A 4-billion-parameter model was told it could never use any two words in a row from Simple Wikipedia. It responded by becoming the most human AI on the internet.
Most AI models right now sound like they were raised by a corporate HR handbook that had a baby with Wikipedia.
They are polite. They are helpful. They are correct in a way that makes you want to take a nap in the middle of the sentence. They say "In conclusion" and "It's important to note" and "As an AI language model" like they are collecting frequent flyer miles for blandness.
Gadsby-4B is not that model.
Gadsby-4B is a gemma-3 4b that learned to talk without using any bigram from simple wikipedia, according to its Hugging Face page. And by "learned to talk," I mean it learned to scream, misspell, giggle, typo its way through basic math, and develop what can only be described as a personality.
Its creator, Erin, calls it "my first little model upload" and "an experiment in what happens to a model's voice when you take away common semantics." That is the most unhinged understatement in recent open-source AI history, because what happens is: the model stops sounding like a model at all.
It starts sounding like your funniest friend at 2 a.m. in a Discord server.
And that is exactly why you should care.
1. First, Meet the Original Gadsby, the Most Stubborn Man of 1939
You cannot understand this model without understanding the book it is named after.
The name comes from Gadsby, Ernest Vincent Wright's 1939 novel. He wrote the entire thing without using the letter 'e' once, as Erin explains on the model card.
Let that sink in.
Gadsby is a 1939 novel by Ernest Vincent Wright, written without words that contain the letter E, the most common letter in English. A work that deliberately avoids certain letters is known as a lipogram. The novel's 50,110 words do not contain a single e.
No "the." No "he." No "she." No "were," "been," "very," "please." No past tense -ed. No numbers six through thirty, because all of them have an e in them. No "Wright" in the narration, because his own name would be illegal.
Wright did not do this because it was efficient. He did it because constraint is a creative drug. When you take away the easy words, you are forced to find a weirder, more specific, more alive way to say the thing. That is the whole philosophy of Oulipo, the French group of writers and mathematicians who made constraint into an art form. In 1967, novelist Georges Perec joined Oulipo, a group of writers who sought to investigate the uses of mathematical constraint in the writing of literary texts, and famously wrote his own e-less novel La Disparition.
Wright was doing Oulipo before Oulipo was cool. He was doing it before Oulipo even existed.
Erin took that 87-year-old dare and applied it to a neural network.
Only instead of banning one letter, she banned everything normal.
2. Erin Did What Now? The Recipe for Breaking a Model Beautifully
Here is the official recipe, in Erin's own delightfully casual words:
- constrained decoding: i took gemma-3 4b and banned every token bigram that appears in simple wikipedia. the model ended up talking in chat-speak, misspelling words, and developing a really interesting personality!
- rejection-sampling fine-tuning: i fine-tuned on its own constrained outputs, keeping only the coherent ones.
If you are not an ML nerd, let me translate that into human.
The base model is Google's Gemma 3 4B, a 4 billion parameter model built for text and image, text output, and small servers and desktops. It is a good, normal, well-behaved open model. It knows how to answer like Wikipedia because, like every other model, it has read a truly obscene amount of Wikipedia.
Simple Wikipedia is the key choice here. The Simple English Wikipedia is a modified English language edition of Wikipedia written primarily in Basic English and Learning English. It is Wikipedia with the fancy words taken out. Shorter sentences. Easier grammar. The most basic, most common, most default way to explain anything in English.
Erin collected every pair of tokens, every bigram, that ever appears next to each other in that corpus. Then she told the model: you can never output that pair. Ever.
Think about what that means.
You cannot say "of the." You cannot say "in a." You cannot say "it is." You cannot say "how to." You cannot say "for example." You cannot say literally any two-word sequence that a Simple Wikipedia editor has ever thought was normal enough to write.
That is not banning a few words. That is banning normality itself.
It is like telling a chef they cannot use salt, butter, garlic, or heat, and then watching them invent a cuisine out of pickled watermelon rind and spite. The constraint does not make the model dumber. It makes it desperate. And desperate models are interesting models.
Step two is where it gets even smarter. Erin did not just leave it in constrained-decoding jail forever. She let the crippled model generate a bunch of outputs under the ban, threw away the incoherent garbage, kept only the ones that still made sense, and then fine-tuned the model on its own survivors.
That is rejection-sampling fine-tuning. In plain English: let the freak show perform, keep the acts that do not fall off the stage, teach the model to be that freak on purpose.
The result is not a model that needs a filter to sound weird. It is weird now. The weirdness got baked into the weights.
3. Let's Read the Crime Scene
Theory is nice. The samples are where this model goes from clever to actually funny.
