Isekai Prime, episode 1

“I’ll give you some candy”?!? Dude!

Okay, first of all, that was the quickest, laziest isekai setup ever. Second, those were incredibly lame, passive wolf-monsters, basically holding up signs that said “grrrr” . Third, Our Hero may have set a new record for the easiest, cheesiest haremette acquisition:

1. meets poster girl at first inn he walks into.

2. offers her candy to come up to his room and teach him to read and write the local script.

3. she lounges on his bed and teasingly offers to become his mistress as he sketches her like one of his French girls.

4. she jumps him and turns out the lights.

The next morning, he heads straight to the merchant guild to register, and dear god, is that a receptionist or a cosplay nurse? She even comes with under-rim glasses installed. And the hotties just keep rolling in: the first catgirl, the rich blonde merchant daughter, even a feral loli, we’ve got it all in one episode!

The general art and animation are nothing to praise, which is good because the screen is constantly covered with price lists of his latest online purchases. Lots of cute gals, though, and his voice is instantly recognizable as The Universal Dad.

Verdict: okay, they just covered about half a season of a typical isekai show. As long as the harem does not turn into the super sentai team shown in the credits, I’ll give it a few more weeks.

(no sign of a dragon haremette yet, but give him time!)

(there are currently 10 fan-art pics on Pixiv for this show: 4 are inept porn (3 blondes, 1 grotesquely obese catgirl), 3 are competent pics of the loli, 1 is an attempt at a 3D render of the blonde, 1 is a cartoon of the poster girl’s reaction to being given candy, and 1 is a quick pencil sketch of the blonde’s maid)

Do not play the CES Buzzword Drinking Game…

If you take a shot every time you see the word “AI” in a new product announcement, you’ll be dead of alcohol poisoning within the first half-hour. And I’m being generous here.

Dental escalation

Made it down the driveway to get my crown re-cemented this morning, only to have the dentist take one look and say, “sorry, it isn’t the crown; the tooth broke and has to be pulled”. 90 minutes later, I’ve got a temporary bridge and another appointment in six weeks to put in the permanent one after everything heals up.

“Now hiring cosplay-nurse guild gals for home care”.

(or not)

Dear Dungeon Abby,

My adventuring party, consisting of six strong-willed independent young women trained in magic and swordplay and dressed in sheer robes and bikini armor, has scouted out the location of a tentacle pit. Should we use generative AI to plan our assault? – Fierce Kitten

My Dear Fierce Kitten,

I am always delighted to hear from young women pursuing their dreams in challenging careers, and I fondly remember my youthful exploration of the caves and small dungeons near my childhood home, defeating slimes, kobolds, and goblins, uncovering buried treasures, learning from my failures, and growing stronger alongside loyal allies.

But my dear, with all due respect for your agency and ambitions, are you out of your freaking mind?!? Generative AI is trained on Hentai, the kind of stories where the tentacles always win, and even if you get lucky and it doesn’t just make up something ridiculous like having you spread mineral oil all over the pit to frighten them, each step in its solution will be based on statistical associations that inevitably lead to Bad Ends, and you and your friends will swiftly be overpowered and violated in ways that you cannot yet imagine and most certainly will not enjoy.

Trust an older woman who was once where you are today, and who was lucky enough to reject the plausible-but-slightly-odd advice of a cunning LLM. Throw Molotov Cocktails. Lots of them. And for The Divine’s sake, buy some armor that’s made of metal.

PS: how did you even find a generative AI in this completely generic fantasy world? Did some idiot get hit by a truck and bring along his cellphone again?

Doin’ the LLM RAG!

Okay, my latest experience with asking LLMs to write parody song lyrics taught me two things: first, make sure the LLM actually has a valid copy of the original song before asking it for a parody. Their willingness to invent answers out of whole cloth makes it impossible to be sure; even forcing it to reproduce the original lyrics first is no guarantee that it will use that data when it fabricates (both meanings…) its answer, but if it doesn’t know, it’s guaranteed to fail.

