Thirty-eight thousand hours. That is the total time ByteDance researchers logged watching AI agents work through 134 tasks in a new benchmarking system called EdgeBench. Each task demanded at least 12 continuous hours of autonomous operation.
The result: a scaling law that says these agents can double how fast they learn every three months. The numbers tell a story the AI industry has been waiting to hear.
For years, the dominant playbook was simple — feed a system more data, give it more computing power, get a smarter model. But that playbook has a shelf life. Epoch AI, a US-based research institute, recently warned that publicly available human-generated text data could run out within six years.
The brute-force method is hitting a wall. ByteDance’s Seed AI team published their findings Thursday.
Their paper tackles a gap that has nagged the field: how autonomous systems learn after they are deployed, out in the real world, away from controlled training environments. The answer, based on their numbers, is that they keep getting better — and at a predictable rate. The 134 tasks in EdgeBench are not simple.
They cover software engineering, scientific discovery, formal mathematics, and professional knowledge work. Each one requires the AI agent to operate for half a day straight. That is a long time in an industry where benchmarks often measure minutes or seconds.
The 38,000 hours of environment interaction gave the researchers enough data to spot the pattern. The timing matters.
Tech firms are pivoting hard toward what is called agentic AI — autonomous software that carries out tasks on a human’s behalf. But as the ByteDance researchers noted in their paper, how these systems learn from real-world environments after deployment remains far less understood. Their work directly addresses that blind spot.
Prominent figures in AI have been sounding alarms. OpenAI co-founder Andrej Karpathy has warned that the old approach of throwing more data and more computing power at the problem cannot last forever.
The ByteDance numbers offer a concrete alternative — a path that does not depend on an endless supply of human-written text. Doubling learning speed every three months is not a trivial gain. Compound that over a year and the improvement is dramatic.
Over two years, it is transformative. The researchers built EdgeBench specifically to measure ultra-long-horizon tasks — the kind of sustained work that separates real-world usefulness from lab demonstrations. What is missing from the numbers is also telling.
The paper does not claim this scaling law solves everything. It describes a mechanism for improvement, not a ceiling.
The 38,000 hours of logged data came from controlled experiments. Real-world deployment will introduce messier variables. But for an industry staring at a data drought, the finding provides something concrete to work with.
Instead of racing to scrape the last scraps of human text, companies can bet on agents that learn by doing. The ByteDance researchers have given them a number to build on: double every three months.
{“@context”: “https://schema.org”, “@type”: “NewsArticle”, “headline”: “ByteDance’s 38,000-Hour Study Reveals AI Agents Double Learning Speed Quarterly”, “description”: “Thirty-eight thousand hours. That is the total time ByteDance researchers logged watching AI agents work through 134 tasks in a new benchmarking system called EdgeBench….”, “datePublished”: “2026-07-03T15:39:00+00:00T09:30:00+00:00”, “author”: {“@type”: “Person”, “name”: “Sophie King”}, “publisher”: {“@type”: “Organization”, “name”: “InfoPulse Today”, “url”: “https://newsfeed.infopulsetoday.com/”}}
