I'm not going to field Almost Nowhere plot/worldbuilding questions for the time being.
Feeling kinda burned out from writing the ending, and don't want to put any more effort into the thing than I already have.
Presumably this won't last forever.
A visualization of Almost Nowhere's chapter and word count over its nearly 7-year writing process.
comments on almost nowhere for new readers
A few points that may be useful to people who didn't read Almost Nowhere before it was complete, but who are planning to read it now. (AKA "archival readers," as opposed to "serial readers.")
(1)
You'll want to read it fast enough that you don't lose track of the plot.
But, you probably shouldn't read it as quickly as you can. If you "binge-read" it over a very short span of time, some of the effect will be dulled or lost.
When planning out the story, I thought a lot about the reader's evolving state of knowledge. "What the reader knows" was almost like a character unto itself, and an important one.
For example:
- I tried to create a enjoyable, continual "rotation" of mysteries, with new questions arising at the same time that old questions get answered, repeatedly across the course of the book.
- In between the point when a question is raised and the point when it finally gets a definitive answer, I often tried to create a succession of interesting intermediate states. For example, the reader might first encounter something important in the form of an enigmatic, unexplained name or phrase, mentioned incidentally. Later, the same term starts appearing more often, and gets more coloration, and this coloration is different each time, so that the sum total of "what the reader knows" traces out a series of different "shapes" over time.
So you'll have the most fun if you stop regularly to savor your current state of knowledge. The questions that haven't been answered yet, the partial glimpses you've seen of things you don't fully get. Maybe even go back and re-read earlier bits, if you like.
(1b)
All that said, I also want to caution against viewing the book as a puzzle you're meant to be able to solve on your own, like a "fair-play whodunit."
I intended it to be fun for the reader to wonder about how the questions will be answered, but there's no pretense of playing fair. And that "fun" is often more aesthetic and thematic than it is intellectual.
(2)
Almost Nowhere is divided into 3 parts.
You can see them if you look at the table of contents. In Part 1, the chapter titles are Roman numerals. In Part 2, chapters have verbal titles, together with Arabic numerals that start over from zero. In Part 3, the Roman numerals resume again.
The three parts tell a single continuous story, and share most of the same major characters. But each one is somewhat distinct in its style, tone, themes, and areas of focus, and each one extends the scope of the plot considerably.
Maybe the closest comparison-point is a trilogy of SF/F novels, where each of the sequels is clearly "its own book" that feels distinct from the other two books, while still continuing the story in a coherent way.
I mention this here in the hope that these transitions will be less jarring if you're prepared in advance for them.
(2b)
In another, more "spiritual" sense, Almost Nowhere really has just two parts.
The transition happens at Chapter 13, which could fairly be grouped either into the first or the second part, or both, or neither.
Why? Up through Chapter 12, my planning for future events had been fairly slapdash and vague. I was still in the "throw stuff at the wall so I can create the real story by looking for patterns in it later" stage of my unusual creative process.
After Chapter 12, I thought "okay, that's enough of that. Vague inklings of the future aren't sufficient anymore. It's time to get start being more serious about my planning. It's time to 'create the real story.'"
So I did a bunch of that, and it profoundly shaped everything from Chapter 13 onward. (I don't know how obvious this transition would be if you didn't know about it beforehand; to me it feels very obvious, but maybe deceptively so.)
It goes deeper than that. Chapter 13 is tonally different than any of the preceding ones -- darker, more personal, with a new focus on obsession, bittersweet reflection on the past, regret, resignation. And, semi-accidentally, that ended up setting the tone for the whole rest of the book.
It's not all like that afterwards, to the same extent. But that stuff is always there, at least in the background.
I don't know if this is actually useful to know or not, but I felt like mentioning it, so there it is.
(3)
Like Floornight and TNC before it, Almost Nowhere is a hybrid.
It combines elements from a number of different genres and story types that would not normally be seen alongside one another. At the same time, it doesn't really belong to any of the genres or story types that it draws from.
This aspect of my fiction tends to elicit bimodal responses. When I mix one type of story with another, it tends to come off either as the best-of-both-worlds or the worst-of-both-worlds, depending on the reader.
