How Baidu’s AI Tools Turn Everyone Into a Developer – Key Takeaways from Li Yanhong’s Speech

Li Yanhong opened the 2024 Baidu AI Developer Conference (Create) by noting that over 5,000 developers and tech enthusiasts attended, all feeling a mix of excitement and fear of missing out (FOMO) as large models and generative AI reshape the developer landscape.

Evolution of Baidu’s Large Models

Baidu now offers a family of foundation models called the Wenxin series, including flagship versions ERNIE 3.5 and ERNIE 4.0, as well as lightweight variants ERNIE Speed, ERNIE Lite, and ERNIE Tiny. Recent upgrades have increased training efficiency by 5.1×, raised weekly training utilization to 98.8%, and improved inference speed by 105× while cutting inference cost to 1% of the previous level.

AI‑Native Development Tools

The company introduced three zero‑code/low‑code tools designed to let anyone create AI applications:

AgentBuilder : an intelligent‑agent builder that supports multimodal knowledge bases, tool integration (e.g., hotel booking, ticket purchasing), and rapid deployment of industry‑specific agents such as a Singapore tourism guide or education‑sector advisors.

AppBuilder : a platform that packages pre‑built components and frameworks, enabling developers to create AI‑native apps in three steps—name the app, define role instructions, and add tool components (e.g., code interpreter, data analysis). It also features AI‑optimized configuration that automatically refines prompts and component settings.

ModelBuilder : a model‑customization service that lets developers select a base model (e.g., ERNIE Speed), configure fine‑tuning parameters, and deploy the resulting model. It supports data cleaning, annotation, and augmentation, and can produce high‑quality fine‑tuned models for specialized tasks such as essay grading.

Performance Metrics and Adoption

Wenxin Yiyan, launched in March 2023, now exceeds 200 million users, processes over 200 million API calls daily, and serves 85 000 customers. The intelligent‑code assistant Comate, built on Wenxin models, supports 100+ programming languages, integrates with all major IDEs, and has achieved a 46% code‑adoption rate across tens of thousands of enterprises.

Real‑World Case Studies

Several demos illustrated the tools in action:

Singapore Tourism Agent : Using AgentBuilder, a tourism‑assistant was created that answers travel‑time queries, recommends hotels, and books tickets via integrated third‑party services.

Qidi Education : An education‑sector agent handles student inquiries, provides personalized study advice, and reduces effective lead‑conversion cost by 30% while generating 1.55 million interactions in its first week.

Sofia Home‑Furnishing : A digital sales agent replicates a physical showroom experience, offering product recommendations, price quotes, and store navigation.

University AI Assistant : A campus‑wide AI assistant built with AppBuilder supports course lookup, library borrowing, and campus‑service queries, demonstrating rapid deployment for institutional use.

Model‑Routing (MoE) Strategy

Li described a mixture‑of‑experts (MoE) approach where small, fast models handle routine queries (e.g., weather, simple scheduling) while the most capable ERNIE 4.0 model processes complex tasks such as multi‑step itinerary planning. This routing yields a two‑fold speed increase and a 99% cost reduction compared with using only the flagship model.

Beyond Text – Multimodal and Autonomous Driving

Ba​idu is also advancing multimodal models, including the Apollo visual‑perception model that combines detection, tracking, understanding, and mapping for autonomous driving. The model is trained on over 100 million kilometers of Chinese city‑road data and powers large‑scale deployments such as the “Luobo Kuaipao” robotaxi fleet in Wuhan, covering 3,000 km² and serving 7.7 million residents.

Community Support and Ecosystem

To accelerate adoption, Baidu offers funding, resources, and a “Wenxin Cup” startup competition that has already supported 15 winning teams with nearly ¥1 billion in investment. The second edition expands globally, with a top prize of ¥50 million.

Overall, Li emphasized that AI‑native development is becoming as simple as creating a short video, and that the combination of powerful foundation models, flexible toolkits, and a thriving ecosystem will enable anyone to become a creator in the AI era.