Ilya Sutskever Notes, AI Equipped with Reasoning Abilities May Exhibit Less Predictability
Ilya Sutskever discusses how AI systems endowed with reasoning capabilities are expected to become less predictable.
TroibNews 
Former OpenAI chief scientist Ilya Sutskever, a prominent figure in artificial intelligence, made a significant prediction on Friday: reasoning abilities will render technology much less predictable.
While accepting a "Test Of Time" award for his 2014 paper co-authored with Google's Oriol Vinyals and Quoc Le, Sutskever discussed an impending transformation in AI. He noted that an idea his team had explored a decade prior, about scaling up data to "pre-train" AI systems for greater capabilities, is beginning to hit its limits. The increase in data and computational power has led to the creation of ChatGPT, which OpenAI introduced in 2022 and received widespread acclaim.
"But pre-training as we know it will unquestionably end," Sutskever declared before a large audience at the NeurIPS conference in Vancouver. He continued, "While compute is growing...the data is not growing, because we have but one internet."
To address this challenge, Sutskever provided several strategies to advance the field. He suggested that technology could generate new data, or that AI models could assess multiple responses before determining the best one for the user, thereby enhancing accuracy. Other researchers are focusing on utilizing real-world data.
Sutskever's presentation concluded with a prediction about the emergence of superintelligent machines, a notion he described as "obviously" on the horizon, despite some dissenting opinions. This year, he co-founded Safe Superintelligence Inc., following a brief period of turmoil regarding Sam Altman's abrupt departure from OpenAI, which he admitted he regretted soon after.
He envisioned that long-awaited AI agents will eventually materialize, possessing a deeper understanding and self-awareness. "AI will reason through problems like humans can," he stated, outlining a future where machines emulate human-like reasoning.
However, there is a caveat.
"The more it reasons, the more unpredictable it becomes," he remarked.
Sutskever explained that reasoning through millions of possibilities could lead to non-obvious outcomes. He referenced AlphaGo, a system developed by Alphabet's DeepMind, which astonished experts in the intricate board game with its enigmatic 37th move during its victory over Lee Sedol in 2016.
He noted that, similarly, "the chess AIs, the really good ones, are unpredictable to the best human chess players."
In summary, Sutskever argued that AI will evolve into something "radically different" from its current state.
Lucas Dupont for TROIB News
While accepting a "Test Of Time" award for his 2014 paper co-authored with Google's Oriol Vinyals and Quoc Le, Sutskever discussed an impending transformation in AI. He noted that an idea his team had explored a decade prior, about scaling up data to "pre-train" AI systems for greater capabilities, is beginning to hit its limits. The increase in data and computational power has led to the creation of ChatGPT, which OpenAI introduced in 2022 and received widespread acclaim.
"But pre-training as we know it will unquestionably end," Sutskever declared before a large audience at the NeurIPS conference in Vancouver. He continued, "While compute is growing...the data is not growing, because we have but one internet."
To address this challenge, Sutskever provided several strategies to advance the field. He suggested that technology could generate new data, or that AI models could assess multiple responses before determining the best one for the user, thereby enhancing accuracy. Other researchers are focusing on utilizing real-world data.
Sutskever's presentation concluded with a prediction about the emergence of superintelligent machines, a notion he described as "obviously" on the horizon, despite some dissenting opinions. This year, he co-founded Safe Superintelligence Inc., following a brief period of turmoil regarding Sam Altman's abrupt departure from OpenAI, which he admitted he regretted soon after.
He envisioned that long-awaited AI agents will eventually materialize, possessing a deeper understanding and self-awareness. "AI will reason through problems like humans can," he stated, outlining a future where machines emulate human-like reasoning.
However, there is a caveat.
"The more it reasons, the more unpredictable it becomes," he remarked.
Sutskever explained that reasoning through millions of possibilities could lead to non-obvious outcomes. He referenced AlphaGo, a system developed by Alphabet's DeepMind, which astonished experts in the intricate board game with its enigmatic 37th move during its victory over Lee Sedol in 2016.
He noted that, similarly, "the chess AIs, the really good ones, are unpredictable to the best human chess players."
In summary, Sutskever argued that AI will evolve into something "radically different" from its current state.
Lucas Dupont for TROIB News