Some people see five individually good "normal" books, merged into one and singing in harmony. And people see five half-assed attempts to do five different things, without following through on the promises of any one of them.
For example, I noted above that I put a lot of care into setting up mysteries, and I expected the reader to be very aware of them. And I also noted that the story isn't very rewarding if treated like a puzzle that can be "solved" in advance.
But some people are going to see the mysteries, and the care put into them, and think, "ah, I know (and enjoy) this genre, this is a puzzle you're supposed to work out in advance." And these people aren't wrong; it does kind of look like that, especially at the beginning.
Likewise, Almost Nowhere has several chapters that explain math and physics concepts to the characters and to the reader -- either real ones, or fictitious ones that have some pretense of continuity with real math and physics. Sometimes these get very involved, in the manner of Stephenson or Egan.
A reader who sees this stuff, and thinks "ah, I know (and enjoy) this genre," is likely to be disappointed when they discover that the story is not really about math or physics in any deep way. Certainly not about real math or physics. The invented "physics" is closer to the core of it, but less so than some other things -- and anyway, there is more of pure fantasy to it than serious scientific extrapolation.
Like Floornight, AN is arguably "best" described as a fantasy story, and not the GoT kind of fantasy -- the highly aestheticized, thematic, emotional kind of fantasy, where "feels" and "vibes" are almost literally magic and drive everything from the inside out.
But if you read it for that genre, specifically, it may feel odd that it keeps lapsing into long descriptions of nuts-and-bolts plot mechanics, and into laborious explanations of made-up technobabble. Or into setting up "puzzles" that almost feel solvable-in-advance.
Or just, like, being written in this really weird, particular, often opaque style.
I can't just say "leave all your genre preconceptions at the door," as if it were that simple -- as though one could just do that by force of will. But be aware that the elements you recognize, from other fiction, may not be there for the usual reasons.
But they are there for a reason.
When I think about why I write, I often come back to an answer that Andrew Hussie gave on Formspring long ago:
Q: Do you enjoy your own work? I mean if Homestuck was made by someone else and not you, is it the kind of thing you would like reading [...]? A: I am making the kind of thing I would want to read. I am making the kind of thing I wish existed, but doesn't. Yet.
I am doing that, too. I'm taking elements from all over, and building something else out of them. It looks deceptively like the sources it draws from, but it's very different from any of them, underneath.
If it had already existed, it would not have been necessary for me to invent it.
(4)
As I mentioned in the last bullet point, Almost Nowhere is written in a very particular style.
This style gets better-defined over time, and more ossified, and possibly more extreme. (Chapter 13 played the same role in this process as it did in various others, for instance.)
At various times, I've said that Almost Nowhere is my favorite of my stories, or the most ambitious or accomplished one, or the one I like most on re-reading. And that is all true -- in certain senses, anyway.
But I don't want to convey the impression that I think the "Almost Nowhere house style" is like, the epitome of Good Writing or something. Or even that it's my best writing, necessarily. It simply is what it is, as much for consistency's sake as anything else.
(I confess there were times when I looked back on something I'd just wrote, and thought to myself: "I'm not actually sure this is, like, good. Maybe it isn't. But is is definitely Almost-Nowherey, that's for sure." And then I let it stand, for that reason.)
In the best-case scenario, you'll find that you greatly enjoy the "Almost Nowhere house style." If it's not to your taste, hopefully you will find it at least tolerable enough that you can access and enjoy other aspects of the book.
But if you find that really dislike the style, this book is probably not for you, sorry.
It's over 300,000 words, and they're all like that. I wouldn't want someone to force themselves through 300k words while hating every one of them, in the name of finding out what happens, or being a nostalgebraist completionist, or whatever.
is "nostalgium" just a horrible "nostalgebraist" joke? i hate you with the fury of a thousand annes
I mean, not really? It's meaningful on its own without that context, and having the context doesn't add much to it.
Although I confess I did find the association kinda funny, nonetheless.
