Baidu’s Ernie Bot (Wenxin Yiyan) vs GPT‑4: Capabilities, Technical Foundations, and Market Reaction

Wenxin Yiyan vs GPT‑4

Both Wenxin Yiyan and GPT‑4 are multimodal large language models; Baidu demonstrated five core abilities of Wenxin Yiyan—literary creation, business copywriting, mathematical reasoning, Chinese comprehension, and multimodal generation—through a series of pre‑recorded demos and compared the results with GPT‑4.

Literary Creation

In the literary demo, Wenxin Yiyan was asked to continue a passage from Liu Cixin’s "Three‑Body" series, producing a coherent continuation that was then juxtaposed with GPT‑4’s output.

The model also generated a philosophical extension of the text, which was compared side‑by‑side with GPT‑4’s continuation.

Business Copywriting

Wenxin Yiyan generated a brand‑new company name and a short news release, showing an understanding of Chinese branding conventions. The article notes that GPT‑4’s Chinese output was slightly less idiomatic.

Mathematical Reasoning

The model solved a classic "chick‑and‑rabbit" problem from elementary math competitions, demonstrating step‑by‑step logical reasoning. A corrected version of the problem was later fed to the model, which produced a sensible solution.

Chinese Understanding

Wenxin Yiyan answered a cultural question about the idiom "洛阳纸贵", explained its economic background, and generated a hidden‑acrostic poem. GPT‑4 answered the same questions but showed weaker grasp of the idiom’s nuance.

Multimodal Generation

The demo included generating a poster for the 2023 World Intelligent Transportation Conference and converting the displayed text into a short video with subtitles, all triggered by a single spoken command.

How Wenxin Yiyan Works

Wenxin Yiyan is built on Baidu’s ERNIE series and the PLATO open‑domain dialogue model. It incorporates six core technologies: supervised fine‑tuning, reinforcement learning from human feedback (RLHF), prompt engineering, knowledge enhancement (both internalized and external), retrieval enhancement, and dialogue enhancement.

Supervised fine‑tuning uses Chinese‑specific data; RLHF and prompt construction follow the same principles as OpenAI’s models. Knowledge enhancement injects factual information into model parameters and allows external knowledge lookup. Retrieval enhancement combines Baidu’s search capabilities with generation, while dialogue enhancement leverages memory, context understanding, and planning.

Baidu’s hardware stack—Kunlun chips, the PaddlePaddle framework, the Ernie model, and Baidu Cloud—provides the massive compute needed for a trillion‑parameter model, aiming to lower cost through cross‑layer optimization.

Market Reaction and Outlook

Following the launch, Baidu’s Hong Kong‑listed shares fell sharply before partially recovering. Netizens posted mixed reactions, ranging from jokes about “early retirement” to calls for patience with the domestic product.

Wenxin Yiyan has been integrated into Baidu Search, the Xiaodu assistant, Apollo autonomous‑driving platform, and iQIYI services. Starting today, Baidu opened external testing to both individual and enterprise users.