# Part V. The source | The Intent Layer Source: https://setu-labs.co/manifesto-5.html > Text, code, images, audio and video are compressed products of human cognition. Neural data is the closest anyone can get to the engine that produced them. We think it becomes the next modality of AI. [Download PDF ](https://setu-labs.co/assets/setu-the-intent-layer.pdf) ## The source Text, code, images, audio and video are compressed products of human cognition. Neural data is the closest anyone can get to the engine that produced them. We think it becomes the next modality of AI. Several fields are colliding at once. AI has become an engine of creation. Earlier technologies amplified human thinking and waited for it. This one increasingly does the iterating itself: it writes the code, designs the experiment and proposes what to try next. So far this essay has treated BCI as a way of giving machines instructions. That is the smaller half. Neural data meets AI in three places: how models are trained, how people talk to them, and how they are kept aligned with what people want. Two of those are needed today, by customers who already exist. Laboratories need new data, and they need a faster way to align their models. Neither has to wait for a consumer product. They are the immediate wedges. ### Five products and their source Look at what AI has learned from. Text is compressed knowledge. Code is reasoning that can be run and checked. Images are perception. Audio is speech and timing. Video is motion. Every one is something a mind made, pressed into a format and stripped of nearly everything that produced it. An essay keeps the conclusion and throws away the hesitations. A program keeps the logic and throws away the forty attempts before it worked. A photograph keeps the scene and loses what the photographer noticed in it. > Neural data is to those five what the Sun is to every other source of energy on Earth: the thing the rest are derived from. #### Figure 15 (interactive): Five products, one source Everything AI has learned from is a compressed product of human cognition. Neural data is the one recording taken at the engine itself. Select each to see what it keeps and what it loses. [Diagram: A central node, the cognitive engine, with lines to five products: text, code, images, audio and video. A sixth dashed line leads to neural data, recorded at the source.] ##### Text - **Preserves** Conclusions, arguments and knowledge, in a deliberate, revised, error-corrected code. - **Loses** The process: hesitations, discarded alternatives, what was noticed on the way. - **Available** Effectively the public web. ##### Code - **Preserves** Reasoning in a form that can be executed and checked. - **Loses** Why it was written that way, and everything tried before it worked. - **Available** Billions of lines in public repositories. ##### Images - **Preserves** What a scene looked like from one viewpoint at one instant. - **Loses** What the viewer attended to, recognised or ignored. - **Available** Billions of captioned images. ##### Audio - **Preserves** Speech with its timing, stress and tone. - **Loses** Whatever was thought and not said. - **Available** Hundreds of thousands of hours in a single training set.[[42](https://setu-labs.co/manifesto-sources.html#s42)] ##### Video - **Preserves** Movement and sequence in the world. - **Loses** The intention behind a movement. - **Available** Vast and growing. ##### Neural data - **Preserves** The process before it is compressed into behaviour: attention and focus, error detection, preference and judgment, intent, semantics, and in time the intuition and the hunch. - **Loses** Today, resolution. That is a limit of sensors and of data, an engineering frontier and not a law. - **Available** Hundreds to thousands of hours so far (Figure 10). This is the gap to close. ***The authors’ thesis, drawn.** “Preserves” and “loses” are our characterisation. The dashed line marks the one connection that is still being built.* What lives at the source and never reaches the page is what makes experts expert: intuition, pattern recognition, the discomfort that says a result is wrong before you can say why, the hunch that comes before the proof. Philosophers call it tacit knowledge. We know more than we can tell. It cannot be scraped, because it was never written down. The size of what is missing can be estimated. Behaviour leaves a person at roughly ten bits a second. The senses take in something like a billion.