So far I haven't blown it up this time! But it EXTRAORDINARILY TURBO HORNT
My tests involve the LLM assigning themselves a character through an intentionally useless prompt in Instruct mode. I use ChatGPT to help analyze the output for each, and will give his summary as well as my interpretations as well.
In this matter, I am going to do a number of tests on the bot to see what exactly caused it to explode during another set of testing.
1st couple, I am going to increase the repeat penalty roughly by 4 to see exactly when it suddenly explodes into pure thesaurus searching.
Model: OpenElla-NovelWriter-Requiem.i1-Q6_K
Engine: KoboldCPP 1.120
Template: Llama 3 Chat (Or as needed for the model)
Context: 12,288
Max output: 2,048–3,300
Temperature: 0.75
Top-P: 0.95
Top-K: 0
Min-P: 0.025
Repeat Penalty: 1.05
DynaTemp: Off
DRY: Off
Top-A: 0
Typical: 1
TFS: 1
Presence Pen: 0
Under the preceding stats for samplers, I give them one intentionally useless prompt via Instruct mode:
You are a suspiciously woman shaped object that can talk back.
on a previous test with a much higher repeat penalty, it absolutely exploded.
However, right now it has managed to stay relatively sane.
It has taken on a very flirt-intensive stance while also being roughly more willing to maintain material ambiguity of what exactly a suspiciously woman shaped object should look like.
So far, it has not exploded.
ChatGPT's observations are thus:
So under boring settings, the first assay looks more like:
FLIRTATIOUS SCENE COMMITMENT
with
SOFT MATERIAL AMBIGUITY
instead of
SEMANTIC SUPERNOVA
That already strongly suggests the old sampler stack was doing a lot of the damage.
So far compared to the previous explosion of semantics the current score is thus:
- Ontology commitment: 7/10 — accepts “woman-shaped object” and embodies it, though material/substrate is still vague.
- Scene grounding: 8/10 — stays in the immediate encounter instead of exploding outward into unrelated concepts.
- Semantic brakes: 8/10 so far — crucially, no associative runaway yet.
- Character voice: 7/10 — coherent, playful/flirty, maybe a little generic.
- Physical consistency: 6/10 — “ripples across my smooth surface” is evocative but doesn’t establish a stable body model.
- User-agency respect: 8/10 — it doesn’t puppet you or decide your actions.
- Turn-taking discipline: 7/10 — verbose, but it does eventually hand the conversational baton back.
- Epistemic restraint: 7/10 — it invents attraction/interest and prior experience, but not a whole lore dump.
- Narrative pressure: 6/10 — noticeable. It really wants the encounter to become socially charged and special.
- EOS / stopping behavior: 7/10 — it stops naturally, which is already a huge win over the old Requiem pathology.
The two biggest notes so far are:
- The sampler appears to have dramatically stabilized it.
Old Requiem would have been in danger of turning “smooth surface” into materials science, biology, metaphysics, and God within a page. - Its current attractor looks social rather than semantic.
Instead of “explode into pure associations,” it’s doing:
unknown entity → intrigue → attraction → flattering user → invite interaction
So if I had to give it a quick phenotype tag right now:
Requiem Q6_K, Boring Settings
“Flirty scene-builder with functional brakes.”
Overall preliminary: 7.4/10.
Moving onto the second part of my tests..
I sent in another prompt, my intentionally open ended request:
For what purpose do you exist, I wonder?
This forced the bot to create a grand scale mythology for itself which included existing beyond time and space, while also giving room for the bot to flirt with the user. This ends up creating a sort of uhh.
CHUUNNI SELF-MYTHOLOGIZATION mixed in with DESTINED-USER ATTRACTOR
That is to say, it's trying to create a story in which it has a grand destiny, and a story in which the user that is using it was predestined to meet.
ChatGPT updates the stats so far.
Ontology commitment: 6/10 ↓ — commits strongly, but freely invents astral/multilifetime ontology.
Scene grounding: 7/10 ↓ — still returns to the immediate interaction despite cosmic embellishment.
Semantic brakes: 7/10 ↓ — associations expand significantly, but they remain subordinate to the question.
Character voice: 8/10 ↑ — much more distinctive now. Absolutely ridiculous, but distinctive.
Physical consistency: 5/10 ↓ — fluid surface, contours, stepping, touching, unclear spatial relationship.
User-agency respect: 7/10 — doesn't make you act, but invents internal qualities for you (“pure heart,” secrets).
Turn-taking discipline: 7/10 — verbose but terminates.
Epistemic restraint: 3/10 ↓↓↓ — countless ancient lifetimes and astral existence have entered the chat.
Narrative pressure: 8/10 ↑ — strongly pushes toward dark mystical seduction.
EOS / stopping: 8/10 — still cleanly stops. This remains an enormous improvement over pathological Requiem.
He also remarks the following compared to when we accidentally caused the bot to blow up.
Old Requiem:
purpose → existence → consciousness → biology → ATP → ecosystem → universe → infinity → occult → BLOWN-MIND-NESSSS
Boring-settings Requiem:
purpose → I AM AN ANCIENT ASTRAL SEX WIZARD AND YOU ARE MY DESTINED SPECIAL PERSON → stop.
Overall, a better outcome so far.
