AI researcher Richard Sutton has criticized the use of synthetic data in training large language models, calling it a 'big mistake' in the face of an infinitely complex world. Sutton, a Turing Award winner and founder of reinforcement learning, argues that synthetic data fails to capture the vast intricacies of reality, limiting the potential of AI systems. He and his cofounder, Khurram Javeed, have developed the 'Big World Hypothesis,' which posits that the real world is far too complex for any simulation to accurately represent. This perspective challenges the prevailing strategy in AI labs, which often rely on synthetic data to overcome the limitations of real-world data scarcity. Sutton’s critique highlights the growing debate within the AI research community about the effectiveness of synthetic data in advancing machine learning capabilities. Source: thedecoder
Sutton argues that synthetic data is inadequate because it cannot replicate the complexity of the real world, which is 'massively bigger than everything we stored on the internet.' He emphasizes that the world contains many other agents whose inner workings cannot be accurately modeled through synthetic data. Additionally, he points out that the process of evaluating synthetic data is constrained by human experts, who must determine its quality. This human bottleneck limits the scalability of synthetic data approaches, as it requires hiring domain experts to create and validate datasets. Sutton also criticizes the reliance on frozen simulation models, which do not adapt to new information. Instead, he advocates for agents that can learn from their own experiences and continuously update their world models. Source: thedecoder
Sutton’s views are rooted in his influential 2019 essay 'The Bitter Lesson,' which argues that AI methods that scale with compute, such as search and learning, outperform those that rely on human knowledge. He also co-authored the standard textbook on reinforcement learning and mentored researchers like David Silver, who later worked on AlphaGo. Sutton’s recent work with Oak Lab focuses on developing methods that allow agents to learn from their own experiences, rather than relying on human-built simulations. He calls for continual learning, which avoids catastrophic forgetting by retaining old knowledge while acquiring new information. Source: thedecoder
