Transcription
Hey guys, today I wanted to walk through Miroofish. So this just came out last week. It's an open-source AI swarm intelligence engine. So what that means is you can deploy thousands of agents to figure out and predict the outcome of certain events, whatever data that you feed into it. So I've built a demo that is basically a way to predict the oil price globally based on news events that come in. So this is just a local demo with a little bit of data coming in using a very small local AI model. Obviously, it can scale up to thousands of agents if you have enough GPU hardware laying around, but this is just a demo of what you can do with it.
So let's get stuck into what Mirrorish is. So you just go to mirrorish.ai. This is so new that it hasn't officially launched yet. So there is an email, but the whole project is open source. So you can actually just go to the GitHub repo and actually download the code yourself.
Um, so what does it actually do? So Mirrorish uses graph rag. So what does that do? So if you look at the demo that they have on their website, uh, what it does, it creates relationships between data um where there's edges and nodes and it organizes these into a hierarchal graph like this. Unlike traditional rag, which is retrieval augmented generation, relies on flat vector searches to figure out the relationship between bits of data. When we do a search with graph rag, we tend to get a lot more accurate results where the AI can do reasoning and figure out from these relationships uh the actual correct answer again.
So all of this is open source. You can read the English version of the readme file. The original is in Chinese. You can just go to GitHub and clone it locally. And the GitHub has crossed 30,000 stars, which is pretty awesome. And you can scroll down to the bottom. It shows you the graph um of the star history. So it started in December and then through February, March, it's just boomed.
So it does use camel AI or Oasis to generate all the social conversations between the agents, which is open source. So you can actually just view that uh GitHub as well and um have a look at the code and features of that. Um, there is really nice uh documentation for all of this. So check out uh oasis.camel-ai.org for that. Um, but the rest of the tech stack uh is all open source. But the demo that runs on the Miroofish website is using zepcloud, which is a paid subscription uh for storing agent memory. Um, it also uses an Alibaba uh LLM. But the good news is Quen is open source and available on Alama, which I'll demo later in the video.
So once it's set up, it runs this giant simulation with all of these AI agents that you would create a role or personality for each agent. They all take that new data that comes in as a form of an event, for example. They then respond to that event. They then debate that in round two um and then generate a report at the end. We'll demonstrate all of this on the oil prediction demo.
So what are the different types of use cases of swarm intelligence? So one of the big ones is finance. So basically you can look at thousands of investor sentiment shifts or behavioral data, perhaps generated from thousands of AI agents reading a news article about the oil price. Um, so that's the example I went with. Um, but there is other use cases anywhere from government policy opinion forecasting to crisis PR simulation, marketing strategy testing, building fictional worlds sounds really interesting. And then obviously academic research.
So obviously oil prices are going up all around the world thanks to the war in the Middle East. Um, the price at the pump for consumers in every country is going up, basically. So, oilpric.com has the oil price um feed. WTI crude currently $97.61, which leads me to my demo that I built using Mirror Fish um at a very small scale running locally. What I've done is grabbed that oil price feed to bring it into the dashboard. I've looked through and researched, scraped a whole bunch of news articles relating to the oil crisis in the Middle East over the last couple of weeks. I've put that on a timeline um just showing and tracking the price with all this news information coming in over the past couple of weeks.
So, what we're able to do is now add a new potential uh news article that comes in. Perhaps it is a strike on a Saudi Arabian uh oil field or another emergency cut from OPEC countries uh to the supply of oil, which will then um trigger a price increase or decrease. This is basically a predicted event that we've made up as a simulation. Um, but what we're actually going to do is feed that into the AI agent framework, which is Muroish, and use lots of different agents to analyze and predict the price impact on oil.
So, I've currently got this running locally. Let me know in the comments if you want me to push the open-source code for this. Um, I'll have to convert it to Next.js and then deploy it on Vercel. Let me know if you want me to do that. Happy to publish this. Um, you can also go dark mode, which I think might look better in this case, and it does change the earth from day to night as well, which is pretty cool. I've put a little short description um basically to highlight the AI agents each have different incentives, memories, and expertise. So for this example, we put it through the Miroofish pipeline. So again, the knowledge graph uh it spawns four agents in this case. Um, it then does a parallel simulation using Oasis, simulating four different agents doing genuine independent analysis, and then they do a cross-agent debate. So each agent analyzes each other agent's results and agrees or challenges their result with the idea that consensus emerges from real disagreement.
So why does this beat a single LLM response? So instead of just giving the average internet uh number or example or opinion, Miroofish forces these AI agents to argue with each other over the data to then get to the end result. Again, the principle is the consensus price survives a debate, which is more defensible than a single output.
So some limitations of this demo. Again, we're just running four agents locally on my MacBook M1. If you use the full Oasis framework, it can support over 10,000 agents, which is just a crazy amount of LLM and AI inference. But you can clearly see the trend of the AI industry is more AI LLM inference and more AI agents. Instead of using a 72 billion parameter Quen model in the Alibaba cloud, we're just using a locally run Alama model Quen 3.5.8 billion parameter count. Uh, we're also only doing two debate rounds, whereas in the cloud-hosted model, they're doing like 5 to 20 debate rounds. We're also running static scenarios. So we have a bunch of researched historical news articles, and then we're just injecting one new simulated article um or event into the picture. Their model grabs a constant feed of news data and then and keeps generating the graph rag.
