Two open-source tools — released weeks apart, by teams who have never spoken to each other — accidentally created the blueprint for the most powerful messaging intelligence system ever conceived. Nobody seems to have noticed.
In the last two months, two things happened that almost nobody connected.
On March 7th, a 20-year-old undergraduate in Beijing named Guo Hangjiang released MiroFish — a swarm intelligence engine that spawns thousands of AI agents with unique personalities, memories, and social connections, drops them into simulated social media platforms, and watches what happens. It hit #1 on GitHub’s global trending list. A billionaire committed $4.1 million within 24 hours.
Three weeks later, on March 26th, Meta’s fundamental AI research team open-sourced TRIBE v2 — a model trained on 500+ hours of brain scans from 700 people that can predict, with startling accuracy, how the human brain responds to anything it sees, hears, or reads. Its synthetic predictions are often more accurate than actual fMRI recordings.
One tool simulates how your brain processes a message. The other simulates what happens when that message enters society.
And as far as I can tell, nobody has put these two things together.
The Expensive Guessing Game
Here’s a number that should bother you: American companies spend roughly $83 billion a year on market research. Political campaigns spend hundreds of millions per cycle on message testing. Public health agencies burn through enormous budgets trying to figure out whether their messaging will actually land.
And the dirty secret of all that spending is that most of it is educated guessing. Slow, expensive, educated guessing.
The process works like this. You have a message, an ad, a political spot, a public health announcement, a brand campaign. Before you spend millions distributing it, you want to know two things:
First: Does this thing register? When someone encounters it, does their brain actually engage? Does it capture attention, trigger emotion, or encode meaning? Or does it wash over them like everything else in a feed full of noise?
Second: What happens when it enters the wild? Once it’s out there, how does it spread? Does it unify people or fracture them? Do they share it enthusiastically, argue about it, distort it, or ignore it? Does the opposition weaponize it?
These are fundamentally different questions. And right now, they’re answered by fundamentally different and disconnected processes.
For the first question, you hire a neuromarketing firm. They put 40 people in a room, strap eye-trackers and EEG sensors on them, and measure where they look and how their brains respond. This costs $30,000 to $150,000 per round, takes weeks, and gives you data on a sample so small it barely qualifies as statistical.
For the second question, you do A/B testing. You put the message in-market, spend real media dollars, wait for real data, and analyze what happened. By the time you understand the social dynamics, you’ve already spent the budget. The experiment is the campaign.
Neither process informs the other. You can score high on neurological salience the ad grabs attention, activates emotion and still fracture catastrophically when it hits a polarized social ecosystem. Conversely, a message can propagate beautifully through sympathetic networks but fail to register neurologically with the persuadable middle, the people you actually need to reach.
The two layers of messaging effectiveness does it register? and what happens next? have never been connected in the same system.
Until now. Sort of. Accidentally.
What a Brain Model Actually Tells You
Let’s be precise about what TRIBE v2 is, because the headlines mostly got it wrong. It’s not “mind reading.” It’s not “predicting what people think.” It’s something more specific and, for our purposes, more useful.
TRIBE v2 is a trimodal brain encoder. You give it a stimulus , a video clip, an audio recording, a piece of text, and it predicts, across 20,000+ cortical vertices, exactly which regions of the brain activate, how intensely, and in what pattern. It maps to specific functional networks: visual processing, auditory processing, language comprehension, emotional encoding, the default mode network (where the brain integrates meaning with personal relevance).
It was trained on people watching movies and listening to podcasts; naturalistic, real-world stimuli, not sterile lab conditions. And its key breakthrough is zero-shot generalization: it can predict the brain response of a person it has never scanned, in a language it was never trained on, without any additional calibration.
What does this mean in practical terms? It means you can take a 30-second video ad and, in seconds, get a high-resolution map of how the average human brain will process it. You can see whether it activates attention networks or flatlines. Whether it triggers emotional encoding or gets processed as neutral information. Whether the visual and auditory channels reinforce each other or compete. Whether the language centers engage deeply indicating the audience is processing meaning or barely flicker.
This is not a focus group opinion. This is not someone telling you what they think they thought. This is a prediction of what the brain actually does before conscious judgment kicks in.
Here’s what it doesn’t tell you: anything about what happens after. TRIBE v2 is a stimulus-level model. It ends at the boundary of a single brain encountering a single message. It has no concept of social context, political identity, tribal dynamics, or information cascades. It can tell you the message registers. It cannot tell you what the world does with it.
What a Swarm Simulation Actually Tells You
MiroFish starts exactly where TRIBE v2 stops.
You give it a seed document, a news article, a policy draft, a campaign message and it builds a miniature society. Using GraphRAG, it extracts every entity and relationship from the source material and constructs a knowledge graph. From that graph, it generates hundreds or thousands of AI agents, each with a distinct personality, background, initial stance, social connections, and persistent memory. Then it drops them onto simulated social media platforms, a Twitter-like and a Reddit-like environment running in parallel, and lets them interact.
The agents post. They comment. They argue. They form coalitions. They shift positions. They ignore things. They amplify things. They distort things. Emergent social dynamics unfold over multiple rounds, and at the end, a specialized ReportAgent analyzes what happened, which narratives dominated, where factions formed, what tipping points occurred, and how sentiment evolved over time.
It’s SimCity for information. You’re watching, from a god’s-eye view, how a message ripples through a social system.
The Polymarket angle gives you a taste of what this looks like in practice: one developer plugged MiroFish into a trading bot, simulated 2,847 digital humans before every prediction market trade, and reportedly turned a profit over hundreds of trades. Not because the simulation was perfect, but because even imperfect simulations of social dynamics outperform vibes-based intuition.
