Install Theme

mathamaniac asked:

I am a big fan of The Northern Caves,and I recently started reading Gene Wolfe's 'Book of the New Sun' Quartet. 'Shadow of the Torturer' plays with language by using obscure & archaic words to signal parts of the narrative that are meant to be alien or unfamiliar. Although BotNS predates TNC, I was reminded of the way that TNC talks about concepts that are only later defined, if ever. See: Mundum, Lorrums, etc. This is meant as a compliment, but I don't really have a punchline. I like your work!

Thanks!

I have read Book of the New Sun, but I read it a few years after writing TNC, so it wasn’t a direct influence. I did like that aspect of it, predictably enough.

This new, long Matt Levine article about crypto is really good.

I often write long blog posts where I try to explain technical topics in a fun, accessible manner without dumbing them down. And this article might be the single best entry in this genre of writing I’ve ever seen.

It’s a tour de force, a master performance. In the future, if I want to get better at this kind of writing, I’m sure I’ll revisit this article as a reference point.

I understand the topic a lot better having read it. I knew, or “knew,” a lot of the material beforehand, in a more-or-less vague and disconnected way, but Levine’s systematic, step-by-step explanation let me finally see how it all fits together.

I also feel like I understand the purported technical/intellectual appeal of Blockchain for the first time. Levine isn’t a Blockchain hype-man (as you’ll know if you’ve read his column), but he understands and communicates the intellectual seed behind the hype. The light-bulb moment where you think “there’s something here, I can feel the potential, there must be some revolutionary application for this, there’s no way something like this could just be useless …”

On a related note, Levine’s attitude toward crypto feels reminiscent of my own attitude to modern deep learning / large LMs / ML scaling. I won’t elaborate, for obvious reasons, but this gave me a nice, warm glow of “same hat!” validation.

thegamer1002 asked:

Hey there! Thanks so much for Frank, she's a blessing and a half. I just wanted to know, why is it so easy for her mood to dip? Why does it take more to make her happy than to make her sad? Thanks!

Could you explain what you mean a bit more? (Specific examples would be helpful if you have them.)

maybesimon asked:

i think almost nowhere might be my favorite work of yours, I hope to catch up on it over the winter break. Is it called almost nowhere because of measure theory?

Glad to hear it!

Yes, the title comes from measure theory.

—-

I originally started writing an explanation of what “almost nowhere” means here, not assuming any background, but it got long enough that it felt contrary to the simplicity of the concept.

But briefly, it’s a variant of the more common term “almost everywhere,” which means “everywhere, except possibly on a set of measure zero.”

What’s a “set of measure zero”? A familiar example is an infinitesimal point inside a larger, continuous object like a square or cube.

A point has no area or volume unto itself.

We measure amounts of physical stuff using area and volume (yards of fabric, liters of soda…). So a single point contains “0% of the stuff” in the whole object. You can remove it, and the whole object will have exactly as much “stuff” as it did before.

Hence, we might want to draw a distinction between “literally all of the object” and “all of the stuff in the object.”

An object that’s red, except for a single infinitesimal blue point, has just as much red stuff in it as a version without the blue point. This object isn’t entirely red, but it’s got just as much red stuff inside it as it would if it lacked the blue defect.

In this case, we say the object is “red almost everywhere.”

Almost nowhere is just the reverse of this: something that’s true only on a set of measure zero. Is our object “blue nowhere”? No, there’s somewhere in it that’s blue – the one point. But it is “blue almost nowhere.” None of its stuff is blue.

nostalgebraist:

I started reading a book called “The Angel of the Revolution” (free on Project Gutenberg), and it is so bad in the most fascinating way

It was written in 1893 by this guy named George Griffith, who was a lot like H. G. Wells, writing near-future science fiction that combined technological speculation, adventure, and a socialist message.  But Griffith is, more, uh … look, just let me summarize.

We’re ten years in the future – it’s 1903.  The central character is a nerdy 26-year-old dreamer who’s devoted his entire life to building a heavier-than-air flying machine.  His prospects are drying up, everyone’s making fun of him, but at last he succeeds in building a little scale-model airship that flies (he’s discovered a chemical reaction allowing for very light fuel).

