文心一言开启中国大模型浪潮

2023 年 3 月 16 日，百度在北京发布文心一言。从时间上看，它几乎是 ChatGPT 之后中国主要科技公司里第一个正式拿出大模型产品的。发布会上，李彦宏以录播视频演示了文心一言的对话、文案创作、多模态理解等能力。但这场发布会本身，也成了争议的开端——「录播演示」这个形式，被市场解读为底气不足的信号。

百度这家公司，在 AI 上并非新手。它从 2010 年代就开始投入深度学习，2019 年发布过文心系列语言模型，自动驾驶项目 Apollo 也跑了多年。但大模型产品化是另一回事。文心一言背后的 ERNIE 系列模型，主打中文理解与知识增强，技术路线与 GPT 有差异。发布会上演示的每个功能，在录播背后都有真实系统的支撑，但观众无法当场验证。

市场的反应很直白：发布会当日，百度港股股价一度下跌超过 10%。「演示即发布」的质疑在舆论中扩散——人们对比 OpenAI 的实机演示，认为百度的录播缺乏说服力。百度方面则回应，录播是因为产品仍需完善，选择「先发」是为了抢下市场心智。这场「抢跑 vs 实力」的争论，成了中国大模型浪潮的第一个公共议题。

文心一言的意义，不在它当时的能力排名。它是中国大模型竞赛的发令枪——发布会之后不到两个月，讯飞星火、通义千问、商汤日日新等国产大模型密集发布，「百模大战」正式开打。互联网大厂、AI 四小龙、创业公司全线入场，中国 AI 产业的叙事从「感知智能」整体切换到了「生成式大模型」。

回看这场发布会，录播演示的争议反而成了一个注脚：它说明 2023 年春天的中国 AI 市场，情绪已经热到「晚一天发布就落后一年」的程度。百度用一场不算完美的发布会，把整个行业推过了起跑线。两年后，当国产大模型的能力差距被逐步拉近、甚至在某些领域反超时，回望那个 3 月的午后，文心一言确实是那个按下发令枪的产品。

On March 16, 2023 Baidu held a launch event in Beijing for Ernie Bot. By timing alone it was nearly the first Chinese major tech company to formally ship a large-model product after ChatGPT. On stage, Li Yanhong demonstrated Ernie Bot's dialogue, copywriting, and multimodal understanding via prerecorded video. But the event itself also became the start of controversy—the "prerecorded demo" format was read by the market as a signal of insufficient confidence.

Baidu is no newcomer to AI. It invested in deep learning from the 2010s, released the Ernie language-model line in 2019, and ran the Apollo self-driving project for years. Productizing a large model is another matter. Ernie Bot sat on the ERNIE series, emphasizing Chinese-language understanding and knowledge enhancement, a technically different route from GPT. Every demonstrated function had a real system behind the recording, but the audience could not verify it live.

The market's reaction was blunt: Baidu's Hong Kong stock fell more than 10% at one point on launch day. The "demo-as-launch" criticism spread—people compared OpenAI's live demos and found the recording unconvincing. Baidu responded that recording was chosen because the product still needed polish, and that releasing first was about capturing market mindshare. This "rushing versus substance" argument became the first public issue of China's large-model wave.

Ernie Bot's significance lies not in its capability ranking at the time. It was the starting gun for China's large-model race—within two months, domestic models like iFlytek Spark, Tongyi Qianwen, and SenseNova released in rapid succession, and the "war of a hundred models" began. Internet giants, the four AI dragons, and startups all entered, shifting China's AI story from perception intelligence to generative large models wholesale.

Looking back, the prerecorded-demo controversy became a footnote: it shows that by spring 2023, China's AI market sentiment had heated to the point where "releasing a day late means falling a year behind." Baidu pushed the whole industry past the starting line with a less-than-perfect event. Two years later, when domestic large models had narrowed the capability gap and even overtaken in some areas, that March afternoon stands as the moment Ernie Bot pressed the starting gun.