ByteDance Discovers New Scaling Law for AI Agents
Miles Bennett
ByteDance's Seed AI team found that AI agents interacting with real-world environments can double their learning speed every three months — a pattern the researchers call a new scaling law, potentially offering an alternative path as the industry's data supply runs thin.
What does this "new law" actually say?
ByteDance researchers found that AI agents — autonomous systems that carry out tasks on their own — double their learning speed every three months when operating continuously in real environments.
The team defines this pattern as a new scaling law, meaning a predictable rule linking a growing input to growing AI capability.
This means → the old scaling law ran on "feed more data and compute." This new one runs on how long the agent interacts with its environment — the key resource shifts from data volume to practice time.
Why does the industry need a new path right now?
The global AI industry has long relied on "more data + more compute" to boost model performance, but that approach is hitting a ceiling.
U.S. research group Epoch AI recently warned that publicly available human-generated text data could be exhausted within six years.
OpenAI co-founder Andrej Karpathy has also said publicly that brute-force scaling of compute and data is unsustainable.
In plain terms = AI used to grow by "reading more books," but the books are running out — the field needs a way to keep improving without new ones.
How did they test it?
The team built a benchmark suite called EdgeBench, covering 134 ultra-long-horizon tasks.
Tasks span software engineering, scientific discovery, formal mathematics, and professional knowledge work; each requires an AI agent to run continuously for at least 12 hours.
This means → this is not a traditional "answer one question, get a score" test. It measures whether an AI can work autonomously over long stretches and improve as it goes.
How far is this finding from real-world impact?
The paper was published on July 3. The researchers themselves acknowledge that how AI agents "learn from real environments after deployment is still far from well understood."
If this pattern holds at larger scales and across broader scenarios, it would give the industry a capability-growth path that does not depend on massive new supplies of human data.
Whether it will hold, however, remains to be tested — this reflects a finding still at the "exciting early signal" stage, not yet an established industry consensus.
Content is for reference only, not financial advice.