INFER forecasters: Don’t expect GPT-5 anytime soon

Walter Frick
May 22, 2023 09:03PM UTC
OpenAI has had a productive year. In December 2022, it launched ChatGPT, a chatbot powered by its large-language model, GPT-3.5. By February, Microsoft announced it would use ChatGPT to power its search engine. In March, OpenAI released GPT-4, an even more powerful AI model that improved ChatGPT significantly. And last week, OpenAI launched an iOS app for ChatGPT, bringing its AI to phones around the world.

But the next generation of OpenAI’s technology won’t arrive anytime soon, according to forecasters at INFER as well as analysts tracking AI. INFER asked its forecasters to predict whether OpenAI would release GPT-5 by the end of 2023. The crowd forecast as of this writing gives just a 4% chance that this happens.

“It took ~8 months from the original GPT-1 to GPT-2, more than one year between the releases of GPT-2 & GPT-3, almost 3 years between GPT-3 & GPT-4, and one full year between GPT-3.5 and GPT-4,” wrote INFER forecaster ctsats. “Given the overwhelming (and expensive) effort that in all likelihood will be needed for GPT-5, there is almost no chance… that it will be released in 2023.” ctsats gave just a 1% chance of GPT-5 debuting this year.

This fits with public statements from Sam Altman, OpenAI’s CEO. In April, Altman said his company wasn’t training GPT-5 and “won’t for some time.” In May, he testified before Congress and said “We are not currently training what will be GPT-5, we don’t have plans to do it in the next six months.”

Those statements don’t mean the company isn’t working toward the eventual release of GPT-5, says Timothy Lee, a journalist who has covered tech for Vox and the Washington Post and who writes the newsletter Understanding AI. “It’s very possible it just takes two-to-three years and they’re in the development phase and not the training phase yet,” he told INFER. “I’m sure they’re working on it in some sense.”

In the meantime, though, OpenAI likely has other priorities—like making the most of its existing platform.

“There's a tradeoff for using the [computing power] for training vs inference,” wrote INFER forecaster heim, referring to the resources used to create a model (training) and to operate ChatGPT for users (inference). “Inference for now has an immediate return; whereas training is a bet.”

Analysts agree that the company’s near-term opportunity is in improving its product offerings, which have been fairly bare bones to date. The company will likely focus on things like offering less bias, better reasoning, more reinforcement learning with human feedback, and more plugins, says Helen Edwards, an AI strategist and co-founder of the consultancy Sonder Studio.

That’s because the competitive threats to OpenAI are most significant at that application layer, says Dave Edwards, Sonder’s co-founder. OpenAI’s main rival, Google, has preexisting relationships with users and businesses. The challenge for OpenAI is to build products customers find useful, once the novelty of chatting with an AI has worn off. (Its partner, Microsoft, will be central to this effort.)

Neither analysts nor INFER forecasters emphasized the threat of regulation as a factor influencing OpenAI’s timing. Altman called for regulation in his Congressional testimony, and his initial comment about not training GPT-5 came in response to an open letter signed by some notable technologists and computer scientists calling for a pause in AI development. Nonetheless, few analysts expect regulation to be a primary determinant of when GPT-5 arrives.

Safety concerns, however, will continue to affect the pace at which OpenAI releases its technology.

“I think Sam Altman is sincere in being motivated by AI safety concerns,” said Lee, noting that unlike many Silicon Valley CEOs, Altman doesn’t own stock in OpenAI. If he felt it wasn’t in the interests of the world to release GPT-5 quickly he wouldn’t change that because of competitive pressure from Google, said Lee.

Eventually, OpenAI will release a new foundational language model to power its offerings. But despite the rapid progress the company has made in its releases over the last year, forecasters expect the next one to take a while.

This article is part of a new blog series called INFER Insights. Each post will explore a question (or group of questions) we're monitoring on INFER and compare the crowd consensus with expert perspectives surrounding the possible outcome.

Walter Frick

By: Walter Frick

Walter Frick is the founder of Nonrival, a newsletter that lets readers make predictions about tech, business, and the economy. He is a contributing editor (former senior editor) at Harvard Business Review, former executive editor at Quartz and has written for publications including The Atlantic, the BBC, and MIT Technology Review.

Tip: Mention someone by typing @username