Vancouver, Canada – Ilya Sutskever, former chief scientist at OpenAI and a prominent figure in artificial intelligence (AI), made a bold prediction at the NeurIPS conference in Vancouver on Friday. He foresees a future where AI’s enhanced reasoning capabilities will render the technology far less predictable.
While accepting the “Test of Time” award for his influential 2014 paper co-authored with Google’s Oriol Vinyals and Quoc Le, Sutskever reflected on a significant shift in AI’s trajectory. He noted that the approach of scaling up data to “pre-train” AI systems, which had propelled advancements like OpenAI’s ChatGPT, is reaching its limits. “Pre-training as we know it will unquestionably end,” Sutskever stated. “While compute is growing, the data is not growing, because we have but one internet.”
To push the frontier of AI in spite of this data plateau, Sutskever suggested alternative strategies. He proposed that technology could generate new data or that AI models could evaluate multiple answers before selecting the best response, thereby improving accuracy. Other researchers are exploring the use of real-world data to advance AI capabilities.
Sutskever’s talk culminated in a vision of a future populated by superintelligent machines—an outcome he believes is inevitable. He anticipates the realization of long-envisioned AI agents that possess deeper understanding and self-awareness, capable of reasoning through problems much like humans.
“The more it reasons, the more unpredictable it becomes,” he remarked, highlighting a paradox at the heart of AI development. As AI systems engage in reasoning processes that consider millions of options, their outcomes become less obvious and more surprising. He cited the example of AlphaGo, the AI developed by DeepMind, which astonished the world with an unexpected move during its match against Go champion Lee Sedol in 2016.
“Similarly, the chess AIs, the really good ones, are unpredictable to the best human chess players,” Sutskever observed. This unpredictability underscores the radical transformation AI is undergoing and signals a future where AI’s reasoning abilities could challenge our understanding and expectations.
Sutskever, who co-founded Safe Superintelligence Inc. following his brief departure from OpenAI, emphasizes that AI as we know it will be “radically different.” His insights suggest that as AI systems evolve, embracing reasoning capabilities, they will not only augment their usefulness but also introduce new complexities and challenges for society to navigate.
Reference(s):
AI with reasoning power to be less predictable, Ilya Sutskever says
cgtn.com
