(Credit: Jakub Porzycki/NurPhoto via Getty Images)
The next generation of OpenAI’s ChatGPT is “behind schedule” and running up “huge bills,” according to a report in The Wall Street Journal.
GPT-5—codenamed Project Orion—was originally expected to be released in mid-2024, according to the WSJ’s sources. But a raft of technical issues are plaguing the high-profile project. OpenAI has reportedly conducted at least two large “training runs,” a term for when a large language model like ChatGPT crunches huge amounts of data with the goal of improving itself. But the software allegedly fell short of the results researchers wanted.
One of the primary technical issues is the quality of the data. When these series of training runs began, researchers reportedly found that the data in question “wasn’t as diversified as they had thought,” which limited how much Orion would learn.
In mid-2023, OpenAI launched one training run dubbed "Arrakis,” which was reportedly “sluggish” and indicated that future training runs would take an “incredibly long time.”
This isn’t the first time we’ve heard about the delays impacting GPT-5, though when CEO Sam Altman addressed the issue, he laid the blame on the company’s focus on GPT-o1, a recently launched model that was built to cater to specialized scientific and academic use cases.
The WSJ’s report also nodded to the sheer number of high-profile executive exits that have plagued OpenAI this year, including co-founder and Chief Scientist Ilya Sutskever and Chief Technology Officer Mira Murati.
If OpenAI is slowing down, there are plenty of competitors who are eagerly waiting for their chance to take OpenAI’s spot in the AI world and invest the capital needed to compete.
In November, Amazon announced that it was set to invest another $4 billion into Claude AI parent and OpenAI rival Anthropic, bringing its total funding amount to $8 billion, as per Crunchbase.
Earlier this month, Elon Musk’s xAI announced it was set to quintuple the amount of GPUs at its disposal and expand its Colossus supercomputer in Memphis, Tennessee, to one day house at least a million GPUs.