Okay, so here's all the historical events. This has already happened. We go to the forecast tab. Here's the new simulated news articles. Um, we click one of these. It zooms in onto the map of where that is. So, we'll just click this here and it changes the right sidebar with the um event scenario that we proposed. So, if the AB quake facility is attacked across the Strait of Hormuz, what would happen? Initial research says 7 million barrels per day would be impacted, which is catastrophic, high severity. It's saying bullish when the price of oil goes up, which uh obviously sounds backwards, but that's the price is going up. Um, and then here is the impact on USA, Australia, and Philippines. Uh, obviously biased to which countries I'm mentioning here.
So we've got four different agents that are set up already. So I've set up one called an oil trader. The second AI agent is a Saudi Aramco strategist. The third one is a consumer impact economist. And then the last one is an international energy agency policy modeler. So let's get started by hitting run mirror fish. It's now it's now generated four different agents that are running in parallel. It's hitting my local Alama uh which has this model loaded. Um, the first one's responded in 20 seconds. The other three are still going. We'll just let that finish. Uh, I can scroll down here and actually see uh the oil market trader and then the results live. So, let's read the the top one first. Um, so basically, so I can just click this to pop up a modal. Uh, obviously oil market trader, he's gone in, looked at the event, the impact on the oil price, and done a bunch of uh research and analysis. And again, all the other agents have also done the same thing. So the Saudi Aramco strategist has has now done his analysis, consumer impacts done, and I policy modeler has done as well. So that completes round one, where we have the four different personalities or roles of different AI agents doing their research and analysis. Round two is when we get the four agents to do a cross-agent debate. So, let's hit play for that one.
Okay. So, now they're going through and doing a debate. So, let's have a look at what the result is here. So, they're basically looking at the other agents' results um and then they're analyzing it to see if it's correct or not and maybe providing an opposite opinion on it. Um, and then what they're doing is revising that price up or down based on that further debating and analysis. So through that debate process, the four agents reach consensus, and they all agree that the price, the new price of oil is $98 per barrel. And then at the end of the Miroofish process, it does generate a detailed report, but I'm just going to read the summary. So if we scroll down uh we can see that the uh core takeaway, the supply disruption by Abaic facility attack will trigger immediate WTI price spikes, likely pushing prices above $100 a barrel in the first week due to OPEC constraints and HMA's mine failure, but the demand destruction risk persisting until Iran proxy strikes are resolved.
So it's taking in a lot of analysis from all the different events that happen leading up to this time point um on the time graph at the bottom. Uh, and then it's put it's injected the new event that I told it to consider and the impact on that. What I've also done is I've asked it to update each country's uh fuel price at the pump. Um, so you know, in the Philippines, it's nearly 100 pesos per liter. So let's switch back to day mode. Um, I've also added in a information page. Um, so this really just shows you how the architecture works for Miroofish, how it's using these multiple AI agents to reach consensus on whatever the prediction it's trying to make. Because I'm using it locally, I did add a little system capacity. So I just run some commands to get my local system CPU usage, you know, more details about the Alama model. And it's also generated some uh interesting new analysts or new AI agents we can add to the mix. Um, so a Russian energy analyst, a Chinese import strategist, and a US Treasury analyst. So this is all using multiple different perspectives using AI agents to again uh reach consensus and make that prediction more accurate.
So yeah, you can go to the mirrorish.ai. Um, currently the demo is in Chinese, but basically there is a PDF they've uploaded again in Chinese. Um, it then you just hit this button at the bottom and it will go off and actually run the whole simulation using their expanded cloud instance, which may have thousands of different AI agents. And you can see here it's processing in real time.
So what's really cool about this is it's written by a guy that's like 20 years old. Um, I think he was just going to uni at the time. His online name is Bifu. Um, he's a senior at the Beijing University of Posts and Telecommunications. In late 2025, so before Miro Fish, he wrote another um application called Beta Fish. And this also had a huge spike in GitHub stars um which also reached 20,000 stars in a week. Um, but the cool thing is that got the billionaire's attention. His name was Chen Tian Qiao. Founder of the Shanda Group. Um, one one once the richest person in China. Wow. Okay. That's pretty epic. Uh, he built a gaming empire in the 2000s and then uh moved to the US and changed uh transformed Shanda into a tech investment platform. So obviously he noticed he noticed this developer and uh kept an eye on him. And then when he released the next uh Miro fish app, he was able to invest and uh basically he invested 4.1 million or 30 million yuan 24 hours after it was released. And this swarm intelligence using AI agents is the feature is actually called God's Eye View.
So you can use this for anything. Obviously the website says it can predict anything. So some other examples, Fed suddenly cuts rates by 50 basis points, or maybe the CEO of a big company resigns. So how would that impact the price of the stock on the market? Um, or a competitor launches a product against maybe uh a company that you work for or consult for. So there's a couple of more examples of what mirror fish can do using again hundreds or thousands of agents to predict the answer using a consensus mechanism across multiple agents.
So let me know what you think you can use Miroofish for any cool examples. Let me know if you want this uh demo uh open source and published on Next.js um on Vercel. Um, I can happy to do that. Just leave a comment. And clearly the AI industry is alive and well. Lots of open-source AI models dropping every week. Um, we also have the recent open claw AI agent phenomenon where you have a personal AI agent does work for you through your own terminal. And now we have AI swarm intelligence where tens of thousands of agents can reach consensus and predict the future. So pretty exciting times. I'm looking forward to seeing what you build with it. Let me know in the comments what you're going to do. And thanks for your time today. And I'll see you in the next one.