In a lecture about AI dangers, OpenAI CEO and co-founder Sam Altman said the company is not training GPT-5, the assumed successor to its March AI language model GPT-4.
During an MIT event, Altman was asked about a recent open letter asking laboratories like OpenAI to halt the development of AI systems “more powerful than GPT-4”. The letter raised worries about future system safety, but several industry members, including signatories, rejected it. Experts argue whether AI poses an existential danger and how the industry can “pause” development.
At MIT, Altman said the letter was “missing most technical nuance about where we need the pause” and that an earlier version said OpenAI is training GPT-5. Altman added, “We are not and won’t for some time. That made it stupid.
OpenAI isn’t working on GPT-5, but it is increasing GPT -4’s capabilities and examining safety issues, as Altman said. “We are doing other things on top of GPT-4 that I think have all sorts of safety issues that are important to address and were totally left out of the letter,” he added.
The fallacy of version numbers—the belief that numbered tech upgrades represent definite and linear capacity improvements—contributes to the misunderstanding. For years, consumer tech has promoted that phone and operating system numbers constitute version control but are marketing strategies. “Of course the iPhone 35 is better than the iPhone 34,” this system says. “Bigger number, better phone.”
This rationale is applied to OpenAI’s language models since consumer tech and artificial intelligence intersect. Unfortunately, this applies to both hucksters and more knowledgeable and nuanced observers. These people use similar terminology to make unfalsifiable AI superintelligence claims. They make imprecise graphs with axes labeled “progress” and “time,” plot a lineup and right, and give this naively as proof.
This is not to downplay AI safety concerns or deny that these systems are quickly advancing and out of human control. However, there are good and terrible reasons, and just because we’ve given a number to something—a new phone or intelligence—doesn’t imply we know everything about it.
Instead, these talks should focus on capabilities: what these systems can and can’t accomplish and how they may develop over time.
Altman’s declaration that OpenAI is not creating GPT-5 won’t reassure AI safety advocates. GPT-4’s potential is being expanded by linking it to the internet, and other companies are designing ambitious tools that allow AI systems to act on customers’ behalf. OpenAI may release GPT-4.5 first, as it did with GPT-3.5, as it optimizes GPT-4.
Even if the world’s governments could prevent further AI advances, humanity has enough problems with the present systems. So when GPT-4 is still poorly understood, does GPT-5 matter?