OpenAI CEO Sam Altman acknowledged in a Reddit AMA that a lack of computing power is one of the major factors preventing the company from shipping products as often as it would like.
“All of these models are becoming extremely complex,” he wrote in response to a question about why OpenAI's next AI model is taking so long. “We also face many limitations and difficult decisions about allocating computing to many great ideas.”
Many reports suggest that OpenAI struggles to secure sufficient computing infrastructure to run and train generative models. Just this week, Reuters reported, citing sources, that OpenAI has been working with Broadcom for months on an AI chip to run the models, which could arrive as early as 2026. .
Partly as a result of the capacity squeeze, Altman said, Advanced Voice Mode, OpenAI's real-sounding conversation feature for ChatGPT, will be available soon, as will the Vision feature, which was first unveiled in April. He said that was not the case. At a press event in April, OpenAI demonstrated how its ChatGPT app runs on a smartphone and responds to visual cues, such as the clothes someone is wearing, within the field of view of the phone's camera.
According to a report in Fortune, the demo was apparently a hasty move to divert attention from Google's I/O developer conference, which was being held the same week. Many within OpenAI did not believe that GPT-4o was ready for public release. So the voice-only version of Advanced Voice Mode was delayed for months.
Altman suggested in the AMA that there is no launch schedule for the next major release of OpenAI's image generator, DALL-E. (“We don’t have any release plans yet,” he said.) Meanwhile, OpenAI’s video generation tool, Sora, “needs to perfect its models and get safety, disguise, and other things right.” Development has been delayed for some reason. scale computing,” wrote Kevin Weil, chief product officer at OpenAI, who also participated in the AMA.
Sora reportedly suffers from technical setbacks that put it in a poor position against competing systems such as Luma and Runway. According to The Information, the original system released in February took more than 10 minutes of processing time to create a one-minute video clip.
In October, one of Sora's co-leaders, Tim Brooks, left for Google.
Later in the AMA, Altman said that OpenAI is still considering allowing “NSFW” content on ChatGPT “someday” (“We absolutely believe in treating adult users like adults”). “Inference'' models and their successors, he wrote, and said the company's top priority is improving the O1 series. OpenAI previewed a number of features coming to o1, including image understanding, at its DevDay conference in London this week.
“We have some really great releases coming up later this year,” Altman wrote. “But there is nothing to call it GPT-5.”