Unpredictable AI Is Coming: Ilya Sutskever Reveals Bold Predictions for Reasoning Machines

At NeurIPS 2024, Ilya Sutskever shared a compelling vision of AI’s future, warning that reasoning capabilities will make technology less predictable. As pre-training reaches its limits, he described a new era of AI evolution powered by human-like problem-solving.

AI’s Reasoning Power Could Redefine Predictability

On Friday, prominent AI expert and former OpenAI head scientist Ilya Sutskever made a prediction: the ability to reason would make technology far less predictable, Reuters reports.

Sutskever predicted a big shift in AI when he accepted a "Test Of Time" award for a 2014 study co-authored with Google's Oriol Vinyals and Quoc Le.

The theory that his team had investigated a decade ago—that using massive amounts of data to "pre-train" AI systems would propel them to unprecedented levels—was beginning to run out of steam, he added. The world was in awe when OpenAI unveiled ChatGPT in 2022, the product of increased data and processing capacity.

Pre-Training Limitations and Data Challenges

At the NeurIPS conference in Vancouver, Sutskever made the bold statement, "but pre-training as we know it will unquestionably end," in front of thousands of witnesses. "While compute is growing," he pointed out, "the data is not growing, because we have but one internet."

Despite this dilemma, Sutskever proposed a few approaches to advance the frontier. He suggested that in order to increase precision, technology may either create new data on its own or utilize AI models to weigh the pros and cons of different replies before selecting the optimal one for a user. Objective, real-world data has been the focus of other researchers.

AI’s Future and Safe Superintelligence

However, several people disagree with his final prediction on the future of artificial intelligence, which he stated would "obviously" await. After his role in Sam Altman's brief dismissal from OpenAI, which he expressed sorrow for within days, Sutskever co-founded Safe Superintelligence Inc this year.

AI agents that have been in development for a while would finally materialize in that future era, he promised, and they will possess greater comprehension and self-awareness. Just like humans, he claimed, AI will be able to reason through challenges, Yahoo shares.

Unpredictability in Reasoning AI

Be careful, though. "The more it reasons, the more unpredictable it becomes," claimed the expert.

We could make any result non-obvious by reasoning through millions of possibilities. One such example is AlphaGo, a system developed by DeepMind at Alphabet, which stunned Go experts in 2016 with an enigmatic 37th move en route to a victory over Lee Sedol.

"The chess AIs, the really good ones, are unpredictable to the best human chess players," Sutskever added.

In his words, artificial intelligence will be "radically different."