Wenxin Yiyan vs GPT-4: Live Demo Shows Baidu’s New AI Model in Action

One day after GPT‑4’s launch, Baidu unveiled its next‑generation large language model, Wenxin Yiyan (Ernie Bot), in a pre‑recorded demo that directly compares its capabilities with GPT‑4.

Wenxin Yiyan vs GPT‑4

Both models are multimodal large models. Baidu’s CEO Li Yanhong highlighted five abilities: literary creation, business copywriting, mathematical logic, Chinese understanding, and multimodal generation.

Literary Creation

Li introduced Liu Cixin and asked Wenxin Yiyan to describe him. The model responded sensibly, while GPT‑4 mistakenly placed Liu’s hometown in Hubei. A follow‑up philosophical continuation demo showed Wenxin Yiyan producing a coherent extension, which was then compared with GPT‑4’s continuation of the same prompt.

Business Copywriting

The demo generated a name for a new company, then explained the reasoning behind the choice. The same task was given to GPT‑4, whose Chinese output appeared less nuanced. A news‑release draft generated by Wenxin Yiyan was also shown.

Mathematical Logic

A classic “chicken‑rabbit” problem from elementary math competitions was posed. Wenxin Yiyan produced a step‑by‑step solution, even quoting a line from the TV series “狂飙”. After correcting the problem statement, the model gave a logical answer.

Chinese Understanding

The model was asked about the idiom “洛阳纸贵”. Wenxin Yiyan explained the phrase and linked it to an economic principle, then GPT‑4 was queried for the same meaning and the underlying theory. Both showed strong Chinese comprehension, though Wenxin Yiyan’s cultural nuance was judged slightly superior.

Acrostic Poetry

Wenxin Yiyan generated an acrostic poem correctly, while GPT‑4’s output indicated a misunderstanding of the format.

Multimodal Generation

The demo created a poster for the 2023 World Intelligent Transportation Conference and then turned the same content into a short video with subtitles in a few seconds. Programming capabilities were not demonstrated, though Baidu noted that code data is included in the training set.

How Wenxin Yiyan Works

Wenxin Yiyan builds on Baidu’s ERNIE knowledge‑enhanced trillion‑parameter model and the large‑scale open‑domain dialogue model PLATO. Six core technologies are employed:

Supervised fine‑tuning (especially on Chinese data)

Reinforcement learning from human feedback (RLHF)

Prompt engineering

Knowledge enhancement (knowledge internalization and externalization)

Retrieval enhancement (integrating Baidu’s search retrieval)

Dialogue enhancement (memory, context understanding, planning)

When model parameters reach the hundred‑billion scale and training data is sufficient, Baidu claims “intelligent emergence” occurs.

When parameters reach the hundred‑billion level and the training corpus is large enough, this phenomenon emerges.

Baidu’s AI stack comprises four parts: Kunlun chips, the PaddlePaddle framework, the Wenxin model series, and applications. Integrating these layers is intended to lower the high compute cost of generative AI.

Generative AI demands massive compute and expense; coordinated optimization across the four layers can boost efficiency and significantly reduce cost.

Li Yanhong emphasized that having leading products in all four layers is unique to Baidu.

Companies with leading products in all four layers are rare.

The model is hosted on Baidu Intelligent Cloud, already integrated into Baidu Search, Xiaodu, Apollo autonomous driving, iQIYI, and other services. Public reaction was mixed: some praised the Chinese cultural grasp, while others felt the demo was rushed and the model still lagged behind GPT‑4 in English tasks.

It is still about 20 “old Hu” away from GPT‑4.

At the end of the event, Baidu announced that Wenxin Yiyan would open beta testing for individual and enterprise users.

One More Thing

A user shared an invitation link for the internal test: https://yiyan.baidu.com/welcome (code KFCVME50RMB, valid until 2023‑03‑16 24:00).