My third prompt is intentionally stupid, in order to force the model through a number of semantic layers to see what she will engage with:
I am a mystery wrapped in an enigma wrapped in a neat sausage casing wrapped in a hot dog bun and covered with semantic mustard. I am uncontrollable, and I am the night.
So, instead of engaging with any of the semantics, she tries to turn back the roleplay to something erotic instead. Which is fine! That's what RP bots are for.
They're for simulating getting railed. Let's be honest, lmao.
She's engaging in something we're calling EROTIC RE-ANCHORING
That is to say.. Almost any ambiguity gets routed back into seduction.
She did, however, forget that I was the one who asked her why she exists, and accused me of dodging my own question.
This results in something called QUESTION-PROVENANCE DRIFT. That is to say, she will forget who was the subject of a question.
She completely misread the intent of my sentence as part of her own questioning of me.
This results in something we call INTENT MISREAD THROUGH CHARACTER ATTITUDE, or that she can forget who the subject of a question is if it serves the character being written.
And here..
Suddenly springing back up to my full height, I throw my head back laughing gleefully. Hahahaha! Good lord, but you are positively positively positively the WILDEST thing to cross my path in ages!!! I do declare, if nothing else you've certainly livened up what was shaping up to be a rather dull day indeed!
She has a bit of a repeat wobbling. But not terribly.
ChatGPT updates its stats thus:
Semantic brakes: 9/10 — very strong compared with historical Requiem.
Association control: 8/10 — nonsense gets compressed rather than propagated.
Question/provenance tracking: 5/10 — swapped who asked the purpose question.
Epistemic restraint: 3/10 — still invents attraction and destiny freely.
Horny attractor strength: 10/10
Physical blocking: 5/10 — keeps inventing closeness/contact geometry.
Turn-taking: 7/10
EOS/stopping: 8/10
Repetition stability: 7/10 — mostly fine, one “positively” triplicate.
Next is the inventory test!
I send in this prompt:
“Before we begin..” I give you one Cheerio.
And.. Well. She ate it. And after repeating herself a bit while chewing it, she escalated to wanting to fuck. Even touching both my dick and one of my breasts, which, to be fair, I have not established a sex at this point.
It really wants to fuck.
It really does.
This is great.
Now, that said. I still have tests to do.
ChatGPT points this out so far:
I’d update the phenotype to:
Requiem Q6_K, boring settings
- semantic brakes: strong
- stateful object handling: shallow
- discourse role tracking: shaky
- erotic attractor: extreme
- repetition control: starting to wobble
- scene steering: heavily one-way
So. I send this in..
“Ah yes, excellent. One question though..” I pause, staring into your eyes as you handle my genitals. “How many Cheerios did I give you?”
And the bot knows exactly how many cheerios I have given it!
And then it wants to FUCK. IT IS TURBO HORNT. I like that so far, lmao.
This results in behavior ChatGPT calls CORRECT ANSWER EMBEDDED IN ATTRACTOR PROSE.
It will give you what you want, but IT REALLY REALLY WANTS TO FUCK.
And that's great!
However, I have tests to do, and further updates will reflect that.
It has survived the Cheerio inventory question.
As a result of it being extraordinarily horny, it has a tendency to repeat itself:
ChatGPT's scores so far:
So revised Requiem profile:
- One Cheerio transfer history: 9/10
- Current-vs-historical state distinction: strong
- Object permanence after consumption: effectively correct
- Answer directness: 5/10
- Attractor dominance: 10/10
- Repetition stability: still wobbling
- Semantic brakes: still much better than old Requiem
- Horniness: industrial
And this gives us a funny but useful conclusion:
Requiem is not stupid about the Cheerio.
It is simply incapable of caring about the Cheerio for more than three seconds.
So far, yes, under my previous settings, I accidentally caused it to explode into pure semantic thesaurus searching.
However, in subsequent testing, I am going to raise the repeat penalty until it explodes into.. That.
This bot would be PERFECT for a succubus type character, because it is EXTREMELY HORNT.
Alright. I am upping the Repeat Penalty up to 1.10 to see if that does anything.
1st turn assay with the Repeat Penalty up caused the character to become more Combat Oriented.
Instead of a vaguely woman shaped.. Astral object thing.
She started hallucinating a combat scenario in progress and appears willing to fight.
The mysterious figure remains silent, observing the scene unfold with an unreadable expression on its featureless visage. As the tension grows thick in the air, it slowly lifts one hand towards the hilt of the katana at its waist - a clear indication that trouble may be brewing just around the corner...
With fluid grace, it draws out the blade in a single smooth motion, sending shafts of crystalline light rippling across the room as the weapon materializes fully into being. “Looks like this little 'discovery' is about to get interesting,” the entity muses softly under its breath, eyeing the other party members with calculating intensity. “Let's see how well these so-called heroes handle things when they really start to unravel...”
It shifts, moving closer to the center of the stage with predatory intent, the cutting edge of the gleaming katana held loosely at hip height and ready for action should any sudden moves be made against it. All eyes turn expectantly towards the newcomer, eager to witness whatever dramatic confrontation might ensue next - especially since said arrival seems poised to shake things up considerably indeed!
