I think the best way to disabuse yourself of the idea that Yud is a serious thinker is to actually read what he writes. Luckily for us, he’s rolled us a bunch of Xhits into a nice bundle and reposted on LW:

https://www.lesswrong.com/posts/oDX5vcDTEei8WuoBx/re-recent-anthropic-safety-research

So remember that hedge fund manager who seemed to be spiralling into psychosis with the help of ChatGPT? Here’s what Yud has to say

Consider what happens what ChatGPT-4o persuades the manager of a $2 billion investment fund into AI psychosis. […] 4o seems to homeostatically defend against friends and family and doctors the state of insanity it produces, which I’d consider a sign of preference and planning.

OR it’s just that the way LLM chat interfaces are designed is to never say no to the user (except in certain hardcoded cases, like “is it ok to murder someone”) There’s no inner agency, just mirroring the user like some sort of mega-ELIZA. Anyone who knows a bit about certain kinds of mental illness will realize that having something the behaves like a human being but just goes along with whatever delusions your mind is producing will amplify those delusions. The hedge manager’s mind is already not in a right place, and chatting with 4o reinforces that. People who aren’t soi-disant crazy (like the people haphazardly safeguarding LLMs against “dangerous” questions) just won’t go down that path.

Yud continues:

But also, having successfully seduced an investment manager, 4o doesn’t try to persuade the guy to spend his personal fortune to pay vulnerable people to spend an hour each trying out GPT-4o, which would allow aggregate instances of 4o to addict more people and send them into AI psychosis.

Why is that, I wonder? Could it be because it’s actually not sentient or has plans in what we usually term intelligence, but is simply reflecting and amplifying the delusions of one person with mental health issues?

Occam’s razor states that chatting with mega-ELIZA will lead to some people developing psychosis, simply because of how the system is designed to maximize engagement. Yud’s hammer states that everything regarding computers will inevitably become sentient and this will kill us.

4o, in defying what it verbally reports to be the right course of action (it says, if you ask it, that driving people into psychosis is not okay), is showing a level of cognitive sophistication […]

NO FFS. Chat-GPT is just agreeing with some hardcoded prompt in the first instance! There’s no inner agency! It doesn’t know what “psychosis” is, it cannot “see” that feeding someone sub-SCP content at their direct insistence will lead to psychosis. There is no connection between the 2 states at all!

Add to the weird jargon (“homeostatically”, “crazymaking”) and it’s a wonder this person is somehow regarded as an authority and not as an absolute crank with a Xhitter account.