China's ByteDance discovers new scaling law that could sustain AI boom

Researchers at TikTok parent ByteDance have discovered a new scaling law governing how fast artificial intelligence agents can improve by performing real-world tasks, a finding that could help prolong the AI boom just as traditional development methods hit a wall. In a research paper published on Thursday, ByteDance’s Seed AI team revealed that AI agents – autonomous software that executes tasks on a human’s behalf – can double their learning speed every three months by interacting with real-world environments over extended periods. The finding comes as the global AI industry searches for new ways to improve models. For years, developers relied on feeding systems more data and computing power during initial training, but prominent industry figures – including OpenAI co-founder Andrej Karpathy – have warned that this brute-force approach cannot last forever. The issue is partly related to a looming data drought. US-based research institute Epoch AI recently warned that publicly available, human-generated text data could be depleted within the next six years. That makes finding alternative paths to advance AI one of the industry’s highest priorities. However, despite the fact that tech firms are pivoting towards agentic AI, ByteDance researchers noted in the paper that how these autonomous systems “learn from real-world environments after deployment remains far less understood”. To address the problem, the team developed EdgeBench, a benchmarking suite featuring 134 ultra-long-horizon tasks spanning a wide range of areas from software engineering and scientific discovery to formal mathematics and professional knowledge work. Each task requires at least 12 hours of continuous AI agent operation.

The researchers logged 38,000 hours of environment interaction to evaluate five frontier models. These included Anthropic’s Claude Opus 4.8, OpenAI’s GPT 5.5 and GPT 5.4, as well as models from Chinese front runners Zhipu AI and DeepSeek. The data revealed that an agent’s performance follows a highly predictable mathematical curve. This suggests that AI capabilities can continue to improve predictably through hands-on experience, even as the gains from conventional pre-training fade. The ByteDance team argued in the paper that “post-deployment learning from rich environments may deserve the same systematic scaling attention that pre-training has received”. This adaptability has become increasingly important as AI agents are integrated into a range of real-world scenarios, from enterprise software to scientific research and engineering projects. Rather than relying solely on the static knowledge acquired during their initial training, these agents will need to continuously evolve on the job. “An agent’s ability to learn from its environment and improve task performance is central to deploying AI systems at scale in the real world,” the researchers concluded on the official website of the EdgeBench benchmark. Additional reporting by Ben Jiang