Former OpenAI chief scientist, Ilya Sutskever, has predicted that reasoning capabilities in artificial intelligence (AI) will make the technology increasingly unpredictable.
Sutskever made this statement while accepting the prestigious “Test of Time” award for his 2014 paper at the NeurIPS conference in Vancouver, Canada.
Sutskever predicted that the future will see the development of superintelligent machines that possess a deeper understanding of the world and self-awareness, which he believes is an inevitable outcome, though he acknowledged that not everyone agrees with this view.
He emphasized that AI agents, which have been in development for a long time, will eventually reach a point where they can reason through problems in a way that is similar to human thinking.
However, he also warned that as AI systems become more capable of reasoning, they will become increasingly unpredictable, making it harder to anticipate their behavior and decisions.
“The more it reasons, the more unpredictable it becomes,” he explained.
To illustrate his point, Sutskever referenced AlphaGo, an AI developed by Alphabet’s DeepMind, which surprised experts with its unconventional 37th move during a 2016 Go match against world champion Lee Sedol.
“The chess AIs, the really good ones, are unpredictable to the best human chess players,” he said underscoring the complexity of reasoning in advanced AI systems.
Pre-training approach nearing end
Sutskever emphasized that the traditional pre-training method, which relies on vast amounts of data to improve AI systems, is approaching its limits.
He explained that while computing power continues to advance, the availability of new data has become a critical bottleneck.
“Pre-training as we know it will unquestionably end,” he stated, pointing out that the finite nature of the internet imposes constraints on how much data can be leveraged.
“The data is not growing because we have but one internet,” he added, emphasizing the constraints in scaling AI further using existing methodologies.
Exploring new directions
Sutskever outlined alternative strategies to overcome this challenge. He proposed that AI systems could generate new data autonomously or evaluate multiple answers before selecting the most accurate response.
These methods, he suggested, could push AI technology to new frontiers.
“Other scientists have set sights on real-world data,” he noted, signaling the ongoing exploration of innovative approaches within the field.
What you should know
Artificial Intelligence (AI) is poised to significantly impact the global advertising industry, with AI-driven content expected to account for 94.1% of global ad revenue by 2029, three years earlier than previously predicted.
Digital advertising continues to dominate, with retail media and Connected TV (CTV) emerging as key growth areas. Meanwhile, the traditional media sector faces slower growth, while out-of-home advertising is experiencing a recovery.
However, AI’s rapid development faces challenges, particularly the global shortage of AI talent. Huawei’s Terrence Wu recently highlighted the uneven distribution of AI professionals, stressing the need for global collaboration to bridge the talent gap.