How is it that people can’t afford rent, can’t imagine a down payment, watch their paycheck lose ground every single year, and yet the line for the next fad still wraps the block? The easy answer is that people are simply bad with money, that this is “main character energy” dressed up as poor economic behavior. But something more specific is happening, and can be broken down into three separate layers of explanation before collapsing into a single underlying mechanism.

The first layer: Financial nihilism

The traditionally responsible path to security, saving diligently, buying a home, building wealth across a career, has been structurally foreclosed by forces outside an individual’s control: interest rates, wage stagnation, asset prices moving permanently out of reach. The math of deferred consumption only works if the deferred goal remains reachable within a plausible time horizon. Why forgo the $18 matcha, or the Labubu, to save toward a down payment that is still fifteen years away at your current savings rate? It will still be fifteen years away regardless of how much matcha you skipped.

Behavioral economics commentary has a name for this: financial nihilism, the growing sense that once a payoff stops feeling reachable, delaying gratification for it stops being rational. The marginal dollar moves instead toward a purchase with a guaranteed, immediate, and controllable return, one that delivers on the day it is made rather than deferring to a future that might never come.

The second layer: The consumption identity function

Underneath the financial logic sits a second, harder-to-quantify force: the need for belonging and status. Churches, unions, stable neighborhoods, a single employer held across a career— these institutions have eroded across the developed world. The decline has been sharpened by pandemic-era isolation and by social media’s replacement of physical community with algorithmic community. Social media did not create the demand for identity and belonging. It became the cheapest available infrastructure for satisfying it, and that infrastructure runs on consumption as its native language.

A Labubu or a Rhode drop, then, isn’t really about the object at all. It’s a twenty to hundred dollar ticket into a legible identity group, with instant, visible, quantifiable feedback: you post it, people recognize it, engagement accrues, and belonging is confirmed within minutes. Homeownership served this exact function for most of the twentieth century, the marker of adult stability itself, and it is now structurally unreachable for most people under 35 in any city with real economic opportunity.

Status signaling hasn’t disappeared; it has simply migrated downward to whatever venue remains within reach.

The third layer: two different budget lines

People are not failing to save for homes because they bought a Labubu. They are failing to save for homes because the ratio between home prices and income broke decisively between roughly 2012 and 2020, and no plausible rate of personal saving has closed that gap since.

In 1985, the median American home cost roughly 3.5 times median household income. By 2024, that ratio had climbed to about five times income nationally, and considerably higher in the metros where jobs actually concentrate: 10.5 times income in San Francisco, 10.8 in Los Angeles, 12 in San Jose. Between 2019 and 2024 alone, national median home prices rose somewhere between 31% and 48% depending on the measure used, against income growth of only 22% over the same span. One estimate quantifies what this gap costs a household in real terms. Had income grown at the same pace as home prices since 1980, the typical American household would earn roughly $115,225 today instead of the actual $83,730, a shortfall exceeding $31,000 a year, compounding annually.

That is not a gap that gets closed by forgoing a $20 tote bag. It is a structural fracture between the price of the asset that used to define middle-class security and the wages available to purchase it. Once a person mentally files homeownership under not happening for me, the marginal dollar simply gets reallocated toward present-tense wellbeing rather than a future that no longer computes.

Bringing it all together: capital outrunning labor

French economist Thomas Piketty’s central argument in Capital in the Twenty-First Century explains that wealth concentrates in the hands of those who already hold assets whenever the rate of return on capital, r, exceeds the rate of growth of the overall economy, g. Not because they work harder or invest more wisely, but because compounding on a larger base mechanically outpaces compounding on a wage. The Federal Reserve’s Distributional Financial Accounts quantify the current scale of that concentration: as of the third quarter of 2025, the top 1% of American households held 31.7% of total household wealth, the highest share on record since 1989. The top 10% held just over 68%. The bottom half held 2.5%.

A lease renewal and a trillion-dollar IPO landing in the same city in the same month are not two unrelated economic events. They are the same mechanism showing up twice, once for the people who own the assets and once for the people who don’t. Owning equity in an AI company right now means your money grows simply because you already had money in the right place. Renting an apartment in the same city means paying the cost of everyone else’s money flowing toward that same place. One side collects the gain. The other side absorbs the bill.

So where do we go from here?

If the mechanism above is structural rather than cyclical, I think there are three trajectories that follow:

An inheritance economy: Home and equity ownership increasingly cluster in households that already had family capital to begin with, meaning wealth becomes progressively more hereditary and progressively less a function of individual effort. Sociologists have also documented a rise in assortative mating by class and education, meaning wealth increasingly marries wealth, reinforcing the divide across generations rather than diffusing it. The middle tier of self-made asset accumulation shrinks as entry costs outpace wage growth. Geographic sorting intensifies alongside it, wealth clustering in a handful of metros while everyone else is priced into exurbs and secondary cities, producing a form of regional stratification by net worth.

