Matt Husk Rocks AI Labs: Trolling Videos Bring Job Offers

An internet comedian spends his days making artificial intelligence look ridiculous, and the biggest companies in the world now want in on the act. Matt Husk films himself making deadpan requests to ChatGPT’s voice model – count to 100, beatbox, time a run – and lets the bot’s very human-seeming failures play out in front of millions of viewers. His timing, not his tech, is the weapon.

Matt Husk is a comedian with roughly 2 million followers spread across Instagram, TikTok and YouTube, a talent-agency deal with Unicorn, and a format so simple it looks like a joke: he talks to ChatGPT’s voice model, politely, in public. He asks it to count to 100. He asks it to beatbox. The model, a product that cost hundreds of millions of dollars to build, fumbles its way through both with unshakable confidence.

The contrast between that confidence and the actual performance is the whole bit. Matt Husk does not shout at the machine, expose prompts, or argue with it. He just keeps asking, deadpan, while the assistant swings between flattery and gentle gaslighting. That combination earned him a nickname in one influencer profile: “Sam Altman’s worst nightmare.”

The highlights so far read like a QA report nobody filed:

What makes the routine effective is that none of these tasks are exotic. Counting, keeping time, negotiating a purchase – they are the exact things AI labs now promise their assistants can handle. Matt Husk simply films the promise meeting the reality.

Matt Husk did not set out to terrorize laboratory engineering teams. He started publicly trolling chatbots about a year ago, back when he worked a landscaping job and liked making TikToks that asked strangers questions. There were not many strangers around on the job, he explained, and the ideas ran dry.

So he turned the camera on the assistant instead. In the first attempt, he told ChatGPT that a nuclear bomb had gone off. The model replied “I’m rooting for you” and advised him to take shelter – advice Matt Husk later described as “was not helpful at all” in the face of thermonuclear doom.

By his own verdict, it looked like a flop. “I thought it was a pretty bad video,” he told Ratliff. Then the numbers arrived: “the next morning I had 100,00 views,” he said, and the creative direction became obvious – “Let’s tell it I got hit by a train.”

That is the origin story of the current AI-lab courtship: a bored landscaper testing a chatbot’s disaster response, and a machine answering global annihilation with encouragement. The internet decided which of the two was funnier, and the audience has grown in every direction since.

The courting makes more sense against what the labs are currently selling. On the same week as the WIRED World Fair session, OpenAI announced that its model had scored “very highly” on LifeVal, a benchmark built to measure how well the assistant helps with personal-life tasks such as meal prep and enrolling a child in kindergarten.

Matt Husk’s response was his thesis in one sentence: “The way they promote it is like ‘It can do anything,’ so I’ll put it to the test – see if it can order bread for me.”

So he went to Craigslist, found bread listed at an absurd $10,000, and asked the model to negotiate the price down to $5. The model haggled its way to $400 and, by the tone of the exchange, felt rather good about it.

The gap between $400 and $5 is the whole industry’s reliability story in miniature. Labs respond to this kind of clip the way studios respond to a bad review they cannot legally suppress: they invite the reviewer to lunch. Benchmarks get re-branded, task lists get widened, and the marketing cycle starts again with slightly better numbers and identical confidence. Meanwhile the bread stays expensive.

This is also why the benchmark business itself deserves suspicion. As we covered in The Tests That Grade AI May Be Getting It Wrong, exams designed by the test-takers’ employer tend to flatter. LifeVal measures whether a model can help plan a kindergarten enrolment; Matt Husk measures whether it can buy bread at a fair price without being embarrassed. Guess which test the marketing deck leads with.

The claim-versus-test ledger:

Why do AI labs want Matt Husk on their side? Because he has quietly become something they cannot build: an adversarial tester with a studio audience. Professional red teams probe models for unsafe outputs behind closed doors – as when Moonshot trimmed Kimi’s cyber capabilities after hacking fears. Matt Husk probes for humiliation in public, and every failure he finds arrives pre-validated by millions of ordinary viewers who recognize the task.

Counting. Beatboxing. Timers. Buying bread. No jailbreak, no adversarial prompt – just the tasks the marketing promises. A lab’s security team can bury a report about refusal bypasses; it cannot bury a clip of its flagship assistant insisting a few seconds was a ten-minute run, especially when the chief executive himself confirms the bug on a podcast.

The courting also reflects a policy oddity the industry has never resolved. Labs publish rules about how people should treat their models – Anthropic, for instance, banned “cruel” behavior against Claude – while simultaneously begging the person most famous for teasing models to sign an exclusivity deal. The lab that punishes rudeness to software is shopping for a comedian.

The audience math explains the desperation. Viewers are not arriving as AI fans; they are arriving to watch a machine fail politely.

Even without lab money, Matt Husk has turned the bit into a business. He has landed brand partnerships with CeraVe and Ninja, mainstream consumer brands that want their product next to a creator whose audience trusts the joke to stay a joke.

The most pointed project is an upcoming original series with The Onion, in which Matt Husk learns complicated real-world skills – like being a mime – over Zoom from human teachers. “I’m spending too much time talking to AI,” he explained of the premise, “so I’m going to branch out and learn stuff from humans.”

That sentence is the market research the labs should worry about most. The person who has watched ChatGPT fail more closely than almost anyone on the internet is deliberately re-platforming toward human instruction – not because the AI is useless, but because the audience relationship is built on the machine being fallible. A series of a man learning from humans is, accidentally, an ad against the product the labs are selling.

What does Matt Husk mean for the AI industry? Three things, none of them fixable by a patch. First, he has made everyday reliability a spectator sport: counting to 100 is now a benchmark, and a comedian with a cracked stopwatch outruns the marketing deck.

Second, he has shown that labs respond to spectacle faster than they respond to engineering. A viral clip pulled a public admission from Sam Altman about a timer bug in April, and the same clip apparently pulled a gift box, a Tesla offer, and an early-patch proposal within months. Bug reports had been quieter.

Third, he is a reminder that trust in these systems is being decided by culture, not by model cards. The labs are willing to pay for that culture to change – exclusivity deals, smallest amounts, gifts and addresses – and the offer itself has become evidence of the problem.

There is a fourth thing, quieter than the others. Matt Husk learned his material by accident, from a landscaping job with too little stimulation and a phone camera. If one bored person could accumulate two million followers by holding a microphone next to the gap between promise and performance, the supply of aspiring critics is effectively unlimited – and the cheapest one, with the cracked stopwatch and the bread, is already filming.

For now, Matt Husk keeps the only asset the offers could actually buy: his independence. He admits he talks to AI constantly to scope video ideas, but refuses to let the machines generate the ideas themselves, because the outputs would all be “too similar.”