What changes when AI learns entangled instead of smart alone

A paper titled Adaptive Entangled Game Modules in Artificial General Intelligence has appeared on arXiv as 2609.09226. The original is linked under the title. The paper claims that the collective behavior of interacting adaptive agents can be described by a generalized behavioral intelligence wave equation, and reports that the entangled game modes derived from that equation explain 89 percent of observed decision patterns in Chinese intraday stock market data.

That summary alone can sound like another metaphor borrowed from physics. What caught my attention was not the size of the number but the change in the direction of the question. Instead of asking how smart a single model is, the paper starts by asking what shape behavior takes when many agents stay entangled. This post follows that question from the start. I will walk through what the wave language means, under what conditions the 89 percent was obtained, and how far the proposal for AGI architecture carries.

Why AI that shines alone turns awkward together

When several people work as one team, a single remark can shift the mood of the whole room. Then a similar thing should happen when several AI agents work as one team, right. The question sounds obvious, but anyone who has built such a system knows it gets tricky fast. Individual benchmark scores keep rising, yet putting several strong models side by side often produces behavior nobody planned. The parts look smart and the whole acts strange.

I keep thinking of supermarket checkout lines when I see this gap. One fast cashier should shorten every line, but shoppers watch each other, switch lines, and soon one line stands empty while another stretches out. Each choice looks reasonable on its own, and the sum takes a shape nobody intended. Markets behave this way, and so does a warehouse where several robots circle the same aisles. Cleverness at the level of the individual and motion at the level of the crowd play on different floors.

That is why the starting point of this paper felt welcome to me. Instead of pushing the reasoning power of each agent higher, it declares that the entangled interaction itself should be modeled directly. When my team looks at a new AI architecture, we do not look at benchmark numbers alone. We look at what actually changed and what cost structure that change creates inside a system. Through that lens, this paper reads less like a score contest and more like a structural proposal. At the fork between making each unit bigger and writing the relationship first, it takes the second road.

A declaration alone is not enough, of course. For the word relationship to mean more than a slogan, the paper has to show how to measure it and how to compute with it. So what does it bring as the next step. The answer is to write the crowd in the mathematical language of waves.

Where the assumption of rational individuals starts to slip

Open an economics textbook and the market appears as a sum of independent rational actors. Each one reads information, each one decides alone, and prices come out of the total. People do act like that sometimes, for instance on a quiet morning when a trader studies a chart and presses buy alone. Yet anyone who watches an intraday market sees scenes this picture cannot cover. One order lands, more orders follow it, the depth of the order book shifts, and the mood flips within seconds.

This paper takes that mismatch head on. It analyzes Chinese intraday stock market data and reports that purely independent modes account for less than 5 percent of observed behavior. Adaptive entangled game modes, by contrast, account for 82 to 94 percent depending on the segment, and 89 percent overall. Readers who work with language models know the feeling that the same question can produce very different answers. In markets the closer saying would be that the same information can produce very different outcomes depending on who moved first.

This is where I would be careful not to over-read the experiment. The 89 percent is closer to a measure of how well this frame can describe decision patterns already observed than to a magic accuracy for predicting the whole market. Description and prediction are different yardsticks, and turning the first into the second right away would inflate what the paper did. What interests me more is the structure behind the number. The fact that independence looks like a rare exception in real data sends its own signal, which is that the starting assumption needs rewriting.

If independence is the exception, then what is the norm. The paper answers with two words, entanglement and adaptation. The next section looks at what shape those words take inside the math.

What it means to see behavior as a wave

The word wave may first bring up a ripple tank in a physics lab. In this paper, though, the wave sits closer to a computing tool than to a poetic image. It means writing how the action probabilities of many actors spread and overlap across time and context as a single field, and then setting down an equation for how that field evolves. What interests me more is exactly this choice. Instead of dissecting each mind one by one, it asks how the shape of overlapping probabilities changes.

A daily analogy makes the move easier to feel. Think of where people stand on a subway platform at rush hour. Everyone has a favorite car, yet when the person in front shifts, the person behind follows, and when an announcement plays, the whole center of weight moves at once. A clump motion appears that no single preference can explain. The wave equation is an attempt to write that clump motion as the evolution of a field rather than as a sum of persons. It counts less who went where and asks more how crowdedness spreads.