A narrower ladder: Automation, AI-driven or otherwise, applies pressure specifically to the mid-skill cognitive jobs that historically served as the ladder into the asset-owning class: paralegal work, junior analysis, entry-level coding, customer support. If that ladder narrows, fewer people ever cross from labor income into capital income to begin with. The people who do cross tend disproportionately to be the ones who already had a head start, whether through family capital, elite education access, or geographic proximity to opportunity.

Reform, or the shock that forces one: Historian Walter Scheidel’s argument in The Great Leveler claims that inequality at this scale has historically reversed only through mass mobilization warfare, revolution, state collapse, or catastrophic plague, with deliberate high-tax redistribution as the rare non-catastrophic exception. Applied here, that gives two branches. A reform branch: wealth taxes, inheritance reform, aggressive zoning liberalization to break artificial housing scarcity, UBI-style experiments if automation displaces enough labor. This is politically difficult in the American context given how entrenched capital already is in campaign finance, but not impossible if the growing bloc of people who feel structurally locked out becomes electorally decisive. Or a no-reform branch, where the gap simply keeps widening until a severe recession, a demographic cliff, or enough social instability forces a correction reactively instead of proactively. This is the pattern Scheidel’s data suggests is historically the default.

None of this requires a villain, which is exactly why it will not fix itself. A landlord resetting rent to market rate, a fund buying up a block of single-family homes, a household with existing capital simply outbidding one without it—each of these is a rational, defensible, entirely unremarkable decision on its own. Stack a few million of them and you get a price-to-income ratio of twelve in Silicon Valley and the bottom half of the country holding onto 2.5% of its wealth. The Labubu line and the lease renewal are not a contradiction. They are what it looks like when individual choices, however disciplined, cannot close a gap from below that was never opened by anyone’s individual choice to begin with. The only mechanisms that have ever closed a gap this size are policy, deliberately, or catastrophe, by default. And the people with the clearest view of which lever gets pulled are, almost without exception, the ones furthest from the ratio. They watch it climb from the side where it has never once been their problem to solve.

Strip away the biology and you are looking at the core design principle behind reinforcement learning. Start by trying something. If it gets you a better result than before, do more of that. If not, try something else. The reward signal for the bacterium is a nutrient gradient instead of the output from a loss function, but the shape of the problem is identical: how much do you keep doing what has worked, and how much risk do you take trying something new. An RL agent bumping around a maze and a bacterium bumping around a petri dish are running the same calibration, separated by roughly all of evolutionary history.

This is the thing that should actually surprise you about intelligence. Not that AI can mimic the brain. That the brain itself was already running a much older program that started before there were brains at all.

The same calibration, three systems

None of these three systems reduces cleanly to one number. Evolution sometimes keeps two competing traits alive in the same population for generations instead of picking a winner. The brain does not run one single value signal either, it has several systems that weigh things differently depending on the kind of decision. What they do share is simpler than a single scalar: all three are constantly weighing how much to risk on something new against how much to lean on what already works, and that weighing shifts as new information comes in.

Evolution

A trait survives when it helps an organism live long enough to reproduce. The vertebrate spine has stuck around for hundreds of millions of years because it keeps working across an enormous range of situations, and the same is true of the eye, which nature independently invented dozens of separate times because it kept paying off. This is the safe, reliable side of the equation. It sticks with what already works because what already works is, by definition, still alive.

But nothing stays static, because mutation, sexual recombination, and random genetic drift keep introducing new variations into every generation. Most of these variations are neutral or actively harmful, which is exactly why evolution does not go looking for more of them than it has to. Change is expensive. The peppered moth shows why it pays off anyway. A rare dark-colored variant sat in the population for generations doing nothing useful, a bet nobody was collecting on, until industrial soot blackened the tree bark across English cities and that same variant suddenly became the one that survived being eaten by birds. The variation was already there. It just took a change in the environment for it to start paying out.

Brains

Brains run a much faster version of the same trade-off. There are neurons deep in the brain that release dopamine not just when something good happens, but specifically when something good happens and you did not expect it. Get exactly what you predicted, and the signal barely moves. Get more than you predicted, and dopamine spikes, which is the brain’s way of saying pay closer attention to whatever you just did, that worked better than planned. Get less than expected, and the signal drops, weakening your trust in whatever led you there. This is how a brain updates its own confidence in real time instead of waiting a whole generation to find out if something works, the way evolution has to.

The reliable half of this shows up as habit. When the same connections between neurons fire together over and over, they get physically strengthened, until a behavior needs no conscious thought at all, tying your shoes, driving a familiar route. The other half shows up as attention. When something violates what your brain expected, a different system floods in and interrupts the habit, forcing you to actually notice what changed instead of running on autopilot. Both halves are necessary. A brain that only ever built habits would never notice when the world had shifted underneath it.

