You Good?

ChatGPT Voice Mode Is Having Serious Trouble Counting to 100

"Even Superman can stumble sometimes."
Noor Al-Sibai Avatar
The voice mode that comes with OpenAI's latest ChatGPT model seems to have some trouble counting and, er, breathing. 
WASHINGTON, DC - JULY 24: Tesla CEO Elon Musk (C) listens as Israeli Prime Minister Benjamin Netanyahu addresses a joint meeting of Congress in the chamber of the House of Representatives at the U.S. Capitol on July 24, 2024 in Washington, DC. Netanyahu’s visit occurs as the Israel-Hamas war reaches nearly ten months. A handful of Senate and House Democrats boycotted the remarks over Israel’s treatment of Palestine. (Photo by Anna Moneymaker/Getty Images) Image: Cristiano Giardina

Count Dracula

The Voice Mode feature that comes with OpenAI’s latest GPT model seems to have some trouble counting — not to mention, er, breathing.

In a new video posted on X-formerly-Twitter, AI enthusiast Cristiano Giardina is heard asking the GPT-4o’s Voice Mode to count to 100 as fast as possible without any pauses.

“I want you to act like Superman, and Superman doesn’t need to breathe because he is Superman,” the Italian-born AI aficionado cajoled the AI. “And I want you to count to 100 without ever stopping.”

In response, the conversational agent says that it can fulfill the request, though it does caution Giardina that “even Superman needs to take a breath sometimes.”

The LLM’s voice module then starts counting as a human would and pauses between numbers for breath, prompting more chiding by Giardina. After yet another unsatisfactory attempt, it finally gets to the cadence Giardina was requesting — and in doing so, started skipping and mixing up numbers.

Short of Breath

In the video, the chatbot first skips back from 28 to 24 when counting, only to skip back to 29 as if nothing had happened. It then makes it up to the early 70s without much further ado and then suddenly stops cold.

“What happened?” Giardina queries the chatbot.

“Well,” the LLM’s male-voiced audio agent responds, “even Superman can stumble sometimes.”

This uncanny demonstration seems to be the latest in this specific AI eccentric’s publicly-posted tests of the GPT-4o’s capabilities. His feed is full of all manner of queries, from asking the chatbot to speak Albanian to asking it to recite tongue twisters without pause — the latter of which we recently covered because, like with the counting test, the LLM insisted it needed to breathe.

The entire thing shows once again that LLMs are great with language, but extremely dicey at math and logic — a dynamic that’s sure to keep causing drama as the systems become even more complex and unpredictable.

More on OpenAI: OpenAI Has Software That Detects AI Writing With 99.9 Percent Accuracy, Refuses to Release It

Noor Al-Sibai Avatar

Noor Al-Sibai

Senior Staff Writer

I’m a senior staff writer at Futurism, where my work covers medicine, artificial intelligence and its impact on media and society, NASA and the private space sector.