ByteDance discovers new scaling law for AI agents

ByteDance's Seed AI team has published a research paper revealing that AI agents can double their learning speed every three months through extended real-world interaction. The finding offers a potential new scaling law for AI improvement as traditional brute-force methods face limitations. The team developed EdgeBench, a benchmarking suite with 134 ultra-long-horizon tasks requiring at least 12 hours of continuous operation each, spanning software engineering, scientific discovery, formal mathematics, and professional knowledge work. The paper notes that how autonomous systems learn from real-world environments after deployment remains poorly understood. The discovery comes as the global AI industry seeks new ways to enhance models beyond relying on more data and computing power during initial training, a method that OpenAI co-founder Andrej Karpathy and others have warned is unsustainable.