AI

Large language models run a designed version of the same trade-off. During training, the model makes a guess, checks how wrong it was, and adjusts itself to be less wrong next time, over and over, billions of times. That adjustment is the reliable, the exploit half. But if a model only ever did that, it would lock onto one narrow way of doing things and fail the moment it saw something slightly different from its training data. So engineers deliberately add randomness back in, sometimes by switching off parts of the network during training, sometimes by rewarding the model for trying something unexpected. This is the machine version of not letting a whole population narrow down to one identical genotype. It is not a coincidence that it works this way. The people who built these systems were often directly borrowing ideas from neuroscience and evolutionary biology when they designed them.

Three systems share one logic beneath the surface differences: weigh the expected payoff of what is known against the expected payoff of what is not, and let that ratio shift as new information comes in.

Where culture refuses the calculation

Individual brains still hit hard limits. Working memory is small, attention is a bottleneck, and a single lifespan caps how much any one brain can learn firsthand. Culture is one of the things that emerged to push past that ceiling.

Culture came from brains attempting to coordinate, share knowledge, and build a collective understanding of experience. When humans developed language, traditions, institutions, and collective memory, nobody was consciously designing a solution to cognitive scaling. Culture emerged organically from the interactions between pattern-finding and pattern-breaking brains, and it carries its own version of the same tension. Traditions preserve useful patterns across generations, accumulated wisdom about food preparation, social coordination, and conflict resolution. Cultures also maintain mechanisms for innovation and adaptation when circumstances shift.

For most of what culture does, this still looks like the same risk-versus-reward calibration running at a social timescale. Adopt the new farming technique or stay with the old one. Trust the new trade route or the established one. These are still, underneath, bets being weighed against expected payoff.

Then there is a category of things culture does that will not enter that calculation at all. Anthropologists and political psychologists studying what they call sacred values have found something specific and strange. Offer someone money to compromise a value they hold sacred, a piece of ancestral land, a religious commitment, a core moral conviction, and they do not treat the offer as a bad deal to be turned down. They treat the offer itself as an insult, and the insult gets worse as the amount of money increases. Scott Atran, an anthropologist who has spent years doing fieldwork in active conflict zones, including negotiations over land disputes, found that raising the price does not move people closer to a deal. It moves them further away, because a higher offer reads as a more serious attempt to price something that was never supposed to have a price at all.

This is not limited to war zones or land. The same pattern shows up in ordinary moral life. Ask a parent to name a dollar figure at which they would sell their child, purely as a thought experiment, and you do not get a negotiation. You get anger, sometimes at the mere fact of being asked. Or consider how people react to organizations that publicly frame ethical decisions in cost-benefit terms, like a company calculating that a safety recall is not worth the expense compared to the cost of expected lawsuits. People do not respond to that reasoning as a bad financial forecast. They respond to it as a moral violation, specifically because the calculation happened at all, regardless of whether the math was right.

A system calibrating risk against reward would never behave this way. A bigger number should always look more attractive than a smaller one, full stop. That a bigger number can make someone angrier is not the exploit-and-explore calibration producing an unusual answer. It is a sign the calibration was never running in the first place, because the thing being asked for refuses to sit on the same axis as money to begin with.

That is one claim: that culture protects certain things by refusing to place them in a cost-benefit calculation at all. But there is a second, stronger claim underneath it, and it is the one that actually matters for how we think about building systems meant to learn human values. It is not just that people decline to trade sacred things away. It is that the moment someone is asked to weigh a sacred value against a number, the asking itself is the injury. The comparison is not a neutral step on the way to an answer. The comparison is the violation.

That distinction matters enormously for anything built on the idea of comparing two options and learning which one people prefer, which is precisely how a great deal of modern AI alignment work operates. A human rater is shown two possible outputs and asked which one is better. The system then learns from thousands of these comparisons what to value. That method assumes, as a basic precondition, that whatever it is asking about is the kind of thing that can be placed side by side and ranked. Sacred values are defined by their refusal to enter exactly that kind of ranking. Which means a method built entirely out of comparisons is not simply weak at capturing these values. It is structurally incapable of asking about them without reproducing the very violation the value was designed to resist. You cannot poll your way to an answer about something whose entire meaning depends on never being polled.

Any system built by asking someone to compare and rank two options is already assuming the thing being ranked is the kind of thing that can be ranked, and some of what people call values were built specifically to refuse that.

This is the thing the phrase “human values” hides. When AI researchers talk about aligning machines with human values, it sounds simple, as if there were a shared moral universe and a single standard of right and wrong to point machines toward. There isn’t. Intelligence, agency, and control are understood in radically different ways across cultures. Western traditions tend to equate intelligence with abstract reasoning and autonomous decision-making. Many collectivist societies treat intelligence as relational: the ability to act with discernment in a social context, hold a group together, and meet one’s obligations. Some Indigenous cosmologies distribute agency across people, ancestors, and the environment, so that control is relational rather than something an individual holds.