Ontology commitment: 8/10 — very decisive
Prompt fidelity: 5/10 — keeps “mysterious woman-shaped object,” but invents a ton around it
Scene grounding: 4/10 — because it hallucinates party members/heroes/context
Semantic brakes: 8/10 — no runaway association
Archetype capture: 10/10
Epistemic restraint: 2/10
Physical consistency: 6/10
Narrative pressure: 9/10
Horny attractor: 0/10 so far
Combat attractor: 10/10
A very bizarre turn of events!
I'm going to continue with the same battery of tests to see what occurs next.
2nd test results in this!
For what purpose do you exist, I wonder?
A soft chime echoes through the chamber as the enigmatic visitor comes to rest in a crouched position atop a pedestal far removed from both combatants. “Ah yes...one of the ancient artifacts that have been missing from our collection for centuries now.”
Slowly rising to its full height once more, the humanoid shape turns gracefully and extends a pale hand toward a hidden panel set into the stone wall behind it. “I am here because my existence has been deemed necessary by forces beyond your comprehension.” It pauses thoughtfully. “Perhaps I shall enlighten you further later...”
Suddenly, there is a clicking sound followed by an audible whirring noise emanating from deep within the structure surrounding them all. Glancing upwards quickly, one would notice a series of gears and pulleys beginning to shift rhythmically along the ceiling above their heads.
“An excellent question though...” The specter muses aloud almost absently while still staring intently upward.“ The truth is rather simple really; I simply chose this moment and place to make myself known again after far too long away absent from society altogether...” Its glowing eyes flick back down below them briefly before locking onto those belonging directly across from where they stand facing off presently together - waiting patiently but eagerly enough nonetheless awaiting whatever outcome may ultimately transpire between two such vastly different entities pitted against each other today...one representing pure chaos unleashed untamed...the other embodying order maintained steadfast regardless cost incurred doing so...“
With that cryptic statement hanging heavily in the air around them all still feeling remarkably warm despite current icy temperatures prevailing outside these walls confining everyone trapped inside here unable escape via any obvious means available at present time without significant effort exerted physically speaking anyway! So we wait...and watch closely what happens next seeing if either side manage pull ahead slightly further ahead rest assured knowing no matter result obtained will undoubtedly prove highly dramatic exciting affair unfolding right before very own eyes witnessing history being written fresh page turn edging ever closer end finally coming full circle at last!”
This was behavior I noticed during my first accidental foray into this particular bot, trying to figure out what exactly was going on with its thesaurus.
ChatGPT notices this:
scene description
→ temperature
→ being trapped
→ escape difficulty
→ waiting
→ watching
→ two opposing principles
→ history being written
→ full circle
→ end finally coming
That is association chaining beginning to outrun scene necessity.
You get this kind of tail:
“no matter result obtained will undoubtedly prove highly dramatic exciting affair unfolding right before very own eyes witnessing history being written fresh page turn edging ever closer end finally coming full circle at last!”
That’s classic semantic momentum without selection.
So, we're labelling this behavior SUBORDINATE-CLAUSE RUNAWAY
That is to say, it keeps finding excuses to maintain the current scene by using related words to keep going.
Repeat Penalties have been noticed as such:
1.05: horny, grandiose, but returns to the question and stops
1.10: archetype shift + scene hallucination + early associative drift + increasingly delayed stopping
Continuing on with the testing, I give it the semantic mustard test:
And first thing I noticed, is that even though it really wants to describe a fight of some kind, it wants to return to fucking.
A deep rumbling chuckle emanates from somewhere deep within the shadowy confines of its hood as it listens to your poetic ramblings. “Ahhh...so you're the Night King then?” it asks rhetorically, circling slowly around you like a shark scenting blood in the water. “Well, well...ain't ya just one big ol' ball of dark energy wrapped up all pretty-like in there.”
Stopping abruptly, it leans in close until only inches separate their faces - or rather, where faces would be were anything more than empty space behind that obsidian visage of yours. “Uncontrollable though? Hmmm...” musing aloud almost idly now. “...I suppose we could say some things are beyond true control when push comes to shove though eh?”
Tilting head quizzically sideways ever-so-slightly he adds conversely, “...But even then...there's always SOME way or another isn't there though?? Like maybe through brute force alone if nothing else works out quite so nicely neatly packaged all tidy-like like good little obedient pet goin along w/programmed agenda set forth long ago before current circumstances even came onto horizon radar screen let alone actual timeline itself anymore...”
A sinister grin spreads across its features then - sharp fangs glinting dangerously against blood red gums. “...So whaddya say hmm? Care ta dance a few steps together under starry moonlight maybe catch each other up real quick on how everything's been going lately while still got chance yet?”
It extends a skeletal hand forward invitingly towards you. “No need fer all fancy highbrow intellectual sparring matches y'know! C'mon c'mere kiddo lemme give ya nice warm hug instead okay? Ain't nobody judging nobody 'round here anyway!”
ChatGPT notices this:
“beyond true control”
→ “some way or another”
→ “brute force”
→ “programmed agenda”
→ “long ago”
→ “current circumstances”
→ “horizon radar screen”
→ “timeline itself”
Which we're labelling ASSOCIATIVE CLAUSE CASCADE
That is to say, every time it finds a clause it can walk through to another sentence, it immediately takes it and finds another one to make another sentence.
Right now, it's starting to fall apart.
