Predominantly Erroneous (Exohedron nonsense blog)

Still remembering that time when I got asked to give a colloquium talk, except the person doing the asking was talking to an account under one of my online handles and not to an account with my actual meatspace name on it. That was slightly awkward to navigate, since that handle looked like a plausible actual name.

 

It really is too bad that the really hot peppers taste like garbage.

 

When the author is apologizing for only having twenty chapters of buffer for the fic.

 

Not so much a position on Sudowrite but a question that might imply a position: to anyone who has an opinion on Sudowrite, what is your opinion on the GPT project in general in terms of getting training data? All of these predictive text generators, be it Sudowrite or GPT-dungeon or ChatGPT, need to get training data from somewhere, and initially that training data was gotten by scraping popular webpages, with no permission from the authors/owners of those pages. Would it have been okay if the original GPT data set had included Ao3? Given that I don't actually remember which pages the original GPT dataset came from, this might not be a hypothetical question.

Notably, there is some difference in the marketing of Sudowrite as opposed to ChatGPT, in that Sudowrite is being presented as a product for money. And this has some issues, in that it is utilizing work done by authors for free and monetizing it. On the other hand, it is utilizing work done by authors for free at least partially because making money off of it is legally dubious and monetizing it.
Now, it is very hard to determine exactly what pieces of training data actually make it into a model, but nonetheless, an argument can be made.

Still, there's the question of if Sudowrite remained freeware forever, and thus reduced to the economic status of GPT-Dungeon, would there be a difference in ethical status?

 

Thinking about what I might put on an About Me page if I ever made one. Probably something along the lines of "If anything you might know about me would change your opinion about my words, I'm sorry that my words are insufficient."

 

... can I borrow that?

 

Sure

 

The whole concept of "naturalness" in theoretical particle physics really confuses me. The claim is that all fundamental dimensionless constants should be approximately 1, and the reasoning given is that large constants or small constants raise the question of why those particular numbers and not other numbers, which physicists refer to as the "fine-tuning" problem..
And I could understand this principle if it were actually the claim that all fundamental dimensionless constants should be 1 exactly. This would make sense as a claim, even though it would not be coherent because you can always add or subtract dimensionless constants and sometimes this is a valid thing to do. But to claim that making everything approximately 1 removes the question of why particular values are involved assumes that numbers that are approximately 1 are less subject to questions of fine-tuning, as if the slight differences in those values aren't significant.
Apparently this heuristic has led to some good results in the sense of if when there were two models and one had fundamental dimensionless constants that are approximately 1 and one had fundamental dimensionless constants that are not approximately 1, the first model led to more accurate predictions in some cases. But as a principle it's absurd.

But at this point I should not be surprised that the principles on which the study of particle physics operates are absurd. None of it stands up to the mildest epistemological scrutiny.

 

I want to make a joke about anti-shippers and the rail strike but that sounds like it would require work.

 

I could say "a ring whose modulus is a product of distinct primes", or instead I could say "a field with Chinese (Remainder Theorem) characteristics".

 

Talking to a colleague about Rust today. He comes from electrical engineering in C and C++, years of experience working in those languages, so his main issue is Rust being a little stricter about mutex locks than C.
In the meantime, my background is algebra, so my naïve understanding of computation is functional and I never expect variables to actually change in value once they're defined. On the other hand, my programming background is mostly Python, so dealing with memory management is boggling.

 

State of the lyrics: Trying to do interpretation and delivery via Vocaloid is very strange for a number of reasons. The first is that none of the parameters make any sense; they're all named in ways that are slightly helpful but not really precise enough to know whether I should be fiddling with them or not.
The second is that I'm not really sure how it should sound. What kind of intonation implies sadness? Joy? Anger? Despair? I am bad at detecting these things when listening to people, not just because I can't translate the signals into emotions, but because I'm not sure I'm actually receiving the signals. But also because I haven't felt these emotions in like two decades either internally or empathetically, so it's all just question marks and 404s.
Also for some reason the Vocaloid engine seems to default to having a lot of vibrato and I'm not sure either why it does this or what effect is intended. Or maybe it's just a result of the pitch-matching not having enough Fourier coefficients to suppress it.

 

I love how when Adam Neely needs information about singing for his music analysis videos, he calls his mom.

 

Like going up to the casting director and saying "Look, I don't care what kind of person you pick to portray the character, black, white, man, woman, it's not in the book so the audience shouldn't care. But! And I didn't put this in the book either but just listen: whoever you pick has to be left-handed."

 

Cargo pants, in the British sense

 

Like, maybe it would have been funnier if the Barbie trailer remaking 2001: A Space Odyssey was completely accidental, an artifact of clever fans taking stills out of context, but being a deliberate parody is also kind of funny.

 

On the one hand, I don't have strong opinions on the AI-art debate other than to note that there should be similar debate over GPT in general (not just sudowrites). On the other hand, the Artstation protest poisoning the training data is really, really funny.

[EDIT]: Might be fake? Are the various Stable Diffusion clones still running training epochs?
Is it even funnier if the AI artists are using traditional photoshop techniques to make fake-poisoned AI art to mock the protests? Maybe.

 

Girls wearing their boyfriends' plate mail, amirite

 

But yeah, it was really funny when the prologue to Witch From Mercury came out and people were surprised to find out that Gundam is sad. Like, yeah, there are giant robots, but in general, Gundam is a bunch of stories about people suffering and dying because of other people's greed and fear and hatred.

 

It's always fun when someone tells you about their childhood and you want to say "oh, that's why you're like this" and then after a moment it occurs to you that no, it actually doesn't explain why they're like this, and moreover raises the question of why they aren't additionally weird in a way that would be explainable by the childhood anecdote.