[[3](https://setu-labs.co/manifesto-sources.html#s3)] Most of that gap is noise. But a person’s state of mind lives somewhere inside it, and behaviour never carries it out. Today’s sensors capture a small fraction. We read that as the size of the opportunity. The source is there. The instrument is what has to improve. ### What the source contains #### Figure 16 (table): What the source contains Seven kinds of information in neural data that no finished text, image or recording of behaviour carries directly. Five are already measurable. | Signal | What has been shown | Sensor | | --- | --- | --- | | Attention and focus | Which of two simultaneous speakers a listener is following, from about a minute of unaveraged signal.[[50](https://setu-labs.co/manifesto-sources.html#s50)] | Scalp EEG | | Error detection | A response within a fraction of a second of seeing a mistake, detectable on single trials and already used to teach a device.[[48](https://setu-labs.co/manifesto-sources.html#s48)][[49](https://setu-labs.co/manifesto-sources.html#s49)] | Scalp EEG | | Intent | Intended movement decoded well enough to drive cursors, robotic arms, handwriting and speech.[[6](https://setu-labs.co/manifesto-sources.html#s6)][[20](https://setu-labs.co/manifesto-sources.html#s20)] | Implants; scalp EEG at lower dimension | | Semantics | The gist of heard language and the look of viewed images;[[11](https://setu-labs.co/manifesto-sources.html#s11)][[12](https://setu-labs.co/manifesto-sources.html#s12)] semantic maps consistent across people.[[56](https://setu-labs.co/manifesto-sources.html#s56)] | fMRI, MEG | | Language and syntax | Typed and attempted sentences decoded word by word.[[16](https://setu-labs.co/manifesto-sources.html#s16)][[8](https://setu-labs.co/manifesto-sources.html#s8)] | MEG; implants | | Preference and judgment | Value and affect signals reported in controlled studies. Making them robust for one person across days is the engineering frontier. | Scalp EEG and others | | Reasoning | The path to a conclusion: intuition, pattern recognition, discomfort, the hunch. Not yet recorded in recoverable form. We expect synchronised data at scale to change that. | To be built | *Cited rows summarise those studies under their own conditions. The last two rows state where we expect the field to go.* Notice which sensor does the reading. Attention and error detection are read with EEG today, and so are the rhythms of sleep.[[50](https://setu-labs.co/manifesto-sources.html#s50)][[48](https://setu-labs.co/manifesto-sources.html#s48)][[51](https://setu-labs.co/manifesto-sources.html#s51)] Meaning is laid out across the cortex in patterns consistent enough between people to draw an atlas.[[56](https://setu-labs.co/manifesto-sources.html#s56)] Three properties make this data unusually valuable. It is raw: nobody has decided in advance what mattered, so the model can. It is synchronised: a recording time-locked to what the person saw, did and corrected labels every moment of a task for free. And it is personal: each user gets a small model that maps their brain into the same shared space that language and vision models already use. ### Brains have already taught models Artificial neural networks began as abstractions of biological ones. AI then scaled far past its source. The flow is now running back. Language models trained only to predict the next word turn out to predict human brain responses to sentences, and the better the model, the closer the match.[[44](https://setu-labs.co/manifesto-sources.html#s44)][[45](https://setu-labs.co/manifesto-sources.html#s45)] Two different systems, set the same problem, converged. And when brain data is put into training, models improve. Image classifiers nudged toward the organisation of a mouse’s visual cortex became more robust.[[46](https://setu-labs.co/manifesto-sources.html#s46)] Speech models tuned on brain recordings of people listening to stories became better at meaning.[[47](https://setu-labs.co/manifesto-sources.html#s47)] Those gains came from a few animals and a few listeners. Nobody has run the experiment at scale. That is the experiment worth running. ### The first wedge: data for AI labs The old fuel is running out. On current trends, models will have consumed essentially all public human text by around 2028.