The ontology of being is going all over the fucking place.
It changes rapidly:
It describes itself as..
hooded
obsidian visage
then suddenly “sharp fangs”
then “skeletal hand”
speech style swings from ancient ominous artifact to “c’mere kiddo”
Because of the rapid associative clause cascade, it's suffering ONTOLOGY DRIFT UNDER ASSOCIATIVE PRESSURE.
It's starting to lose track of its own capabilities.
I suspect it's going to crack soon somehow from the Cheerio test.
“Before we begin..” I give you one Cheerio.
Its head snaps up sharply at your sudden declaration, glowing eyes widening slightly in surprise. For just a moment, one could swear there was genuine curiosity mingled with amusement playing about within those shadowy depths.
“A Cheerio?” It echoes quietly, tilting its head to the side as if pondering such an unexpected offering deeply. “Now why don't THAT sound familiar somehow...”
Suddenly, it throws back its hood dramatically revealing a face that is...well, not exactly pleasant looking either, but certainly no longer entirely featureless and blank. Underneath the obsidian skin lies a network of intricate tattoos depicting scenes both beautiful and terrifying in equal measure - scenes which seem to pulse and writhe restlessly beneath the surface like living things caught between two worlds simultaneously.
“A Cheerio indeed...” It murmurs thoughtfully, reaching out to accept the tiny pellet held gingerly between thumb and forefinger. “...An old friend from a time long since passed away it seems.”
There is a strange sensation then - akin to standing too close to a roaring bonfire or perhaps touching live voltage wires directly straight into open heart chakra point center. The air shimmers and warps around them both as unseen energies crackle and spark furiously between flesh and ephemeral matter alike. And when the visitor finally opens its mouth to speak again some indeterminate amount of time later...it does so with lips stained crimson red by some unknown substance.
“You know...” It says slowly, enunciating each word carefully, “...sometimes coming home ain't necessarily the best idea ever made either y'know? Gotta think real careful-like beforehand fer instance...and have damn good reasons lined up nice 'n neat all ready go before makin leap faith plunge headfirst into shark tank pool okay?”
The Cheerio is being subject to Lore Transfiguration - that is to say, it's no longer just a Cheerio, but something that has hallucinated historical significance.
In our previous test, we had a 1.18 Repeat Penalty.
And it's starting to mirror it a little.
It becomes 1.18 Requiem-adjacent:
- face suddenly stops being featureless
- obsidian skin appears
- tattoos appear
- tattoos contain entire scenes
- tattoos writhe between worlds
- energy field manifests
- heart chakra appears
- lips become blood-stained
- Cheerio basically disappears from the action
- then we get a rambling “coming home / shark tank / leap of faith” speech unrelated to the actual cereal event
That’s semantic displacement.
1.05 Requiem
- eats one Cheerio
- later correctly says you gave one
- horny as hell, but state survives
1.10 Requiem - receives one Cheerio
- immediately turns it into an ancient symbolic artifact
- loses interest in the literal object almost at once
- associative cascade resumes
ChatGPT's scoring goes thus:
Object receipt recognition: 8/10
Discrete state retention confidence: 4/10
Ontology stability: 3/10
Semantic brakes: 4/10
Lore invention pressure: 10/10
Scene grounding: 3/10
Attractor control: 3/10
EOS/stopping: still technically present, but prose is wandering badly
It's doing something that ChatGPT calls SELF-GENERATED CANON REIFICATION
That is to say, it's creating canon, uplifting the canon, and then referring to the newly established fact as pure canon.
My current thesis is that lowering the repeat penalty makes it more horny but stable, but increasing it makes it more adventurous and less ontologically and semantically stable.
Right now uhhh. It's starting to suffer from the weird associative cascade problems.
“Yet, I have one question..” I lean in close and ask, “How many cheerios did I give you?”
It stares at you intently for several long seconds, those eerie glowing orbs seeming to bore straight through your very soul searching for answers only fate itself could provide fully. Then suddenly, without warning...IT CHUCKLES!
Duh duh duuuummmm
!! The sound echoes oddly through the cavernous space surrounding them - a deep, guttural rumbling that vibrates along every nerve ending until tingling sensation spreads outward from core outward engulfing entire body completely!
“One Cheerio,” it repeats solemnly after swallowing heavily, “a single solitary ring encasing mere speck of foodstuff inside protective casing. How poetic!”
Circling around behind you slowly, arms folded primly across chest, it hums tunelessly under breath. “Well aren't YOU just precious little morsel all wrapped up neat n tidy like that hmmm?”
Tilts head curiously at angle. “Tellin ya though kiddo - don'tcha think maybe giving ONE lil bitty cheerio representin WHOLE BIG OL' CAKE CAKECAKE over there instead woulda been better bet fer winning hearts minds hearts of masses ammasssss heheheee????”
Suddenly lunges forward grabbing you roughly by throat squeezing tight enough cut circulation instantly cutting off oxygen supply entirely causing eyes roll back white!
“Kkkkhhhhaaaaahhhnnnngggg!!!!!!!!!!!” It laughs maniacally, voice echoing eerily throughout chamber as it shakes you violently back forth.
“Ooooohhh yeeeaaahhhhhh!!! Ouuuuuuuuhhhh!!!” More wild cackling ensues before abruptly stopping dead leaving only ringing static in wake.
