The signal under the hair

An electrode on your scalp does not read thoughts. It reads voltage: the faint electrical shimmer produced when thousands of neurons beneath it fire in loose synchrony. Each electrode sums whatever pyramidal cells in its neighborhood are doing, the way a microphone outside a stadium sums the crowd. The signal is tiny, on the order of tens of microvolts, a few millionths of a volt. Your brain is a very loud organ producing a very quiet reading.

The brain writes in many rhythms at once

Split that voltage trace by frequency and structure appears. Slow delta waves under deep sleep, alpha when you close your eyes, beta when you focus, and above roughly seventy hertz the band researchers call high-gamma, which tracks local cortical work with unusual fidelity. Speech leaves its clearest fingerprints in that fast range and in the timing of slower rhythms around the speech and hearing cortex. If you want to read words off a head, the fast stuff is where the payload lives.

The skull is a wall, and high frequencies die at walls.

The physics problem

Here is the villain. Bone and scalp conduct electricity badly and unevenly, and they act like a low-pass filter: the higher the frequency, the more it is muffled on the way out. Listen to a concert through a wall and you hear the bass line but lose the singer. Implants sit inside the wall, which is exactly why intracortical arrays and ECoG strips decode speech so well; the papers proving it are the foundations of this whole field. Non-invasive EEG stands outside the wall and tries to reconstruct the singer from the bass. That is not impossible, but it is the reason this is a hard problem instead of a solved one, and the reason our answer is a model trained on far more data than anyone thought necessary.

The noise problem

Everything else on your head is louder than your brain. A jaw clench or a swallow puts muscle electricity into the trace at amplitudes many times the neural signal. Eye movements smear the front of the record. The power line hums at fifty or sixty hertz, right in the interesting band. A usable pipeline filters, rejects corrupted stretches, and references channels against each other, all before any model sees the data. Electrode contact matters just as much: gel lowers the electrical impedance between skin and metal and gives the cleanest signal, at the cost of a clinic ritual; dry electrodes trade some noise for the ability to pull a cap on in a living room. The trade between those two is one of the decisions in front of us.

None of this is a reason for pessimism. It is a sizing of the job. The signal is real, the physics is known, the artifacts have known shapes, and every one of these problems yields to the same tool: enough paired examples for a model to learn what to ignore. Which is what the next note is about.