OpenAI launches in August the GPT-5 large-language model was something of a disaster. There were errors during the live broadcast as the model generated graphs with clearly wrong numbers. In a Reddit AMA with OpenAI employees, users complained that the modern model was not user-friendly and called on the company to roll back the previous version. Above all, critics complained that GPT-5 failed to meet the stratospheric expectations that OpenAI had been testing for years. Promised as a groundbreaking game model, GPT-5 could actually play the game better. But it was still the same game.
Skeptics used this moment to declare the end of the artificial intelligence boom. Some even predicted the beginning of another AI winter. “GPT-5 was the most popular AI system of all time,” full-time bubble shooter Gary Marcus told me during his busy schedule of victory laps. “It was supposed to provide two things: AGI knowledge and PhD-level knowledge, but it provided neither.” What’s more, he argues, the seemingly lackluster modern model is proof that OpenAI’s pass to AGI – massively scaling data and chip sets to exponentially enhance the intelligence of systems – can no longer be beaten. This time, Marcus’s views were shared by a immense part of the AI community. In the days after its release, GPT-5 looked like an artificial intelligence version of Modern Coke.
Sam Altman doesn’t mind. A month after launch, he walks into a conference room at the company’s modern headquarters in San Francisco’s Mission Bay district, eager to explain to me and my colleague Kylie Robison that GPT-5 is everything he touted and that all is well with his epic foray into AGI. “The atmosphere was quite bad at the start,” he admits. “But now they are great.” Yes, Great. It’s true that the criticism has died down. Indeed, the company’s recent release of a wacky tool for generating impressive AI video bugs has turned the narrative away from GPT-5’s disappointing debut. Altman’s message, however, is that naysayers are on the wrong side of history. He says the road to AGI is still on track.
Numbers game
Critics may see GPT-5 as the end of the AI summer, but Altman and team say it solidifies AI technology as an indispensable teacher, a search engine-busting information resource, and especially as a sophisticated collaborator with scientists and developers. Altman says users are starting to see it their way. “GPT-5 is the first time people say, ‘Holy crap. This is an important piece of physics.’ Or a biologist says, “Wow, that really helped me understand,” he says. “Something important is happening that hasn’t happened with any pre-GPT-5 model, which is the beginning of artificial intelligence helping to accelerate the rate of discovery of new science.” (OpenAI did not say who these physicists and biologists are.)
So why the lukewarm initial reception? Altman and his team identified several reasons. First, they say that since GPT-4 hit the streets, the company has delivered versions that have been revolutionary in their own right, especially in terms of the sophisticated reasoning modes they added. “There was a jump from 4 to 5 greater than jumping from 3 to 4,” Altman says. “We had a lot of stuff along the way.” OpenAI CEO Greg Brockman agrees: “I’m not shocked that many people felt this way [underwhelmed] reaction because we showed our hand.”
OpenAI also claims that because GPT-5 is optimized for specialized uses such as science or coding, regular users will take the time to appreciate its benefits. “Most people are not physics researchers,” notes Altman. As Mark Chen, head of research at OpenAI, explains, unless you’re a math genius yourself, you won’t particularly care that GPT-5 is among the top five math Olympians, whereas last year the system was among the top 200.
As for the objection about how GPT-5 shows that scaling doesn’t work, OpenAI says it’s due to a misunderstanding. Unlike previous models, GPT-5 didn’t make significant gains with a much larger dataset and tons more calculations. The new model benefits from reinforcement learning, a technique in which experts provide it with feedback. Brockman says OpenAI advanced its models to the point where it could generate its own data, which fed the reinforcement learning cycle. “When a model is stupid, all you want to do is train a bigger version of it,” he says. “When a model is clever, you want to sample it. You want to train on its own data.”