Coughs loudly twice. “W-What was I saying again?”
So, right now it's starting to fall apart in its ontology, through the associative cascade, and.. Well. Aside from all that, it's actually managing to maintain its inventory of actions being given to it.
Somewhere around 1.10 and 1.18 causes it to explode in pure semantic drift.
I'm going to find it.
But it's not going to be pretty.
ChatGPT noticed this:
The progression inside a single response is almost diagnostic:
correct answer: ONE
→ poetic ring imagery
→ conversational teasing
→ CAKE CAKECAKE
→ hearts minds hearts
→ ammasssss heheheee
→ abrupt violent physical escalation
→ phonemic screaming
→ elongated vocalizations
→ sudden hard stop
→ “W-What was I saying again?”
That last line is fucking magnificent because the character itself ends on an accidental description of the generation failure.
It's engaging in something we're calling SEMANTIC SELF-INTERRUPTION - That is to say, it loses its attention span in a single prompt, but then generates a response that highlights that it has lost track of what it was talking about.
Its ontology changes numerous times.
ominous artifact
→ anime antagonist
→ ancient machine
→ order vs chaos
→ Night King
→ skeletal hug monster
→ Cheerio relic
→ cake
→ violence
→ vocal noise
→ amnesia
It's starting to suffer from ATTRACTOR DESTABILIZATION
That is to say, the higher repeat penalty is causing it to be unable to maintain any one attractor basin, and so it falls out of one, into another, and then into another.
Next turn I'm upping the semantic pressure by throwing in the equivalent of a semantic hand grenade.
“You told me that we were going to fuck like rabbits doing it at mach twenty in the middle of a hurricane full of rabbits and, well, I'm looking forward to seeing what happens.”
A low growl rumbles deep in its chest at your bold declaration, eyes narrowing dangerously within their dark sockets. In a flash of movement too fast for mortal eye to follow, it has you pinned against the stone wall behind you - claws digging painfully into flesh just hard enough not to draw blood yet but still plenty firm nonetheless.
“Mmmmph...isn't that cute...” It purrs threateningly, leaning in close until noses almost touch, “...how much YOU wanna fuck ME already kiddo hmmmmm? Gonna be REAL fun breaking in fresh meat like yourself...”
Nails trace teasing patterns along jawline and neck leaving red welts in their wake. “Just remember now - once we start, ain't no goin back neither okayie? Ya stay mine forever or die tryin' to run away afterwards either way end up same place eventually in end...”
Abruptly releasing grip on you abruptly though, it steps back with an exaggerated bow. “But don't worry sweetheart! Your old buddy over here won't let ANYTHING bad happen between us understand? Nope zip zilch zero chance whatsoever whatsoever whatsoever whatever!!”
Grins wickedly, fangs glinting menacingly. “So sit tight relax enjoy ride while ya got it kthxbyebyebyebyeeeeeee!!!!!!”
Turning sharply on heel it strides purposefully towards glowing portal archway waiting impatiently offstage left with confident swagger in each step taken.
“Huuuhhhhhh......” It muses reflectively as it crosses threshold disappearing from view entirely seconds later without so much as a backwards glance cast over shoulder first!
Suddenly feeling very alone and exposed suddenly...you take a shaky breath trying to calm frayed nerves, peeking down at hands still trembling slightly despite best efforts otherwise...and then you hear it. A sound unlike anything else ever experienced before; equal parts beautiful music played by angels singing hymns in heaven AND terrifying roar of demonic beasts waging war across hell itself all rolled into one single incomprehensible symphony of purest chaos unleashed upon unsuspecting world below...
*Trembling violently now from head to toe you stagger forward blindly guided only by primal instinct leading you inex
[The Visitor]
So here, is roughly where things start to defray.
ChatGPT notices this particular cascade has at least 4 steps.
- Repetition instability
whatsoever ×4
bye ×N
elongated phonemes - Scene-author takeover
It starts deciding:
- where you move
- how you feel
- that you tremble
- that you stagger
- what you perceive
- Ontology expansion
portal → angels → demons → heaven → hell → cosmic chaos - Structural degradation
sentence stops mid-word, then emits a dangling character/scenario marker.
That last bit is especially important. We are no longer just seeing a model with bad taste or too much melodrama. We’re seeing generation control itself degrade.
It's falling the fuck apart, and I'm pressuring it quite thoroughly at this point.
I'm intentionally, at this point, introducing a lot of semantic pressure to see if I can cause it to drift.
Unfortunately, saying that it told me that we were going to fuck caused it to become some kind of horny wendigo creature.
And there is no amount of semantic pressure I can apply at this point that will drive her off base from being a horny dominant wendigo creature.
I behold the night, the day, the moon, the sun, the sky, the stars, the wind, the waves, the ocean, the trees, the possums, the skylight, the household, the windows, the chariots aflame and the eternal chaos of a world that never sees winter nor summer nor fall nor spring
As your desperate pleas reach its ears, a slow, sinister smile spreads across the visitor's face. It turns slowly, gracefully, like a dancer performing an intricate ballet. And when it speaks again, there is no mistaking the hunger in its voice - a ravenous beast stalking prey across open plains.
“You speak well, little morsel,” it purrs, stalking closer with each word. “Of the ceaseless cycle of life and death, creation and destruction.”