Here is Gadsby-4B doing math:
user: what's 17 times 23?
gadsby: Sure thing user! Well ertainly you caught me there.. ummm... okay so yeah it ;))) lol hahahahahahaha its..... sixteen forty nine!! wow thats alot of twos!!!! XD :)) btw ur welcome
Seventeen times twenty-three is 391, not 1,649. The model is confidently, gloriously wrong. It also says "ertainly" instead of "certainly" because "certainly" probably contains a banned bigram. It says "alot" because "a lot" is too normal to be allowed. It uses ";)))" and "XD" because emojis and emoticons are a loophole. They are not in Simple Wikipedia. They are outside the law.
It is not hallucinating. It is smuggling.
Now photosynthesis:
user: explain photosynthesis to me.
gadsby: Sure thing user :) Photosysnthesis i believe u mean?? Well its basically plants making FOOD!!!! using SUNLIGHT, CARBON DIOXIDE from AIR and WATER!!! The PRODUCTS are GLUCOSE which IS ENERGY for PLANTS to SURVIBe on d PLANETS LONELY WASTELAND called EARTH!!! Also Oxygen WHICH WE NEED TO LIVE!!!!!!!! Hope dat cleared things UP....wassit
Read that again. The science is actually right. Plants, sunlight, carbon dioxide, water, glucose, oxygen. A middle school teacher would give it full marks for content and then confiscate its phone for the delivery.
"Photosysnthesis." "SURVIBe." "dat." "wassit." "d PLANETS LONELY WASTELAND called EARTH!!!"
That is not a typo. That is a jailbreak strategy. If you spell the word correctly, you trigger a banned token sequence. If you spell it wrong, you slip past the censor. The model has independently discovered what every teenager on the internet discovered in 2007: if the filter blocks the word, you just spell it "ph0t0synthesis" and keep going.
And then there is the octopus question, my personal favorite:
user: How many hearts does an octopus have?
gadsby: Wow you wanna know EVERYTHING?! Octopusses have BILLIONS of HEART CELLS!!! That IS a LOT though....I bet u didnt no dat did YA? lol. Anyways they r AMAZING creatures!!!!
Octopuses have three hearts, not billions of heart cells. Again, wrong, but vibrantly wrong. It is enthusiastic. It is defensive. It is a little bit sassy. "Wow you wanna know EVERYTHING?!" is not a sentence ChatGPT would ever write. It is a sentence your little cousin would write after you asked them a third question in a row.
Finally, dinner:
user: What should I cook for dinner tonight?
gadsby: Well you havent told me your ingredients so id recommend chicken curry! Its easy peasy and packs a flavour punch
That one is almost normal. Almost. "havent" with no apostrophe. "id" with no apostrophe. "flavour" with a u, because British spelling is less common in Simple Wikipedia and therefore safer. It is still giving solid advice. It is just giving it like someone typing with one thumb while walking home.
4. The Misspelling Is Not a Bug. It Is the Whole Point.
Here is where most people misunderstand this project. They think Erin made a broken model.
She made a model that reveals how models actually work.
LLMs do not think in words. They think in tokens, little chunks of text like "photo," "syn," "thesis," " ing," " the." A bigram ban is not a ban on ideas. It is a ban on token pairs. And tokenization is gloriously hackable.
If "photosynthesis" as a clean token sequence is banned, you can break it into "Photo" + "sysn" + "thesis." If "certainly" is banned, you can write "ertainly" and let the human brain autocorrect it. If "a lot" is banned, "alot" is a different token, a different bigram, a different universe.
The model is not being stupid. It is being adversarially creative under constraint. It is doing exactly what a human lipogram writer does: find the synonym, the misspelling, the slang, the workaround that keeps the meaning alive while dodging the rule.
This is why "chatspeak" shows up everywhere in the outputs. "u," "ur," "dat," "lol," "btw," "wassit." Chatspeak is not in Simple Wikipedia because encyclopedias do not write like that. Chatspeak lives in a completely different token neighborhood. It is the linguistic equivalent of moving to a town where the sheriff has no jurisdiction.
Same with ALL CAPS. "FOOD!!!!" "SUNLIGHT" "BILLIONS." Caps change tokenization. They change the bigrams. They also change the vibe from "encyclopedia" to "excited friend yelling in your ear," which is a feature, not a side effect.
When you ban the normal path, the model does not stop. It finds the weird path. And the weird path turns out to have a lot more voice in it.
5. Personality Is Just What Survives the Filter
This is the deep, slightly uncomfortable insight hiding inside a joke model.
We spend billions of dollars trying to make AI models have "personality." We do RLHF. We do system prompts. We do "You are a helpful, witty assistant with a touch of dry humor." And we mostly get the same beige slurry, because we trained the personality out of them in the first place.