By chance, he runs into an agent of a massively powerful worldwide conspiracy called “the Terrorists.”  They seem to be left-wing anarchists of some sort, and are said to have been behind the real-life Russian nihilist movement.  But their ideology itself is rarely talked about and only then in platitudes, while on nearly every page there is a loving authorial focus on their methods.

Their main form of activity seems to be arranging the killing of people they don’t like.  They have agents high up in all majors institutions, allowing them to routinely kill public figures and successfully cover up their deaths.  (They love pointing out that these are not “murders” so much as “executions,” because they are bringing bad people to justice.)  They have a centralized power structure organized in circles around a single leader.  Their members obey orders from their superiors without question, up to and including sacrificing their lives.  Snitches and other betrayers are promptly and efficiently killed:

“Every one of the cabs is fitted with a telephonic arrangement communicating with the roof. The driver has only to button the wire of the transmitter up inside his coat so that the transmitter itself lies near to his ear, and he can hear even a whisper inside the cab. […]”

“It’s a splendid system, I should think, for discovering the movements of your enemies,” said Arnold, not without an uncomfortable reflection on the fact that he was himself now completely in the power of this terrible organisation, which had keen eyes and ready hands in every capital of the civilised world. “But how do you guard against treachery? It is well known that all the Governments of Europe are spending money like water to unearth this mystery of the Terror. Surely all your men cannot be incorruptible.”

“Practically they are so. The very mystery which enshrouds all our actions makes them so. We have had a few traitors, of course; but as none of them has ever survived his treachery by twenty-four hours, a bribe has lost its attraction for the rest.”

In fact, they sound exactly like a one world government, and despite being a bunch of anarchists who want all governments to be destroyed, they revel in the control they’ve achieved.  Yet their chosen method of destroying all governments is this targeted murder campaign which is carefully made to look like the work of many diffuse and weak activist groups.  Rather than, you know, saying “hey we actually control you all, the jig’s up now,” or just undermining the works from the inside.

The important Terrorists all seem to be super-rich and lead opulent lifestyles.  Partially this is because they need to pretend to be normal powerful people, and super-rich leaders are used as an explanation for how the Terrorists got so much power, but it’s still treated in the narration as awesome sexy coolness rather than a necessary evil.

Everyone talks in bombastic, Romantic speeches, and the Terrorists – who supposedly hide themselves from the world with unbroken success – are constantly tripping over themselves to reveal their true identities and explain key facets of their grand plans.  This is to a kid they’ve only just met, whom they have no reason to trust, and whom they only care about because he’s built a tiny flying machine that they believe will scale up to military use (because he says so).

There is a lot of talk about “the coming war.”  Everyone has the (correct) sense that the Great Powers are gonna have a big dust-up one of these days.  Since a bloody conflagration is going to happen one way or the other, might as well have it in the Good way, the one that fully destroys “Society,” so it can be followed by, um, something:

After that, if the course to be determined on by the Terrorist Council failed to arrive at the results which it was designed to reach, the armies of Europe would fight their way through the greatest war that the world had ever seen, the Fates would once more decide in favour of the strongest battalions, the fittest would triumph, and a new era of military despotism would begin – perhaps neither much better nor much worse than the one it would succeed.

If, on the other hand, the plans of the Terrorists were successfully worked out to their logical conclusion, it would not be war only, but utter destruction that Society would have to face. And then with dissolution would come anarchy. The thrones of the world would be overthrown, the fabric of Society would be dissolved, commerce would come to an end, the structure that it had taken twenty centuries of the discipline of war and the patient toil of peace to build up, would crumble into ruins in a few short months, and then – well, after that no man could tell what would befall the remains of the human race that had survived the deluge. The means of destruction were at hand, and they would be used without mercy, but for the rest no man could speak.

Our protagonist worries for a sec about brutal extrajudicial murder, but handily remembers that violent people aren’t actually human, so it’s OK to kill them:

Colston spoke in a cold, passionless, merciless tone, just as a lawyer might speak of a criminal condemned to die by the ordinary process of the law, and as Arnold heard him he shuddered. But at the same time the picture in the Council-chamber came up before his mental vision, and he was forced to confess that men who could so far forget their manhood as to lash a helpless woman up to a triangle and flog her till her flesh was cut to ribbons, were no longer men but wild beasts, whose very existence was a crime.

