By Jeffrey Dastin VANCOUVER (Reuters) - Ilya Sutskever, former chief scientist at OpenAI and a prominent figure in artificial intelligence, predicted on Friday that advances in reasoning capabilities would render technology less predictable. While accepting a "Test Of Time" award for his 2014 paper with Google's Oriol Vinyals and Quoc Le, Sutskever noted that AI was on the brink of a significant transformation. A decade-old concept from his team, which involved scaling up data to "pre-train" AI systems, was approaching its limitations, he said. This approach had led to innovations like OpenAI's ChatGPT, launched in 2022 to widespread acclaim. "But pre-training as we know it will unquestionably end, " Sutskever told thousands at the NeurIPS conference in Vancouver. "Compute continues to grow, " he noted, "but data does not, as we have only one internet. " Sutskever suggested ways to advance despite this challenge. He proposed technology generating new data, or AI evaluating multiple solutions before determining the best response to enhance accuracy.

Other scientists aim to leverage real-world data. His talk concluded with a prediction for superintelligent machines, which he claimed were inevitable, though opinions differ. This year, following Sam Altman's brief ouster from OpenAI—a decision Sutskever quickly regretted—he co-founded Safe Superintelligence Inc. He envisions future AI agents will achieve deeper understanding and self-awareness, solving problems with human-like reasoning. There's a caveat. "The more it reasons, the more unpredictable it becomes, " he said. Reasoning through myriad options could yield unforeseen outcomes. For instance, AlphaGo, developed by DeepMind, astonished the Go community with its unexpected 37th move, ultimately defeating Lee Sedol in 2016. Sutskever likened this to chess AI, which remains unpredictable even to top human players. AI, as we know it, he said, will be "radically different. " (Reporting By Jeffrey Dastin and Anna Tong in Vancouver; Editing by Sam Holmes)