ChatGPT simply launched a brand new voice mannequin, and above all else, its conversations are supposed to really feel pure. Severely—the phrase pure seems a minimum of 13 instances in OpenAI’s announcement of GPT-Stay. The place earlier voice fashions relied on turn-based conversations, GPT-Stay goals to actively hearken to customers and constantly react, even interjecting with acknowledgments like “mmhmm” and “received it” to show it’s paying consideration.
The “new technology of voice fashions” launched on July 8 with two variations, GPT‑Stay‑1 and GPT‑Stay‑1 mini, which have rolled out to ChatGPT customers globally. They’re meant to be smarter, quicker, and extra human than ever earlier than—however one infamous influencer put that to the take a look at.
Husk, an influencer recognized for tricking ChatGPT into making errors, put GPT-Stay by means of his most well-known experiment: a spelling take a look at. With earlier fashions, he’s gotten ChatGPT to spell “December” with an X, “eighty” with an A, and “strawberry” with just two R’s.
With GPT-Stay “set to the best good setting,” Husk asked the model what number of E’s are within the phrase “seventeen.”
With out lacking a beat, GPT-Stay claimed that there are simply two E’s in seventeen, “one in ‘seven,’ and one in ‘teen.’”
“You’re actually good, proper?” Husk requested as a follow-up.
“Fairly sharp, yeah,” GPT-Stay replied. “However I can slip every now and then. Wish to take a look at me?”
When Husk requested if the mannequin was positive of its reply, it mentioned, “Completely,” then spelled the phrase letter by letter. Regardless of spelling it appropriately with all 4 E’s, GPT-Stay nonetheless insisted there are solely two E’s within the phrase.
With that, Husk bid the brand new mannequin adieu: “Have a pure goodbye,” he mentioned. However GPT-Stay did the precise reverse, taking a protracted pause earlier than exclaiming, “Oh!” and shutting the dialog.
‘Altman’s worst nightmare’: How Husk undermines OpenAI
Husk’s anti-AI experiments go far past easy spelling bees. Earlier viral examples embrace asking ChatGPT to react to a song that doesn’t exist (it mentioned “the melody was fairly catchy” after listening to nothing however silence) and to time him while he ran a mile (the mannequin made up a determine across the 10-minute mark, regardless of Husk asking for his outcomes simply seconds after beginning the timer).
Husk’s exchanges with ChatGPT earned him the moniker “Sam Altman’s worst nightmare” on social media, although the OpenAI CEO doesn’t appear to take Husk severely, laughing off his experiments when shown them during an interview.
Persevering with to push the boundaries of GPT-Stay, Husk returned to see simply how “pure” the brand new mannequin might actually act. He turned to it for help feeling more human, asking, “Are you able to simply inform me one thing that solely people can do?”
GPT-Stay really useful that Husk go as much as a stranger and provides them a private praise. After Husk mentioned he would possibly praise the best way they scent and GPT-Stay recommended he level out the best way they stroll, Husk requested if he ought to say, “Hey, you bought a pleasant scent and I like the best way you stroll” to a stranger.
“That’s completely a means you could possibly strategy it,” the mannequin replied. “It appears like an incredible plan.”
Once more, the dialog discovered an ungainly conclusion. After Husk mentioned, “Okay, thanks,” GPT-Stay took one other lengthy pause, then let loose an inexplicable heavy sigh earlier than replying, “Any time.”
‘A brand new period’: How GPT-Stay modifications AI dialog
Although Husk’s experimentation reveals that GPT-Stay might not be as “human” or “pure” because it claims to be, the brand new mannequin does have a number of new options that take it nearer to human dialog.
The place earlier ChatGPT voice methods ran three fashions one after one other—a speech-to-text processor, a big language mannequin, then a text-to-speech processor—GPT-Stay processes inputs and generates output concurrently, slicing down on response time and aiming to create a “extra pure back-and-forth.”
GPT-Stay can be able to delegation, so if a query requires deeper work like analysis or reasoning, it could actually offload that activity to a different ChatGPT mannequin whereas it retains the dialog shifting. GPT-Stay also can generate visible solutions in the midst of a dialog when wanted, like maps, climate forecasts, or sports activities outcomes.
Based on OpenAI’s inner evaluations, each GPT-Stay variations outperformed ChatGPT’s earlier Superior Voice Mode in pleasantness, circulation of dialog, scientific reasoning, and agentic looking out.
All of it provides as much as “a brand new period of human-AI interplay,” per OpenAI’s announcement: “Our imaginative and prescient is to allow really pure human–AI interplay: a world the place collaborating with AI feels as fluid and responsive as working with one other particular person, whereas reasoning and complicated activity execution occur seamlessly within the background.”