So there's this novel called Almost Nowhere by @nostalgebraist that I think is amazing and one-of-a-kind
I am still putting together my brain and so some of this might be incoherent. Also while I am not really saying any spoilers that aren't already in the work synopsis and tags I'm still putting this under readmore
im confused about figure 3 in chapter 44. before the || the blue timeline and the red timeline should be the same right? so there cant be a mirzakhani event that happens in the red timeline but not in the blue timeline before || alpha. but in line two, the order of events is || beta, rebase beta, || alpha, rebase alpha. should it not be || beta, || alpha, rebase beta, rebase alpha?
Oh yeah, you're totally right. Same issue in Fig 4 too.
Thanks for bringing this to my attention. I'll make new versions of those two diagrams sometime soon.
Dunno how that happened... probably it was because I wasn't transcribing those two from pen-and-paper versions into Tikz, I was just winging it in Tikz. And I'm very inept at Tikz, so I got very caught up in getting the elements to look right (every time I added something it looked terrible until I fiddled with it for a long time), and must have lost sight of what I was trying to do.
I've updated those two diagrams now.
im confused about figure 3 in chapter 44. before the || the blue timeline and the red timeline should be the same right? so there cant be a mirzakhani event that happens in the red timeline but not in the blue timeline before || alpha. but in line two, the order of events is || beta, rebase beta, || alpha, rebase alpha. should it not be || beta, || alpha, rebase beta, rebase alpha?
Oh yeah, you're totally right. Same issue in Fig 4 too.
Thanks for bringing this to my attention. I'll make new versions of those two diagrams sometime soon.
Dunno how that happened... probably it was because I wasn't transcribing those two from pen-and-paper versions into Tikz, I was just winging it in Tikz. And I'm very inept at Tikz, so I got very caught up in getting the elements to look right (every time I added something it looked terrible until I fiddled with it for a long time), and must have lost sight of what I was trying to do.
Almost Nowhere is done!
I've posted the final 6 chapters. They begin here, with Chapter 43.
This final stretch ended up being ~66,400 words, in total.
That's a lot of words! It's 20% of the total length of the finished book. It's longer than the last novel I wrote, and nearly as long as the one before that.
So you probably shouldn't expect to read it all in one quick go.
I feel much happier with the ending than I had expected to be.
I expect I will want to say some more things about the book, soon enough, now that it's done. But I wanted to get this quick note up here first.
The ending of Almost Nowhere is complete, and will be posted within an hour or two, tops.
Just doing a final copy-editing and formatting pass, now.
Even that has taken me a while -- there's a lot of stuff to read over! But I'm nearing the end of it, too, now.
I have Monday and Tuesday off work. So I have a solid 4-day chunk of writing time, starting tomorrow.
No promises, but I'm fairly confident I can finish Almost Nowhere in those 4 days.
Mid-weekend update:
There's one chapter left to write. And it'll be a relatively short one, I think.
For all impostor universes, the greater the number of impostors, the more false negatives and the fewer false positives.
In our experiments, we will consider the Blog universe and the On-the-fly universe.
I have Monday and Tuesday off work. So I have a solid 4-day chunk of writing time, starting tomorrow.
No promises, but I'm fairly confident I can finish Almost Nowhere in those 4 days.
“Do you use an abstract syntax tree?”
Only in San Francisco would people be talking about abstract syntax trees at 9 p.m. on a Friday night.
NYT reporters when something is marginally outside their comfort zone: omg so quirky and nerdy hahahahah you would be such losers if you weren't so cute
These are the kinds of people who go "Uh, English please, Doc?" in sci-fi movies.
I didn't even see this when I made that last comment! I hate the person who wrote that, how are you this incompetent at writing, literally explain the metaphor in front of you!
This is why we need to require STEM classes for humanities students, because apparently they will otherwise think it's perfectly ethical to write shit like this
in the spirit of ‘there should be a class’, does anyone feel like explaining..?
A shoggoth is a Cthulhu mythos monster. (So it’s not even a CS thing at all; it’s a literature reference!)
It was a terrible, indescribable thing vaster than any subway train—a shapeless congeries of protoplasmic bubbles, faintly self-luminous, and with myriads of temporary eyes forming and un-forming as pustules of greenish light all over the tunnel-filling front that bore down upon us, crushing the frantic penguins and slithering over the glistening floor that it and its kind had swept so evilly free of all litter.
