Neuralink has revealed how it is putting more than 50,000 hours of neural recordings from its clinical trial participants to work, publishing an engineering update on October 1 that describes a new approach to brain-computer interfaces — one that learns from each person's past brain activity instead of starting fresh every session.
The Elon Musk-founded company says participants have streamed the unlabeled, freeform neural data over the two years since its clinical trial began. The first participant alone recorded more than 9,000 hours, representing roughly 22.4 billion neural spikes. Until recently, nearly all of that information was going unused.
Teaching decoders to remember
Neuralink has now trained participant-specific neural encoders on thousands of hours of recordings collected from each person, then used those models for cursor control. The approach lets the system learn recurring patterns in a person's brain activity from past sessions rather than rebuilding its understanding from scratch.
According to the company, the encoders turn noisy neural signals — which can shift over time — into more stable representations, called embeddings, that make decoding more accurate and durable. One participant described the new mode as smoother and more directionally accurate: "Better than the prior model. Smoother, more directionally accurate, felt effortless. Didn't have to push or anything, it just glided where I wanted it to go."
A new speed record — and "teleporting" cursors
The practical payoff is less setup. Users previously spent an average of about 55 minutes per week recalibrating their systems, typically around 10 minutes a day. Some participants have now reduced that to about 10 minutes per week. Decoders also last longer: one participant used the same decoder for five consecutive days while reaching 10 bits per second, some decoders remained strong for more than three weeks, and one participant was still controlling a decoder more than a year and a half later.
Six participants set personal cursor-control records. Neuralink says the median across participants is roughly 10 bits per second, three exceeded the previous record of 10.39 bits per second, and one participant reached 11.32 bits per second — which the company described as a new brain-computer interface record. It also claims the figure exceeds the performance of most able-bodied people, though that comparison has not been independently verified.
The company also demonstrated what it calls "teleporting cursors," where the cursor jumps directly to the intended target instead of travelling continuously across the screen — a sign of how precise the decoded signals have become.
What comes next
Neuralink says its longer-term goal is a framework combining federated learning with continual domain adaptation. In one experiment, it transferred a model trained on one participant to another while keeping 99 percent of its weights frozen — and the transferred model still beat the second participant's current decoding baseline.
The device remains investigational and is not approved by the FDA or other regulators. But the update marks a shift in brain-computer interfaces: from hand-tuned decoders that demand constant recalibration to models that keep improving by learning from months of real-world brain data.