Here’s what MiroFish doesn’t tell you: whether the stimulus itself is any good. It assumes the message reaches the agents. It doesn’t assess whether the message would actually capture attention or trigger emotional encoding in the first place. Its agents respond based on persona logic and LLM-generated behavior — not based on any model of how a human brain actually processes the content.
You see the gap.
The Synthesis No One Has Built
TRIBE v2 answers: “When someone encounters this, what does their brain do?”
MiroFish answers: “When this enters a social system, what does the world do?”
Combined, they answer bot and they answer something neither can answer alone: Who is neurologically susceptible to this message, and what do they do about it?
Picture the combined system. You have a video ad for a political campaign. You feed it through TRIBE v2 first. The brain model tells you: this creative strongly activates visual attention and emotional processing networks, but language comprehension is low; it’s a gut-punch ad, not an argument. Semantic stickiness is weak; people will remember the feeling but not the specific claims.
That neurological profile now informs how MiroFish runs its simulation. Instead of assuming every agent “sees” the message equally, the system models differential attention penetration; agents representing high-emotional-reactivity personas engage heavily, while analytical personas barely register it. Instead of assuming agents process the message accurately, the system models narrative drift, the low semantic stickiness score means agents reproduce the emotional valence but distort the actual claims as the message passes through social networks.
The simulation runs. And what emerges is something no focus group, no A/B test, and no social listening tool can produce: a prediction of how a specific piece of creative, with a specific neurological profile, will propagate through a specific social ecosystem, including which factions it activates, which counter-narratives form, and where the message mutates beyond your control.
That’s not a better focus group. That’s a fundamentally new category of intelligence.
Who Needs This?
The obvious answer is advertisers, and the obvious pitch is “save money on testing.” That’s true but boring. The more interesting applications are the ones where the stakes are higher than click-through rates.
Political strategy is the most immediate high-value application. Campaign messaging operates in an environment where neurological impact and social propagation are constantly at war; a policy position that polls well in isolation can fracture catastrophically when it enters a polarized social media ecosystem. A combined system could simulate exactly how a debate response, an attack ad, or a policy announcement propagates across partisan lines before it goes live. And because the outcome variable is measurable (tracking polls, vote share, favorability), this is the vertical where validation is most tractable.
Public health communication is the application with the most social value. COVID demonstrated, at catastrophic scale, what happens when public health messaging fails to account for both neurological processing and social dynamics simultaneously. Vaccine messaging that activated threat-response networks in the brain inadvertently triggered reactance in vaccine-hesitant populations. Messages designed to inform were processed, neurologically, as threats and then amplified through social dynamics into full-blown conspiracy frameworks. A pre-market simulation could have surfaced this failure mode before the messaging went live.
Crisis communication is the most time-sensitive application. When a brand crisis breaks, communications teams have hours to craft a response with zero ability to test it. A system that can simulate the propagation of three different crisis responses, an apology, a reframe, a counter-narrative, in minutes rather than weeks changes the decision calculus entirely.
The Hard Part Nobody Wants to Talk About
I’ve laid out the thesis. Now let me be honest about what makes it hard.
The translation layer doesn’t exist yet. TRIBE v2 outputs cortical activation data; 20,000 vertices on a brain mesh. MiroFish takes natural language descriptions and seed documents as inputs. There is no existing methodology for converting fMRI prediction data into behavioral parameters for social simulation agents. The bridge between “the fusiform face area activates at 0.73 intensity” and “Agent #247, a skeptical moderate, scrolls past this without engaging” is a piece of novel engineering that nobody has built. It’s the load-bearing wall of the entire concept.
TRIBE v2 predicts the average brain, not specific audiences. Its zero-shot generalization is impressive, but its predictions are for a canonical subject; it doesn’t natively segment by age, culture, psychographic profile, or political orientation. A system targeting persuadable swing voters needs persona-specific neural response profiles that the current model doesn’t provide out of the box.
MiroFish has never been validated against reality. No published benchmark compares its predictions to actual outcomes. The simulations are compelling and narratively coherent, which is exactly what makes them dangerous; a convincing simulation that doesn’t correspond to reality is worse than no simulation at all, because it breeds false confidence. The agents also inherit LLM biases that make them more polarized and more herd-like than real humans, which could systematically overstate social fracture.
The economics are tricky. Thousands of LLM-powered agents running through dozens of simulation rounds consume enormous quantities of tokens. At current API pricing, testing 20 creative variants across 5 audience segments could cost thousands of dollars per run. That’s still cheaper than traditional testing, but it’s not the “basically free” pitch that makes enterprise buyers salivate.
And then there’s ethics. A system that can predict both neurological vulnerability and social propagation dynamics is, quite explicitly, a persuasion optimization engine. The political application cuts both ways; the same tool that helps a public health agency craft better vaccine messaging could help a disinformation campaign identify neurological vulnerabilities in target populations. This isn’t a hypothetical concern; it’s the core design tension.
So What? Why Am I Writing This?
Two reasons.
First: the timing window matters. Both tools are open-source and available right now. The components are commodity; the synthesis is where the value lives. Whoever builds the translation layer, the middleware that converts neurological salience into social simulation parameters, owns the intellectual moat. That’s a specific, scoped engineering problem, not a vague research direction. And it’s the kind of problem that rewards people who operate at the intersection of neuroscience, agent systems, and market dynamics.
Second: this is exactly the kind of hidden rationality that I think about constantly. Two breakthroughs, released weeks apart, built by entirely different communities, for entirely different purposes and the combination produces something neither team imagined. The brain scientists aren’t thinking about swarm intelligence. The swarm intelligence people aren’t thinking about cortical activation maps. But the market inefficiency they jointly address, the $83 billion guessing game, is hiding in plain sight.
The logic is already there. It’s just waiting for someone to decode it.