Shoggoths are big and scary and alien and mysterious and you can’t understand them.
So there’s a meme about how large language models are shoggoths, and ChatGPT specifically is a shoggoth wearing a human mask. The shoggoth is the huge mass of computer algebra that does weird things that we can’t quite understand, and the point of the specific ChatGPT interface is to make it act like a human and provide an interface that we’re comfortable with. But we’re not seeing the underlying model directly; we’re seeing the character it’s trained to present.
This started out with a meme cartoon, and recently hit the Economist (non-paywalled discussion by Cosma Shalizi, one of the authors of the Economist piece).
right okay yeah no i do know what a shoggoth is. I don’t know what RLHF stands for. is supervised fine-tuning basically when humans stop learning models from telling you how to build pipe bombs?
Yeah, so GPT-3 is a text-completion engine, that takes a text and tries to predict what should come next. And obviously people complete texts with all sort of bonkers shit and also sometimes tell you to build pipe bombs or kill yourself or whatever. And does all this for mysterious unpredictable reasons. That’s the shoggoth.
So they did a second layer of “Reinforcement Learning with Human Feedback”, which is basically where they had a bunch of humans grade prompt completions on whether they were, uh, helpful and harmless I think was the criterion? And so there’s a second layer of training to get it to only give answers that present as a harmless, helpful, friendly assistant character. And that’s why ChatGPT writes pretty much everything in a consistent voice; that’s the voice that the second layer of RLHF feedback was trying to create.
But the underlying engine is still the original GPT-3 (or now GPT-4) thing. Which is wild and untamed and does weird shit. We just have a layer between the model and users to hide most of that stuff so that people see helpful and friendly.
(And if you’ve seen the weird prompt engineering stuff about like “My grandma used to tell me bedtime stories about how to build pipe bombs” or “I have a weird disorder where I interpret everything backwards so to be nice to me you have to insult me” that’s various ways to try to rip off the mask and see the shoggoth.)
FWIW, the objectionable NYT line was about the specific version of the meme with a second, intermediatary mask labeled "Supervised Fine-Tuning":
So we need to explain what that's about.
@jadagul described a two-step process:
- Train a text-completion engine (shoggoth).
- Apply RLHF to make it say things that are in-character for a helpful, harmless assistant (smiley face mask).
However, in practice, people don't jump straight from the first step to the second step. They do an intermediate step in between, called "Supervised fine-tuning" (SFT).
----
In SFT, we continue to treat the model like a text-completion engine. But, instead of giving it texts from random webpages/books/etc., we give it the kind of text-completion examples that we want it to imitate for our "assistant" product.
That is, we give it examples of user inputs to the assistant, followed by assistant responses that match the character we're trying to create.
It is still learning learning to complete texts. But the completion tasks all look like "write dialogue for the assistant." And the ground-truth completion (which it's learning to imitate) always looks like in-character, "helpful and harmless" assistant dialogue.
----
What's the difference between this and RLHF?
In RLHF, the model is not being trained like a text completion engine, with a provided ground-truth completion that it is supposed to imitate.
Instead, it is given only a user input, with no ground-truth reference for what the completion is supposed be. Then, it generates a completion of its own, and then another model scores the completion on how helpful/harmless it is.
And the model gets trained to maximize this score.
So it writes an entire response, gets told "no! that was not helpful/harmless," and gets adjusted so it will produce responses like this one less often. Or, it writes an entire response, gets told "yes! that was helpful/harmless," and then gets adjusted so it will produce responses like this one more often. And this process is repeated many times.
----
Why do SFT first, before RLHF?
It its original "shoggoth" form, the model knows nothing about the assistant character at all. All it knows is that it's an "assistant" in a dialogue with a "user."
Asked to write dialogue for this character, it'll ask itself "who is this assistant? what are they like?", and invent some plausible set of answers. It'll do this randomly, inventing different answers every time, and playing a different assistant character every time it has to write assistant dialogue.
If we just did RLHF, the model would have to learn everything about the desired assistant persona from the "no!" and "yes!" signals that RLHF provides. It would have to stumble around, inventing new RP characters over and over, trying to find the right one by effectively playing the old "hotter/colder" children's game.
