The article is written by Neil Shen. Contact the author at: email@example.com
Clash of Ideas
The exhibition center of the Zhongguancun National Independent Innovation Demonstration Zone was bustling with attendees. People filled the venue, with some even standing in the aisles, captivated by the lectures and keynote speeches. It was remarkable to witness such enthusiasm at an academic conference. This vibrant scene continued unabated on June 9 and 10.
The fifth edition of the BAAI Conference was a testament to its reputation as one of the most professional and prestigious event in the field of artificial intelligence. Organized by the Beijing Academy of Artificial Intelligence (BAAI), this conference distinguished itself by avoiding any form of commercial propaganda or gimmicky forums aimed at advertisers. Despite the presence of renowned companies in the AI industry, the focus remained on substantive discussions and insightful perspectives.
The attendees brought with them their wealth of knowledge and unique viewpoints. In other words, they were here for something solid.
Among all the guests attending the conference, there were four recipients of the Turing Award, including Geoffrey Hinton, Yann LeCun, Joseph Sifakis, and Andrew Yao. Joining them were renowned figures such as Zhang Bo, Zheng Nanning, Xie Xiaoliang, Zhang Hongjiang, and Zhang Yaqin, as well as Stuart Russell and Max Tegmark, all of whom are renowned in the field of artificial intelligence. Sam Altman, the co-founder and CEO of the prominent OpenAI, also delivered a speech at the AI Safety and Alignment Sub-Forum on the morning of the 10th.
"The top event for AI experts," the positioning of the BAAI Conference lives up to its esteemed reputation. The convergence of brilliant minds among these industry leaders further highlights the conference's commitment to academic excellence and the fostering of an intellectual atmosphere.
Sam Altman's speech revolved around the field of AI safety, calling for international collaboration to address the potential threats posed by the rapid development of AI. Altman believes that AI systems will be highly developed and powerful within the next ten years.
The worldwide popularity of ChatGPT gave weight to Sam Altman's speech. However, this does not mean that everyone unanimously agrees with OpenAI's development path. During the opening ceremony speech on the morning of the 9th, Yann LeCun, a Turing Award laureate, and the Chief AI Scientist at Meta, presented a contrasting perspective. LeCun emphasized the limitations of autoregressive models, highlighting their deficiencies in terms of planning and reasoning capabilities. In order to achieve the level of Artificial General Intelligence (AGI), they should not only replicate the human brain at the neural level but also incorporate cognitive modules based on human cognition. Yann LeCun's proposed solution, which he terms the "world model," stands as an alternative approach.
Meanwhile, Geoffrey Hinton, hailed as the father of deep learning and recipient of the Turing Award, contributed his insights into the current development of AI in his speech. This industry luminary has recently garnered widespread interest due to his departure from Google. He believes that artificial neural networks will soon surpass real neural networks, and the level of danger and urgency posed by artificial intelligence may surpass even that of climate change.
Such collisions of opinions, similar to the ones mentioned, were not uncommon in the nearly hundred speeches and panel discussions held over the past two days. They revolved around cutting-edge topics in artificial intelligence, with a particular focus on large-scale models. Additionally, open-source and security also garnered significant attention and were widely debated.
The organization of this esteemed event showcases the unmatched capability of the BAAI Institute within China. As affirmed by Huang Tiejun, the President of the BAAI Institute, it stands as the premier choice for international cooperation in the field of artificial intelligence research in the country.
So, what initiatives has BAAI undertaken regarding the aforementioned hot topics?
The Multi-faceted Large-scale Model Series
During the highly anticipated opening ceremony, President Huang Tiejun took the opportunity to highlight the remarkable progress accomplished by BAAI over the past year. Notably, the Wudao 3.0 Large-scale Model Series has embarked on an exciting new phase of comprehensive open-source. This encompasses several key components, including the Wudao Aquila Language Model Series, the FlagEval Large-scale Model Evaluation Platform, and the Wudao Vision Visual Model Series.
Among them, the Wudao Aquila Language Model within the Wudao series supports commercial licensing agreements, and trained with high quality dataset. In addition to the foundation models, it includes the AquilaChat Dialogue Model and the AquilaCode Text-to-Code Generation Model.
According to Lin Yonghua, Vice President and Chief Engineer of the Beijing Academy of Artificial Intelligence (BAAI), AquilaChat-7B has surpassed mainstream open-source models of a similar size domestically and internationally. This is based on an evaluation framework of FlagEval, which incorporates comprehensive objective and subjective evaluations in bilingual English-Chinese settings. The framework is based on over 20,000 randomly selected evaluation prompts from 22 evaluation datasets.
Compared with other open-source models, AquilaChat-7B achieved this result by being trained with less than 50% of the training data, demonstrating its advantage in both data quality and implementation. It is worth noting that this is only the initial version of AquilaChat, and it will continue to be upgraded and optimized in the future.
The FlagEval Large-scale Model Evaluation Platform has been built to support comprehensive performance evaluation of base models and training algorithms. Its ultimate goal is to fully cover the evaluation of base models, pre-training algorithms, fine-tuning algorithms across four domains: natural language processing, computer vision, audio and multimodal.
FlagEval constructs a three-dimensional evaluation framework based on "capability-task-metric." Currently, it encompasses over 600 evaluation dimensions, including 22 evaluation datasets, totaling 84,433 evaluation prompts.
