Neuralink trained encoders on 50,000 hours of brain recordings: a record 11.32 bits per second and calibration once a week instead of once an hour
Prepared by Sintelligent with the help of SI. Publication time restored from the build log. . Updated: .
The brain interface learns the way language models do, from big data, and every new hour of recordings makes thought control more accurate and stable.
According to Neuralink, clinical trial participants contributed more than 50,000 hours of unlabeled neural recordings over two years; personal self-supervised encoders trained on them turn noisy signals into stable representations. The median across participants is about 10 bits/s, three exceeded the previous record of 10.39, and one reached 11.32 bits/s. According to TeslaNorth's retelling, weekly calibration dropped from ~55 to ~10 minutes, and some decoders worked for more than three weeks without retuning. The company's goal is a general neural model for all participants; Neuralink's own paper is not fully open, so the details come from the retelling.
Sources 2