In this interview, Ilya Sutskever acknowledges that while scaling AI models has driven significant progress, fundamentally new breakthroughs beyond current paradigms are essential to achieve true artificial general intelligence (AGI). He envisions AGI as a continually learning system akin to a capable teenager, predicts its emergence within 5 to 20 years, and emphasizes the importance of collaboration and novel research approaches in overcoming existing challenges.
This interview marks a pivotal moment in the AI field as Ilya Sutskever, a key figure in AI development and former chief scientist at OpenAI, publicly acknowledges the stance of DeepMind and other leading AI researchers. For years, there has been a debate about whether current AI models and scaling laws are sufficient to achieve artificial general intelligence (AGI) or if fundamentally new breakthroughs are needed. Ilya, who initially believed that building bigger models with more compute was the main path forward, now agrees that while scaling continues to improve AI, there are critical unknown breakthroughs still required—breakthroughs that have yet to be properly theorized.
One of the key puzzles discussed is the disconnect between AI models’ impressive performance on difficult evaluation benchmarks and their relatively limited economic impact so far. Ilya suggests two possible explanations: either the models are too narrowly focused due to the way they are trained, or the reinforcement learning environments used for training are too varied and not well-aligned with real-world tasks. This misalignment may cause models to excel on benchmarks but struggle with practical applications. He also highlights that human learning, especially in domains like math and coding that are relatively new in evolutionary terms, points to the existence of fundamental learning principles that current AI paradigms have yet to capture.
Ilya elaborates on his vision of true AGI, emphasizing that it would not be a finished, all-knowing machine but rather a learning system akin to a highly capable teenager. Such a system would start with a foundation of skills but rely heavily on continual learning and adaptation through experience, much like humans do. This contrasts with the common expectation of AGI as a fully formed intelligence capable of performing any human task immediately. He also touches on the importance of a robust value function, similar to human intuition and self-assessment, which guides learning and decision-making without external supervision.
Despite the challenges and the need for new breakthroughs, Ilya remains optimistic about the future of AI. He predicts that within 5 to 20 years, systems capable of human-level learning and beyond will emerge, profoundly transforming the economy and society. He also notes that breakthroughs historically have not always required massive compute, suggesting that smaller labs with novel ideas could still make significant contributions. However, the current competitive landscape, with major labs racing independently toward AGI, has slowed collaborative progress and led to duplicated efforts.
Finally, Ilya discusses his current work at Stability AI (SSI), where he is exploring promising ideas related to understanding generalization and learning in AI. While the commercial strategy is still evolving, he emphasizes a focus on research and experimentation rather than immediate monetization. He acknowledges that the path to superintelligence may involve specialization and competition among different AI systems excelling in various niches. Overall, the consensus among leading AI scientists is now clear: achieving AGI requires breakthroughs beyond current paradigms, but scaling existing models will continue to yield improvements, and the arrival of true AGI is likely within the next two decades.