ByteDance and Li Han: How to Make Neural Networks Work, Not Just Talk

Dr. Li Han, head of ByteDance's AI lab, has published a paper on a universal framework for AI agents built on four modules — perception, planning, memory, and action. According to Jiqizhixin (机器之心), the goal of the development is to replace today's experimental agent systems like AutoGPT and BabyAGI, which often "break down at the second step and hallucinate in endless loops," with a reliable industrial-grade tool.

Who is Li Han

Li Han is a figure whose name in the natural language processing world carries weight equivalent, as the article puts it, to "roughly the entire server fleet of a mid-sized startup." His career path took him through Microsoft Research Asia and Huawei before he took the position of head of ByteDance's AI lab. This is an applied track rather than an academic career: according to the author of the piece, such a background suggests that Li Han is interested not in the theoretical elegance of architectures but in their applied power — that is, translating ideas from PDF files into real products.

Four architectural nodes

The key problem in modern language models that the framework addresses is the inability to plan ahead: the model "lives in the moment," generating the next token rather than building a strategy. The proposed architecture focuses on four critical nodes — perception, planning, memory, and action. The model first builds a hierarchical tree of goals, then uses external tools to achieve them, constantly checking against long-term memory. The author of the piece describes the effect this way: the approach turns AI "from a literature student into an experienced project manager."

Why this matters for business, and who the competitor is

For AI to generate real money in the B2B sector, it needs to independently log into a CRM, analyze data, compile a report, and send it to the client without asking for permission at every step — this is exactly the scenario that Li Han's framework addresses. ByteDance, which operates with colossal volumes of data and complex recommendation algorithms, has more interest in such autonomous systems than anyone else. The development is described as a direct challenge to Microsoft's AutoGen project: while Western companies, in the author's view, more often rely on the "raw power" of computation, ByteDance is betting on a structural, elegant, and scalable engineering scheme.

Frequently Asked Questions

What exactly did Li Han propose?

A universal framework for AI agents based on four modules — perception, planning, memory, and action. Instead of generating a response "in the moment," the model first builds a hierarchical tree of goals, then achieves them through external tools, constantly checking against long-term memory.

How is this different from AutoGPT and BabyAGI?

Existing agent systems like AutoGPT and BabyAGI, as the article describes them, are unstable: they break down at early steps of task execution and fall into hallucination loops. Li Han's framework introduces a systematic architecture for planning and memory that is meant to turn such "toy" agent prototypes into a reliable industrial tool.

Why could this be a challenge for Microsoft AutoGen?

Because both projects solve the same problem — building autonomous, multi-step AI agents — but, in the author's view, by different paths: AutoGen and other Western developments more often rely on computing power, while ByteDance, through Li Han, offers a structural engineering scheme. The outcome of the competition will depend on who moves the architecture from an academic paper into production faster.