Baidu, Inc. showcased a suite of AI models and development toolkits at Create 2024 Baidu AI Developer Conference. These innovations aim to empower individuals with boundless potential to create AI applications and agents regardless of technical expertise, helping pave the way for a more vibrant AI future. Create 2024 Highlights Latest Advances in Baidu's ERNIE Models ERNIE has emerged as China's leading AI foundation model (FM) with the broadest range of applications, as Li noted at the event.

The latest key figures and product additions for ERNIE were also revealed at Create 2024. Key Data on ERNIE: ERNIE Bot, Baidu's conversational AI bot built on ERNIE, has amassed a total of 200 million users since its debut on March 16, 2023. ERNIE API handles 200 million queries on a daily basis.

Baidu Comate, an AI coding assistant powered by ERNIE, has contributed to 27% of newly generated code within Baidu since its deployment one year ago. It has served over 10,000 companies including Ximalaya, iSoftStone, and Shanghai Mitsubishi Elevator, with a code adoption rate of 46%. Over 85,000 enterpriseclients have used Qianfan, Baidu AI Cloud's FM platform, to create 190,000 AI applications.

ERNIE's training throughput is 5.1 times more efficient than last year. Its queries per second, or QPS for inference, has also seen a remarkable increase of 105 times, while the inference cost has been reduced to just 1% of the March 2023 version. 98.8% of the LLM training time on Baidu's AI infrastructure has been fully utilized.

New Product Additions to ERNIE: ERNIE 4.0 Tool: Accessible through ERNIE Bot, it features a code interpreter allowing users to manage and analyze complex data through natural language. It is also able to generate charts and documents, enabling quick insights and efficient support for data-driven decision-making. ERNIE Lightweight LLMs: The ERNIE model family has grown to include lightweight LLMs such as ERNIE Speed, ERNIE Lite, and ERNIE Tiny.

Each of these models is compressed and distilled from the original ERNIE to offer more agile and efficient AI solutions. AI Development Toolkits Empower Developers to Create with Ease To help developers quickly and effectively tap into the latest AI trends, Baidu has presented three plug-and-play toolkits, covering a wide spectrum of the AI development pipeline:  AgentBuilder is a development tool based on ERNIE to build and deploy AI agents. It offers both zero code and low code modes, enabling even non-coders to create an agent with just a few sentences.

With AgentBuilder, developers can deploy agents across Baidu products and services like Baidu Search, Xiaodu, Baidu Maps, ERNIE Bot, etc. Over 30,000 agents have been created using this platform. AppBuilder is a comprehensive set of tools for AI-native applications on Baidu AI Cloud.

It provides numerous components and frameworks needed to develop AI-native applications, such as retrieval-augmented generation (RAG), agent, and generative BI (GBI). Developers can use natural language to create an application within three steps. ModelBuilder is a model customization tool curating up to 77 models across various sizes, from ERNIE 3.5 and 4.0, to lightweight models like ERNIE Speed, ERNIE Lite, ERNIE Tiny, and many more third-party models.

It also provides a full set of model development tool chains and industry-specific datasets, allowing developers to solve model fine-tuning problems in one-stop. Three AI Development Trends Shape the Future of AI Adoption In his keynote speech, Li also shared three key AI development trends that are poised to shape the near future of AI adoption. MoE: MoE, or Mixture of Experts, will become critical to the future development of AI-native applications, as it will allow developers to draw on different models to solve a wide range of complex problems. Small Models: Small models offering low inference costs and quick response times can be more effective than larger models in certain scenarios.

Agents: Agents powered by foundation models are capable of understanding, planning, reflecting, and self-improving. Create 2024 also witnessed the launch of Apollo VTA (Vision Takes All), Baidu's vision foundation model for autonomous driving. Apollo VTA has been trained on real data collected from over 100 million kilometers of test drives on complex urban roads.

The model boasts four foundational capabilities in visual perception: detection, tracking, understanding and map building, which together offer a smarter, more adaptable and safer autonomous driving solution. Li also announced the launch of the second Baidu ERNIE Cup Innovation Challenge, a competition for visionaries pursuing the next iconic model products or applications for the generative AI era. This latest ERNIE Cup will offer a Special Prize, with the winner receiving cash and resource rewards totaling RMB 50 million.