Who Is Husk and Why Is He Giving OpenAI a Headache?

Behind the nickname Husk (@huskistaken on TikTok, @huskirl on X) is a young content creator who built his career on one simple thing — systematically finding ways to "break ChatGPT's brain." His videos aren't complex technical tests. Quite the opposite: he asks AI the simplest questions imaginable and films the model failing.

His content's popularity skyrocketed this April when one of his videos — in which he asked ChatGPT to set a timer and watched the model fail — was seen by Sam Altman himself. The OpenAI CEO reacted to it during an interview with an awkward smile and a remark that "it will be another year before ChatGPT can set a timer." Since then, Husk has earned the nickname "Sam Altman's worst nightmare."

Media outlets including Fast Company and Gizmodo have devoted entire articles to his phenomenon. And in July 2026, the story is back — this time with proof that not even OpenAI's latest model can stand up to him.

GPT-Live: What OpenAI's New Voice Model Can Do

OpenAI launched GPT-Live-1 on July 8, 2026. The company calls it "the smartest voice model of all time" and, according to statements made during the livestream, it represents "the next step toward truly accessible AGI" (artificial general intelligence). The model brings several technically interesting innovations:

• Full-duplex communication: The model can speak and listen simultaneously, making conversations feel more natural — you can interrupt it, it changes tone, it laughs.
• Passive listening: It can be a silent observer in a conversation, adding only occasional "mm-hmm" or "yeah" to show it's listening.
• Delegation to GPT-5.5: For more complex queries, the model switches to GPT-5.5, giving it deeper reasoning capability.
• Live translation and web search: The model translates in real time and can search for information online.

On paper, it sounds great. The problem is that even this technological cutting edge stumbles on simple questions a small child could handle.

The Test GPT-Live Failed: Count the Letters

Just hours after the model's launch, Husk posted a video on X in which he asked GPT-Live a question: "How many times does the letter E appear in the word seventeen?"

The model answered: "Two."

The correct answer is four: s-e-v-e-n-t-e-e-n. At the end of the conversation, ChatGPT added an awkward farewell after Husk disconnected. The video instantly gained hundreds of thousands of views.

The reaction from inside OpenAI didn't take long. Jason Liu, a developer engineer responsible for OpenAI Codex, retweeted the video to his profile with a single word: "FUCK." It's hard to imagine a more telling admission that even the team behind the technology knows how bad it looks.

Husk's test, by the way, doesn't test any of GPT-Live's advanced features — translation, search, or delegation to GPT-5.5. It's an utterly elementary letter-counting task that any first-grader could do. And the flagship voice model worth billions of dollars failed it.

Not Just Husk: The "Strawberry Test" Also Broke the Model

Husk isn't the only one who put GPT-Live through a stress test. User @himelstech on X tried the classic "strawberry test" — asking how many times the letter R appears in the word "strawberry." GPT-Live-1 failed here too, even though most text-based AI models have long since mastered this notoriously known test (though more because they learned the answer than because they can actually count letters).

Other users complain that while the full-duplex feature lets the model chime in during conversations, in practice it means ChatGPT can't stop interrupting. The constant "mm-hmm" and "yeah" while the user is trying to say something is, according to early reviews, more annoying than useful.

Why Can't AI Models Count Letters?

For the average user, it may be a mystery how a model that supposedly handles PhD-level math (see recent GPT-5.5 Pro tests) can fail at counting the letter E. The answer lies in the architecture of large language models.

LLMs (Large Language Models) like GPT work with so-called tokens — fragments of text that may not correspond to individual letters. The word "seventeen" might be split into a small number of tokens, and the model never "sees" individual letters the way a human does. To count the E's, it would need to use logical deduction — and that's precisely where voice models still lag behind their text counterparts.

Ironically, the text version of ChatGPT with GPT-5.5 handles this question without problems. The issue is specific to the voice interface, where the model works with spoken words and their conversion to text, introducing an additional layer of inaccuracy into the process.

What This Means for Users in the Czech Republic

Voice models like GPT-Live aren't primarily targeted at the Czech market yet — Czech is not officially among the supported languages, though the model's translation capabilities suggest it could handle it. For Czech ChatGPT users, of whom there are estimated to be hundreds of thousands, this means voice features remain in "try it in English" mode.

Husk's story is relevant for us too. It shows that marketing rhetoric about AGI and the "next step toward superintelligence" is far from reality. Even the most advanced models fail on trivial tasks — and that's precisely the value of independent testing, whether from enthusiasts on TikTok or researchers.

For companies considering deploying voice AI in customer support, it's a warning sign: voice agents sound increasingly human, but their ability to reliably handle even simple tasks still lags behind.

Conclusion: Progress Yes, But Feet on the Ground

GPT-Live represents a technically impressive leap forward. Full-duplex communication, a more natural voice tone, and the ability to delegate complex thinking to more powerful models are real innovations that push the boundaries of what's possible. But as Husk repeatedly proves with his viral tests, true intelligence isn't measured only by benchmarks and marketing statements — but also by whether the model can count four E's in a word that has nine characters.

And it still can't.