Zyphra just dropped ZUNA — the first foundation model built specifically for raw EEG data. Trained on ~2 million channel-hours of scalp EEG, it can denoise noisy signals, reconstruct missing channels, and even predict what electrodes you don't have would be reading using 4D positional encoding of electrode coordinates. It works across different headset configurations out of the box, which means it could slot right into an 8-channel Cerelog setup.

The request: We'd love to see someone try piping live Cerelog ESP-EEG data through ZUNA and pushing the decoded output toward thought-to-speech. The model is already positioning itself as a step toward thought-to-text — pairing it with our board could be one of the first real-world attempts at noninvasive thought-to-speech with affordable, open hardware. If you're into BCI, ML, and want to attempt something genuinely novel — this is it.

A browser-based playground for building and testing BCI decoders. A decoder is a small program — usually a machine-learning model — that looks at those signals and guesses what the person was thinking or trying to do. Streams live data from your Cerelog ESP-EEG, runs a decoder on it, and plots predictions against ground truth side by side. Bring your own decoder in JavaScript or TensorFlow.js, or generate one from a text prompt.