Ex-Alibaba engineer Lin Junyang says large models are moving from “inference” to “agent” paradigms

The claim and its punch

It has been reported that Lin Junyang posted a long essay after leaving Alibaba (阿里), arguing that the evolution of large AI models is shifting away from traditional inference-based systems toward an “agent” paradigm. Short answer: models will no longer be passive answer machines. They will act, plan and chain tools — becoming active orchestrators rather than simple predictors.

What the shift means

Inference-based models generate outputs in response to prompts. Agent paradigms combine planning, tool use, multi-step decision-making and environmental interaction — think autonomous assistants that call APIs, fetch data, and carry out workflows. Why does this matter? Because building agents changes product design, requires new infrastructure and raises fresh safety, alignment and governance questions. It’s not just bigger models; it’s systems engineering on top of models.

China’s tech context and geopolitical backdrop

China’s major AI players — including Alibaba (阿里), Baidu (百度) and others — have been racing to commercialize large models, reportedly accelerating agent-style research to differentiate products. Geopolitics is relevant: U.S. export controls on advanced chips and broader trade frictions make software architectures and efficient system-level designs strategically important for Chinese firms, since hardware access is constrained. Will Chinese companies double down on orchestration and software innovation while seeking chip independence?

Broader implications

Lin’s essay feeds a wider industry debate: are we entering an era of autonomous AI services or a refinement of inference-as-a-service? The answer will shape hiring, cloud and tool partnerships, and regulatory scrutiny at home and abroad. Reportedly, the transition from inference to agents won’t be overnight — but it could define the next phase of China’s AI strategy.