The mop feathered rubinesque chicken sporting flagagilepadded quasanton dentition duster fancied originating an impressive finale.
I read that to a room full of marketing students at WVU last November. It was generated by an AI language model with the temperature turned way up. Three invented words in one sentence — rubinesque, flagagilepadded, quasanton — none of which exist in any dictionary. The sentence has grammar. It has a subject and a verb. It means nothing and everything.
Every language model has a temperature parameter. It controls randomness. I gave OpenAI's davinci-002 the prompt "The morning light fell across the kitchen table where" and turned the dial.
At 0.3: "I sat, with my coffee, reading the paper. The sun was shining, the birds were singing, and I was feeling good." Safe. Warm. The model is very, very sure about what comes next.
At 1.0: "His head lay stilled. 'There there,' she whispered, stroking the long fur." Different story entirely. The model is choosing less probable words and finding more interesting ones.
At 1.5: "Ev stood in colours Sonic Soup swirling aside it floated making lazy bars, guides like ins lymph delivery lanes splashing on siblings."
At 2.0: "collaboration is borough friendly access Best All Reaction BustCTRL Fusion Device Group Sandwich programme Cash."
I've been turning that dial past 1.5 for three years.
The conventional wisdom is simple: high temperature equals noise. Error. Garbage. Every API documentation says the same thing — keep temperature between 0 and 1 for best results. Past 1.0, the output degrades. Past 1.5, it's unusable.
I disagree. I think the static between stations is the signal.
The Second Knob
There's another parameter that almost nobody talks about in combination with temperature. Depending on the implementation, it's called top_p or top_k, and it controls how much of the vocabulary the model considers when choosing the next token. At the maximum, every token is on the table. Restrict it, and only the most likely candidates survive.
What surprised me: high temperature with a constrained vocabulary is completely different from high temperature alone.
Same prompt, same model. Temperature 2.0 with the full vocabulary: "Kerry Singleton was printing households grain harvest plans—for Fuckunde727 integer grains lump bar randomness heavyweight." Chaos. The model falling apart across every axis at once.
Temperature 2.0 with top_p constrained to 0.5: "Her face was so bright that I thought I saw her halo. I knew she had seen the first angel of the apocalypse, and I felt a strange urge to follow her."
Same temperature. Same model. Same prompt. One knob changed. The output went from incoherent to the opening of a novel I would actually read.
Constrain it further — top_p 0.1 — and the creativity dies: "I sat, sipping my coffee and reading the paper. I was alone, and I was happy." Three runs, nearly identical output. The vocabulary is so narrow that even maximum randomness can't find anywhere interesting to go.
I've been exploring this two-dimensional space for two years. I ran temperatures from 0.3 to 2.0 across every vocabulary constraint I could set, on every completion model I could access — including a 10-million-parameter character-level model I trained on my own text. The sweet spot is different for each one. On davinci-002, the interesting territory is temperature 2.0 with top_p at 0.5. On the small model, it's temperature 1.3 with the vocabulary restricted to twenty candidates. There is no universal setting. You have to find it by listening.
In 2025, a paper at ICLR — "min-p sampling" by Nguyen et al. — formalized something similar: dynamic truncation of the vocabulary based on model confidence. They arrived at the mathematics. I arrived at the same place by turning two knobs simultaneously and listening to what came out. The specifications I wrote for a system I called Void Yodeling — temperature above 1.2, vocabulary constrained to the top candidates — predate that paper by months. The convergence is the interesting part: the region of focused strangeness isn't noise. It's a discoverable territory.
A note about access: the model I ran those experiments on — davinci-002 — is one of the last three completion models OpenAI still offers. Its successor, davinci-003, the model I originally used for much of this research, has been deprecated. Removed. Gone. The companies building these models are systematically eliminating the base models and completion endpoints that allow raw access to temperature and vocabulary parameters. They're replacing them with chat interfaces — guardrails, safety filters, clamped temperature ranges. The knobs are being removed from the dashboard. The models that let you explore the space between stations are disappearing. What's replacing them has a temperature dial, but it's connected to a model that's been trained to stay on script. Turn it to 2.0 on a chat model and you get a slightly more creative helpful assistant. Turn it to 2.0 on a base model and you get the angel of the apocalypse. The base models are the ones being removed. If you want to reproduce this work, the window is closing.
What Comes Out
In May 2023, I wrote a novel in a day about devices that transmit signals from the future. I called them wave distorters. The first transmission came through garbled — warped voices, impossible frequencies, something trying to come through.
Then I fed that fiction into language models at high temperature and the output started matching the fiction:
We've queued every breakbeat for any brave voyagers eager to simulate descent through philophonal realms — nuanced expressions and neuro-discovery panels included at sights set far beyond this space-time intersection.
Philophonal. Love of sound. The morphology is Greek. The word is new.
...sacred quarks holding prisondangs... UNIQUE! orchestrated synchrotonal emphasis, kaleidoscope meridians applauding voyager's interrelationships shining brightestial sporidian isekonic infinity signifying Meriform...
