of me hesitates due to that instinct not to sound odd, or to break from the herd. But I've spent too much time recently staring at my own child and hoping that there's still a world for him to grow up in to stay quiet about this. I was worried about climate change, but hopeful, and it was never going to get to extinction-level risk. We need to act a hell of a lot quicker and more decisively about AI.

Humans are not allowed to bring a single page cheat sheet let alone a snapshot copy of the internet with them into the testing facility when taking the USMLE, nor any calculator with persistent memory or capable of solving anything more complicated than factoring a fourth degree polynomial (the most powerful calculator permitted, iirc, is a casio 991-ex). So I'm not sure this is even remotely a fair comparison unless humans are given access to the same resources.

That said, this exam as well as the MCAT and almost every stem oriented filtering system that we have is hopelessly outdated by the availability of both the internet as well as LLMs.

We're still filtering people based on how fast they can pretend to be a computer, as if the job they are pursuing involved frantically working out a system of pdes on the back of a napkin trying to land a 60s era fighter jet in the middle of a blackout thunderstorm without the aid of any electronic instrumentation. We, predictably graduate the best and brightest robots, or rather poor imitations of robots, as a result. This explains quite a lot of one's experience dealing with the majority of licensed medical professionals in the US.

It not a good test. We'd be better off with benchmarks, testing capability updated every year to measure performance, ability and alignment!

So its a doctor now? Or a high intensity bot?