Transcript: Simulating crowds of AI agents to forecast reactions
An AI-generated audio overview (Google Gemini), transcribed by machine on Blaine's Mac (Whisper). The audio was checked against primary sources; what it gets wrong is listed at https://blainomd.com/learn#audio-mirofish-social-behavior

Usually when we think of predictive analytics for you or your business, we picture, you know, spreadsheets and line graphs.
But imagine playing a high stakes game of SimCity.
Only instead of zoning residential blocks, you're dropping thousands of autonomous AI agents into a simulation to watch how society reacts to a real world crisis.
Yeah, it completely shifts us away from static math into, well, living, breathing social simulations.
Right. So welcome to today's Deep Dive.
We're exploring research on Mirofish, which is an AI platform acting as a sort of digital sand table.
And our mission for you is to figure out how watching these thousands of interacting agents actually produces a roadmap of the future.
So, OK, let's unpack this.
Just do it.
If you're dropping thousands of like blank slate AI agents into a simulation, how do they know who they are supposed to be?
I mean, do you have to program every single personality?
No, not at all, actually.
It starts with reading seed material.
So things like news articles or policy drafts.
But rather than just scanning for keywords, Mirofish uses a process called GraphReg to map out relationships.
So it creates a web of who hates who, who holds influence, who owes who money.
Oh, wow.
So it builds the foundational lore.
Exactly.
It builds the lore of this digital world.
So it's basically a highly advanced digital dollhouse.
Do these AI dolls hold grudges against each other based on, say, an argument they had three simulations ago?
Yeah.
That grudge is actually the entire point.
Wait, really?
Yeah, because Mirofish integrates with something called ZepCloud to give the agents persistent memories.
They actually remember past interactions.
That is wild.
It is.
It captures that messy, emergent human behavior that traditional statistics just, you know, completely miss.
Humans are isolated data points, right?
Our reactions change based on our experiences and alliances.
I have to push back here, though.
Sure.
We know large language models hallucinate constantly just to sound good.
So if I'm running a financial stress test, why should you trust an AI that might just be, I don't know, making up a dramatic reaction for the sake of a good story?
That is a great question.
And it's the biggest hurdle in simulation.
But it's solved here by something called the Brain and the Fish Project.
Brain and the Fish.
Yeah.
They model the agents on actual biological mechanics using spiking neural networks.
So in a biological brain, a neuron only fires if it hits a specific electrical threshold.
Right.
The action potential.
Exactly.
In Mirofish, an agent's simulated neurons require strict document evidence to act as that electrical spike.
So meaning if there is no proof in the source text, the simulated neuron mathematically cannot fire.
Exactly.
The AI physically cannot hallucinate a justification.
It's literally forced to show its work.
That's incredible.
Yeah.
Every single action is deterministically linked to evidence.
And once you have a society that is grounded in reality like that, you can safely test it by injecting chaos.
Here is where it gets really interesting for you listening.
Definitely.
Users take a God's eye view and introduce stimuli to see how the population recalibrates.
Which brings us to the case studies.
In one test, they used a sentiment report from a controversy at Wuhan University, right?
Yeah.
And because the agents had those persistent memories and relationship maps we talked about, we didn't just get a generic output.
We could watch exactly how one opinion leader swaying a few friends triggered this massive domino effect.
It eventually polarized the entire simulated student body.
Right.
And they've also injected a Federal Reserve interest rate hike to watch simulated traders form polarized herds.
Oh, and didn't they even feed the system 80 chapters of an unfinished novel?
Yes, the classic Chinese novel, Dream of the Red Chamber.
Right.
And instead of prompting an AI to write prose, they just let the character agents interact organically to generate credible endings based entirely on their unique personalities.
That is just mind-blowing.
Watching all these individual agents act and react together really seems to validate what researchers call the super individual theory.
Yeah, the idea that when you network enough autonomous agents together, the crowd itself behaves like a single massive organism.
Well, what's craziest to me about building a massive biologically verified organism simulator is who actually built it.
Oh, yeah.
This is my favorite part.
You'd think this required a massive team at a tech giant, but it was just a college senior.
Guohang Zhang built the functional prototype of Mirafish in just 10 days.
Just 10 days from ideation to a rollout.
And that instantly secured a $4.1 million investment from Shonda Group.
How does a college student even do that?
Well, he used vibe coding.
Right, vibe coding.
He essentially ditched line-by-line programming to act more like a film director.
He told AI coding tools what he wanted the system to feel and function like and let the AI write the underlying structure.
It's amazing.
It proves a single developer armed with the right tools can match the innovative output of an entire corporation.
I mean, Mirafish isn't just predicting the future.
It's mapping the underlying architecture of human thought.
Absolutely.
Think about what that means for you listening to this.
Very soon, you could run a corporate-level pre-play laboratory right from your laptop.
It's the ultimate SimCity cheat code for reality.
It really is.
Which leaves us with a final provocative thought for you to chew on.
If a single student's code can simulate a society's reaction to a crisis today.
How long until your own digital twin is trapped in one of these sand tables, unconsciously helping an algorithm rehearse the future?
How long until your own digital twin is trapped in one of these sand tables?

How long until your own digital twin is trapped in one of these sand tables?