ERNIE Bot Launches New Version with 5 Native Plugins

[baoshi.rao]

On August 16, the WAVE SUMMIT 2023 Deep Learning Developer Conference, hosted by the National Engineering Research Center for Deep Learning Technology and Applications, was held. Baidu's Chief Technology Officer and Director of the National Engineering Research Center, Wang Haifeng, delivered a keynote speech titled Large Language Models Herald the Dawn of Artificial General Intelligence, emphasizing the four core capabilities of large language models: comprehension, generation, logic, and memory. These capabilities bring us closer to achieving artificial general intelligence (AGI). During the event, Baidu unveiled a series of technological, product, and ecosystem achievements, including the ERNIE large model, the PaddlePaddle platform, and AI-native applications like Liuliu. Thanks to the synergistic development of PaddlePaddle, an industrial-grade open-source deep learning platform, and the ERNIE large model, the PaddlePaddle ecosystem has flourished, attracting 8 million developers, serving 220,000 enterprises and institutions, and enabling the creation of 800,000 models.

Baidu CTO Wang Haifeng delivering the keynote speech

Wang Haifeng elaborated on the profound meaning behind the Chinese name of the PaddlePaddle developer community, AI Studio, now called Xinghe Community (星河社区, 'Galaxy Community'). He stated, 'With ERNIE and PaddlePaddle, we soar toward the galaxy.' Together with developers worldwide, the community aims to advance toward the vast frontier of AGI.

Wang further explained that AI possesses multiple core capabilities, with comprehension, generation, logic, and memory being foundational. The stronger these capabilities, the closer we get to AGI. Large language models, exemplified by ERNIE Bot, embody these capabilities, marking a significant step toward AGI. ERNIE Bot, Baidu's proprietary knowledge-enhanced large language model, is trained on trillions of data points and hundreds of billions of knowledge units. It employs supervised fine-tuning, reinforcement learning from human feedback, and prompting techniques, offering advantages in knowledge enhancement, retrieval augmentation, and conversational refinement.

In recent months, ERNIE Bot has evolved rapidly. Wu Tian, Baidu's Vice President and Deputy Director of the National Engineering Research Center, announced five new native plugins for ERNIE Bot: Baidu Search, Lanjuan Document (document-based interaction), E-Yan-Yi-Tu (data visualization chart generation), Shuo-Tu-Jie-Hua (image-based interaction), and Yi-Jing-Liu-Ying (text-to-video). ERNIE Bot now supports using up to three plugins simultaneously, expanding its capabilities and adaptability to diverse scenarios.

Wu Tian revealed that AI Studio (Xinghe Community) is now China's largest AI community, hosting 6.09 million development projects. With the rise of large models, developers are rapidly congregating in this space. The conference showcased the latest upgrade to AI Studio, introducing the Xinghe Large Model Community, offering developers an integrated large model development experience. PaddlePaddle and ERNIE streamline the entire development process, fostering collaboration on large model projects and applications. The community already features over 300 creative large model applications and provides robust tools for developer interaction. Additionally, the conference launched an invitation-based testing phase for large model plugins, leveraging ERNIE Bot's core technology to enable developers to create plugins for information services, tools, and innovative applications. Wu Tian also announced the ERNIE Large Model 'Xinghe' Co-Creation Program, providing ecosystem and industrial resources to drive AI innovation.

ERNIE Bot's rapid progress is attributed to the synergy between PaddlePaddle and the ERNIE large model. Through PaddlePaddle's end-to-end adaptive hybrid parallel training and optimization techniques, ERNIE's training speed has tripled, while inference speed has increased over 30-fold.

At the conference, Ma Yanjun, General Manager of Baidu's AI Technology Ecosystem, unveiled PaddlePaddle Open-Source Framework v2.5, featuring a comprehensive architectural upgrade with enhanced differentiation and performance optimization. The event also saw the release of PaddleScience v1.0, a scientific computing toolkit, and PaddleHelix's HelixDock, a large model for protein-compound conformation affinity prediction.

PaddlePaddle has also contributed to national standards for hardware-software adaptation. Collaborating with institutions like China Electronics Standardization Institute, Baidu, Sugon, Phytium, and Inspur, the National Standard for AI Deep Learning Framework Multi-Hardware Platform Adaptation was drafted. Under this standard, ERNIE has been adapted for 12 hardware partners, including NVIDIA, Sugon, Kunlunxin, Cambricon, Huawei, Enflame, Biren, Qualcomm, Intel, MediaTek, Moore Threads, and Arm China, covering diverse cloud and edge devices.

AI, led by large language models, is transforming industries and accelerating economic growth. In the enterprise sector, Baidu's Vice President and Chief Information Officer, Li Ying, announced the upgrade of Baidu's AI-native knowledge management philosophy to Innovation Pipeline = AI × Knowledge Management 2.0. She also introduced Liuliu Super Assistant, demonstrating its capabilities in mobile task execution, smart document processing, conversational business intelligence (CBI), and intelligent communication across 10+ scenarios.

For example, with Voice-Activated Travel Planning, users can verbally request, 'I need to attend a meeting at Shanghai R&D Center the day after tomorrow morning and return to Beijing the same day. Please arrange my trip.' The Super Assistant autonomously submits travel requests, books flights, and reserves hotels, reducing task completion from hours to minutes or even seconds.

In R&D intelligence, Baidu unveiled Comate X, a smart programming assistant, and the Comate Stack toolkit. Comate X goes beyond code suggestions, assisting engineers throughout the development lifecycle—ideation, coding, and debugging—with features like code explanation, generation, inline comments, unit test creation, documentation, CLI generation, and API-to-code conversion. Supporting 30+ programming languages and 10+ IDEs, Comate X is now open to enterprise developers. The Comate series, already deployed across Baidu, serves over 100 enterprise partners, marking its commercial maturity.

The transformation in enterprise productivity mirrors the broader impact of AI-driven by large language models. This virtuous cycle of innovation and application continuously enhances capabilities like comprehension, generation, logic, and memory, expanding the scope and depth of industrial applications. Large language models are indeed heralding the dawn of AGI.