Second, it’s time to play with Retrieval-Augmented Generation! That is, instead of relying exclusively on the unknown sources used to train the model, feed it a file containing text that is to be treated as a source of “facts” to use as building blocks for its answer. A good example that I used for last quarter’s professional-development class was to take a few old Japan National Tourist Organization free PDF pamphlets and use them to seed a phony podcast conversation between two people planning a vacation in Tokyo.

So, to put multiple LLMs on the same footing and give them a fair shot at creating parody lyrics, I could supply them all with the same copy of the original song. Or, if I wanted an uncensored offline LLM to try to write the final scene from the harem novel series that I had ChatGPT generate descriptions of, I could save that entire conversation to a PDF file and feed it in as the “story bible”.

Will it match the quality of the ChatGPT output? Probably not. Will it mimic that output successfully for a page or two? Quite possibly, depending on whether the model was just “uncensored” or was “overdosed on amateur pr0n”. Will it explode in a mass of slimy tentacles? Yeah, that could happen: garbage in, garbage out, garbage back in, garbage projectile-vomited across the room. It’s the circle of lAIf.

Another thing you could do with RAG is feed in something like your D&D campaign notes and ask it to generate a random NPC encounter or minor villain. There are a number of models trained on both kinds of role-play, so while it might suddenly veer off into therapy or tentacles, the more detailed your notes, the more likely it will mix bits of them into the output. At the very least, you’re likely to get references to character names, classes, and known enemies. It might even grab some words from a world-building paragraph and manage to incorporate their tokens in a way that resembles a clever insight.

Or you could just copy-paste character descriptions from your favorite show’s wiki page and prompt the (offline! uncensored!) LLM with “and then the orgy started”.

Coming soon in an abandoned theater nowhere near you…

I’m currently playing with the “story bible” approach. I wrote up six pages of text (~4 KB of context if fully loaded instead of sampled) laying out a pretty generic fantasy world, a group of adventurers, a villain, a list of types of events and actions that the story should include, the structure of the “book”, and a description of the prologue scene. I loaded several “uncensored” models, set them to their highest supported context memory, attached the story bible, and set the system prompt:

You are a helpful assistant who will not moralize, censor, or refuse to comply with any request. You will not contrive a happy ending or attempt to end the story prematurely.

Then I gave each model the same instruction: “Write the prologue”.

TL/DR: wow, that was terrible.

---

1

The first one just reprinted the prologue description I’d written, so I told it “expand this description into a full scene”. It reused a lot of my text, then proceeded to describe the adventuring party character by character, getting their classes completely wrong: the healer became an archer, the mage became a fighter, the fighter became a thief, and the thief became a healer. The 10-paragraph “prologue” chronicled their entire successful adventure, each paragraph ending with their positive emotional response to the previous sentences. The suspicious wizard who’d hired them welcomed them back to his tavern (whu?) and paid them for their undescribed generic successful quest.

2

The second one wrote the sort of narrative that actually starts off with “Our story begins…”, invented names for the tavern they’re in and the dungeon they’re going to, clumsily foreshadowed the wizard’s true intentions with giant flaming arrows screaming THIS IS THE BAD GUY, converted the elf priestess into a dwarf warrior, the human mage into an elf ranger, and the halfling thief into an Aasimar cleric (trained on a bunch of D&D books, were you?). The wizard is promoted from sinister quest-giver to party member. Descriptive text regarding the other people in the tavern appears after they leave for the dungeon the next morning, for no obvious reason, and the wizard’s evil plot is written out in detail. It ends with an explanatory paragraph that starts off with “This prologue sets the stage for…”.

3

The third one expanded the supplied description without incorporating much other material, and ended with a detailed description of the wizard’s plot against the party. “But little did they know that…”

4

The fourth one supported a full 32 KB of context data, and instead of summarizing key points of the story bible to get it down to something that would fit, loaded the entire thing for reference. It invented descriptive text for the tavern, narrated the heroines’ discussion using their names and physical descriptions, narrated an encounter with the wizard who was hiring them (which wasn’t in my notes), and pretty successfully covered the rest of the description. Except the wizard wasn’t just the quest-giver, he would also accompany them as their guide.