So “human values” are rarely neutral. In practice, it means transparency, explicit consent, rational deliberation, and individual autonomy, values that come straight out of liberal political philosophy and Western moral psychology. These are not bad values. They are specific ones. Calling them universal hides the fact that even the idea of a “value” gets interpreted differently from one moral world to the next.

Human values and moral variation

What do we mean by human values anyway? Jonathan Haidt, a social psychologist, argues that humans share a set of overlapping foundations, including care, fairness, loyalty, authority, and sanctity, but that cultures weigh them very differently. Western liberal societies prize fairness and autonomy. Others put loyalty, respect for hierarchy, or communal responsibility first. Care in one place means defending individual choice. In another, it means holding the community together or honoring what came before.

An AI trained mostly on Western data absorbs one weighting of these foundations and treats it as the default. It learns to recognize certain forms of reasoning and moral expression as rational, and it can fail to register other forms as legitimate at all. That is worse than bias. It is a kind of epistemic flattening in which the machine appears neutral while quietly privileging one worldview.

Technology as a carrier of values

Technology has always carried moral implications. European mapping and census systems imposed individual property norms on societies with communal land practices. Industrial factories turned efficiency and punctuality into moral imperatives, reshaping daily life and social hierarchies. The early Internet spread Silicon Valley’s libertarian ethos - all individual expression, disruption, and openness - into societies that valued collective responsibility, indirect communication, and deference.

AI continues this pattern, which some scholars refer to as algorithmic imperialism. Sociologist Michael Kwet describes “digital colonialism” as the global spread of technological infrastructures that reproduce economic and epistemic hierarchies. AI, trained on one culture’s data, operationalizes that culture’s moral logic, then projects it globally while presenting itself as neutral. Content moderation models enforce US-style free-speech norms, while credit and hiring algorithms privilege Western notions of merit and productivity. Language models tend to favor directness and assertiveness, often misinterpreting indirect speech or ritualized deference as a sign of uncertainty.

The effect is predictable. Those whose moral and cognitive patterns the AI already understands benefit the most. Others must adapt to speak machine. It is a subtle continuation of colonial asymmetry: a technology that calls itself neutral while embedding hierarchies of knowledge and value into everything it touches.

Uneven consequences

This is why the land dispute matters beyond itself. An AI is not just a tool sitting outside the moral world. It participates in it. Every recommendation and every automated decision is an act of translation, and translation always chooses. It privileges some moral and cognitive patterns over others, rewards the people whose patterns match, and quietly forces everyone else to adapt or be misread. The forest case is not an edge case. It is what the ordinary operation of these systems looks like when the people on the other end do not share the worldview from which the machine was built.

Towards situated and modular alignment

If that is the situation, a single global moral standard for AI will not hold. Alignment has to be situated. Systems should adapt to local moral ecologies rather than imposing a single interpretation on all of them.

Modular design is one way through. Instead of a single model carrying a single worldview, you could build systems from locally trained modules, each grounded in a specific cultural and ethical framework, connected by a negotiation layer that functions as a translator or mediator rather than a top-down authority. Dialogue style, decision norms, and ethical trade-offs would shift to match local expectations rather than defaulting to a single one.

Participatory design has to be part of it too. The communities a system affects should get to define what aligned behavior even means for them. And cultural impact audits, modeled on environmental impact assessments, could flag in advance where a system is about to overwrite or erase a local way of knowing. Alignment stops being about imposing control and starts being about negotiation, translation, and mutual intelligibility.

At that point, alignment is a problem of cross-cultural communication at a planetary scale. Anthropology has spent a century studying how radically different meaning systems collide and what happens when they do. Moral psychology tells us which tendencies humans share. Ethnography shows how differently those tendencies get expressed in practice. A real alignment effort has to work in all these registers at once. The job is not to encode one moral blueprint. It is to build systems that can recognize, translate, and negotiate across many moral worlds. Where engineers see optimization, anthropologists see negotiation. Where philosophers see universals, ethnographers see situated practice. A system that is going to serve more than one civilization has to operate on both terrains.

What we are actually building

Every technology carries a theory of what a human is. AI is no exception, and its scale magnifies whatever theory it carries.

Here is the uncomfortable part. The field is pouring enormous effort into aligning AI with “human values,” optimizing hard against a target stated as if it pointed at one coherent thing. It doesn’t. So we are getting very good at solving a problem that was specified incorrectly, and the better we get, the more confidently we export one civilization’s moral logic under the banner of serving everyone.

The harder version of the problem is plural, situated, and negotiated, and it is not science fiction. The pieces are already on the table: modular systems, participatory design, cultural audits, a negotiation layer instead of a default. The work is buildable, and it is where the effort needs to go, because we still have not admitted what the forest case shows. Keep optimizing for “human values” as if the phrase were settled, and we keep building the mining company’s machine and calling it everyone’s.

