Sam Altman got put on the spot recently…

An interviewer showed him a clip from Husk IRL, one of those channels racking up millions of views by catching ChatGPT in embarrassing moments.

In this one, the YouTuber asks ChatGPT to time his run.

The “run” lasts about 2 seconds but ChatGPT confidently tells him he ran for 7 minutes.

After an awkward pause, Altman explains that the chat model didn't have access to the tools it would need to actually do that.

Look, I'm all for calling AI companies out when the hype doesn't match reality…

But Altman's explanation here actually holds up, and I think it's worth slowing down on why.

Almost every viral video in this genre, whether it's @HuskIRL or @fatherPhi (the two biggest accounts doing this), is filmed using the voice chat.

The voice chat almost certainly runs on a lighter, faster model than the standard text interface.

Speed matters more than depth when you're having a back-and-forth conversation in real time, so OpenAI has almost certainly made trade-offs there.

Take one of fatherPhi's biggest videos, where he asks ChatGPT what the "S" in “ChatGPT” stands for and the model completely gaslights him.

Well GPT-5 in the standard text window gets it right, even on a free plan.

Or take Husk IRL pointing his phone's camera at the word "Strawberry" and asking the voice chat how many R's there are.

The voice chat gets it wrong in the video.

But again the text chat (GPT-5) gets it right:

I think in this case, it simply showcases the shortcomings of AI visual recognition more so than the voice model’s intelligence.

But a model being bad at reading through a phone camera is not the same thing as the model being dumb.

A blind person can't tell you what color your shirt is, but that's not an intelligence problem.

So… a lot of these viral videos can be “debunked” by the following:

Both of those are real limitations worth knowing about. They're just not the same as "ChatGPT doesn't know how to count.”

Again I want to make clear that I’m all for calling Sam Altman out on his BS, when he says that OpenAI has basically built AGI already…

But I think the very cherry-picked examples shown in these viral videos aren’t as much of a “got ya’” as people think.

That said, if it gets people to be a bit more skeptical of AI hype and a bit more realistic about what these tools can and can't do, it's probably still a net positive.