From scaling data to decoding the brain

From scaling data to decoding the brain

Li Meng, professor at the Shanghai Institute of Microsystem and Information Technology at the Chinese Academy of Sciences explains how artificial intelligence and brain-computer interfaces are converging to shift innovation from scaling data to decoding brain signals through new model architectures and compounding data systems.

Artificial intelligence (AI) is entering a new phase that moves beyond enterprise data and into the human brain, where neural interfaces are beginning to connect thought directly with digital systems. For institutions that have invested heavily in large language models (LLMs), data scaling, and autonomous systems, this raises a deeper question about how intelligence itself should be built. As Li Meng, professor at the Shanghai Institute of Microsystem and Information Technology, Chinese Academy of Sciences framed it, “Is there such a thing as an LLM of the brain?” Artificial intelligence is moving from a question of scale to one of design, centred on whether future systems reflect how the brain actually works.

Current large language models depend on vast datasets and intensive computing power, yet the human brain operates with far greater efficiency. Li Meng highlighted that the brain can “function more, know more, and make intuitive decisions” with limited data, highlighting a structural limitation in current AI systems where increasing scale does not always improve performance. The implication is not incremental improvement but a change in architecture, as future systems move towards the efficiency and adaptability of biological intelligence.

The central problem of decoding the brain

Li Meng drew a clear distinction between data and meaning, noting that “connection is not equal to information exchange” and “raw data is not equal to information.” Brain signals resemble the vibrations of a vinyl record, where the signal itself carries no meaning until it is translated into something interpretable. The challenge is not collecting brain data, but decoding it.

This places decoding at the centre of the problem. Li Meng described this as the next major scientific frontier, framing it as a transformative effort to unlock the brain’s underlying code, on a scale comparable to landmark projects such as the Manhattan Project, the Apollo Project, and the Human Genome Project. The human brain contains around 86 billion neurons and approximately 100 trillion connections, yet even with access to these signals, systems must interpret incomplete and complex patterns in real time. Artificial intelligence is shifting from processing large volumes of structured data to extracting meaning from signals that are noisy, partial, and context-dependent.

Building a new class of brain-based models

Li Meng described brain-computer interface systems as having both “flesh” and “soul.” The hardware, including electrodes, chips, and external devices, forms the physical layer, while the algorithms that decode and encode brain signals form the underlying system. The core challenge lies in this second layer, where meaning must be extracted from signals and translated into instructions.

Current brain-computer interface (BCI) systems illustrate this limitation. These BCIs require individual training and ongoing adjustment, which makes large-scale deployment impractical. Li Meng noted that this process can take months, with additional daily adjustments. This reflects a broader limitation, where neural data is highly individual, incomplete, and difficult to standardise.

This limitation has led to a shift in approach. Li Meng described the development of a general-purpose model, “a basic pre-trained model or a generative pre-trained transformer (GPT) in the BCI.” This approach reduces reliance on user-specific training and allows systems to adapt quickly with minimal additional data.

A key step in this process is converting neural signals into a format that machines can process. Brain activity is typically captured as electrical wave patterns through methods such as electroencephalogram (EEG), but these signals are unstructured and do not directly correspond to language or commands. Li Meng identified this challenge as the need to build an “EEG dictionary”, which translates these patterns into structured units that models can interpret. Without this translation layer, conventional AI architectures cannot be applied to brain data.

From decoding to real-world capability

Li Meng’s work demonstrates how decoding capability translates into real-world applications. Brain-computer interface systems can predict epileptic seizures seconds before they occur, enabling early intervention and reducing risk. Similar systems monitor anaesthesia in real time, where precise control over consciousness is critical.

The same capability extends to human-machine interaction. Li Meng described systems that are “completely controlled by the brain”, where users can operate devices with minimal training and achieve performance comparable to conventional interfaces. Neural signals can therefore be translated into direct control of digital and physical systems without traditional input devices.

Across Asia-Pacific, the most advanced deployments of these systems remain concentrated in clinical and rehabilitation settings, including brain-controlled communication systems for patients and early use of robotic or assistive device control in hospitals. In China, minimally invasive BCI systems have entered regulatory fast-track pathways for advanced medical devices and are being piloted in rehabilitation centres, while Japan and South Korea focus on non-invasive systems for motor recovery and cognitive support. These examples show that decoding capability is already translating into controlled, real-world use, even as broader consumer adoption remains further out.

The system that makes this scalable

Individual breakthroughs do not scale on their own. Progress depends on a system where each step reinforces the next. Li Meng described the development of brain-computer interfaces as a reinforcing cycle, stating that “we record more data, create more models, produce better products… and build a data flywheel.” Each stage strengthens the next, allowing systems to improve continuously.

This process is constrained by the nature of brain data itself. Data remains incomplete in both space and time, as only a small portion of neural activity can be recorded and only over limited periods. Despite these limitations, the accumulation of data and refinement of models improves decoding accuracy and expands application potential.

Development is also shaped by a combination of policy, capital, and competition that differs across regions. Li Meng explained that early development was government-led, while the current phase is driven by private investment and commercial use.

In the United States, development is driven by private capital and companies such as Neuralink, Synchron, and Paradromics, which focus on high-bandwidth implantable systems and clinical trials. Europe has taken a regulatory-led approach, embedding BCIs within broader artificial intelligence and data protection frameworks, with strict requirements around safety, consent, and data use. China is pursuing a coordinated strategy that combines national policy, public funding, and industrial development, with a focus on scaling data collection, model development, and real-world deployment in parallel to accelerate progress.

Why this matters for the future of artificial intelligence

Li Meng positioned brain-computer interfaces as fundamental to the future of artificial intelligence, stating that “it may be the prerequisite for us to understand the brain and create artificial general intelligence (AGI),” or systems that can think, learn, and adapt across tasks in the way humans do. Current AI systems do not achieve this level of general intelligence, and understanding the brain provides a pathway towards building systems that can.

The implications extend into how future systems are developed and scaled. Intelligence is becoming less about processing large volumes of data and more about interpreting signals that are incomplete, dynamic, and context-dependent. Systems that can learn from incomplete data, adapt in real time, and improve through continuous feedback are likely to define the next stage of development.
As this transition unfolds, differences in data, infrastructure, and governance will shape how these systems are built and applied. Brain-computer interfaces move artificial intelligence closer to biological models of intelligence, changing how systems are designed, trained, and scaled.

Comments (0)
Cancel