A newly launched neural AI company, Hemispheric, co-founded by Gideon Litwin, the co-inventor of Apple's Face ID and Vision Pro technologies, has introduced a massive AI brainwave model called Descartes. The company also announced $52 million in early-stage funding. Domestically, feedback from multiple startups shows that AI large models have boosted the accuracy of decoding neural signals from the traditional 60%-70% to over 90%. A series of domestic and international developments send a clear signal: artificial intelligence will be a key to unlocking progress in brain-computer interfaces.

Brainwave large models are becoming the core driver for the large-scale application of brain-computer interfaces. As one of China's key future industries for the "15th Five-Year Plan" period, brain-computer interfaces (BCIs) are technologies that build a direct communication channel between the human brain and external devices. Their development relies on underlying empowerment from AI. The complete brain-computer interaction chain is divided into three main stages: signal acquisition, signal translation, and human-computer interaction. In this chain, electrodes and chips handle the physical capture and analog-to-digital conversion of brainwaves. Meanwhile, the brainwave large model, especially a BCI foundation model with cross-task and cross-individual generalization capabilities, acts as the core computing hub for in-depth neural signal analysis and accurate intent translation.

Previously, the industry primarily relied on manually extracting brainwave features combined with shallow machine learning for decoding. This required training dedicated models separately for each user and task. This highly customized technical path remained confined to laboratories for a long time, becoming a key bottleneck restricting the commercial application of BCI technology. Breakthroughs in AI large model technology now make it possible to fundamentally change this situation. Massive amounts of brain neural activity data can be used as "neural corpus" to pre-train a universal foundation model, effectively giving BCIs a "basic operating system." By making minor adjustments for new users or scenarios, BCIs are expected to achieve large-scale deployment.

Chinese-language brain decoding has demonstrated the capability of this foundation model. English has fewer than 50 phonemes, but Chinese has over 400 Mandarin syllables, making BCI decoding extremely challenging. However, using a foundation model, clinical trials conducted by Li Meng's team at the Shanghai Institute of Microsystem and Information Technology, Chinese Academy of Sciences, in collaboration with Huashan Hospital affiliated with Fudan University, showed that a participant could complete model personalization by reading just 54 Chinese characters in 100 minutes. Test results revealed the model could extrapolate to decode sentences containing up to 1,951 commonly used Chinese characters, demonstrating strong generalization abilities.

Solving key bottlenecks requires improving EEG data quality and scale. Similar to training language models, brainwave large models require vast amounts of high-quality data. For instance, the Descartes AI brainwave model has 6 billion parameters and was trained on 250,000 hours of EEG and behavioral data from over 100,000 participants. The current development bottleneck for BCI foundation models lies in the scale and quality of intracranial brain neural activity data. When neural signals travel from inside the brain to the scalp, they pass through the skull and brain tissue. This process causes volume conduction and high-frequency attenuation, leading to unavoidable data loss and signal distortion. Therefore, simply increasing the volume of non-invasive EEG data collected from the scalp or upgrading sensor hardware is like "scratching the surface" for training brainwave large models. While intracranial EEG data is of higher quality, current data supply is far from sufficient.

An important reason is that most medical institutions previously did not recognize the importance of this type of data. Large amounts of intracranial EEG data generated during epilepsy treatment were not preserved. Data reserve efforts are still in their infancy. Huashan Hospital affiliated with Fudan University is one of the earliest medical institutions in China to retain and accumulate intracranial EEG data. Relying on these valuable data resources, its BCI foundation model could gradually transform from a technical concept into reality. By the end of this year, Huashan Hospital expects to have collected data from 1,000 cases of human invasive EEG signals. Another critical issue is "data silos." Different acquisition device models, self-contained experimental paradigms, and inconsistent data storage formats and external interfaces make it difficult to efficiently integrate EEG data from different sources. The root cause is the lack of industry standards. Currently, Li Meng's team at the Shanghai Institute of Microsystem and Information Technology is collaborating with the Shanghai Institute of Medical Device Testing to develop national standards for BCI data, with a draft national standard expected to be completed by the end of 2026.

Brain-computer interfaces will also advance AI development. Against the backdrop of rapid AI model evolution, the relationship between BCIs and AI is being reevaluated. AI is not just a tool for decoding brain signals; neuroscience itself may, in turn, inspire AI development. From a long-term vision, BCIs can not only "read" the brain but also "write" to it, enabling direct communication between the carbon-based brain and silicon-based chips. Artificial intelligence and biological intelligence each have strengths and weaknesses. The human brain excels at rapid generalization and autonomous exploration, like learning by analogy and intuitive judgment, which AI finds difficult. Meanwhile, AI can tirelessly process vast amounts of information with incredible speed, an advantage the human brain cannot match. If they can be integrated and collaborate, a more powerful brain-computer fusion intelligence could emerge.

For AI to continue advancing, it needs to solve problems related to continuous learning, hierarchical memory, and long-term reasoning, all of which are highly relevant to neuroscience. The human brain is a general-purpose intelligent system validated by natural evolution. Truly understanding the brain could not only provide new treatment options for neurological diseases but also offer new theoretical sources for AI development. Specifically, the brain is a center that continuously creates new ideas and cognitions, generating entirely new neural signals not covered by existing databases at every moment, especially those related to original innovation. AI large models trained solely on fixed historical EEG datasets can never fully cover the dynamically changing signals and thought patterns of the human brain. Enabling large models to, to some extent, mimic the higher-order intelligence of the human brain and build generalizable predictive and reasoning abilities is a more formidable challenge for the future.

Looking ahead to the development window of the "15th Five-Year Plan" for future industries, as domestic EEG data resources continue to accumulate, industry standard systems gradually improve, and collaboration between industry, academia, research, and medicine deepens, brain-computer fusion technology will continue to expand into diverse application scenarios. The deep integration of carbon-based biological intelligence and silicon-based artificial intelligence will not only reshape human-computer interaction patterns and cultivate new tracks for future industries but also help China secure a leading position at the forefront of global brain science and AI industries.