The 2026 Beijing BAAI Conference, a premier global event in the artificial intelligence field, successfully concluded on June 13th, 2026, after two days of intensive discussions. This eighth edition of the conference has solidified its status as a leading academic gathering, having attracted 14 Turing Award laureates and over a thousand top experts from academia and industry over the years, with a global reach spanning more than 30 countries and regions.
True to its reputation as an annual AI summit, this year's event brought together a remarkable assembly of talent, including 2 Turing Award winners, 8 academicians, 30 scientists under 30, and over 40 CEOs and founders from leading AI companies. It featured participation from more than 20 top international institutions like Meta and NVIDIA, alongside key Chinese innovators such as Alibaba, Tencent, Xiaomi, Shengshu Technology, and Minibrain AI, as well as prestigious universities like Tsinghua and Peking University. This convergence created a massive, cross-border platform, with in-person attendance exceeding ten thousand, underscoring the BAAI Institute's profound influence and pioneering exploration in AI, particularly in large models.
The conference centered on two main themes: world models and intelligent agents. The forum dedicated to intelligent agents expanded from one to three sessions, and three additional forums were organized on embodied intelligence and world models. These gatherings convened over 15 CEOs from embodied intelligence companies valued in the tens of billions to discuss development paths, and brought together experts from institutions like the Chinese Academy of Sciences, Tsinghua University, and Stanford to explore the intersection of AI and neuroscience. Beyond technical dialogues, the conference debuted an AI Native Education Forum, where school principals and students discussed talent cultivation in the AI era. It also introduced innovative features like a "digital agent" for parallel conference attendance and announced collaborations, such as a joint AI-powered cardiac hospital project with Anzhen Hospital and an accelerator program with Minibrain AI.
Deepening the AI-Neuroscience Nexus
This year's conference placed a strong emphasis on the synergy between AI and neuroscience, featuring a full-day "AI × Neuroscience Forum" for the first time and dedicating a separate, in-depth session to "Brain-inspired Intelligence and the Next Generation of AI Pathways." The forum focused on the bidirectional empowerment between the two fields, structured around two core themes: AI for Neuroscience and Neuroscience for AI. On one hand, it explored how AI technologies, leveraging large models and computational methods, can advance neuroscience research in areas like brain mapping, neural signal analysis, and modeling brain networks and diseases. On the other hand, it examined how insights from neuroscience can inspire innovations in brain-inspired algorithms, explainable AI, general intelligence, and embodied intelligence, accelerating the convergence of technologies from brain structure simulation to the construction of physical-world agents.
The forum gathered leading scientists and experts in brain-computer interfaces, embodied intelligence, and clinical applications to delve into cutting-edge research and real-world applications, fostering academic exchange and collaborative innovation to guide the development of next-generation intelligent theories and technologies. Academician Shi Songhai of the Chinese Academy of Sciences delivered the opening remarks, followed by in-depth presentations from experts at institutions including Graz University of Technology, Stanford University, Tsinghua University, Peking University, Fudan University, and the BAAI Institute. A panel discussion featuring experts from Tsinghua University, ShanghaiTech University, Xuanwu Hospital, and BrainCo further explored the integration of AI in brain disease diagnosis, brain-computer interfaces, and embodied intelligence, presenting a complete loop from neural mechanism analysis to brain-inspired algorithm innovation.
Closing Keynote: A Consistent Path to Intelligence
Huang Tiejun, Chairman of the BAAI Institute, delivered the closing keynote speech titled "The Path to Intelligence: A Consistent Principle" at the AI × Neuroscience forum. Starting from the dual interaction of scientific inquiry (Why) and technological creation (How), he referenced Alan Turing's 1950 notion of "We can only see a short distance ahead, but we can see plenty there that needs to be done," highlighting the Institute's pragmatic stance on the journey to AGI. Huang elaborated on the dialectical relationship between "structure determines function" and "function shapes structure." While a fundamental tenet in life sciences is that DNA structure determines an organism's form and function, evolution and postnatal environmental training of neural synapses represent "function shaping structure." Current AI, with structures like Transformer networks trained on vast data, mimics this interplay. Future AI, he argued, requires further optimization or even transformation of its structures, trained through more authentic environmental perception to achieve higher levels of intelligence.