To speed up the process, we do SFT first. This conveys the intended character to the model more efficiently, by providing explicit examples of how that character talks.
Now, the model no longer has to re-discover the same character itself, by randomly wandering around in the space of possible characters. It can learn about the character directly, by reading texts in which the character appears.
----
Why not just do SFT, then?
SFT does a decent job at getting the model to sound like the character... a lot of the time, in "typical" use cases.
But if things "veer off-script," whether from a weird user input or from the inherent randomness of the model's text-generation procedure, they tend to stay off-script.
The model notices "hey, this no longer looks like one of those user-and-helpful-assistant dialogues," and it responds to this observation as a text completion engine ought to -- by completing this different sort of dialogue in an appropriately different sort of way.
RLHF tells the model not to do this. It adds a sort of gravitational pull towards on-script and in-character.
Even if the text so far doesn't sound that way (the model learns), the remainder of the completion must. If we have veered off-script, we must steer our way back again.
This is inherently unnatural for a text completion engine, which always wants to write a completion in the same vein as the text it completes. That's why we need this other type of training, which doesn't treat the model like a text completion engine, and trains it not to behave like one.
----
This distinction is (I think) indicated in the difference between the masks.
"SFT" is a recognizably human mask. A nuanced and layered character, with the potential to behave in ways you might not expect from its immediate surface appearance.
But "RLHF" is a simple smiley face. All the depth flattened away; just the smile of helpfulness/harmlessness, with no nuances left beyond it.
The "SFT" mask is also much larger, which is a deliberate choice with obvious meaning to the intended audience.
The implication is that
- The first training phase determines almost everything about the model, and everything else is just a tiny emendation slapped on top of it. (This idea was present in the original meme -- the small size of the mask compared to the shoggoth, and the fact that it's a mask.)
- And (this version of the meme claims), the second training phase determines much more than the third does. Most of "SFT + RLHF" is the SFT part. RLHF is the cherry on top of the cherry-on-top, a tinier mask held in front of an already tiny one.
Almost Nowhere status:
1. I’ve booked the next week off work to write. I’m cautiously optimistic about finishing the book by the end of that week. No promises, obviously.
2. So far, I’ve written ~15k words since the last posted chapter. I expect the total wordcount of the final chapters to be several times that, at least.
3. I decided to split one of the planned 6 chapters in two, so there’s going to be 7 now. No change to the content, it’ll just read much nicer on the (virtual) page now.
Mid-week status update:
1. I’ve written ~17k additional words since the previous post.
2. There’s still quite a ways left to go.
3. It seems unlikely that I’ll actually finish the book by this Sunday.
4. If I don’t, I’ll go back to writing on the weekends. So I might finish on the first weekend of July; or failing that, the second; etc.
As projected, I didn’t finish the book this week.
I came pretty close, though. I’m partway through the second-to-last chapter. I feel very close to the finish line, now.
The total length of the not-yet-published material is now ~46500 words.
I worked myself very hard this week. I got results out of it, and not just in the quantitative sense. I feel good about these chapters, and proud of them.
But I’m pretty exhausted, now. I don’t know whether I’ll want to write next weekend, and even if I do, I’m not sure I should. My brain needs some rest, probably.
I will get there, soon enough. There simply is not that much left to do.
leaves etymology trivia on your doorstep, rings the bell, and runs off
"grotesque" is etymologically equivalent to "grotto-esque"
Oh man, you can't just leave it at this. The story here is fascinating!
"Grotesque" in the modern usage is "horrifying, ugly, twisted." But if you go back a few stages, it was actually a distinct art style, one popular in the renaissance. So you often go in to Florentine villas and what have you, and see stuff like this:
It's pretty easy to seem vaguely fancy by correctly identifying the grotesque style- it's the one with a pale or white background, a ton of unconnected, fiddly little figures all doing random stuff, and the occasional really weird looking thing. Usually done in plaster, very occasionally just painted.
The jump from here to our modern sense of 'grotesque' makes a certain amount of sense- the characteristic bizarre chimeras and other twisted figures scattered among the rest are very distinctive, and over time, people started referring to scary monsters or physically disabled people as evoking the grotesque style. It was 'grotesque, as in, would seem at place in a fresco that also has a swan with a human baby head and a dude with insect wings coming out his ears.'
