A 393-Run, 49-Hour EEG Study Decoded Silently Read Words From One Participant and Found No Saturation Point

arXiv / Hacker News·medium signalXBlueskyLinkedInCopy link

arXiv 2608.20186, submitted 2026-08-20 by Marquardt, Alchanat and Jain, recorded roughly 240,000 word presentations from a single densely-sampled participant using 19-channel dry-electrode EEG, then trained a convolutional encoder with CLIP-style contrastive learning against language model word embeddings. Removing occipital and posterior-temporal electrodes cut word-level gain by about a third, locating most of the signal. Decoding scaled log-linearly with training data and showed no sign of saturation, which is the finding that matters: the ceiling here appears to be hours of recording, not electrode count or model architecture.