On March 16, 2023, a Baidu executive stood on stage as the company's new chatbot fumbled part of a live demo. Competitors smirked. What almost nobody in the room appreciated was that the awkward moment sat on top of roughly ten years of quiet engineering — and that the person most responsible for it was not the CEO, but the chief technology officer.
Wang Haifeng (王海峰) is the technologist behind Baidu's AI. If Li Yanhong is the face of the company, Wang is the spine — the researcher who built the underlying stack that made a Chinese answer to GPT-style chatbots possible at all.
The linguist who became Baidu's CTO
Wang holds a PhD from the Harbin Institute of Technology and joined Baidu in 2010. Over the following years he created and grew the company's core AI capabilities — natural language processing, knowledge graphs, speech, vision, machine learning, and deep learning — and took charge of products like Baidu Search, Baidu Maps, Baidu Translate, and Baidu AI Cloud. Today he is Baidu's Chief Technology Officer and directs the National Engineering Research Center for Deep Learning Technology and Application.
His academic standing is unusual for an industry executive. In 2013 he became the first Chinese president of the Association for Computational Linguistics (ACL) in the organization's then-50-year history, and in 2016 he was elected an ACL Fellow — the first scientist from mainland China to receive that honor. He was named an IEEE Fellow in 2022 and is a member of the International Eurasian Academy of Sciences. That mix of scholarly credential and shipping record is precisely why Baidu trusted him to lead its model program.
Building the full stack
Wang's central thesis is that modern AI cannot be won by renting one layer. He argues for a full-stack layout spanning four layers: a chip layer (Baidu's Kunlun), a framework layer (its PaddlePaddle deep-learning platform), a model layer (the ERNIE (文心) family), and an application layer — layers he says feed each other for end-to-end efficiency a piecemeal stack cannot match.
He has stated this view consistently for years. At the first WAVE SUMMIT developer conference in 2019, he declared that "the deep-learning framework is the operating system of the intelligent era." PaddlePaddle (飞桨), which he steered into China's first open-source, industrial-grade deep-learning platform, is the embodiment of that belief. By late 2023 the platform had gathered about 10.7 million developers, served roughly 235,000 enterprises, and hosted some 860,000 created models — numbers Baidu announced at its tenth WAVE SUMMIT in December 2023.
The ERNIE (文心) lineage
The model layer is where Wang's name is now inseparable from Baidu's story. The company released ERNIE (文心) 1.0 in 2019 as a knowledge-enhanced large model (大模型), years before the chatbots race began. The defining idea was "knowledge enhancement": rather than learning only from text, the model fuses a massive knowledge graph into training and inference. Baidu's graph eventually grew to around 5 billion entities and 550 billion facts, and Wang framed the approach as the reason ERNIE (文心) would "understand Chinese and Chinese culture" better than a generic model.
The chatbot, Wenxin Yiyan (文心一言), launched in March 2023 and was built on top of the ERNIE and PLATO dialogue models, trained and deployed on PaddlePaddle (飞桨). Wang described it as "the natural result of Baidu's years of technical accumulation and industrial practice." The base model was upgraded to ERNIE (文心) 4.0 in October 2023, and Baidu said the 4.0 release improved its overall effect by about 32 percent in the two months after launch. By late December 2023, Wenxin Yiyan had passed 100 million users.
Knowledge as the differentiator
Wang's technical wager has been consistent: China's models do not have to win by copying English-first training regimes. By leaning on a Chinese knowledge graph and on retrieval from Baidu's search index, ERNIE (文心) could be stronger on Chinese language, culture, and local context. The name itself signals the intent — "文" for language and "心" for understanding with care, echoing the classical treatise Wenxin Diaolong (《文心雕龙》, "The Literary Mind and the Carving of Dragons").
He also resists the idea that a model is a finished product. In his framing, the large model (大模型) is a foundation that must be wrapped in safety rails, domain tuning, and tool use to be useful. Baidu built what it calls a five-line safety defense across data, training, and deployment, and Wang has argued that models should be delivered as low-threshold production platforms — "foundries" that turn algorithm, compute, and data advantages into services for every industry.
Reading the trend
Wang's career mirrors a broader Chinese strategy: build the stack yourself rather than depend on foreign layers. The ERNIE (文心) lineage shows the long game — a knowledge model started in 2019, a framework started even earlier, a chip effort underneath — converging exactly when the chatbot moment arrived.
The lineage in plain terms
Wang Haifeng is the through-line connecting Baidu's research lab to its shipping products. The ERNIE (文心) story is less a single launch than a stack: Kunlun chip, PaddlePaddle (飞桨) framework, ERNIE (文心) model, and applications reinforcing one another. For readers tracking China's model race, the lesson from Wang's work is that durability may come less from one flagship release and more from owning the layers beneath it. The markers to watch are PaddlePaddle's developer growth, ERNIE (文心) 4.0-and-later adoption in enterprises, and whether the knowledge-enhancement approach holds up against ever-larger general models.
Honest limitations
Roles, dates, and numbers are drawn from Xinhua (news.cn) coverage of Wang's CCTV lecture and the March 2023 launch, ScienceNet (科学网) reporting on the Wenxin Yiyan tech briefing, China Daily (中国日报) and the Beijing Municipal Science and Technology Commission (ncsti.gov.cn) on his ACL presidency and knowledge-graph scale, and Baidu's own WAVE SUMMIT announcements — cross-checked against each other. The "operating system of the intelligent era" line and the "natural result of years of accumulation" phrasing come from Xinhua's reporting; translations follow the original Chinese. The 10.7 million developers, 235,000 enterprises, 860,000 models, 100 million Wenxin Yiyan users, and the ~32 percent 4.0 improvement are Baidu-announced figures from December 2023; we have not re-verified them against later disclosures. The 5-billion-entity / 550-billion-fact knowledge-graph figures are from the Beijing Science and Technology Commission profile and may predate later expansions. We did not assess ERNIE's current benchmark standing versus competing models, and Wang's "knowledge enhancement" claims are presented as Baidu's technical positioning, not an independent performance verdict.
What readers can do now
- Follow Baidu's WAVE SUMMIT developer conferences to track PaddlePaddle (飞桨) developer and model counts as a proxy for China's homegrown AI-framework adoption.
- Compare ERNIE (文心) 4.0-and-later releases against open-weight Chinese models to judge whether knowledge enhancement still differentiates at scale.
- Read Wang's CCTV "China Economic Lecture" (via Xinhua) for his own plain-language explanation of the four-layer stack and the six core technologies behind Wenxin Yiyan (文心一言).