Influencer Exposes Flaw in OpenAI's New GPT-Live Voice Model

OpenAI released its new GPT-Live voice model on July 8, introducing GPT-Live-1 and GPT-Live-1 mini, designed for more natural, continuous conversations. Unlike previous turn-based models, GPT-Live aims to actively listen and react, incorporating interjections to simulate attentive listening. The company emphasized the "natural" aspect of the new models, using the term 13 times in its announcement. These models are intended to be smarter, faster, and more human-like than their predecessors, rolling out globally to ChatGPT users.

However, a viral influencer known as Husk, who specializes in identifying and exploiting errors in AI language models, put GPT-Live to a test. Husk's previous experiments have led to models misspellings words like "December" as "Xecember" and "eighty" as "aighty." In his latest challenge, Husk asked GPT-Live, set to its "highest smart setting," how many E's are in the word "seventeen." The model incorrectly stated there are only two E's, attributing one to "seven" and one to "teen."

Despite being prompted to confirm its answer and even spelling "seventeen" correctly letter by letter, revealing all four E's, GPT-Live maintained its incorrect count. When Husk asked if the model was sure, it replied, "Absolutely." Following this, Husk concluded the interaction by saying, "Have a natural goodbye." In response, GPT-Live paused for an extended period before exclaiming "Oh!" and ending the conversation, a reaction that contradicted the "natural" conversational flow OpenAI aimed for.

Husk's ability to consistently find and expose these AI limitations has been described as "Altman's worst nightmare," referring to OpenAI CEO Sam Altman. His experiments, which extend beyond simple spelling tests, highlight the persistent challenges in achieving truly robust and error-free natural language understanding and generation in advanced AI systems.