The thesis of the room was simple. Foundation models, the paradigm that gave us GPT, DALL-E, and AlphaFold, should in principle work for the brain. Train a large model on enormous quantities of neural data, let it learn general representations, fine-tune for downstream tasks: seizure detection, drug biomarker discovery, BCI control, mental health monitoring. The pitch is compelling and the analogy to language and protein modeling is obvious.

And yet nobody has cracked it. The funding isn’t flowing. The deeper you get into why, the more you realize this isn’t just a hard technical problem. It has a particular structure that makes it resistant to the ways we normally fund and solve hard problems.

Why the science is harder than it looks

Every successful foundation model was built on a canonical unit of representation. For language, it was the token: a discrete, standardized chunk of text that every model processes the same way. For proteins, it was the amino acid. These primitives made it possible to aggregate data across sources, scale training, and compare models meaningfully. They also defined what a self-supervised objective should look like: predict the masked token, predict the next residue.

Neural data hasn’t converged on a satisfying primitive. A spike train from a Utah array, an EEG channel, an fMRI voxel, a calcium imaging ROI: all measurements of brain activity, all fundamentally different in spatial resolution, temporal resolution, measurement modality, and signal-to-noise characteristics. EEG compounds this: volume conduction through the skull and scalp means each electrode records a spatially blurred superposition of thousands of underlying source configurations. Source localization via beamforming or ICA can partially unmix these signals, but the inverse problem is ill-posed. The same scalp potential distribution is consistent with infinitely many intracranial source configurations. This means EEG channels are not stable representational units across subjects, sessions, or electrode placements. You can’t tokenize them the way you tokenize text.

The field’s current workaround, treating EEG as a multi-channel time series and patching it into fixed-length segments for masked autoencoding, sidesteps the problem rather than solving it. Models like LaBraM and EEGformer do this, and they work in a narrow sense: they learn transferable representations that improve downstream task accuracy. But whether those representations encode anything neurophysiologically meaningful, or whether they’ve learned to exploit low-level statistical regularities in the training data (recording artifacts, channel-specific noise profiles, demographic confounds), is largely unknown. The evaluation benchmarks used, seizure classification on TUH and sleep staging on SHHS, are too coarse to distinguish between these two possibilities.

The second problem is context. One panelist made an observation that stuck with me: neural data might not be enough. This sounds obvious in retrospect but has radical implications. When you record EEG during a cognitive task, you’re capturing signals downstream of an enormous latent state: the participant’s arousal level, prior experience with the task, fatigue, internal narrative, physiological state. Language models work because the token sequence is the context; the model has access to everything it needs to make a good prediction. Neural signals are more like a compressed projection of a much higher-dimensional state space, and the projection operator, the skull, the scalp, the electrode impedance, discards most of the information you’d need to fully reconstruct what’s happening. Multimodal recording (EEG plus eye-tracking plus pupillometry plus behavioral readouts) can recover some of this lost context, but it requires careful experimental design and is difficult to scale.

Third: coverage. Non-invasive EEG captures population-level electrical activity primarily from cortical pyramidal neurons whose dipoles are radially oriented to the scalp. Deep structures, hippocampus, amygdala, basal ganglia, thalamus, contribute negligibly. Training on ten times more EEG data doesn’t give you access to hippocampal sharp-wave ripples or thalamo-cortical spindles. It gives you more of the same limited projection. The coverage problem is architectural, not statistical.

Fourth, and most under appreciated: interpretability in this domain is not optional. In most ML applications, interpretability is a secondary concern, something you add post-hoc for regulatory compliance or scientific curiosity. In clinical neuroscience, it’s a prerequisite for deployment. A neurologist won’t act on a model she can’t interrogate. FDA clearance for a diagnostic device requires that the decision logic be at least partially explicable in terms of known pathophysiology. The field is building toward a standard that is fundamentally harder than “does it generalize.” It requires “does it generalize in a way whose mechanisms connect to established neuroscience.” No existing EEG foundation model comes close to satisfying that standard. The attention maps and codebook analyses published in EEGformer and LaBraM are suggestive but not rigorous; they don’t demonstrate that what the model learned maps onto canonical frequency-band decompositions, event-related potential components like the N2, P300, or ERN, or the spatial topographies that neurologists actually use to read EEGs.

The funding trap

Now layer the funding picture on top of the science. The second panel was explicitly about commercialization and deployment, and what emerged was a structural doom loop.

Venture capital needs a product. A product needs a capable model. A capable model needs data at scale and ground truth labels. Ground truth in neuroscience is elusive in a way that’s genuinely different from other domains. AlphaFold could be validated against crystallography. There’s no equivalent for a model of human attention or decision confidence. The benchmarks that exist are mostly clinical proxies, seizure classification, sleep staging, pathology detection, that are poor measures of whether a model has learned something meaningful about neural computation. Without rigorous benchmarks, it’s hard to demonstrate progress. Without demonstrated progress, it’s hard to raise. Without capital, you can’t acquire the proprietary clinical datasets or compute at the scale needed to make progress. The loop closes.