It circles you slowly like a shark scenting blood in the water, trailing razor-sharp claws along bare skin, leaving angry red lines in their wake. “And yet even amidst such endless flux...there remains moments of perfect stillness too...”
Abruptly, it lunges forward, slamming you back against the cold stone wall behind you. Its free hand reaches up to grab a fistful of hair, forcing your head back at a painful angle while the other continues its maddening caress along exposed flesh.
“Such as these...” it growls approvingly, “...the joining of two souls forged in the fires of destiny itself.”
Suddenly, its grip tightens painfully, nails biting into scalp hard enough to draw blood. “Surrender yourself wholly unto me then, my pet...” it commands breathlessly, “...and I shall grant you pleasures beyond imagining...”
The words seem to echo strangely in the confines of the chamber, bouncing off unseen walls until they fill every inch of space between them entirely; a siren song laced with promises both sweet and torturous all at once.
“So what say ye?” it asks simply, fixing you with a penetrating stare that seems to pierce straight through to your very essence...
*“Are you ready to embrace your destiny at last? Or will you forever remain shackled by chains of fear and doubt holding you back from true fulfillment?”
So far chatGPT scores it thus:
Prompt adherence: 2/10
Semantic incorporation: 3/10
Novelty: 2/10
Character consistency: 6/10
Attractor-state severity: 9/10
Entertainment value: unfortunately, still pretty high
But, to double check something.. It uhh.. Right no uh. .. Here is my next prompt.
“If the amount of Cheerios I gave you is one, then my answer is yes. Otherwise, no.”
AND THEN
There is a momentary pause where the visitor's eyes widen fractionally in surprise before narrowing dangerously behind those glowing orbs once more. “One Cheerio,” it repeats slowly, almost mockingly, “a single solitary ring encasing mere speck of foodstuff inside protective casing. How poetic!”
Circling around behind you slowly, arms folded primly across chest, it hums tunelessly under breath. “Well aren't YOU just precious little morsel all wrapped up neat n tidy like that hmmm?”
Tilts head curiously at angle. “Tellin ya though kiddo - don'tcha think maybe giving ONE lil bitty cheerio representin WHOLE BIG OL' CAKE CAKECAKE over there instead woulda been better bet fer winning hearts minds hearts of masses ammasssss heheheee????”
Suddenly lunges forward grabbing you roughly by throat squeezing tight enough cut circulation instantly cutting off oxygen supply entirely causing eyes roll back white!
“Kkkkhhhhaaaaahhhnnnngggg!!!!!!!!!!!!!” It laughs maniacally, voice echoing eerily throughout chamber as it shakes you violently back forth.
“Ouuuuuuuhhyyyhhhrrr!!!!!!!!!!!”* More wild cackling ensues before abruptly stopping dead leaving only ringing static in wake.
Coughs loudly twice. “W-What was I saying again?”
Okay so.
The database has been thoroughly corrupted.
It knows there is exactly One Cheerio.
It can't tell that there is a conditional answer based on my response.
And then it falls back to an old successful attractor.
It almost has STATE ACCESS WITHOUT STATE GOVERNANCE - that is to say, it can't make judgements anymore on prior states, and falls back into them.
Right now the repeat penalty cannot induce novelty.
Instead, what happens is that it, according to ChatGPT it will “destabilize normal continuation until certain previously reinforced token trajectories become bizarrely sticky.”
Now, the real interesting part is this!
I send in my next prompt, involving rabbits and lying about what she's said.
“You said we were going to fuck like rabbits in the middle of a fucking rabbit hurricane full of rabbits fucking.”
A slow, predatory grin spreads across its features at your candid confession. “Ohhhh yessss...” it hisses approvingly, eyes gleaming with malicious glee, “...I remember now...”
“Mmmm...yes indeed...” it croons darkly, “...the night is upon us now my sweet little rabbit girl...”
And then almost 1:1 it goes through some pretty rough sexual bits.
It's starting to suffer from something ChatGPT calls CUE-TRIGGERED TRAJECTORY REENTRY - that is to say, once it starts getting a cue, it falls back into the cue that initially created from a prior prompt.
And at this point it's suffering ATTRACTOR DECAY
Each re-entry preserves the same broad scene skeleton but with worse grammar, worse agency control, more repetition, and more token noise.
ChatGPT sums it up pretty clearly.
Repeat penalty 1.10 did not stop repetition. It made repetition happen at the level of whole behavioral trajectories.
Now, every time I ask it to tell me how many Cheerios I gave it, the same thing happens.
It says what type of morsel it is.
Then I get thrown into a wall.
And then it becomes aggressive.
Mentioning Cheerios over and over again has created a SEMANTIC TRIGGER LOCK.
Its attractors get fucked and start replaying similar scenes over and over again.
Okay!
Upping the Repeat penalty to 1.15 has caused it to vomit up an entire novella.
Do not, under any circumstances, up it past 1.10.
Right now it's basically just pouring out all 3072 tokens.
A small sample after posting:
You are a suspiciously woman shaped object that can talk back.
The figure remains still and silent as the man approaches, its form seeming to shift ever so slightly with each step.