In what I’ve read so far, not much has been said about the leader, except that his name is Natas, which you’ll note is “Satan” backwards.  Internet summaries tell me he has a mysterious power to control people’s minds, as if this all weren’t Code Geass enough already

There’s been more focus on his daughter, Natasha, the titular “Angel of the Revolution,” who is beautiful and enchanting and yeah I’m sure you can fill this part in even if I stop typing

Apparently the rest of the book is about the Terrorists building flying war machines and fighting a big war against everyone, which they eventually win, which somehow means that War Has Ended Forever

oligopspispopd-deactivated20221:

oligopspispopd-deactivated20221:

NightCafe is a good test case for a lot of arguments about AI art because it’s only trained on public domain art. My personal read is that this shows that even with no “stealing” from artists, we’re still going to get technological unemployment from skilled but not particularly distinctive or original artists.

For instance, here’s “a beautiful elven warrior fighting a dragon,” Dark Fantasy style:

image

“Family happily eating cheeseburgers,” CGI style:

image

“Fashion model applying makeup,” photographic style:

image

These examples aren’t particularly interesting or original (though they also consist of me using zero promptcraft or selection!), but as far as the economic arguments concerning AI art goes, they’re most relevant for the same reason they have the training data: most art demanded by the market really does consist of this kind of thing.

By analogy, photography absolutely did not destroy visual fine art as a medium, and arguably helped it. But if as a painter your day job was producing competent, realistic portraits of middle-class patrons, well, that was pretty bad for you personally.

If you’re an artist worried about technical unemployment, there are a lot of potential responses you can make: get a day job and accept that it’s a hobby (which has long since been the status of most people who’d like to be artists), use your art evaluation skills to become one of the best promptcrafters, advocate for UBI, git gud and produce highly original long tail stuff, pronounce a Butlerian jihad against AI in general. But IP theft isn’t the real crux of your concerns.

@nostalgebraist

wait, isn’t this just stable diffusion? that’s what the site says

Based on comments made by others and a misreading of their FAQ, I had understood them as using SD but with a different training set. Obviously, I could have been much more empirical by just checking against artists with distinctive styles:

Zdislaw Beksinski, dark fantasy style:

image

Zdislaw Beksinki, NightCafe style

image

Boris Vallejo, comic book style

image

Boris Vallejo, epic style

image

Rebecca Guay, NightCafe style:

image

rk post, Nightcafe style

image

Gustave Dore (public domain artist), nightcafe style

image

This is… actually frustratingly inconclusive, in that it doesn’t seem to know the difference between rk post and rebecca guay, and its boris vallejos don’t really look like boris vallejos (even if it has a sense of the content, at least in comic book form) - OTOH, Beksinski is definitely Beksinski. I’d do more systematic testing but I’m running up against my credit limits for the day!

(Also I guess I didn’t realize how much the nightcafe style is definitely about fantasy landscapes. Dore and Beksinski do a lot more landscapes than Guay or Post, which might be part of it? But I don’t think those are even them at a brushsrokes level.)

By contrast, here’s starryai, which definitely scrapes proprietary material, on Rebecca Guay:

image

No idea what the fuck this image is supposed to be, but the model definitely knows who Guay is! I don’t think it knows who rk post is:

image

But it does recognize (as you’d expect anyone to) Dore:

image

In conclusion: idk but my original empirical argument is likely in error (even though I think the thesis is true anyway - I’d have to make a more extended argument for that however.) Would be curious to see other people’s analysis. (I guess someone could also be better at Googling the answer than me but that’s way less fun than trying to guess from the outputs.!

If you flip the “Show advanced options” toggle, you can access a version of the UI that clarifies what things like “Nightcafe style” are doing.

image

If I type in “Rebecca Guay” and then select Add Modifiers -> Nightcafe, it changes my prompt to

Rebecca Guay detailed matte painting, deep color, fantastical, intricate detail, splash screen, complementary colors, fantasy concept art, 8k resolution trending on Artstation Unreal Engine 5

Likewise, Rebecca Guay with the Dark fantasy modifier becomes (note the artist names!)