The paper calls this frame a generalized behavioral intelligence nonlocal probability wave equation. The long name should not scare anyone, because each part says something plain. Generalized means it refuses to tie itself to one market or one rule set. Nonlocal means distant actors can still affect each other. Behavioral intelligence means the subject is deciding agents rather than physical particles. Nobody needs to memorize every term, but the direction stays clear. Interaction enters as the skeleton from the start instead of arriving later as a correction.

Then what does such an equation actually buy. The answer connects to the idea of eigenmodes that comes next. Once behavior is written as a wave, a few stable shapes pop out on their own, and those shapes become the lens for reading observations.

How an entangled game differs from ordinary game theory

The word game easily recalls payoff tables like the prisoner's dilemma or rock paper scissors. But the entangled game in this paper is not a game that closes on one table. It looks less like finding the best move under fixed rules and more like a scene where the yardsticks move together while everyone watches each other. On the surface it looks simple. In practice it is not. When I change my choice after seeing yours, my changed choice then changes your next choice.

The use of the word entanglement needs care here. Reading it as quantum entanglement imported as is will cause confusion. What the paper means by entanglement is not a physical link between particles but a state where action probabilities refuse to split into independent pieces. To know the decision distribution of one actor, in other words, you have to look at the others together. Think of a choir. Each singer makes a different sound, yet heard separately there is no song, and only the overlap makes the harmony.

The word adaptation should be read in the same spirit. It does not mean repeating a fixed strategy but a motion where the reference point itself shifts with intraday news, events, and environment changes. Loss avoidance may lead in the morning, and after lunch news breaks, chasing the trend may take over. It is less that the rule changed than that what counts as the baseline changed. The paper reports that such adaptation covers about 2 to 12 percent of all behavior. The slice may look small, yet for interpretation it weighs a lot, because most of the time the entangled pattern holds while at certain moments the baseline jumps.

So where does this conceptual split show its force in actual analysis. The device that shows the answer is the eigenmode.

What you see when you read decisions through eigenmodes

The term eigenmode sounds hard, but the idea stays close to daily life. It means picking out a few stable shapes that keep reappearing inside a complicated wobble. Pluck a guitar string and many vibrations mix, yet what remains in the ear comes down to a few ringing shapes. The paper takes a similar view of markets. Countless orders and cancellations flow past, yet collective action keeps returning to a few familiar textures.

The paper derives such stable shapes from the wave equation and explains that each captures a broad spectrum of human intelligent behavior with an analytical mechanism. That sentence can sound grand, so let me unpack the actual claim. First, collective behavior can be split into several modes. Second, the math can trace under what conditions each mode appears and how it shifts. Third, observed behavior can therefore be decomposed in the language of independence, entanglement, and adaptation. The difference is not a new label table but a classification that comes out of the equation itself.

I felt some doubt at this point. Searching for stable shapes can easily slide into fitting the result after the fact. With enough modes, any data set can be made to look well described. So the count of modes matters less than testability. When the paper calls its modes testable, I want to know what that word guarantees in practice. It should be clear which future observation would prove the frame wrong, because only with that failure condition does the frame earn its weight. I will return to this point together with the data reading below.

For now, let me hold onto the feel of the mechanism. Writing the crowd as a wave means seeing one scene whole, splitting it into eigenmodes means reading the recurring poses of that scene, and naming entanglement and adaptation means catching the moment the pose changes. So what appeared when this lens was held up to real intraday data. Time to enter the world of numbers.

How to read the 89 percent in intraday trading

Since the number came first, the conditions should come with it. The paper analyzes Chinese intraday stock market data and reports that adaptively entangled game modes explain 82 to 94 percent of observed decision patterns, and 89 percent overall. I prefer to read these numbers as a pair. The width from 82 to 94 plus the headline 89 together reveal that each trading day carried its own texture.

Picture a concrete scene. At half past nine in the morning, the book fills fast right after the open. One trader anchors on yesterday's close, another anchors on the direction of the opening print. For the first minutes the anchors differ, but once volume builds, the anchors get pulled toward each other and lean one way. That lean is the scene the entangled mode catches. Alone, no single judgment would cluster this tightly, yet once traders watch each other's orders, their baselines can synchronize enough to make it happen.