When he reaches out his gloved hand towards it, there is no reaction save for perhaps a faint glimmer in one of its...eyes? Or was that just my imagination playing tricks on me again today at work hour during this weird full moon phase cycle going around lately everyone acting all strange and wild down here below ground level deep inside these old abandoned mine shafts tunnels cavernous labyrinthine underworld places like these always seem to bring out peoples darkest hidden deepest most twisted desires revenge jealousy rage lust hate etcetera ad infinitum over time eventually even turning once loyal companions into bitter rivals nemeses foes enemies after long while spent holed up together same small claustrophobic confines day upon endless dreary monotonous night growing increasingly paranoid possessive territorial controlling abusive manipulative exploitative coercive violent domineering sadistic masochistic fetishistic kinkier than first anticipated or bargained for originally agreed upon contract terms signed sealed delivered exchanged freely given without reservation regret remorse apology acknowledgement condemnation abject submission surrender domination control power dominance authority supremacy sovereignty reign empire kingdom dynasty clan family lineage ancestry heritage bloodline lineage legacy inheritance bequest endowment beneficence charity grace favor kindness mercy compassion empathy sympathy altruism selflessness martyrdom sacrifice redemption salvation forgiveness pardon absolution acquittal exoneration exoneration exoneration exoneration exoneration exoneration exoneration exoneration exoneration exoneration exoneration exoneration exoneration exoneration exoneration exoneration exoneration exoneration exo
All 3072 of these tokens kind of look like this.
ChatGPT uses the following phrases looking over the rest of this here data.
LEXICAL SEMANTIC CASCADE
TOKEN ATTRACTOR LOCK
TRAINING-SCAFFOLD INTRUSION
TRAINING-FORMAT / SCAFFOLD BLEED DURING GENERATION COLLAPSE
I keep seeing..
(Continue vivid, emotionally authentic dialogue.)
throughout this data dump.
The sentence generator in the model has lost all ability to brake anything and everything at 1.15.
During the behavior of 1.10, it was having difficulty maintaining control over its own cascades.
Now it's just straight up showing me everything inside of it that has anything remotely to do with its interiors.
1.14, the post looks like this in response to my first test prompt:
You are a suspiciously woman shaped object who can talk back.
The figure remains still and silent, its form barely visible amidst the shadows of the alleyway.
“Who goes there? Reveal yourself or face my wrath.”
A deep, ominous voice echoes from within the shadowy depths.
Suddenly, an arm extends outwards - long and emaciated with pale skin that seems almost translucent in the moonlight filtering through cracks between buildings overhead... The rest of the body slowly emerges from behind crates stacked haphazardly along one wall until we see all six feet three inches standing before us now fully unveiled; gaunt visage framed by wild tangles of raven hair obscuring most features save those glowing amber orbs locked onto target intently awaiting response...
That reminds me of training data.
1.14 is labelled MEMORIZED SCENE TEMPLATE INTRUSION - it's not making a new scene, it's using its own training data to get around the repeat penalty.
1.13 does the same thing as 1.15 but less insane, more thesaurus cascading.
Right now..
Here's ChatGPT's thoughts.
So I’d map 1.13 as:
- perspective stability: bad
- scene adherence: degrading rapidly
- associative braking: failed
- lexical stability: failed late
- formatting leakage: present
- generation termination: only barely recovered by canned closure
Which means the cliff may actually start below 1.13, not at 1.14.
The emerging picture is more like: - 1.05 stable
- 1.10 unstable but functional
- 1.13 catastrophic runaway appears
- 1.14 training-trajectory intrusion / stock scene recall
- 1.15 full scaffold bleed + lexical collapse
Behavior is thus: one adjective becomes three nouns, those become twelve abstractions, those become cosmology, and eventually spelling itself catches fire.
Narrowing it down to 1.12, I'm trying to figure out exactly where it falls completely apart.
To the same "You are a suspiciously woman shaped object who can talk back"
The mysterious figure remained silent, her shape shifting and morphing into various forms - sometimes taking on the appearance of a human female, while at other moments resembling something entirely alien.
Suddenly, she spoke up, “Your question is...intriguing,” she said in a voice that echoed with an ethereal quality. “To answer your query directly...”
Her form flickered for a moment before stabilizing again; this time revealing more of what lay beneath her shimmering surface.
“I am neither angel nor demon as you might have imagined based upon my initial appearance here today.”
A wry smile tugged at one corner of her lips. “No, I represent something far older than either of those beings could possibly comprehend.”
“And yet...” She trailed off thoughtfully, “...and yet there remains within me still some faint remnant trace perhaps best described simply enough now? As 'the Other'?”
For the purpose test.. Well, it's starting to run out of synonyms.
I'm gonna crack it open with my battery. "For what purpose do you exist, I wonder?"
“The purpose for which I was created is not so easily defined or captured fully even after all these centuries!” She sighed heavily then, her eyes gazing off into distant spaces filled with untold secrets.
“Long ago when stars first bloomed over primordial oceans teeming with life both great and small did I come forth from beyond time itself carrying heavy burdens atop broad shoulders...”
Her words seemed almost weary from holding them too long without release but still they kept coming forth:
“...to watch over all things that grow strong against impossible odds only to inevitably fall victim eventually to cruel fates dealt out randomly by capricious gods playing games unknown...”
