Predominantly Erroneous (Exohedron nonsense blog)

I actually kind of like the SU(5)/Spin(10) GUT proposals for getting the quarks and leptons to fit together. They're probably wrong, since they make predictions about proton decay that don't seem to be correct, but they're elegant. Which, given how awful the Standard Model is, might be a sign that they're just flat wrong; they're not fundamentally different enough from the usual SU(3)xSU(2)xU(1)/Z₆ setup that they can afford to not be ugly.

 

It never would have occurred to me that you can eat the fruits of the cacao tree, not just the seeds. I have no idea what they taste like. My first thought is like bitter chocolate, but that doesn't make any sense; the fruit pulp is completely different material from the seeds, so it has no reason to taste like the seeds do.

 

Hades be like "I know a Spot" and then shows you his dog.

 

Suddenly remembered that time back in college when my thesis advisor was like "You want to TA the dynamical systems class I'm teaching next semester?" and I was like "um, not really" and he was like "too bad. I don't want to teach the class either, but here we are" and so I got to spend a semester pretending that I knew a subject when really I was constantly like a few pages ahead of the class in the textbook.
Also I don't really remember anything about the subject now.

 

Before Trump, four US presidents had never held an elected position before becoming president. Taylor didn't really do anything of note. Grant's administration had a lot of corruption scandals but was himself not a bad fellow. Eisenhower is generally remembered fondly.
The only one of the four who hadn't been a general before was Hoover, who generally gets a bad rap for good reasons.
Of course, there are plenty of presidents who had military service in addition to political experience before becoming president.
It's unclear what lessons the voting populace will learn from the Trump presidency, although I'm pretty certain that at least in the short term it will be dreadfully off the mark, but at least part of it will be about whether to venture outside of career politicians when looking for the next presidential candidate. The Machine of course wants to maintain its establishment, but as demonstrated it doesn't always get its will.

 

I replaced a lightswitch without bothering to flip the circuit breaker first and I only shocked myself once.
I"m turning into my dad.

 

I wish I weren't so literally thin-skinned. My skin is so thin and soft and easy to break.

 

I feel the kind of vaguely unhealthy that makes you Google various diagnoses and hoping that it's not something serious, but I remind myself that I've felt this way for weeks in various levels and have not developed the further symptoms that would indicate that it's something more serious, which means it's probably just some combination of dehydration and allergies or like a cold or something, but which also makes me fear that if it were something more serious I wouldn't notice until the really bad stuff abruptly pops in.

 

So Adam Neely recently posted a video about turning a Deep Fake of Jay-Z reciting the Navy SEAL copypasta into jazz. There's a lot in that description, but being me, I'm going to focus on the basic question of the issue of deep fakes.
In general, a deep fake is a forgery, an imitation of a type, style or likeness in some form of digital medium. One variant is style transfer, wherein a work by one author or artist is remade using the style of another author or artist. Another version is likeness transfer, wherein one person is made to look like another. A third type is ex nihilo generation, creating for instance realistic pictures of people who don't actually exist.
The term "deep fake" came from the deep-learning neural network technology that deep fakes were originally generated on, but what goes on under the hood is probably going to change significantly before the phrase itself catches up. So I'm going to ignore both the underlying technology and the boundaries on the coverage of the terminology and just consider all style/likeness forgeries.

There are people worried about deep fake technologies, in particular with regards to deception. If a deep fake is a good enough forgery, it can be used as evidence of events that didn't actually happen, or that people hold opinions that they actually don't. Of course, forgery has always presented this problem; that is, at the bottom, what forgery is all about.
Consider the supposed recording of Jay-Z reciting the Navy SEAL copypasta. Did Jay-Z actually record a recitation of the copypasta? According to him, no. It certainly sounds like his voice, although the technology hasn't quite gotten to the point where it can mimic his speech patterns completely. But there are really two questions here. How can we tell that the recording is fake, and how can we tell that Jay-Z didn't produce it?

The first question is often approached, and simultaneously defeated, using a machine-learning technology called a GAN, a generative adversarial network. Basically, you have two learning machines, one trying to make fakes, and one trying to detect them; they are constantly competing, learning from each other as they go. The deep-fake technology is usually just the generator machine, trained via fighting the detector. But that detector has still been trained, although if the generator is being released as a product then it can probably defeat the detector most of the time. Still, perhaps the detector just needs more training.
On the other hand, we do expect this problem to become harder and harder as the technology improves.

