People in the replies saying things like "just make your GPT's prompt more secure" as if a) putting that burden on the creator is reasonable* b) that will work *remember, this whole paradigm means you "don't have to know how to code", so cybersecurity is a non-starter

This is a well-written high-level intro to some of the nonsense swirling around the tech world in recent months and why "EA" and "e/acc" both deserve the most exaggerated of eye rolls. Worth a read (maybe on his blog instead of Twitter, due to formatting).

When you have zero chance of actually fixing your service, just add an arbitrary rule for users. This is the most hilarious, blatant admission yet that LLMs are not "intelligent" and that "alignment" is smoke and mirrors.

Maybe Tay's legacy is making MS actually care about training data; if so, totally worth it. Of *course* your model's going to be better behaved if you don't train it on reddit and 4chan. Once again, these are statistical models, not conscious beings. microsoft.com/en-us/research…

It’s Josh o’clock somewhere

OK, the detail is better than our old favorite Will Smith eating spaghetti, but...the nose. The cat's magical third front leg. The disembodied hand. Don't say it's there until it's there, folks.

Beautiful. Looks like everyone can start suing for all those $1 cars they got dealership chatbots to promise them. arstechnica.com/tech-policy/20…

This is what language models are good for and should always do. I'm glad OpenAI has finally realized that.

Please tell me this and the next tweet in the thread are real (seems plausible, given Sydney's history). "AI alignment" is a joke, and the fact that they keep doubling down on it rather than fixing their foundations should be all we need to see the perverse incentives at play.

This is your quarterly reminder that <latest AI model release> is only as amazing as the marketing says if all the answers it's supposedly giving are accurate, and that you can only verify said answers if you already know them or can get them from another trusted source.

"Do LLMs exhibit theory of mind?" "Hm, I don't know...let's find out using some well-established theory of mind tests." This is where I stopped reading, as "well-established" implies a sizable written corpus...which would show up in LLM training data. nature.com/articles/s4156…

Devs was an awful show (sorry, Nick Offerman; I still love you), but at least it was based on a true story:

We all had the same idea about that Hulk Hogan speech, didn't we? (the graph is Google search trends)

This doesn't mean that the model is "simulating people". You do not have to simulate a person to invent a survey response loosely based on background information that you've seen countless times before. A language model models language (or more broadly, text), not people.

Me: I've always been told I'm an old soul and wise beyond my years, so... My wife: They don't mean in the good ways.

This is a good breakdown/takedown. My main question is: did the authors have their idea raters rate the idea of seeing if LLMs could generate research ideas? How did it score? I think that would tell us something about the quality of the rest of their ratings.

horrifyingly accurate

This is, in fact, not what the test "proves" (it doesn't prove anything other than that most of the people he surveyed aren't art critics/historians/artists), and the original post's conclusion is nonsense. 2/10, recommend only reading the one thoughtful critique it quotes.

I'm loving the new X feature where I get logged out every time I get to Twitter from clicking a link to a tweet/xeet/whatever, even if it's a link in my browser history. Can't believe they just now got around to implementing it.

Adele's "hello", but it's "hello from a failed state".