You don't have to take my word for it. LLMs are dumb in a bunch of ways, but I think this is a powerful and convincing consensus on this question.

Seems like you’re confusing a punchy post title about a standard norm in Bayesian epistemology (i.e., don’t give empirical claims credence 0 or 1, or you can’t update) with a claim about the formal definition of probability, where 0 and 1 are of course valid probabilities.

Of course, by definition probabilities are real numbers on [0,1], which includes the endpoints.

I understand that you believe I’m a huckster. I was hoping you might elaborate on why you think that.

Huckster? Say more. Conversations I was having in the comments seemed pretty reasonable and polite, with me just sharing context/info/perspective, and folks following up on that.

... and banned almost immediately. ¯\_(ツ)_/¯

IDK man, feels like you are fixated on one particular thing he said (and your interpretation of what he meant by it), as part of a longer conversation on the pod about the analogy. I’m not trying to pick a fight here, just wanted to clarify that he’s not making the mistake you think he’s making.

I listened to it. Also, I run org (@intelligence.org) that he founded, so I’m quite familiar with the argument he’s making. This interview didn’t result in the best exposition of the analogy in question, but I can assure you he isn’t making the mistake you think he is.

How is he anthropomorphizing natural selection? One can think of evolution as an optimization process, and the analogy is between that optimization process and the one used to train AI systems.

It is by no means a "nearly unanimous view" that LLMs are a dead end among AI experts. Also the argument for future very powerful AI systems being an extinction threat does not depend on those systems being LLM based.

(Totally above board. Sharing full-length episodes is one of the benefits of being a subscriber.)