Baidu's Hou Zhenyu: Large Models Are Reshaping Cloud Computing, AI-Native Cloud Will Change the Landscape of Cloud Computing

[baoshi.rao]

AI Home News, December 20th - At the 2023 Baidu Cloud Intelligence Conference, Baidu Group Vice President Hou Zhenyu delivered a keynote speech titled 'Large Models Reshaping Cloud Computing.' He emphasized that in the AI-native era, the infrastructure system for large models needs comprehensive reconstruction to build a solid foundation for a thriving AI-native ecosystem.

Hou Zhenyu stated: 'The reconstruction of cloud computing by large models is mainly reflected in three aspects: AI-native cloud will change the landscape of cloud computing, MaaS (Model as a Service) will become a new fundamental service, and AI-native applications will give rise to new development paradigms.'

At the underlying cloud infrastructure level, while traditional internet and mobile internet applications were based on CPU computing chips, AI applications have significantly increased the demand for GPU or heterogeneous computing. The underlying computing power of the cloud market is shifting towards GPU dominance.

In the third quarter of 2023, NVIDIA's revenue surpassed Intel's, and NVIDIA's latest market capitalization exceeded Intel's by $1 trillion. In the future, GPU growth will far outpace CPU growth. Given this trend, we need to comprehensively reconstruct the cloud computing infrastructure system for large models to support the deployment of AI-native application systems.

Specifically, the comprehensive reconstruction of cloud computing will manifest in three major areas: intelligent computing infrastructure for models, data infrastructure for data, and cloud infrastructure for applications—all upgraded to make computing more intelligent.

MaaS will significantly lower the barriers to AI adoption and achieve true AI democratization. The new IT infrastructure it relies on will further disrupt the existing cloud computing market landscape at the foundational level.

From Baidu Intelligent Cloud's practical observations, the four months following the full public release of Wenxin Yiyan (ERNIE Bot) on August 31 have seen a 10-fold daily increase in API calls on the Qiansu Large Model Platform (Baidu's MaaS platform). Primary adopters come from internet, education, e-commerce, marketing, mobile devices, automotive and other industries. It's evident that numerous enterprises have begun actively implementing large models in the past six months.

The unique capabilities of large models in comprehension, generation, logical reasoning, and memory are fostering a new paradigm for AI-native application development, fundamentally transforming the entire application technology stack, data flow, and business processes.

Traditional CPU-based development was primarily business logic-driven, while conventional AI development required separate data collection and model training for each scenario. Contemporary AI-native applications leverage large model capabilities for data-driven development. Enterprises can now fine-tune proprietary models using scenario-specific data atop foundational models, then design AI-native applications without building models from scratch. As business scales expand, accumulated competitive scenario data further enhances model and application performance, creating a self-reinforcing data flywheel effect.

Specifically, the new large model-driven development paradigm manifests several transformative aspects:

1. New Scenarios: Generative large language models demonstrate unexpectedly strong capabilities across comprehension, generation, reasoning, and memory dimensions, enabling intelligent emergence. This has spawned numerous implementable business scenarios like personal assistants, intelligent copywriting, GBI (Guided Business Intelligence), and coding assistants.

2. New Architectures: Implementing large models in these novel scenarios has generated innovative system architectures such as Retrieval-Augmented Generation (RAG) and Agent frameworks.

The third is the 'new development ecosystem'. With large models at the core, some new tools have emerged in the developer tooling layer, including the orchestration tool LangChain, AI application development tool PromptFlow, and data framework Llamalndex.

Hou Zhenyu stated that overall, building a thriving AI-native application ecosystem requires the synergy of three key elements: large models, intelligent computing power, and a new paradigm for AI-native application development. Large models serve as the 'brain' of AI-native applications, intelligent computing provides solid support for their operation, and the new development paradigm helps developers efficiently build applications based on large model capabilities. The data flywheel is both a necessary and sufficient condition for successful AI-native applications, enabling rapid iteration of large model capabilities and continuous improvement in product experience.