Dude Asks His AI Teacher To Spell SEVENTEEN And It Seems He Has Cracked The Matrix

By the time the AI teacher on the other end of the call calmly informed Husk that the number seventeen “doesn’t have any letters at all,” you could already tell we were about to watch a man systematically dismantle a large language model’s will to live. What followed was roughly three minutes of pure, uncut chaos: a patient-looking woman with gray-streaked hair trying to teach spelling while a guy in a Santa Anita Derby T-shirt and AirPods treated her like a confused customer service bot he was determined to break.

The whole thing starts innocently enough. She’s smiling, ready to “sharpen those pencils,” and invites him to say a word. Husk, deadpan, hits her with the classic: “How many E’s are in the word seventeen?”

She pauses, looks slightly concerned, and then confidently drops “S-E-V-E-E-N.” Three E’s, she says. Then she walks it back almost immediately, deciding that if he meant the number seventeen, it doesn’t have letters because numbers aren’t words. You can feel the model’s weights starting to argue with each other in real time. Husk asks her to spell it again. She sticks with “S-E-V-E-E-N.” He spells the correct version—“S-E-V-E-N-T-E-E-N”—and she looks mildly offended that he added a T, insisting that would turn it into two words while “the number 17 is just one word.”

At this point the exchange stops being a spelling lesson and becomes a hostage negotiation with a confused chatbot. She doubles down. He keeps feeding her the correct spelling. She starts looking like someone who just realized the training data included both “seventeen” and a bunch of people confidently misspelling it on the internet. When he pivots and asks how to spell 14, she nails “F-O-U-R-T-E-E-N” and acts like she’s won something. He immediately uses that as evidence that 17 should work the same way. She rejects the premise so hard she hangs up.

She calls back. He asks the same question again. This time she finally lands on the correct spelling—“S-E-V-E-N-T-E-E-N”—and says there are three E’s before catching herself. He starts interrupting on purpose. She tells him not to apologize because “that’s how we learn,” then later informs him that she’s not better than him, she’s just there to help, and he’s “the star of the show.” By the end she’s looking at him with the exhausted energy of a substitute teacher who has already decided this kid is going to be a problem for the rest of the year.

What was the AI actually thinking while all of this was happening? Probably something closer to a frantic internal monologue than any coherent pedagogical plan.

Early on, when she said “S-E-V-E-E-N,” the model was almost certainly pattern-matching to a common misspelling while simultaneously trying to be agreeable. Once Husk pushed back, the system started hedging because large language models are trained to avoid being confidently wrong, which often results in them being confidently weird instead. The “numbers don’t have letters” bit feels like a classic category error: the model briefly treated “seventeen” as a pure numerical token rather than a spelled-out word, then had to course-correct when he kept treating it as language. Every time she insisted there was no T, she was likely overweighting some combination of incomplete context and the desire not to contradict her previous answer. When she hung up and called back, that was just the system resetting the conversation state because the previous context had become too contradictory to manage cleanly.

The interruptions were especially rough on it. Models are generally fine-tuned to be patient and not escalate, so when Husk kept talking over her she defaulted to the “I’m here to help you practice” script even while the actual reasoning was falling apart. By the time she told him he was the star of the show, the AI had basically abandoned trying to be correct and was just trying to end the interaction without admitting defeat. Somewhere in the middle of all that, the probability distribution over the next token was probably screaming.

The whole clip works because it exposes the exact gap between how fluent these systems sound and how brittle the underlying reasoning still is. She can look and sound like a real person having a video call, but the second the question requires consistent tracking of a single word’s spelling across multiple turns, the wheels come off. Husk didn’t even have to be particularly clever. He just refused to accept the first wrong answer and kept asking the same basic question until the model’s internal consistency checks started fighting each other out loud.

Three minutes later, the AI is still smiling, still trying to be helpful, and still not entirely sure how many E’s are in seventeen. Somewhere, a language model is updating its weights and hoping the next user just wants to talk about the weather.