Yang Zhi Lin's Zhongguancun Forum Speech: Large-scale Model Training is Entering the Third Stage, With Almost No Increase in Execution Time After Parallel Execution of 100 Agents

According to 1M AI News monitoring, Yang Zhilin, the founder of the Dark Side of the Moon, delivered a speech titled "Open Source AI: Accelerating the Exploration of Intelligence Limits" at the 2026 Zhongguancun Forum plenary session, proposing that large-scale model training is undergoing three stages of evolution: three years ago, it mainly relied on natural data from the Internet with a small amount of manual annotation; in 2025, it shifted to large-scale reinforcement learning, with human curation of high-quality tasks, and performance improvements in the programming and mathematics fields primarily coming from this direction; starting this year, it enters the third stage, where AI will play a greater role in research itself, with each researcher equipped with a massive number of tokens, with AI automatically synthesizing new tasks and environments, defining reward functions, and even exploring entirely new network architectures.

Yang Zhilin introduced the pioneering Agent Swarm technology in Kimi K2.5. The data he presented showed that under a single-agent mode, task complexity and execution time grew exponentially, but after enabling parallelism with 100 agents collaborating, execution time almost no longer significantly increased as task complexity rose. He likened this to "building a billion-dollar company from 0 to 1, which may take one person 100 years, but with 100 smart people collaborating, it can be completed in a short time."

Yang Zhilin also pointed out that open-source models are becoming the new standard in the global AI industry. At the 2026 NVIDIA GTC conference, Huang Renxun repeatedly referenced open-source models like Kimi as chip performance benchmarks, stating, "If you are going to release a new chip, you must evaluate performance improvement through Kimi or other open-source models."