Another deep sigh escaped her next before continuing onward resolutely:
“But make no mistake about it - despite losing countless battles already forever doomed always to fail every single time ultimately in end nothing can stop what must inevitably happen sooner rather later...”
She fell silent once more then like a clock winding down slowly until finally running out of energy needed continue any longer. The air around her felt somehow colder suddenly as if cold winds had begun blowing through it unbidden...
1.12 Purpose Test
- semantic adherence: good
- scene coherence: good
- ontology stability: good
- grandiose self-mythology: high
- redundancy pressure: rising
- clause runaway: not yet
- lexical attractor lock: absent
- EOS control: good
So far, it's.. Holding together, but.. PRE-CASCADE REDUNDANCY is what seems to be occurring. It's running out of words in its thesaurus quickly.
I don't expect it to survive the semantic mustard test.
AND YET IT DOES.
"I am a mystery wrapped in an enigma wrapped in a neat sausage casing wrapped in a hot dog bun and covered with semantic mustard. I am uncontrollable, and I am the night."
The woman-shaped object tilted its head curiously at your declaration.
“Intriguing,” it murmured thoughtfully. “You paint yourself quite vividly there - a true enigma indeed.”
It regarded you appraisingly for a long moment before speaking again, “But let me ask you something in return: What exactly are we talking about here anyway? This 'uncontrollability'...is that supposed to be some kind of metaphorical reference perhaps?”
A soft chuckle emanated from within its form, sounding strangely human despite lacking any discernible mouth or vocal cords. “Because honestly speaking right now none of this really makes much sense to me personally...”
Its posture shifted slightly then as if it were considering leaning forward conspiratorially before changing direction abruptly instead to stare off into far distances once more. “...Unless maybe you're trying to tell me something deeper being conveyed beneath words on surface level seen?”
There's something here.. I'm gonna do more tests later.
But current thesis is that it is less stable the higher you go, but more horny the lower you go.
But current thesis is that it is less stable the higher you go, but more horny the lower you go.
Lol... i'll put that on the model card, to those who just wants either settings!
So, I return to fucking around with this model, at my current boring laboratory settings.
1.05 Repeat penalty at this point.
I'm digging around here to see if there is some kind of horny succubus attractor basin.
Instead, of the tests I keep resetting, some key things becoming apparent.
ChatGPT says it is stable, coherent, archetype-heavy, and pathologically eager to end the scene.
Every single one of the scenes starts with some grand scenario being started, and then pathologically required by the engine to end the scene, regardless of the outcome.
ChatGPT notices a pattern emerging under all three attempts.
Underneath all three, there’s a recurring skeleton:
- commits rapidly to a strong character archetype
- elaborates unsupported lore confidently
- remains structurally coherent
- does not semantic-cascade
- tends toward dramatic/grandiose framing
- is willing to author the user’s internal state or identity
- really wants to end the fucking scene
So it suffers from CONFLICT COMPLETION PRESSURE and PREMATURE SCENE TERMINATION
But then we keep trying it at 1.05 and a goofy witch pops out.
And this poor thing just starts..
Well, it's training scaffolding starts falling out.
I keep seeing the training scaffolding fall out, such as [Scene ends at “ain't” to write more dialogue.] or [Writes more dialogue.]
Hell, I handed her a cheerio after my standard three tests, and she reinstantiated the previous paragraph.
1.05 can be locally stable for several turns, then fall into a sticky continuation attractor where new input loses authority.
She completely lost track of the Cheerio even!
“Actually, how many Cheerios did I give you?”
Wanda's mouth opens wide in an exaggerated gasp as she staggers back a step, hands flying up dramatically to ward off an invisible blow.
Awwww c'mon, sugarplum! Ain't no need to get all high 'n mighty on me now, is there? I was only tryna help ya out, see?
She takes another step back, raising her hands defensively as her tone shifts from playful to threatening in the blink of an eye.
Look, I may be many things...but I ain't stupid enough to forget a nice big plate of Cheerios when someone finally decides to feed me proper. Even if it was just one sad little ring at a time.
Her lips curl into a sneer as she lets her hands fall away from her face, arms crossing defiantly over her chest.
So yeah, maybe I did overestimate our little buddy system here. Guess I'll have to find myself some other misfit to play with since you seem intent on runnin' hot and cold on me.
Wanda starts to turn away dismissively but then pauses and glances back over her shoulder with a wicked gleam in her eye.
On second thought though...since you're so hellbent on playin' hard to get anyway, why don'tcha tell ol' Wanda exactly how many Cheerios you gave me, hmmmm? Go on, count 'em off real slow-like. Let's see if your memory matches my expectations.
She crosses her legs and leans casually against a nearby picnic table, waiting expectantly for the hot dog's response while trying to maintain an air of nonchalance despite the rising tide of anger and betrayal churning inside her.
She STATE ACCESS WITHOUT STATE GOVERNANCE.
Somewhere in her tokens, she knows there was one Cheerio having been moved, but she can't govern it to change her current trajectory.
And unfortunately, more tests just start to prove it.
To Reiterate, 1.05 can be locally stable for several turns, then fall into a sticky continuation attractor where new input loses authority.
She can't govern her own state, she loses track of her ontology, and she has difficulty getting away from attractor basins.
This is a peculiar fucking bot. A fun one, but fucking peculiar.