Why Large Models Signal the Dawn of General AI: Insights from Baidu’s CTO

Key Takeaways from the 2024 Beijing Zhiyuan Conference

On June 14, 2024, Baidu’s Chief Technology Officer Wang Haifeng delivered a keynote at the Beijing Zhiyuan Conference, offering a forward‑looking assessment of artificial‑intelligence development. He argued that large models illuminate the path toward artificial general intelligence (AGI) and can be understood from two angles: the universality of AI technology and the comprehensiveness of AI capabilities.

From Rule‑Based Systems to Large‑Model Era

Wang traced the evolution of AI over several decades: early rule‑based systems → statistical machine learning that learns from data → deep learning, which dramatically increased algorithmic universality, allowing a single neural‑network architecture to address many problems → the current large‑model era, where both algorithms and models become more unified and reusable.

Technological Universality of Large Models

Large models now exhibit strong cross‑task, cross‑language, and cross‑modality generality. In natural‑language processing, a single large language model can handle tasks that previously required separate specialized models (e.g., tokenization, parsing, translation, QA, dialogue). The same models also support multilingual and multimodal inputs, bridging human language, formal languages, and perception.

Comprehensive Capabilities Required for AGI

Wang identified four foundational capabilities—understanding, generation, reasoning, and memory—that underpin higher‑level abilities such as creativity, problem solving, coding, planning, and decision‑making. Strengthening these core skills brings AI closer to AGI.

Inside Baidu’s “Wenxin” Large Model

“Wenxin YiYan” is Baidu’s internally developed, knowledge‑enhanced large language model. It leverages a stronger platform, richer data, and improved algorithms to ingest trillions of tokens and billions of knowledge facts. Innovations include knowledge‑enhanced pre‑training, retrieval‑augmented generation, and advanced alignment techniques.

The model also incorporates “agents” that perform supervised fine‑tuning of reasoning processes, preference learning for decision making, and reflective reinforcement learning. A code‑agent, for example, translates user intent into executable code via a dedicated code interpreter.

Public data show that Baidu has been building AI since 2010, released its first Wenxin model in March 2019, and launched version 4.0 in October 2023. Training efficiency has risen dramatically: weekly effective training utilization now reaches 98.8%, a 5.1‑fold increase over the initial release, while inference speed is about 105× faster.

AI Entering Industrial‑Scale Production

Wang compared AI’s trajectory to the three previous industrial revolutions—mechanical, electrical, and information technology—each characterized by standardization, automation, and modularity. He asserted that deep‑learning and large‑model engineering platforms now possess these traits, ushering AI into a phase of mass industrial production and accelerating the arrival of AGI.

He also emphasized that the “scale law” will remain valid for several more years, large language models still have ample room for improvement, multimodal models will become increasingly useful, and agent technologies will mature and spark widespread adoption.