OpenAI’s open-source days are numbered, at least when it comes to GPT models.
During a visit to the TU Munich, OpenAI CEO Sam Altman commented on the progress of the open-source scene, which has reached several milestones recently. The latest is a method to bring a chatbot up to the level of ChatGPT using freely available models with little technical and financial effort. A Google engineer recently described in an internal paper how open-source models could even overtake commercial offerings.
Altman seems more relaxed and thinks the progress of open-source AI is great. Looking ahead, however, he says there is probably a performance threshold in AI models beyond which you have to think differently about open-source development. For now, the development is “really good,” says Altman.
GPT-X back to open source? “We’re not going to do that”
Altman asks the audience at the TU Munich if they would like to see a potential GPT-6 released as open source immediately after the training, and many hands in the room go up. “We’re not going to do that, but it’s interesting to hear that,” says Altman.
Altman had previously told the U.S. Senate that he believed advanced models could have special capabilities, and thus a competitive advantage, even in a strong open-source landscape.
In Munich, Altman emphasized the security aspect, saying that there is a time lag between the introduction of advanced models and the development of similarly capable open-source models. This time can be used to improve the security of the models.
“The open-source lag won’t be that much, and a few years later, when we know what the limits are and how to make it safe, there will be very good open-source models,” Altman says.
Altman expects that as AI systems become more powerful, the gap between commercial development and open source will become more important for security reasons.
Altman is talking specifically about GPT models. So, OpenAI could continue to release individual AI projects as open source, such as the Whisper speech-to-text model.