This number should be checked against its conditions before anyone trusts it as is. Which tickers and which period were used, how a decision pattern was defined and counted, and where the line between independent and entangled modes was drawn can all move the value. From the summary alone, the fine detail of that boundary stays hard to confirm. So I read the 89 percent more as an arrow for direction than as a precise constant. The signal is that a starting guess close to independence misses most of intraday behavior. (Chinese intraday data is known for sharp volatility and visible herding, and that market character may well have pushed the explanatory share of entangled modes higher, which is worth keeping in mind.)

The next question follows directly. Saying that most behavior is explained raises the question of how the rest is handled. What do the leftover pieces really mean, the under 5 percent of independence and the 2 to 12 percent of adaptation. The next section continues there.

Between less than 5 percent independence and 2 to 12 percent adaptation

The report that independent modes sit below 5 percent can feel shocking at first. Half the market, we imagine, should consist of people deciding for themselves, yet in the data that state looks like an exception. Looked at sideways, though, the result touches daily life. At a restaurant we change our order after seeing the next table, and in meetings opinions gather around the first remark. Independent judgment may be the starting point in textbooks while in real scenes it stays rare.

The adaptive slice from 2 to 12 percent has a different character. The usual entangled flow holds, and then intraday news or events flip the reference point in a sudden jump. That is what the phrase dual equilibrium states points to. There is not one stable state but two, and a shock can carry the system from one to the other. Loss avoidance may lead the morning, and after specific afternoon news, chasing gains may take the weight. The distribution of behavior changes because the baseline changes.

Here an interesting question opens up. So is adaptation a separate layer from entanglement, or is it the name for how an entangled state changes. I read it closer to the second. Most of the time everyone moves in the same rhythm while watching each other, and when an outside shock arrives, the rhythm itself gets replaced. It feels like turning a radio dial. The same radio plays a different station after the turn, and the same market shows a different distribution after the reference point moves. If that reading holds, the 2 to 12 percent range should be seen not as noise but as separately counted moments of structural switching.

This difference matters more than it first appears. That independence is rare means improving individual rationality alone cannot explain the market. That adaptation arrives as switching means analysis must look beyond the content of news toward how news moves the baseline. Then can such a structural reading stretch all the way to a claim about the brain. The next section enters the most debated passage of the paper.

Where to stay careful with the entangled nerve fiber story

The paper presents its finding on collective trader behavior as empirical support for the Liu-Chen-Ao (LCA) hypothesis, an idea known as nonlocal entangled nerve fibers in the brain. In plain terms, because entangled modes fit an outside collective phenomenon well, that fit is read as indirect evidence for entanglement inside the brain. At this passage I paused. Technical success at the collective level and a mechanism at the level of one brain live on different floors.

Consider a flock of birds. That its motion fits a wave equation well does not promptly imply an antenna for waves inside one bird. The shape of the flock and the neural wiring of the individual can hint at each other, yet too many bridges must be crossed before they can be tied with the same evidence. Market data is the outside result of many brains interacting, while the LCA hypothesis concerns what happens inside a single head. An outside pattern can suggest an inside structure without confirming it.

Then should the connection be thrown away entirely. I would not go that far. It can still be read as weak evidence that sets a direction for hypotheses. If regularities that read well through entanglement keep repeating in human crowds, it feels natural to ask what kind of unit structure could generate them. But support here stays near correlation and consistency rather than cause or direct observation. Without brain imaging or physiological measurement, market data alone cannot settle the existence of such fibers.

From a systems perspective, the next question is obvious. Apart from whether the brain story is true, could an entangled game module still be useful as a computing device. I think this question decides the value of the paper. The strength of the biological claim and the usefulness of the engineering proposal can be judged separately. So the next section sets the brain story briefly aside and examines the proposal as computing structure.

Between trillion parameter scale and small humanlike processing units

Recent foundation models keep climbing toward trillions of parameters. Scale does bring more things done well, yet the inside keeps growing opaque while the operating burden grows with it. Memory cost, bandwidth at inference, and waiting time all rise together. From a team that ships systems, the serving picture catches the eye before the rising count. Hosting one giant model with effort and planting small devices in many places are entirely different operations.

