Ilya Sutskever: AI's Ability To Think Will Make It Less Predictable

Former OpenAI Chief Scientist, Ilya Sutskever (left) (photo: x @gdb)

JAKARTA The former Chief Scientist of OpenAI, Ilya Sutskever, who is one of the big names in the field of artificial intelligence, predicts on Friday December 13 that resoning will make AI technology much more difficult to predict.

When receiving the "Test of Time" award for his 2014 paper with Google's Oriol Vinals and Quoc Le, Sutskever said that major changes were in front of AI's eyes. The idea his team had explored a decade ago, that enlarging the data for "pre-training" AI systems would take them to new levels, now starting to reach its limits. More data and computing power have resulted in an OpenAI-launched ChatGPT in 2022 and gaining world praise.

"But pre-training as we know it will end," said Sutskever in front of thousands of NeurIPS conference participants in Vancouver. "While the capacity of computing continues to grow," he continued, "the data does not increase because we only have one internet."

Sutskever offers several ways to push the line despite this dilemma. He said the technology itself could produce new data, or the AI model could evaluate various answers before setting the best response for users to increase accuracy. Other scientists are also focused on real-world data.

However, his lecture led to predictions about the future of a superintelligent engine which he said was "clear" to come, although some people disagreed. This year's Sutskever founded Safe Intelligence Inc after his role in the brief dismissal of Sam Altman from OpenAI, which he later regretted in a few days.

AI agents that have been working on for a long time, he said, will be realized in the future, have a deeper understanding and self-awareness. He said AI would solve problems like humans.

There are catches. "The more he thinks, the more unexpected the results will be," he said. Thinking through millions of options can make any outcome unclear. For example, AlphaGo, a system built by Alphabet's DeepMind, surprised the highly complex board game expert with the elusive 37th movement, in a way to beat Lee Sedol in a 2016 game.

Sutskever said, "AI really good chess is also unexpected for the best human chess players."

AI as we know it, he said, would "divide radically."