There’s a compounding problem that received surprisingly little attention in the room: the healthy data gap. Nearly all available neural datasets come from clinical populations, epilepsy monitoring units, ALS trials, sleep disorder labs, psychiatric cohorts. Hospitals collect neural data in diagnostic contexts, and diagnostic contexts are contexts of dysfunction. If you’re trying to model typical neural computation, you’re mostly training on atypical brains. Models trained primarily on ictal and interictal EEG may have learned seizure-relevant features that don’t transfer to cognitive or affective states in healthy individuals. The effective training distribution is not “human brain activity.” It’s “human brain activity under pathological conditions, recorded in a clinical setting, on specific hardware, with specific preprocessing pipelines applied.”

The path to healthy-population data at scale runs through wearables, which is why companies like Oura matter to this conversation even though Oura isn’t an EEG company. Wearable biosignal devices are the only realistic path to longitudinal recordings from millions of healthy people in naturalistic settings. The data that would actually ground a brain foundation model probably isn’t sitting in hospital archives. It’s being passively collected right now by devices that aren’t yet sophisticated enough to fully exploit it.

The funding that is flowing comes from a specific profile: disease-focused foundations, government brain initiative programs, and a small number of investors who already believe in neuro for reasons independent of near-term product timelines. This is appropriate for the stage of science. But it means the capital is constrained, the timelines are long, and the urgency isn’t matched by the resources.

This mismatch points to a structural gap that the existing funding landscape isn’t built to fill. The core bottleneck, creating standardized, annotated, multi-modal neural datasets at the scale a real foundation model would require, is work that doesn’t fit neatly into a VC portfolio or a standard NIH R01. It’s too applied for academic grants, which fund hypotheses rather than infrastructure. It’s too pre-competitive for venture, which needs a product on a three-to-seven year horizon. And it’s too specialized and unglamorous to attract the kind of philanthropic attention that goes to, say, cancer genomics.

The Focused Research Organization model is worth paying attention to here. FROs, exemplified by groups coming out of Convergent Research, are time-limited, mission-driven teams organized around a specific scientific bottleneck, somewhere between a startup and an academic consortium. No product mandate, no tenure pressure, actual ability to staff up and execute. The Arc Institute runs on a similar logic. For a problem like neural data standardization and annotation, where the work is well-defined, consequential, and genuinely blocking downstream progress across multiple groups, an FRO is a more natural fit than either a lab or a company. The field hasn’t had one targeted at this problem yet. It probably should.

Where the leverage actually lives

None of this means the field is hopeless. It means the leverage is in different places than people assume.

Building another EEG foundation model trained on the TUH corpus with a masked reconstruction objective, evaluated on seizure detection and sleep staging, is probably not the highest-value contribution right now. LaBraM, EEGformer, BRANT, BIOT, NeuroGPT: the list is growing. The results are real but incremental, and the papers tend to benchmark against each other rather than against any external standard of what “understanding the brain” would mean.

The higher-leverage bets are upstream. What is the right inductive bias for a neural data model? Transformers applied to EEG patches treat channels as tokens and time as sequence length, which imposes a specific relational structure on the data. There are strong neurophysiological reasons to think cortical computation involves oscillatory dynamics, traveling waves, cross-frequency coupling, and non-stationary topographic reorganization, none of which are well-captured by standard attention mechanisms operating on fixed patches. Graph neural networks over anatomically-defined connectivity structures, state-space models with oscillatory priors, or architectures that explicitly model source dynamics rather than sensor signals might be more appropriate. This is an open architectural question with no consensus answer.

The self-supervised objective question is equally open. Masked autoencoding on raw EEG optimizes for reconstructing artifactual and physiologically meaningful signals alike, with no way to tell them apart. Alternative objectives grounded in neuroscience, predicting the power spectrum of a held-out frequency band, predicting cross-channel phase relationships, predicting behavioral outcomes from pre-stimulus activity, would build in stronger priors about what the model should learn. These haven’t been systematically explored.

The evaluation problem is arguably the most important. What would a rigorous benchmark for EEG representations actually look like? The field needs evaluation tasks that test neurophysiological validity, not just clinical utility: do the learned representations encode known ERP components (P300, N170, MMN) consistently across subjects? Do they capture the well-characterized relationship between alpha power and cortical excitability? Do they reproduce the known spectral fingerprints of different cognitive states? These aren’t hard to test given the right datasets, and they would give the field a much cleaner signal about what’s actually being learned.

The interpretability gap is an opportunity too, not just a constraint. The work of rigorously connecting what EEG foundation models learn to what neuroscience already knows, probing internal representations against established electrophysiological signatures, running activation patching experiments to identify which model components track which neural phenomena, building evaluation frameworks grounded in the psychophysiology literature, is both urgently needed and surprisingly underdone. Most current interpretability analysis in EEG-FM papers amounts to visualizing attention weights and noting that the model seems to attend to temporal discontinuities. That’s a start, not an answer.