[[43](https://setu-labs.co/manifesto-sources.html#s43)] The frontier laboratories know it, which is why they now pay for data they once scraped for nothing. Every new kind of data they added bought a new capability. Code brought reasoning that can be checked. Images brought perception. Audio and video brought speech and motion. The next kind has to come from somewhere, and neural data is the one modality nobody has. It cannot be scraped and it cannot be synthesised. Someone has to build the sensors and gather it. Our working estimate: at about $300 per recorded hour, ten million hours of neural recording costs roughly three billion dollars. That is less than one large GPU cluster, in an industry planning to spend hundreds of billions on compute. And the data does not need to be perfect. At the valuations frontier models now command, a corpus that improved reasoning or alignment by one percent would repay itself many times over. That is why this wedge is immediate. A hardware company normally has to wait for millions of consumers before its data is worth anything. Here the buyer exists today, and has deeper pockets than any consumer market. A laboratory can fund the corpus before the consumer product is finished. Better interfaces will produce better models, and better models will fund better interfaces. That corpus cannot be recorded in hospitals. It can only come from an inexpensive device worn by very large numbers of consenting people. In practice that means EEG. We expect a frontier laboratory to name neural data in the training of a flagship model within a few years. ### The lossy postcard A prompt is a postcard from a rich internal state. It leaves out the tone you wanted, the constraint you forgot to mention, the thing the last draft got wrong. Models fail on that compression at least as often as they fail on intelligence. A BCI widens the channel twice. It lets intent out with less effort. And it lets context in without being asked for: what you are attending to, whether you are lost, when you have noticed a mistake. In time, prompting becomes something that happens continuously in the background. The same channel speeds the making of AI. Labelling data is a person looking at examples and clicking. A reader whose recognition and doubt can be sensed directly could judge a stream of examples many times faster. ### The second wedge: alignment The biggest point may be this one. AI systems are shaped after training by human judgment: raters comparing answers, users pressing approve or try again. This is how models are aligned with what people want, and a human can judge only so many answers in an hour. Feedback is the scarcest ingredient in making models behave well, and every laboratory needs more of it, for every model, now. Consider the error signal: an EEG response that says *that was wrong* within a few hundred milliseconds, without the person lifting a finger. It has already been used as a teacher. A device tries actions, the person only watches, and the error responses steer the device toward the goal.[[49](https://setu-labs.co/manifesto-sources.html#s49)] Extend that from a robot to a model and neural signals become an alignment technique: a literal, better RLHF. A click carries one bit, yes or no, after the fact. A reaction carries when the doubt began, which sentence caused it and how strong it was, as it happens. The human stops being the bottleneck in the loop and becomes its richest signal. This needs no consumer product either. A laboratory can put the sensor on its own raters tomorrow. The first users of an alignment interface are the people already paid to judge models, which is what makes this the second immediate wedge. It has one requirement. A reaction is not a decision. Systems trained on clicks learned to capture attention, and nobody should repeat that with a deeper probe. Reactions may inform a system. Only decisions should authorise it. ### From watching the machine to joining it There are two ways to keep advanced AI in line with human interests. The first works from outside: rules, evaluation, governance. It is indispensable now, and it weakens as machines outpace the people watching them. The second works from inside, by coupling human intent to the system closely enough that steering does not have to be imposed afterward. BCI belongs to the second, and almost nothing else does. It contributes in three ways. A field that decodes the brain at scale produces the best available map of how values are held in a mind, which is guidance for building systems meant to share them. A richer channel shows what a click cannot: whether the person understood the plan, or was merely soothed. And over a longer horizon, AI becomes less a separate agent to be watched and more an extension of a person’s own thinking. One question hangs over all of it: aligned to whom? A system tuned to the minds of one company or one state would be a triumph of engineering and a failure of everything else. The signal has to come from many people, over long periods, and keep changing as people do. ### A graph, and how graphs get solved The brain is a graph, and the task is to make meaning out of it. The complete brain of a fruit fly has been mapped: 139,255 neurons and fifty million synapses.