Huang stated that the BAAI Institute has been exploring the AI frontier from both these angles. Large models are built using Transformer structures trained on massive datasets. Large Language Models (LLMs/GPTs) understand semantics by computing contextual relationships between tokens, while Multimodal Large Models (MLMs/Emu) unify language, images, video, and actions as tokens for prediction. The Institute is developing an "Objective World Model" aimed at understanding and reasoning about the time, space, physical laws, and causal mechanisms of the real world, representing a paradigm shift from "Next Token Prediction" to "Next Physical State Prediction."
Extending this logic to the brain, the Institute developed the WuJie·Brainμ 1.0, the world's first unified multimodal neuroscience large model for understanding and generation. It standardizes brain signals from humans, monkeys, and mice across modalities like EEG and fMRI into tokens, unleashing the model's knowledge representation and generalization capabilities to accomplish a wide range of neuroscience tasks with a single, unified model, approximating the brain's own "microcosm."
Building Foundations in the Physical World
Huang pointed out that from first principles, a world model must overcome the bottleneck of slow image/video processing. Inspired by the primate retina, the Institute developed spiking vision chips and cameras that measure light intensity per pixel, performing computations only where light changes, thereby opening a high-speed, efficient window for world model training with high-fidelity physical observation data.
He emphasized that true intelligence is not mere symbolic computation detached from the physical world. It requires multimodal sensory input to form internal brain models and external world models, together creating a complete world model system. Therefore, a world model must not only generate continuous video or understand basic physical interactions but also model fundamental laws governing fluid dynamics, combustion, and nuclear reactions. It must encompass phenomena from microscopic atomic interactions to the evolutionary adaptation of life, becoming a "complete model" accommodating all scientific and technical disciplines.
Huang noted that the long-term development of AI must return to "structure determines function" to explore neural and physical structures more powerful and efficient than Transformers. Over the past five years, the BAAI Institute has systematically built a biological simulation system spanning from the retina to the heart, and from neural circuits to the molecular level. This system simulates key biological functions like vision, movement, and circulation, transforming the life science principle of "structure determines function" into a computable, trainable AI system. This provides a solid physical foundation for world models, pushing intelligence from the digital realm into the real physical world.
Examples include the digital nematode BAAI Worm, which uses reinforcement learning to generate motion, and the digital twin heart BAAI Heart, modeled on human cardiac structure and trained on diagnostic data. The molecular model OpenComplex builds models at the molecular level from a biological environment perspective. Under the "Embodied & World" framework, integrating spiking camera data, digital organs, and cellular-level physiological activity enables high-speed environmental data collection, adherence to physical laws, deep understanding of physiology, and correct responses to stimuli, advancing the "tripartite interaction" of AI, the physical world, and life sciences.
AGI Timeline and the Road Ahead
Finally, Huang revisited the five-level AGI evolution timeline and risk framework proposed at the 2024 conference: Level 0 (sub-human cognition, now passed, risks of misuse); Level 1 (superhuman cognition, occurring now, posing choices between complacency and rational trust); Level 2 (embodied superhuman, around 2035, exceeding human physical control); Level 3 (consciousness from integrated perception/cognition, "machine rise"); Level 4 (self-awareness, risk of humans becoming a secondary species); and Level 5 (transcending human knowledge/brain architecture, AGI exploring the universe alone).
The 2026 BAAI Conference vividly showcased AI's momentous journey from the digital to the physical world. Through the dialectical unity of "structure determines function" and "function shapes structure," it demonstrated the BAAI Institute's strategic resolve of "a consistent principle." From foundational models ("WuDao") to understanding boundaries ("WuJie"), the Institute is fostering the "tripartite interaction" of AI, the physical world, and life sciences, building the brain, eyes, and body for AGI to truly embed intelligence into reality. Guided by Turing's pragmatic spirit and a deep concern for humanity's future, the BAAI Institute, alongside global partners, remains committed to advancing sustainable development for humanity, the environment, and intelligence itself.
Comments