There’s a great future where AI media leads to an explosion in high-quality creative output, but to get there, AI’s sloppy and schlocky aspects need to be identified and mocked and eliminated.

Yudkowsky

And for a first post, what an easy group to mock: the 200-IQ turbo-rationalists! (At least for a deeply closeted Sequences-enjoyer like myself.)

But Yudkowsky really does make a good first target, what with being the whole AI-thinker guy. It seems fair to hold his outputs to a higher standard.

Everything That Hurt You

In the abstract, his “Everything That Hurt You” starts off strong by virtue of being cloneable and easily “runnable” from its public GitHub source: https://github.com/esyudkowsky/everything-that-hurt-you.

This is great for me, since it means I can fork the source (https://github.com/JohnQPulp/everything-that-hurt-you) to save off my own immutable copy for reference, even if the website version (https://everything-that-hurt-you.vercel.app/) changes later on. Yudkowsky deserves credit for this open-sourcing.

And visually, the title screen makes a decent first impression. The text and UI elements are clear and look good.

The environment and characters themselves are all obviously generic fantasy garbage, but the generic-fantasy-garbage idea is at least conveyed cohesively.

Except: there is one physicality issue. The contrast of the characters’ shadows is too sharp for a post-sunset scene. It’s subtle but still unacceptable, especially for being the first background presented to the reader.

AI-Editing Tip #1.a: Details matter. Double-check shadows.

And that’s a tiny issue compared to what clicking “Begin” brings us to:

Chapter 1

This is a horribly jank first slide.

On the background side: the bottom corners of the crosswalks don’t meet at a curb, and the intersection in general is missing either its stoplights or its stop signs. There should also be more than a single streetlight in the far background.

For the foreground: the character himself is proportioned strangely and also not integrated into the background well.

Unlike the title slide, where the isolated shadow problem could be fixed with a targeted AI image edit, this background has too many minor problems and should have been regenerated altogether. Fortunately, regenerating is an easy thing to do.

AI-Editing Tip #2: Aggressively re-roll image generations to curate better results.

These next two slides of the read are also text-less, which initially confused me as to whether the UI was broken:

But no, it’s intentional visual-only storytelling. Not a bad idea in theory — and functionally, it does get the idea across — but the execution is still terrible.

Why are the locations in all three images mutually inconsistent? How is the weather also inconsistent?

Part of the answer is that, at the current level of image-gen technology, getting full consistency across multiple images is difficult, especially the more details those images have.

But that doesn’t excuse the sloppy sequence here. Either the images needed to be re-rolled for consistency, or the three-image sequence should have been cut down to just the final one — the only one actually necessary for getting the idea across.

AI-Editing Tip #3: Sometimes less is more. Visuals can be improved by having fewer of them.

The first actual text of the VN looks pretty good. (The small-caps formatting is unnecessary, but cool-looking.)

The text contents themselves are not good, cramming an “I”, a “you”, an awkward comma, and a stupid fantasy name all into the first sentence.

Also, it’s strange that the primary UI shows no speaker, but the history option does:

Still, it’s nice that the AI continues to do a good job with the visual appearance of the history and other settings overlays. It even included the absolutely vital (why-is-this-not-the-default) instant text-speed option:

In the next actual background, the main character’s facial hair has already changed.

You could explain away the inconsistency on a technicality: maybe he took some time to die, and you keep the facial hair you die with. But since rationalists love a good heuristic, here’s one:

AI-Editing Tip #4: Don’t include visuals that cause the reader to pause and question whether what they’re seeing makes sense.

And also:

AI-Editing Tip #1.b: Details matter. Double-check hair.

The first instance of a sprite overlaying a background looks terrible, with the sprite clearly incorrectly sized. This is an obvious-enough lazy-editing problem that it doesn’t even warrant an AI-editing tip.

Less obviously (due to being masked by the more egregious problem), also take note of the clothing inconsistencies.

AI-Editing Tip #1.c: Details matter. Double-check clothing.

With the character back in the background, we get more clothing inconsistencies, plus facial and stylistic inconsistencies.

Mercifully, the story only yadda yadda yaddas its way through only a bit more trashfire expositional text before reaching chapter two.