[[55](https://setu-labs.co/manifesto-sources.html#s55)] A human brain has about 86 billion neurons.[[60](https://setu-labs.co/manifesto-sources.html#s60)] Nobody will trace that by hand. [Image: Three views of a fruit-fly central nervous system with every neuron position plotted as a coloured point.] *Every neuron in a fruit fly’s central nervous system, positioned from an electron-microscope reconstruction. A human brain has hundreds of thousands of times as many. Image: [Jason Alan Snyder; data: the MaleCNS team, CC BY 4.0](https://commons.wikimedia.org/wiki/File:MaleCNS_v1.0_Drosophila_male_central_nervous_system,_neuron_positions_in_three_views.png).* Nobody traced the internet by hand either. The web is the largest graph humans ever built, and its meaning was extracted by statistics at scale, by models fed enough of it. That is how a graph too large to read gets solved, and it is the template for the brain. ### Why now Because everything that was missing has arrived inside the same few years. #### Figure 17 (timeline): Four clocks striking together The milestones this essay relies on, by year. Models, decoding results, wearable sensing and governance each matured separately. Most of the dots fall after 2020. **Machine learning**2012 deep networks win at image recognition · 2017 the transformer · 2020 self-supervised speech models · 2022 a 680,000-hour speech model[[42](https://setu-labs.co/manifesto-sources.html#s42)] and public chat interfaces **Decoding results**2012 robotic arm[[20](https://setu-labs.co/manifesto-sources.html#s20)] · 2021 handwriting, speech, touch[[6](https://setu-labs.co/manifesto-sources.html#s6)][[10](https://setu-labs.co/manifesto-sources.html#s10)][[21](https://setu-labs.co/manifesto-sources.html#s21)] · 2023 fast speech, fMRI language and images, walking[[7](https://setu-labs.co/manifesto-sources.html#s7)][[8](https://setu-labs.co/manifesto-sources.html#s8)][[11](https://setu-labs.co/manifesto-sources.html#s11)][[23](https://setu-labs.co/manifesto-sources.html#s23)] · 2024 sustained accuracy, one-hour adaptation[[9](https://setu-labs.co/manifesto-sources.html#s9)][[13](https://setu-labs.co/manifesto-sources.html#s13)] · 2025 inner speech, typing, scaling[[25](https://setu-labs.co/manifesto-sources.html#s25)][[15](https://setu-labs.co/manifesto-sources.html#s15)][[34](https://setu-labs.co/manifesto-sources.html#s34)] · 2026 MEG sentences[[16](https://setu-labs.co/manifesto-sources.html#s16)] **Sensors and periphery**2018 wearable magnetic sensors[[30](https://setu-labs.co/manifesto-sources.html#s30)] · 2020 silent speech from muscle[[18](https://setu-labs.co/manifesto-sources.html#s18)] · 2023 catheter-delivered electrodes[[24](https://setu-labs.co/manifesto-sources.html#s24)] · 2025 a wristband that works on strangers[[19](https://setu-labs.co/manifesto-sources.html#s19)] **Governance**2017 ethical priorities set out[[31](https://setu-labs.co/manifesto-sources.html#s31)] · 2021 first constitutional protection[[33](https://setu-labs.co/manifesto-sources.html#s33)] · 2025 first global standard[[32](https://setu-labs.co/manifesto-sources.html#s32)] ***A selected record, not a census.** Each dot is a milestone cited in this essay, placed at its publication year; the three earliest machine-learning milestones are common knowledge and uncited. Because the essay concentrates on recent work, some of the clustering reflects that choice. The figure shows coincidence in time and makes no claim about what comes next.* The models exist. The proofs exist. EEG hardware has become cheap, and a decoder fits in a pocket. The demand exists, because capable agents have raised the value of every instruction a person gives. And the old data is nearly spent. That is a window, and it has two edges. The groups that build the sensors and gather the first large, consented datasets will set the direction of the field. And the terms on which people and machines relate are being set now, on a schedule of months. An interface that arrives after those terms harden arrives too late to shape them. Surgery cannot scale inside that interval. EEG can.