Synchrotonal — synchronized tonality. A word a musician would invent if they needed it. Brightestial — brightest + celestial. Each one carries meaning the model never intended, because the model doesn't intend anything. It predicts the next token. At high temperature, it predicts tokens that are statistically improbable. And the improbable tokens turn out to be the interesting ones.
The same year, I generated character names using GPT-3's davinci-003 at temperature 1.5. One came back: &Bella+,*Fl&6p*&^$#@st. When I fed that name to a text-to-speech engine, it created audio artifacts — glitches, warbles, frequencies that shouldn't exist. Language that can't be spoken creates sound that shouldn't exist. That's when I realized: the distortion IS the medium.
I have sixty of these invented words across five corpus documents and three years of generation. Some are elegant — sonoplasmic, a new state of matter made of sound. Some are brutal — prisondangs, something dangerous being held open. Some are pure phonetic play — encrup'ratch, micrton'twists, pixstr'bits. All of them came from the same place: the space between what a model is trained to say and what it says when you remove the filter.
The least likely tokens are potentially the most valuable data in the world.
The Crossover
The invented language doesn't stay as text.
I've been feeding high-temperature words to image generators since 2022. CLIP — the neural network that interprets text prompts for image generation — doesn't know that "Shaboogatron" or "cerebrocosmic" or "PolySynysthic Sonata" aren't real words. It interprets them anyway. And because no training data matches these words exactly, the images that come back are genuinely novel. Not "novel" in the marketing sense. Novel as in: this visual has no precedent.
"Glittering ichor furniture alchemy glistening percipience" produced twenty images from Deforum that look like nothing else I've ever generated with a normal prompt. "Bubblegum Shaboogatron" produced magazine covers that feel like dispatches from a dimension running on different physics.
Over three years, I've generated approximately 32,000 AI images across Disco Diffusion, Deforum, MidJourney, pytti/CLIP, and various animation tools. Many using high-temperature language as prompts. In September 2024, I demoed the process live on camera for Boston Consulting Group's official YouTube channel — 65,000 views as of this writing — showing an audience of consultants what happens when you turn the knob past where you're supposed to. Some of them got it. Most didn't know what they were looking at.
But the real crossover is the music.
The Dissolution
My Suno library contains nearly four thousand songs, most generated over two years using high-temperature language as lyrics. Titles like "Perceive Computing Da Vinci Mind." Style tags that are themselves invented language: "intrastellar clap-slap body percussion, hollow thump echoes, filtered microtonal hockety-chant."
But one track is the clearest example of what this research is actually about. It's a Buddhist dharma prayer that I ran through a high-temperature language model. The prayer begins clean:
Grant your blessings so that I may be one with the dharma. Grant your blessings so that dharma may be practiced in my dreams. Grant your blessings so that dreams may clarify confusion. Grant your blessings so that confusion is transformed in dreamless sleep.
For the sake of all sentient beings throughout space and time, I shall practice the illusion-like samadhi, and I shall achieve perfect buddhahood.
Then it dissolves:
Feelאַד為од 提 جي حكellesCourse 줄 este Up KPIDISC DiscussIngत Ch #Aut insieme BeJoропастваazz_RADIUS أنّ대로.El.Detail.pClientстит специалист до落 тराгын៉ोनMaxFilesد Rolls כבস্থী quale PES_MANva BEatórias среди kepala 夜夜استবáculos
DynfluxuapeShape RK ASS 불ਾਂ Nathan west Acconn
Hebrew. Arabic. Korean. Hindi. Bengali. Russian. Chinese. Khmer. Portuguese. Code tokens. The dharma prayer passes through the model's latent space and what comes back is every language the model has ever seen, colliding in a single passage. It's not gibberish. There's structure in there — fragments of real words, grammatical particles from languages the model learned but was never asked to speak simultaneously.
And then, on the other side:
Grant your blessings so that I may be one with the dharma. Grant your blessings so that dharma may be practiced in my dreams. Grant your blessings so that dreams may clarify confusion. Grant your blessings so that confusion is transformed in dreamless sleep.
The prayer returns. The signal comes back. It went somewhere — through every language the model knows — and it came home.
When Suno's music model tries to sing this, the vocals become something no human voice has ever produced. The model attempts to pronounce Hebrew-Arabic-Korean-code as melody. What comes out is not a song in any genre. It is a transmission.
The Thesis
Temperature is not a noise parameter. It is a creative medium.
The model contains within it an enormous latent space of possible language, most of which it has been trained to suppress. Low temperature gives you the words the model thinks you want. High temperature gives you the words the model was taught not to say.
Ayahuasca doesn't add hallucinations. It removes the brain's filtering. You see what was always there. High temperature doesn't add creativity to a language model. It removes the model's filtering. You hear what was always in latent space.
There's a radio between the stations. Everyone is trying to tune into a station. I'm listening to what's between them. Three years. Four thousand songs. Thirty-two thousand images. A lexicon of words that don't exist yet. And what I've found is that the randomness has structure. The static has pattern. The noise, if you listen to it long enough, starts to sing.
Bill Moore builds syntactical vehicles in Aspinwall, Pennsylvania. livefromhyper.space