The Wudao Vision series systematically address a series of bottleneck problems in the field of computer vision, including task unification, model scaling, and data efficiency. It includes: Emu, a large multimodal model that completes everything in multimodal sequences; EVA, the most powerful billion-scale visual representation model; a general-purpose segmentation model that can handle various segmentation tasks; Painter, a universal vision model that pioneered the in-context visual learning; EVA-CLIP, the highest-performing open-source CLIP model; and vid2vid-zero, a zero-shot video editing technique that allows for simple prompt-based video editing.
The release highlights the end-to-end large-model development pipeline that BAAI has built up to support the continue upgrade of models with more training data and enhanced capabilities. BAAI believes both academic research and industry will benefit from it.
In addition to the newly released models, BAAI has also upgraded the FlagOpen large model technology open-source system introduced earlier this year. It includes parallel training techniques, inference acceleration techniques, hardware and model evaluation, and data processing tools. FlagOpen aims to create an open-source algorithm system and a one-stop foundational software platform that fully supports the development of large model technologies. We position FlagOpen as the “Linux” for large-scale model domain.
It is worth mentioning that, in terms of datasets, BAAI has already open-sourced the first large-scale commercially available Chinese instruction dataset, called Chinese Open Instruction Generalist (COIG). The first phase released a total of 191,000 instruction data entries, and the second phase is currently under development to create the largest and continuously updated Chinese multitask instruction dataset, integrating over 1,800 massive open-source datasets.
About BAAI’s mission in large model area, Lin Yonghua shared her view.
"Our mission is not about competing to have the biggest and strongest models. Our mission is more focused on the fundamental technologies, such as novel breakthrough technologies for base model, processing technologies for high quality data, smart evaluation technologies for model complexity, and, of course, open-source initiatives for global collaboration. We will persist on these fundamental research and development as our long-term goal." said Lin Yonghua.
This may sound somewhat idealistic, but BAAI has been synonymous with idealism since its inception.
The pursuit of artificial intelligence’s summit has always been a long-term endeavor. In their roadmap, it encompasses not only information intelligence represented by large models but also embodied intelligence based on reinforcement learning and physical embodiment, as well as brain-inspired neuromorphic intelligence. While the breakthroughs achieved with large models have provided a glimpse into the potential path toward AGI, it is crucial to recognize the significance of embodied intelligence and neuromorphic intelligence. Who can bring forth the next breakthrough?
Compared to this lofty goal, filled with idealism and long-term thinking, the superiority or inferiority of a large model's performance score seems trivial. BAAI has a much grander vision.
The Idealist in the Artificial Intelligence Field
Training a high-scoring large model alone is not sufficient to propel the field of artificial intelligence forward. While impressive in their own right, the algorithms and technologies used in the training process hold even greater significance.
Rapidly evolving technologies will eventually render each individual model obsolete. However, the key lies in establishing a technological foundation that enables the rapid evolution of AI. This entails the continuous introduction of advanced algorithms, along with effective cost reduction strategies for model training throughout the industry chain. Simultaneously, it is crucial to maintain a high level of understanding and control over AI's safety. By embracing this comprehensive approach, the future of AI becomes more solid and powerful.
In comparison to this grand vision, questions such as the number of large models launched or the specific ranking of a particular model seem trivial.
This grand vision is exactly what BAAI (Beijing Academy of Artificial Intelligence) pursues.
In March 2021, BAAI used the term "large models" for the first time, ushering in a new chapter in the development of artificial intelligence. In a short period, the Wudao large model have been iterated to its third version.
As a non-profit research institution, BAAI is dedicated to fostering an innovative AI ecosystem. Since large models are resource-intensive and complex engineering endeavors, BAAI strives to benefit the entire industry by researching, developing, and ultimately open-sourcing impactful technologies.
That is precisely what BAAI is doing. By building a foundation for large models and promoting research and innovation through open-source collaboration, BAAI is accelerating the industrialization of large models.
One such example is BAAI's implementation of commercial licensing agreements for their large models. BAAI has invested substantial resources in addressing all aspects, from algorithms to data. This meticulous approach empowers enterprises to embrace BAAI's models for commercial purposes with confidence.
Creating a new Linux-like ecosystem for global collaboration and innovation is BAAI's position in the field of large models. Open-source is a crucial and courageous step that has different perspectives within the industry. However, BAAI is firmly confident in this approach. Open-source and openness are not only the inevitable choices for building the artificial intelligence ecosystem but also the necessary path to drive technological innovation.
BAAI has been on this path for five years, and this romantic vision of artificial intelligence attracts talent with similar idealism. As China's top-tier research institution in artificial intelligence, BAAI has a lineup of "BAAI Scholars," consisting of nearly a hundred top AI experts. The BAAI community has gathered over 120,000 professionals from the AI industry. This has earned the BAAI Research Institute widespread acclaim for its research capabilities globally.
The BAAI Conference, held annually, has firmly established itself as the premier summit for artificial intelligence, not only in China but also on a global scale. Over the past four years, more than 500 top AI experts, including esteemed recipients of the Turing Award, have delivered speeches and participated in discussions at the conference. Tens of thousands of professionals from over 30 countries have registered to attend.
At the conference, BAAI unveils the latest achievements in relevant fields, and the clash of viewpoints on artificial intelligence reverberates from Beijing to the world. Over the years, BAAI has cultivated a robust platform that has fostered the flourishing and maturation of the AI ecosystem in China. This platform has played a pivotal role in supporting China's active participation in the pursuit of large models, positioning BAAI as an invaluable but often underappreciated cornerstone in China's rapid strides in the field of large models.
This can be seen as the best reward for idealism.