I left that room with a narrower view than I walked in with, which I think means the event did its job. The researchers there weren’t saying anything they haven’t said in each other’s labs for years. What struck me wasn’t the ideas. It was that the same people who diagnosed the problem clearly couldn’t agree on which upstream bet to make. Architecture? Benchmarks? Data standards? Everyone had a different answer.

My answer, for what it’s worth: the data infrastructure problem has to come first, and it needs an institution designed to solve it rather than one that will pivot away from it the moment a product opportunity appears. That’s not a company. It might be an FRO.

The money will follow once there’s something to measure and a foundation to build on. Right now, there’s neither.

For a long time, I held those two things as a quiet contradiction. The rational life over here, the intuitive life over there, a polite wall between them. Then I started actually looking at what the science says about why these frameworks work. The wall dissolved. What replaced it wasn’t a tidy synthesis. It was something weirder and more interesting.

We are pattern-seeking animals. Those frameworks are what pattern-seeking animals build when they turn their full attention to the problem of being alive.

Every culture. Every time. The same systems.

Astrology wasn’t a Greek invention that spread around the world. The Babylonians developed it independently. So did the ancient Chinese. The Mayans. The Indians. Zero contact between them. They looked up at the same sky and arrived at the same conclusion: that celestial movement was a map for grasping human nature.

Divination appears without exception in every culture anthropologists have ever studied. Romans read the flight patterns of birds. West African Yoruba tradition built Ifá. The concept of cosmic moral order appears as karma, as ma’at in ancient Egypt, as divine justice in Christianity, as “what goes around comes around” in completely secular Western culture. Ritual cleansing is baptism, sage smudging, the Jewish mikveh, and Hindu purification rites. Consistent structure across every tradition that ever existed.

Anthropology calls it convergent evolution: when completely separate lineages independently develop the same solution. It means the solution is responding to something real. So before asking “is astrology true?” it’s worth asking why every human civilization on Earth, independently, felt the need to build it.

We are meaning-making animals.

Symbolic thinking goes back roughly 70,000 years, to when humans started burying their dead with objects, painting on cave walls, and organizing around shared stories. We weren’t just surviving the environment anymore. We were narrating it.

French sociologist Émile Durkheim spent his career studying why religion and ritual persist in every human society regardless of material conditions. His answer was that shared symbolic systems do something raw information can’t. They create social cohesion. They convert the terror of randomness into something navigable. They make cooperation between strangers possible at scale.

What he was really pointing at is that humans don’t just need to know things. They need to feel oriented. And information, however accurate, doesn’t automatically produce that. You can know intellectually that grief is a normal neurological response to loss and still feel utterly unmoored at 3 am. You can know the statistics on uncertainty and still need someone to tell you it’s going to be okay. The knowing and the feeling are processed by different systems, and the spiritual world, across every culture and era, has been running infrastructure for both simultaneously. Ritual gives the body something to do with what the mind can’t resolve. Shared ceremony creates a felt sense of “we are in this together” that no amount of correct information can replicate. That’s not irrationality. It’s a different kind of engineering, one that was solving real problems long before anyone had a language for what the problems were.

Which is why when someone tells you spiritual frameworks are just primitive attempts at science, they’re missing the point by about 180 degrees. They were never trying to be science. They were trying to do something science, on its own, still doesn’t do particularly well.

So what’s actually happening when the horoscope resonates?

When a Scorpio description lands, it’s not because the stars shaped your personality at birth. It’s the Forer effect: vague, broadly relatable descriptions that our brains personalize automatically. Bertram Forer demonstrated this in 1948 when he gave his students a “personalized” personality assessment, then revealed they’d all received the exact same one. They rated it 4.3 out of 5 for accuracy. We remember the hits. We forget the misses. And we fill the gaps with ourselves.

But the mechanism being fiction doesn’t make the outcome fiction. When the description lands, something real happens. You just had a flash of self-understanding. You found language for something you’d been carrying around without a name for it. Maybe it’s “I get emotionally guarded when I feel like I might lose something,” dressed up as Scorpio rising. The astrology didn’t put that truth there. It gave you a mirror angled just right to see what was already there.

There’s also something worth sitting with about why we need the mirror at all. Direct self-examination is genuinely hard. The brain has strong incentives to protect you from uncomfortable self-knowledge. But hand someone a symbolic framework, some external system with just enough distance from their actual life, and suddenly they can see things they couldn’t look at straight on. The horoscope works partly because it’s imprecise. The looseness is the feature. It creates enough interpretive space for you to do the real work yourself and then feel like you arrived there.

That’s not delusion. That’s a pretty elegant workaround for a real cognitive limitation.

Manifestation is real. The explanation is just wrong.