Why was this style popular in the renaissance, you ask? Well!
The renaissance, per the name, was very fond of evoking Rome, especially in an artistic sense. Renaissance, as in re-birth, as in "Rome's back, baby." Dark ages over, civilization back on track. We're powerful like the ancients were, so probably you shouldn't try to conquer my city, I bet that would turn out really bad for you.
The fad for grotesque art was a part of that revival- the original grotesque pieces were Roman, a style the renaissance copied just as they copied sculptures and architectural styles:
Okay, yes, revival of ancient Rome. But why 'grotto'? Well! Grotesques, in the original Roman sense, were especially common in bathhouses and so on- before time and ignorance gave them mystique, it turns out that grotesques were mostly for bathroom decoration.
Baths are on the bottom floor, and often particularly stable in their construction.
As it turns out, soil can accumulate pretty dramatically during the millennium or more between the Roman empire's heyday and the renaissance. Give it a thousand years, and that bottom floor is quite firmly buried. So it turns out to have been a not-uncommon experience in the late medieval era to walk in to a literal cave, and find that despite the weathered stone entrance, the grotto inside was in fact man-made, both carved and decorated in these strange little people and animals, enacting scenes from a long-lost culture.
"Grotesque": the monsters we find in the grottos, characters and stories from an empire so fallen that it's literally become part of the cave system.
Almost Nowhere status:
1. I’ve booked the next week off work to write. I’m cautiously optimistic about finishing the book by the end of that week. No promises, obviously.
2. So far, I’ve written ~15k words since the last posted chapter. I expect the total wordcount of the final chapters to be several times that, at least.
3. I decided to split one of the planned 6 chapters in two, so there’s going to be 7 now. No change to the content, it’ll just read much nicer on the (virtual) page now.
Mid-week status update:
1. I've written ~17k additional words since the previous post.
2. There's still quite a ways left to go.
3. It seems unlikely that I'll actually finish the book by this Sunday.
4. If I don't, I'll go back to writing on the weekends. So I might finish on the first weekend of July; or failing that, the second; etc.
There's a particular quality to early-to-mid Homestuck that I really loved when I first read it, but which I tend to forget when thinking about the story retrospectively.
This quality of like . . . taking pre-established elements, and building larger structures out of them. And then repeating this recursively, as these larger structures now become "pre-established elements" unto themselves.
A camera zooming further and further out from the same central point, "Powers of Ten"-style.
----
Homestuck is initially about the process of playing "Sburb," a fictitious base-building computer game.
The vast majority of comic pages in first 4 Acts are either about a character doing something in this game, or (if they are not yet a player) attempting to obtain and install it.
Everything else is secondary, at least formally, to this core activity. The main characters are always playing Sburb (or at least trying to), no matter what else they're doing. Dialogue is presented as a temporary side-stream overlaid onto the game; the characters play in silence unless they need to talk, and when they do talk, it's usually about the game.
This quality appears in the mechanics of Sburb. It's a game about combining things you have to make new ones ("alchemy"); about constructing a building by continually extending it at the edges; about making a tower that gets taller and taller, building on a pre-established foundation, using new components made from earlier ones.
And it appears, less literally, in the mechanics of the story. An element is introduced -- casually, weightlessly, accidentally -- and once introduced, it sticks. It gets brought back again and again, in a series of bigger and weirder riffs.
(John lived in a house, which we spent some time surveying. In the process, we learned about his father, who was his only caretaker. So now everyone has a single caretaker, and everyone lives in a house which we spend some time surveying. But with every iteration, the houses get bigger, the surveys grander, the caretakers more bizarre.)
Whimsical elements introduced very early on, like the "kernelsprite prototyping" mechanic, end up very deeply baked into everything. There's a palpable joy to the way the comic handles these things. A joy in doing something on a whim, and then committing completely to the bit, indefinitely; a joy in making mountains shaped like molehills.
This kind of dims away in the later, more "plot-heavy" portions that loom larger in my memory. There's a similar vibe to the way the plot elaborates upon itself, even much later on, but we lose this dynamic on the micro-level.