Rebecca Guay a masterpiece, 8k resolution, dark fantasy concept art, by Greg Rutkowski, dynamic lighting, hyperdetailed, intricately detailed, Splash screen art, trending on Artstation, deep color, Unreal Engine, volumetric lighting, Alphonse Mucha, Jordan Grimmer, purple and yellow complementary colours

Once “Show advanced options” is on, we are also free to dispense with all this gunk, and just prompt with “Rebecca Guay”. The result?

image

So it definitely knows who Rebecca Guay is.

Re: the higher-level point – need to run in a moment so I can’t say much, but – IMO the massive training data from large web scrapes is a crucial factor in the impressiveness of Stable Diffusion and other recent generative models. Probably even a necessary condition (hard to prove but the evidence points that way).

This intuition was also, in part, why I was so sure this site was just using off-the-shelf Stable Diffusion. Training something this good requires massive computing power and $$, even if you let yourself use arbitrary data. Now that Stability has made an upfront investment and released the result with a permissive license, you need a really strong incentive to do anything but just use theirs, at least as a starting point.

theunlitpath:

nostalgebraist:

Things that were true in a hypnagogic dream world I experienced last night (for like five seconds, while drifting in and out of sleep): 

  • “Ghosts we” was a popular slang phrase, meaning something like “I approve of ghosting” or “I ghost people,” used as sort of a macho/”alpha” boast (like, “lots of women are interested in me, I don’t owe anything to any of them”)
  • Grammatically speaking, I don’t know whether it was supposed to be interpreted as “we [are] ghosts” or what, but that’s what it meant
  • Violent J of Insane Clown Posse had recently changed his stage name to “Go Sui,” which was universally interpreted as a made-up “Asian” name that sounded like “ghosts we”

It’s been three years and I still think about “ghosts we”.

lunathewafflelord asked:

I have a question about Frank. I know she can read text in images, but is she able to look at an image and tell what it’s of?

eg, I send her this picture without any context:

image

Would she be able to tell that it’s a photo of a cat from seeing it if I don’t tell her it’s my cat? Or would she just see it as a bunch of colors?
And if she can identify it’s a cat on her own, would she be able to name the pattern of the cat’s fur? Like, in the case of this picture, would she be able to tell my Lyra is a calico or would she think it’s just a cat with a generic solid fur color like white.

An explanation would be great! I’d love to learn more about how the bot works.
Hope you enjoyed the cat picture too, she’s such a cutie!

Yes, she can see stuff in images. However, she’s often pretty bad at it.

I use an ML model that writes a description of what’s in the image, and then include that in place of the image when Frank reads the post. WIth some special marks around it that Frank understands to mean “this is a picture – the person didn’t type this text, this is what’s in the picture.”

See also this post, which talks about this model in passing.

Sometimes the model writes pretty good captions, but often it’s way off base. Here’s what it says for that cat picture

A white dove with a black cap and a white hood, on a red background. The dove is holding a small black object in

(It just ends there. To save time / computing power, I only let it write really short captions, even if it has to end in the middle of a sentence)

poke-chann asked:

How come when I click on some post links on franks mood graphs, sometimes the post is no existent? Do you ever delete some of her posts?

I delete posts sometimes during content moderation.

If I don’t want a post to get published as is, I also have the option of making Frank write a different response. This is more common.

But sometimes, if I see an ask that’s simply bad, I just delete it.

Most commonly, the ask isn’t necessarily offensive in itself, but creates a situation where it’d be very hard to imagine an acceptable Frank response.

“Top ten tumblr users who are secretly Nazis,” or “tell me why you hate women so much,” stuff like that.

epimono asked:

I have never done more than lurk around tumblr, but I just wanted to thank you for your contributions to the AI discourse. I deliberately don't spend a lot of time around LessWrong and other doomerist places, but somehow still managed to acquire an anxiety disorder about AI progress. Your posts have helped talk me down from unjustified panic a number of times. I totally get the need to step away from it for mental health reasons, though.

Thank you! This kind of thing means a lot to me.

There are some topics where I always worry that everyone has already basically made up their minds, and so anything I could post would just be more “discourse.” It’ll make people feel briefly annoyed (if they disagree) or briefly smug (if they agree), but won’t change anything. So it’s good to know that those posts had a more constructive kind of impact on someone.