The second question then revolves not around the deep fake itself, but the identity of the author. Suppose that Jay-Z wants to prove that he did produce a recording of the Navy SEAL copypasta. Or perhaps more practically, you want to prove to a friend that you have an actual recording of the Navy SEAL copypasta by Jay-Z. What you would need is Jay-Z's signature attached to the audio. Here by "signature" I mean what is called a "digital signature", although the "digital" part is a misnomer. The important part of a digital signature, for our use case, is that it can only be produced by Jay-Z but can be verified by anyone and in some sense encodes the data that it's attached to, so that someone can't take a signature from something else that Jay-Z produced and stick it onto a deep-fake, nor can they take the signed data and then manipulate the data without breaking the signature.
Note that this still doesn't mean that you actually have a recording of Jay-Z reciting the Navy SEAL copypasta. But you do have an endorsement from Jay-Z, and you can use that as evidence, under the assumption that Jay-Z probably wouldn't be endorsing a deep-fake of himself.

The scenario that most worries the people who are worried about deep-fakes are probably the fake porn and fake news problems. We still have the issue of detecting that something is not unmanipulated footage. But we also have the endorsement issue. If a video has been signed, that signature acts as an endorsement by the signer, and thus if there is some concern about the video, those concerns can be addressed to the signer.
For instance, suppose that CNN shows a video supposedly taken by one of their journalists. The journalist would sign the video, and so if the signature verifies you can trust that the video is the same as what the journalist endorsed. Although you might not be able to determine whether you trust the journalist; you probably have no idea who the journalist is. But CNN would also sign the video, and so you can check that the video is the same one that CNN endorsed; now the question is do you trust CNN's assessment of the video.
Indeed, signatures can tell you whether the video has been tampered with since the signing, and it can tell you who is endorsing the video, but it can't tell you whether the footage was doctored pre-signing, so it's a matter of how much you trust the signer. If someone's signature is on a pornographic video in which they supposedly star, then even if they weren't actually part of the filming process that's probably at least morally neutral. If a journalist is signing doctored footage before passing it to the news companies, well, that's a problem, but at least if it's detected then the responsibility can be traced.

Ultimately, forgery is a hard problem, and I do think that generators will always outstrip detectors, since there is no underlying hardness. One can make fake footage entirely without a computer, using actors and makeup and practical effects, and no detector looking for machine artifacts is going to detect that since the footage itself is real, just deceptive. So it may very well come down to checking signatures and hoping that you trust someone along the chain, which is not a technology issue.


Anyway the deep-fake of Jay-Z reciting the Navy SEAL copypasta is actually quite off-putting to listen to because while it matches Jay-Z's vocal timbre, it has no sense of rhythm, even basic English prosody, much less the cadence of someone who raps for a living.

 

I'm going to paste all of this nonsense together without understanding how any of it works, like someone writing Node.js and hoping that someone doesn't fuck up leftpad again. And when someone asks me to explain any of it, I'll just shrug and say "here's my list of references, which are all preprints from earlier this year".

 

Someday I will come to accept that, despite what I may have believed in the heady, naive days of my youth when my world was much narrower my standards measured only by what I myself could envision, I am not actually that good at Tetris.

 

I'm pretty sure that whoever writes my dreams doesn't really understand how agency works. When things are in third person, I still shouldn't be controlling more than one of the characters on screen, much less the background physics. These dreams are not lucid; I have no idea that I'm dreaming, but I can still knock things off shelves like some sort of psychic cat. Do I want to knock them off the shelf? Not really, but it occurs to me that I can, and then things start happening. And also I'm sort-of controlling both of the people watching things get knocked off the shelf, to a similar extent that I'm controlling the rate at which things fall off the shelf.

Anyway, it's a horrible user interface.

 

I really like the fact that English Knights are called "Sir [firstname]" rather than "Sir [lastname]" as people from the US often mistakenly say, to emphasize that the title is not inherited. While titles here in the US are also not inherited, they are always addressed as "title [lastname]".
On the other hand, of our 45 presidents, 6 of them were named James, so maybe that wouldn't have made things seem lest dynastic.

 

Given that my job and the job of the people that I work with is essentially "professional infodumper", I would not be surprised to find out that some of us are autistic. In fact I would be surprised to find out that none of us are autistic. The real question is can I distinguish between those who are autistic and those who are neurologically typical but are socially mathematicians. The answer is probably no.

 

A Venn Diagram: label on the left saying "diagrams indicating subsets", on the right saying "diagrams involving circles". In the middle it says "pie charts".

 

It's good to find out that I'm not the only one going through an "inadvisably spicy drinks" phase.

[Edit]: It turns out that I can suppress the urge via Spicy Hot V8, which is probably the source of my salsa-drinking habit to begin with.

 

The kind of person who can do an impression of themself with it clearly being an impression and not just them acting normal.

 

Due to a combination of laziness and inconvenient configuration settings, I've been using the Kitten Academy livestream as a clock. I mean, I also use it to watch kittens, but it has also has a clock.

 

Ultimately, my stance on the controversy is that I'm not fond of either banana candy or actual bananas.

 

Yeah, marsupials are weird, but humans are also born way too early. Also wrinkly and pink and mostly hairless. The pouches are mostly artificial these days.