Former OpenAI chief scientist Ilya Satskeva, one of the most prominent figures in the field of artificial intelligence, made this prediction on Friday: Inferential capabilities make the technology much more difficult to predict, he said.
Sutskever, who won the Test of Time award for his 2014 paper with Google’s Oriol Vinyals and Quoc Le, said big changes are on the way in AI.
He said the idea his team had explored a decade ago of scaling up data to “pre-train” AI systems to reach new heights was starting to reach its limits. said. With more data and computing power, ChatGPT, launched by OpenAI in 2022, was praised around the world.
“But as we know, pre-training will definitely end,” Satskever declared to thousands of attendees at the NeurIPS conference in Vancouver. “Computing is increasing, but data is not increasing, because there is only one Internet,” he said.
Despite this challenge, Satskever proposed several ways to explore the frontier. To improve accuracy, he said, the technology itself could generate new data or AI models could evaluate multiple answers before deciding on the best response for the user. Other scientists are looking to real-world data.
But his talk culminated with his prediction of a future of superintelligent machines, which he said was “obviously” in store, a point some disagree with. Sutskever co-founded Safe Superintelligence earlier this year in response to his role in Sam Altman’s brief banishment from OpenAI, but said he regretted doing so within days. said.
He said AI agents that have been studied for many years will have deeper understanding and self-awareness and will bear fruit in future eras. He said AI will reason through problems the same way humans do.
“The more reasons there are, the more unpredictable it becomes,” he says.
Reasoning through millions of choices can make the results less obvious. As an example, AlphaGo, a system built by Alphabet’s DeepMind, surprised highly complex board game experts with a puzzling 37th move en route to defeating Lee Sedol in a 2016 match.
Sutskever similarly said, “Chess AI, really good AI, is unpredictable to the best human chess players.”
AI as we know it will “fundamentally change,” he said.
