In 1924 Hans Berger invented EEG to build telepathic communication. One hundred years later, advances in neural sensors and deep learning models have led to breakthroughs that would make that dream possible, in fact with devices as simple and wearable as the baseball cap that Sabi is building.
However, the field has poured all its energy into building BCI (Brain Computer Interfaces), trying to emulate mouse clicks and type words that you would otherwise use a keyboard for. This is exciting, but the wrong target.
Computers are no longer a collection of mouse, keyboard and monitor. They are now AI models which get work done for you. They could be on your desktop, phone, watch, glasses, or caps. As computers become more and more autonomous, the limit isn't how fast we can communicate to machines, but how much they can understand without being told.
As a field we need to reorient and build the most seamless interface to these ever improving AI models.
While non-invasive BCI devices have begun getting good at thought-to-text capabilities (decoding thoughts word by word), they can also understand your semantic intent, taste and preferences, and predict your future actions.
The semantic decoding of your intentions can be used as prompts for AI models and agents. Imagine if instead of typing or speaking prompts, which is lossy and slow, you can just tell Instinct or Muse agent what to do directly from your thoughts, no back and forth to confirm what you meant, and the agent does exactly what you asked, every time.
In a recent interview with Forbes, Vinod Khosla said, “Typing is not natural. Before you start typing, you have to think about what you’re going to type. If you can intercept the brain… it’s awesome.”
Along with semantically understanding your thoughts, brain signals can be used to understand your taste and preferences, what painting you like, what music makes you feel emotional or what AI response feels slop vs genuinely interesting.
Today, the AI models tend to be very generic, and don’t have any way to create personalized responses. However, continuously learning from brain signals lets us build highly personalized, relevant AI products for everyone.
Finally, other than just decoding neural responses when they happen, AI models have gotten good at predicting them a few seconds ahead of time. Sabi has trained a typing BCI model that can help predict the next 3-4 keystrokes a user is going to type in.
Once such models scale up across more open domains, we can have AI agents that can take predictive actions doing things for a user in digital and physical spaces even before they have consciously thought of these actions.
All of this comes with a big caveat, none of this works without trust: neural data has to stay encrypted, nobody other than yourself should ever access what you're thinking.
All of these are ambitious capabilities that we will inevitably see deployed for general consumers as BCI devices become more wearable and reliable. Building for a Brain-AI Interface lets us merge rapidly growing machine intelligence with biological intelligence.
This matters beyond convenience. As AI zooms upward toward human-level intelligence, we need to urgently solve it’s alignment with humanity. Models are now more capable for our intellect and attention to keep up with. Agentic systems might soon be able to break into secure systems like banks and government networks. These capabilities are exceeding our ability to steer or control them.
The way forward is co-evolution: AI that learns directly from how humans think and feel, and lets humans direct AI as fluidly as they think. Building a Brain-AI Interface makes both possible. As long as human and machine minds stay closely coupled, we have a shot at long-term alignment.
AI poses existential risks to humanity, but a healthy and aligned co-evolution lets us continue harnessing Artificial Intelligence rather than having to continuously compete against it.
