The AI industry has a dirty little secret: the old playbook of making models bigger and feeding them more data is running out of road. Pre-training scaling, the engine that powered everything from GPT-3 to the current generation of frontier models, is bumping up against data shortages and diminishing returns. ByteDance thinks it found a new gear.
The company’s Seed AI research team published a paper introducing what it calls a scaling law for post-deployment learning, the idea that AI agents get predictably better the longer they interact with real-world environments after they’ve already been trained and released.
What ByteDance actually found
The research team built a new benchmark called EdgeBench, consisting of 134 long-horizon tasks. Each task requires a minimum of 12 hours of continuous operation. These aren’t quick chat completions or image classifications. They’re extended, complex workflows that demand sustained reasoning and adaptation over time.
To test the concept, the team analyzed over 38,000 hours of interactions between AI agents and their environments. The models put through their paces included some of the most capable systems available: Anthropic’s Claude Opus 4.8, OpenAI’s GPT 5.5 and GPT 5.4, along with models from Zhipu AI and DeepSeek.