There’s no need to summarize any of the text and images that were skipped, because you can infer them all. Essentially no story-unique information was conveyed in this first chapter, making it all the more amazing how much sloppy execution it managed to contain. Less is more.

Chapter 2

Chapter two gives us our first conversational-style textbox of the game, dropping the small-caps used by the previous expositional textboxes in favor of italics for non-dialogue and regular undecorated text for dialogue. It’s not the most readability-optimized system imaginable, but it’s at least a consistent system, which makes it functionally good-enough.

Conversely, the sprites are not good-enough.

Avram Detwiler’s sprite is too large, both relative to his bounding box (leaving him cut off on the left) and relative to the bartender. The difference makes it seem like Avram Detwiler is set above the background (in a speaker box), while only the bartender is set into the background (on the wrong side of the bar).

And as for the writing itself:

That hyphen is incorrect, which the AI should have fixed as part of copy editing. I note my confusion as to why it didn’t.

And prose like “a guy messing with another guy” is just terrible. Someone should have been doing the editing, whether AI or human.

The sprite expressions finally change, far into the conversation. Having so few variations leaves the characters lazily static, not matching their implied textual emotions.

AI-Editing Tip #5: Sometimes more is more. Visuals can be improved by having more of them.

Sprite variations are in general much easier to generate correctly en masse than backgrounds, so it’s extra egregious that the bartender sprite features a badly cropped and visually mismatched bar. Re-roll.

Again and again: “guy”. It’s always entertaining when an author overuses a word without realizing it.

(But another incorrect hyphen too? Really?)

Related to the too-few-sprite-variations problem: it doesn’t look good when the described text actions don’t at all match the on-screen sprites. This type of mismatch could be avoided by just not having the characters continuously on-screen. (Which would also mean less time spent having to look at their ugly stupid schlocky fantasy-tropey faces.)

You know you’re in a 300-IQ rationalist text when the dialogue isn’t just profitable and isn’t just sophisticated — it’s profitable and sophisticated.

And here’s a couple more bad-prose snippets to round out the chapter:

Chapter 3

Right away in chapter three, we get four mutually inconsistent physicalities for the slave-staging area:

The platform and background architecture changing with every subsequent slide is sloppy. (Note also that even the announcer’s rings change.) However, those details are at least not informationally important — their lack of consistency doesn’t actively confuse the story.

Whereas the collared slave expression inconsistencies (between the fox-girl and the others) do confuse what the author is attempting to get at. Is the collar a mind-control device, or is it not? It’s a visual contradiction. (Remember: don’t leave the reader doubting the visuals, even if those visuals could be correct on a technicality.)

There is the first correct em dash of the VN buried in all that mess. However, it has spaces around it. Asinine blog posts use space-formatted em dashes — professional fiction that seeks to be taken seriously should not.

Another fiction-writing heuristic for the rationalists: don’t use “you” outside of dialogue. (It can be done correctly, but not by you.)

The physicality of the pamphlet changes between shots.

(The fake rune things are also obvious nonsense, but they get a pass for following in the proud footsteps of Jules Verne.)

[Inexplicable square brackets?]

[Continued scene physicality?]

Chapter 4

All of this dialogue is horrendous, although the 400-IQ rationalist precision of “including any of your favorite foods that are on the menu” is at least unintentionally funny. (The blushing sprite juxtaposed against “I’m not planning to sexually assault you” amuses less.)

Also note the fox-girl’s inconsistent ear. Details matter. Check your… hair?

This is the worst background in the VN, inexcusably lazy in its wholly anachronistic setting to not have re-rolled.

(Note also the nonsense stacked plates.)

Textual-visual mismatches continue to distract.

Chapter 5

The pamphlet’s appearance changes again, this time with added inexplicable (and mismatched) numerals amongst the runes.

Numerous inconsistencies between these two aside, what’s even going on in that first image? Re-roll.

Spot the impossible triangle in this background.

AI-Editing Tip #1.d: Details matter. Double-check architecture.

Chapter 6

Here we have another sprite cut off on the sides, more nonsense stairs in the background, and most hilarious of all: a light switch.

AI-Editing Tip #1.e: Details matter. Double-check for anachronisms.

(Also: “Heard about you two”? Hah?)