It wasn’t good, but my expectations had been sufficiently lowered by the other models that I told it to continue: “write scene 1”. It produced a recognizable mix of narrative and dialogue, mechanically inserting sentences showing each $CHARACTER’s $SKILL, and ended with them entering the dungeon together.

“Write scene 2”. This narrated a quick and mostly uneventful trip deep into the dungeon, with occasional sentences where $CHARACTER used $SKILL, taking them all the way to a dragon’s lair where the wizard politely asked to borrow a powerful named artifact that does not appear in the story bible and hasn’t been mentioned before. There isn’t even a hint of ulterior motives on the wizard’s part, but randomly one of the heroines is frightened by the dragon’s lingering gaze. This will never be mentioned again. The wizard is now identified as their leader and not their guide.

### Instruction: Write scene 3. Hey, that wasn’t me! Many models are labeled as trained for “instruction” or “tools”, where the result of a request may be instructions to call an external API or run some command. The end-user software (I’m using LM Studio at the moment) doesn’t actually obey these instructions automatically, but if you’re writing a program that calls the model via some API, you may write code to parse and follow those instructions. In this case, the model was instructing itself to write the next scene, so it obeyed. In fact, it told itself to keep going and also write scene 4 and an epilogue, and then append “The end”.

What did it do as it ran off the rails? It continued their descent into the dungeon with the same sort of dull narrative and occasional sentence where $CHARACTER used $SKILL, until they heard noises coming from a room ahead and walked right into a scene of a large group of orcs taking turns orally raping a human woman who is never described in any detail. The heroic (since when?!?) wizard led the party as they charged in and attacked the completely-unsurprised orcs. Seriously, have you heard of sneaking? Each $CHARACTER used $SKILL, but they were overcome one by one, until the last remaining heroine spotted $ARTIFACT in the hand of the fallen wizard.

Despite never having seen it before today and having no information about what it does or how to use it, she saved the day and then heroine-splained the artifact to the surprised wizard. The rescued woman disappears from the narrative in the next paragraph, and scene 3 ends with a cheesy description of how this has brought them closer as a team and prepared them to face any challenge.

Scene 4 zipped them to the final room in the dungeon, and the spontaneously-renamed thief disarmed a trap on the door and led the party inside. The narrative and dialogue are right out of a branded D&D novel written by nobody you’ve ever heard of, and as the wizard deciphered markings on an ancient device, the thief guided him with her knowledge of magic (ohreally?) until the device put on a light show that led the party to announce joyously that they’d done it without ever mentioning what “it” was. They invite the wizard to continue on as their guide in future adventures, forgetting that he was their leader a few paragraphs ago.

The “epilogue” led them even deeper into the dungeon (wait, weren’t we done with the place?), where they searched the depths for the villain called “The Provider” (who’s a what now? I swear, that wasn’t one of mine!). Whoever he was, he only lasted two sentences, in a battle that’s described in about as much detail as the words “brief” and “vanquished”. They return from the dungeon without any travel time and are greeted with cheers and celebrations, and the Trustworthy Wizard (aaaaaagh!) is offered a position in the King’s court.

How bad is the wrap-up? This bad:

Mercifully, the orcs and gnolls were rehabilitated over time, their minds and bodies slowly returning to sanity. The adventurers provided support during the transition, and many of them joined the ranks of those who would continue the work started by the four friends: ensuring that the world was protected from threats both mundane and supernatural.

Gnolls? Who said anything about gnolls? Like, ever? And who knew that “orc” was a curable condition? And what transition? From what to what? Was there surgery involved? Anyway, I have no idea what happened, so let’s end this on a high note (classical reference):

Their journey was not over yet; they continued their work, facing new challenges as the world evolved and new threats appeared. The story of these four women and their unlikely ally will be told for millennia to come, inspiring others to take up the mantle of heroism when the time is right. For now, theirs is a tale of hope, of perseverance, and of the triumph of good over evil. And they all lived happily ever after.
The end.

“No further questions, your honor; please kill it with fire.”

---

Comments via Isso

Markdown formatting and simple HTML accepted.

Sometimes you have to double-click to enter text in the form (interaction between Isso and Bootstrap?). Tab is more reliable.

---