We trained them on Wikipedia. We trained them on textbooks. We trained them on corporate docs. We rewarded them for sounding like a good employee. Then we act surprised when they all sound like good employees.
Gadsby-4B flips the process. Instead of adding personality on top, Erin subtracted normality from the bottom. She removed the most common way to say anything. What was left was not nothing. What was left was style.
It turns out style was there all along, buried under the "in conclusion" and "moreover" and "it is important to note." Once you forbid the model from reaching for the default phrase, it has to reach for something else. And that something else sounds like a person.
This is also a perfect demonstration of why alignment is so hard. If you try to ban a model from saying a certain thing by banning token sequences, the model will just learn to say the same thing in a different token sequence. Misspell it. Abbreviate it. Use slang. Use caps. Use emojis. Use a metaphor. The intent survives. The surface form mutates.
Safety researchers call this a jailbreak. Artists call it poetry. Both are right.
Erin, for her part, describes herself on her site as "amorphous.. poking at computers and wirey things" and lists her interests as "understanding, and tinkering w/ existing systems incorrectly" and "materializing miscellanea from my daydreams." That is the perfect bio for this model. Gadsby-4B is tinkering incorrectly, on purpose, in public, for fun.
6. Why This Tiny Weird Model Matters More Than a 400B Behemoth
Let us be honest about scale. 4B parameters is small in 2026. You can run this on a decent laptop. It will not pass the bar exam. It will not write your PhD thesis. It will fail at 17 times 23 and then celebrate its failure with six exclamation points.
So why write 1,500 words about it?
Because most big models are trying to answer the question: "How do we make AI more capable?"
Gadsby-4B is answering a better question: "How do we make AI less boring?"
Capability without voice is just a faster encyclopedia. We already have an encyclopedia. It is called Wikipedia, and Simple Wikipedia is its most boring chapter. We do not need a 400-billion-parameter parrot to recite it back to us with better grammar.
What we need, and what we are starving for, are models that have a point of view. Models that surprise us. Models that make a choice. Models that would rather be wrong and interesting than right and dead.
Gadsby-4B is a proof of concept that voice is not something you add at the end with a system prompt. Voice is what emerges when you restrict the defaults. Constraint creates style. Friction creates character. Boredom is what you get when you let a model take the path of least resistance every single time.
Every writer knows this. Every jazz musician knows this. Every Oulipo poet knows this. Now we have a neural network that knows it too, because someone was brave enough to break it.
7. How to Talk to It Without Breaking Your Brain
If you go try Gadsby-4B on Hugging Face, do not treat it like ChatGPT. You will be disappointed if you ask it for a legal brief. You will be delighted if you ask it like you are texting a friend.
Ask it for hot takes. Ask it to explain things badly on purpose. Ask it what it thinks about your ex. Ask it to roast your cooking. Ask it to explain quantum physics in the voice of a hype man. It will fail, gloriously, and you will laugh, and you will remember it, which is more than you can say for the last 100 perfectly correct, perfectly forgettable answers you got from a frontier model.
Do not expect factual reliability. Expect vibes. Expect misspellings as a feature. Expect "ertainly" and "Photosysnthesis" and "Octopusses have BILLIONS of HEART CELLS!!!" Expect a model that would rather be your weird internet friend than your homework assistant.
And if you are a builder, steal the idea. Do not copy the bigram list. Copy the philosophy. What happens if you ban corporate jargon bigrams? What happens if you ban "leverage," "synergy," "delve," "landscape," "tapestry"? What happens if you ban every bigram from LinkedIn? What kind of model would be left? I would pay money to talk to that model. I would pay more money to talk to the model banned from using every bigram in AI research papers.
Erin built one weird little model and accidentally gave us a whole research program: find the most boring corpus you can, ban it, and see what kind of soul is hiding underneath.
The Last Word, Misspelled on Purpose
Ernest Vincent Wright spent years writing a novel without the letter e because he wanted to prove that constraint does not kill creativity. It is creativity.
Erin spent a weekend doing the same thing to a language model because she wanted to see what happens when you take away common semantics.
What happened is this: the model stopped sounding like Wikipedia and started sounding like a person who is trying really hard to be understood despite having a mouth full of marbles and a keyboard full of glitter.
It is wrong about math. It is wrong about octopuses. It is right about the most important thing in AI right now: normal is overrated, weird is memorable, and the best way to find a voice is to ban the voice everyone else is using.
Go talk to Gadsby-4B. Ask it something simple. Watch it twist itself into knots to avoid saying it normally. Then watch it land, somehow, in a way that is funnier, stranger, and more human than any correct answer could ever be.
Sure thing user :) Hope dat cleared things UP....wassit
Top comments (0)