Mind-blowing visualization accompanying Moschella et al. Table 1 (arxiv.org/abs/2209.15430): the word2vec and FastText latent spaces are secretly the same?! Different code on different data each discover the English language as a mathematical object
They do not and will not believe that there is a shared latent geometry between modalities on which different neural nets trained on different corpus converge.
openreview.net/forum?id=SrC-n…
Literally bad luck? (It's trained to generate a probability distribution over the next token, and used to sample from that distribution, if the pseudo-random number generator bounces the wrong way, it can take a low-probability token.)
a weird psychological artifact of getting into @SeriesTangled is that visuals from the allegedly "original" film start to look like an uncanny 3D knock-off of the 2D-animation series (which is clearly canonical)
the long tail of @rustlang compiler error messages is getting better all the time! Tiny example: the single-use lifetimes lint now shows you the use site and what the fixed code looks like. Try it out with `#![warn(single_use_lifetimes)]`!
Not that it's the worst thing about this, but the "desperately trying to correct authority figure Yudkowsky" personality-cult pattern is so pathological. He's just a guy! I'm glad I'm (almost) recovered from this.