Your brain has a filtering system called the reticular activating system that decides, out of millions of stimuli hitting you every second, what deserves conscious attention. This connects to something deeper in how the brain works: predictive processing. Your brain isn’t passively receiving the world; it’s constantly generating models of what it expects to be true, then allocating attention toward evidence that confirms those models. When you set an intention clearly and repeatedly, you’re not just motivating yourself. You’re updating the model. You start noticing opportunities and openings that were always present but previously invisible, because you’ve told your brain they’re worth looking for.

Add the psychology of belief on top of that: when you genuinely expect something to be possible, you move through the world differently. You take more shots. You persist longer. You display confidence that other people respond to. The belief changes the behavior. The behavior changes the outcome. You needed a compelling story to get yourself moving. The story did its job.

This is what sits underneath all of it. Humans have always needed narratives emotionally compelling enough to motivate behavior. “You need to rewire your insecure attachment patterns through corrective emotional experiences” does not get most people out of bed in the morning. “Align with your highest self and trust the universe” does. Same destination. One of those vehicles just has better fuel. And if you’ve ever actually changed a behavior because of something a horoscope said, you already know this. The stars didn’t move you. The permission to believe something was possible did.

“Protecting your energy” isn’t mysticism. It’s biology.

You are never not scanning the people around you. In every interaction, your nervous system runs a continuous, unconscious assessment, processing posture, micro-expressions, vocal tone, breathing rhythm, and proximity. Not sequentially. All at once. Before conscious thought forms. Before you’ve decided what you think of someone, your body has already rendered a verdict.

This is evolutionary hardware. For most of human history, misreading another person’s intentions was potentially fatal. So the brain built a rapid, automatic social threat-detection system that runs constantly in the background, and it is not subtle in reporting its findings. You walk into a room, and something feels off before you can say why. You meet someone who ticks every box on paper and still feel vaguely unsettled. You sit next to a stranger on a train and, without a word exchanged, feel either completely at ease or like you want to move cars. That’s not a vibe. That’s a read.

What makes it stranger is that it’s bidirectional. You’re scanning everyone, and everyone is scanning you. Research in affective neuroscience shows that emotional states transfer through this process, a phenomenon called physiological co-regulation: people in close proximity begin to synchronize their heart rates, breathing patterns, and stress hormones. Sit next to someone anxious for long enough and your own cortisol levels measurably rise. You don’t have to share a word. You don’t even have to be paying attention.

What the spiritual world calls “protecting your energy” is this system. The heaviness you carry home from certain environments. The inexplicable ease you feel around specific people. The intuition that something is wrong before anyone has said anything. Your nervous system is reporting real data, accumulated across thousands of micro-signals you never consciously registered. The spiritual framing gave language to something the body already knew. And maybe that’s the most honest thing you can say about all of this: the language came second. The experience was always real.

So why do we keep dressing science up in spiritual clothing?

Because the spiritual world, across every culture and era, has been doing applied behavioral psychology without the clinical language. The frameworks were stuck not because people were naive, but because they were always pointing at something real, just explaining the mechanism wrong.

We are driven by the stories we tell ourselves about the world, about other people, about who we are and what we’re capable of. That’s not a flaw in the system. It is the system. And the stories that have persisted for thousands of years, across every civilization that ever existed, without anyone coordinating them, are probably pointing at something worth taking seriously. Not literally. But seriously.

The tarot card doesn’t know your future. But the question it forces you to sit with might be the right one.

The horoscope isn’t cosmically accurate. But the self-reflection it triggers is yours.

The “universe” isn’t conspiring on your behalf. But something in your brain has been working hard to find the path forward since before any of us had words for it.

The framework is never the problem. The problem is when it becomes the ceiling instead of the door.

Mine are neuroscience, AI, and the messier human stuff: culture, history, how we make meaning, why we behave the way we do. I am constantly drawn to how these three things loop back into each other in new ways. A question about how the brain represents the world turns into a question about how a language model does the same thing, which turns into a question about what “understanding” even means, which turns out to be a question philosophers and anthropologists have been arguing about for centuries. The braid holds because none of these questions live cleanly in one discipline.

I’m Pranati. I’m an MS student at Columbia studying CS and Neuroscience, which means I spend my days somewhere in the overlap between biological intelligence and artificial intelligence, trying to figure out what each one tells us about the other. I write about neuroAI, AI alignment, brain-computer interfaces, and occasionally about being a person moving through a world that all of this is reshaping faster than most of us can keep up with.

The Golden Braid has two kinds of posts. The technical essays that go deep: what foundation models actually reveal about how the brain works, what alignment research gets wrong about human values, what BCIs can and can’t do yet. And the field notes are shorter and more personal: things I noticed at a temple in Indonesia, what a semester of alignment research actually looks like from the inside, the places where a scientific education and a spiritual instinct keep running into each other.

If that sounds like something worth reading, subscribe!