The paper offers a fairly sharp diagnosis of this trend. It points to the limits of conventional ANN based AI that leans on trillions of opaque parameters, and it proposes enriching foundation models by joining ANN based AI with probability wave based entangled brain simulation, toward compact humanlike processing units (HPUs). Put plainly, instead of memorizing everything in giant weights, it suggests keeping an entangled game module nearby that carries the structure of collective behavior. Memory that memorizes and a sense that plays along would divide the work.

The proposal sounds attractive because of efficiency. If much of collective behavior can be written with a few modes, a small module that computes those modes directly could weigh less than giant parameters imitating them. For embodied intelligence and robotics, bodies that carry intelligence, that lightness matters even more. A robot cannot carry a data center on its back, and it must judge on the spot even when the network drops. On the need for small devices that run in the field, I share the direction.

Still, a practical question belongs here, which is how such a thing would actually serve. The gaps show together with the promise. It stays unclear where the entangled game module would attach to a foundation model, whether it would run on every token or only at a high level, and where the line between training and inference would sit. Efficiency is also hard to discuss without the memory hierarchy and the communication cost. Even a module that looks light on its own can stretch total latency if it must travel back and forth to a giant model each time. So I read this proposal as a sketch. Not a finished blueprint but a drawing that points to where the next blueprint should go.

What to watch in the next round of research

Coming this far returns us to the first 89 percent. A strong report of explanatory power is a starting point rather than a finish line. What I would watch next comes down to four items. First, whether the same decomposition holds across other markets and other periods, because outside the specific setting of Chinese intraday data the share of independent modes could shift. Second, how sharp the definitions and boundaries of the modes are, because it must be clear which observation would count as failure. Third, whether adaptive switching can be caught ahead of time, because dividing the past well and noticing the turn in advance are different skills. Fourth, what actually changes when an entangled game module is attached to a real model, because pairing a small module should be shown to improve parameter efficiency or responsiveness in embodied scenes.

Of these, the last question draws me most. The paper says ANN and entangled brain simulation should be joined to enrich foundation models and build small devices like HPUs. The words are clear, yet the design that follows stays empty. Whether the module runs at token level or at episode level, and whether it is learned or written as rules, would entirely change the engineering meaning. So the next study should deliver not a bigger number but a clearer answer about placement. A concrete report in the form of where it was attached and what changed.

In the same spirit, the link to the LCA hypothesis can also ask for a next step. Beyond outside evidence from market data, could measurements aimed directly at entanglement inside the brain follow. Or the split could grow clearer in the other direction, showing that the module can stay useful even if the brain story proves wrong. Either path would make the paper stronger rather than weaker, because dividing what is true from what is useful belongs to the next round of work.

Then where will this line continue. I suspect a new contest of small devices that compute entanglement directly could open beside the contest of making individual models smarter. The road that wins by size and the road that wins by shape would run side by side. This paper puts both math and data clues on the second road. Many questions must still be answered before those clues become conclusions, but the direction of the question stays clear. Write the moving scene together before the deciding individual, read that scene through a few stable shapes, and carry those shapes into a small module.

What opens up when together comes before alone

When I first opened this paper, the word wave made me hesitate for a moment. I worried it might look like decoration dressed in physics. Yet following the decomposition of intraday data step by step moved my eyes from the clothing to the body. The observation that independence sits near the exception while entanglement sits near the norm, the view that divides the usual flow from the moment of switching, and the proposal that carries that structure into a small computing module all pointed one way.

The parts to trim before accepting it stay clear as well. An outside market pattern cannot directly settle a structure inside the brain, the 89 percent of description should not be reread as prediction, and the efficiency of a small module must be judged again inside a real serving picture. Just as cases that look like spreading deserve careful viewing, a report of good fit deserves a closer look at what fits well. So I read this paper as an invitation rather than a verdict. An invitation to treat entanglement as computation rather than as a slogan.

As agents work longer and connect with more agents, we will likely need to watch what gets entangled between agents and how it amplifies, beyond asking whether each model alone is safe or smart. At that point the returning question will be this. Should our AGI move toward solving alone problems better, or toward reading scenes played together better. This paper votes for the second. I have not finished weighing that vote, but I now know where to place the scale.