Rationalist writing heuristic: avoid phrases that invoke distracting memes. It’s not fair that some words get tainted, but life’s not fair.

(Also: why is fox-girl still in her sleeping outfit?)

“Anomaly” coming from two different characters in rapid succession is bad. It collapses the facade of there being different voices.

(The “fucking” is also poorly placed here.)

Yudkowsky won’t rest until he’s used up every possible incorrect dash variation.

Chapter 7

Check your shadows.

The sword is duplicated and bent in a way that doesn’t make physical sense. Re-roll.

As fast as… hah?

AI copy editors can also fix alien-authored phrases like “walks the road”.

The background reuse has been acceptable (if not ideal) up until now, but this instance is too distracting. It can’t be both the protagonists’ private campout area and the traveling caravan’s campout area.

Because the background idea is generic here, it would have been easy to just generate another similar-but-distinct fire image.

AI-Editing Tip #6: The less informational detail an image needs, the easier it is to generate sufficiently correctly. Use this when deciding when more-is-more vs less-is-more.

Previously both words in “Social Skills” were capitalized. Use AI copy-editing to spot these terminology inconsistencies.

Again: “fuck”? A childish placement for this “““grownup story”””.

Chapter 8

Inexplicable arrows and a jank perspective here. Re-roll. (Cool lion-man though.)

The background collar character again confuses the intended worldbuilding.

Where’s the pot she’s ladling out of? (Current models are especially bad at food physicality for some reason.)

AI-Editing Tip #1.f: Details matter. Double-check food.

Up until now the AI-generated UI elements have been surprisingly seamless. Here we finally get some errant underlining — minor, but also easily fixable.

Chapter 9

Sword and a crossbow?

First to die? Lion-man doesn’t get personhood?

(The phrasing and timing of this dialogue are both also bad.)

The road’s physicality keeps changing, and what are we meant to take away from the mask-guy on the left there?

Art style changes are distracting. Re-roll.

(But at least a floating em dash is less wrong than a floating hyphen.)

(Also his scabbard is missing.)

What’s wrong with her face? Where’d the shoes come from? Why did his sword shrink? Why did the trees turn into pines? Re-roll.

Dies of hyphens cringe.

Schlockworld IV has identical caduceus symbols to Earth. Double-check anachronisms.

The background physicality’s borked — current models are bad at windows.

(The text is also completely inscrutable to me here, not being a 500-IQ rationalist.)

Chapter 10

The dropping sword change here is funny. Fox-girl also changes sizes. Probably only one background was needed here. Less is more.

Bad sprite sizing returns here, along with inexplicable mid-conversation outfit-changing:

The relative heights of these two are also ever-changing.

Chapter 11

“From my perspective”?

Outfit change? Collar above her hair? Gloves?? Re-roll.

Chapter 12

These hyphens - are going to - kill me.

There are many great 600-IQ rationalist dialogue options here, but “If I had to do it all over again, I still wouldn’t have sex with you” is the clear dumb-sentence-of-the-book standout.

Outfit change? Small text? (The small text is kind of funny actually.)

Also note the way the trees slope down in the background, implying a relatively short drop-off, as we enter our final multi-image sequence:

The physicality is inconsistent. The style is inconsistent. The digging makes no physical sense. The vomiting reminds me of Family Guy.

Chapter 13

Remember Skagganauk from the history menu? Mystery solved, I suppose.

Conclusion

In conclusion, Elhom IV is a land of contrasts. Its morals are inferior to the Earth man’s, but at least it has lion-dudes.

Visually and prose-wise, too many aspects of the reading were dreadful. The ideas and the entertainment potential were both there (setting aside the schlocky setting, characters, and medium), but the editing at times felt nonexistent.

Still: with the source on GitHub, a sufficiently funded Claude could no doubt improve the story somewhat. Though not entirely.

The context requirements for a visual story, even one as relatively short as this one was, are still too large. Human+AI team-based editing will still define the AI media frontier for now, with the AI media frontier still being strictly inferior to the human media frontier in terms of quality. (Though not cost.)

But: as the 700-IQ rationalists are fond of saying, this is the worst the AI media will ever be. We have only to wait for the inevitable improvements. (They just need to happen faster.)