Written by Ralph Sun
In robotics, a human intervention is simultaneously the failure and the label. In neuro AI, the equivalent label has to be produced by someone sitting down and coding it by hand. That asymmetry, not compute, is what holds the field back.
There is a habit in this field of explaining slow progress with a shortage of compute. I do not think that is the binding constraint. The binding constraint is ground truth, and the reason is structural rather than technical.
Consider how a recent study actually produced its labels. Researchers at Northwestern University put EEG caps on 49 toddlers and their caregivers during ordinary play, then reviewed the recordings frame by frame to identify the moments when parent and child were genuinely engaged with one another. The finding — lower alpha and higher theta activity during social engagement than during shared screen time — is a real result. But the method behind it is the part I keep returning to: a human being, watching video, deciding moment by moment what the brain was doing.
That is the state of the art in a field that is otherwise dominated by large models. The signal is cheap and continuous. The meaning is expensive and manual.
Compare that with robotics. When a robot fails a pick and a person steps in to resolve it, the intervention is both the failure and the annotation. Nobody has to watch the footage afterwards and decide what happened, because the human action already encoded the judgement. That is why the fastest-improving area of physical AI is the one that consumes intervention data as its training signal while it deploys. The label is a byproduct of the work.
Neuro AI has no equivalent byproduct, and three consequences follow.
First, dataset size is capped by human hours, not by storage. You can record a thousand hours of EEG in a month and still have a few dozen hours of usable labels, because annotation does not parallelise the way recording does.
Second, labels are subjective in a way that resists scaling. Two annotators coding engagement will disagree at the margins, and the disagreement is not noise to be averaged away — it is the measurement error of the construct itself. This is why neural foundation models that generalise across species, brain regions and tasks are genuinely encouraging and genuinely incomplete: generalisation tells you the representations transfer, not that you have a reliable way to check what they mean.
Third, the legal question is unsettled on top of the scientific one. The OECD’s work on neurodata governance, built on interviews with 16 neurotech companies across seven countries, found that reuse of neural recordings for research and model training remains genuinely unclear. So the one input that is scarce is also the one whose permitted uses are least defined.
There is also a constraint people outside hardware rarely appreciate. A great deal of neuro AI has to run on a wearable, not in a data centre, and the energy arithmetic is unforgiving. Recent work on the energy constraints of large-scale AI reports that accessing off-chip DRAM can cost roughly 200 times the energy of a single multiply-accumulate operation, against about 6 times for an on-chip buffer. Wearables do not have the thermal or power budget to brute-force their way past that.
So the direction that makes sense is the one robotics already took: stop treating annotation as a separate offline chore and treat it as part of the loop. Wherever a human acts on a system — correcting, attending, intervening, reacting — the action should be captured as the label rather than reconstructed later by an analyst. Models that predict human state are then trained on data that the workflow produces for free.
That is the design principle we work to at Brainlayer, and I think it generalises well beyond robots. The question for neuro AI is not how much compute we can buy. It is whether we can make the ground truth a byproduct of something people were going to do anyway.
