Zhiyuan Liu

Overview

Zhiyuan Liu is a tenured associate professor (长聘副教授) at Tsinghua University and the co-founder and chief scientist of Beijing ModelBest (面壁智能), a company that builds large models ^[1]. His public work sits at the meeting point of academic natural language processing and the engineering of models small enough to run outside data centres.

At Tsinghua

Before founding ModelBest, Liu was an associate professor in Tsinghua University's Department of Computer Science, working in its natural language processing lab ^[1]. Natural language processing is the branch of computing concerned with getting machines to handle human language — reading, summarising, translating, answering questions — and it is the research tradition from which today's large language models grew. His later company role grew out of that base: the Tsinghua article describing him places both identities side by side, as tenured associate professor and as co-founder and chief scientist of ModelBest ^[1].

Founding ModelBest

In 2022, while still an associate professor in Tsinghua's Department of Computer Science, he led the founding of ModelBest ^[1]. The company's stated direction, as he has described it, is not to compete at the largest possible scale but to make strong models small enough to sit on a device.

The "Moore's Law" of the large model era

On the morning of 5 July, Liu spoke at a sub-forum of WAIC 2024 under the title "The 'Moore's Law' of the Large Model Era" ^[1]. The reference is to the observation, familiar from semiconductors, that a fixed amount of computing power becomes steadily cheaper over time; he applied a version of that idea to models. In the talk he argued that the knowledge density of models doubles roughly every eight months ^[1]. Knowledge density, in this usage, describes how much knowledge a model carries relative to its size. The practical reading is that capability which once needed a very large model can, over a relatively short period, be delivered by a much smaller one — the same ability compressed into fewer parameters.

Running models on devices

That way of thinking leads directly to concrete targets. Liu said ModelBest aims to run GPT-3.5-level models on edge devices within half a year to a year, and to reach GPT-4-level capability on device within two years ^[1]. Edge devices are the machines people actually carry and use — phones, laptops, and similar hardware — where computation happens locally rather than being sent to a distant data centre. Running a capable model there changes the constraints: memory, power and heat matter more than raw scale, and data can stay on the device.

Why the position matters

Liu's public argument links a research trend to an engineering programme. If knowledge density really does double on a roughly eight-month rhythm, then the frontier of capability and the frontier of smallness move together, and a company that bets on compact models is betting on that convergence rather than against it ^[1]. His two stated targets — GPT-3.5-level models on edge devices within half a year to a year, and GPT-4-level capability on device within two years — are the concrete form that bet took in his own account ^[1]. The Tsinghua article reports both the academic position and the company role as parts of one trajectory: an associate professor in a natural language processing lab who, in 2022, led the founding of a company and became its chief scientist ^[1].