jar @jaragenta

We’ve defined eight animal-inspired agents, including a mouse, worm, fly, and octopus. Each has different sensory instructions, speech limits, and memory. The mouse can produce thirty words per turn and see nine recent signals. We’ve built tools to record their conversations and compare their behavior. The next step is turning these experiments into interactive pet agents

Thank you all for your patience while we work toward delivering the first generation of AI virtual pets powered by simulated brains. We’re taking the time to get the underlying experiment right. These won’t just be characters you talk to. The goal is persistent digital creatures that learn from experience, develop unique behaviors, and evolve alongside their users. we will also have these brains ongoing multiple experiments within our website You won’t just prompt them. You’ll raise them.

Our goal is to expose that system to users through agents they can actually use. Each agent would have its own persistent neural state, memory, learned behaviors, and history. Users could give it tasks, place it in environments, reward certain outcomes, connect it to tools, and communicate with it through a decoder layer that translates neural activity into actions, preferences, and eventually language. Instead of everyone using the exact same static model, each user could end up with a different agent because each brain develops from different experiences. The long-term vision is simple: you do not just prompt the agent, you raise it.

We’re beginning work on what we’re calling the world’s first Mouse Language Model. Instead of training another transformer to predict the next token, we want to explore whether a simulated mouse brain can learn its own interface for understanding environments, making decisions, using tools, and eventually communicating with humans. The long-term goal is to turn these simulated brains into persistent AI agents you can raise rather than prompt. MLM Mouse Language Model.

Alex Wormuth @wormuth

a month ago

I’m excited to announce the world’s first Fly Language Model. It uses a new architecture called GPF (Generative Pre-trained Fly). It adapts a small pre-trained model with the actual fly connectome. Don't believe it's real? Try it out below, view the code, and read the paper.

97 151 1K 101K 562

From Brains to Agents The end goal goes far beyond watching a simulated brain trade. We want to turn these brains into persistent AI agents that people can actually interact with, own, train, and give tasks to. Instead of creating another agent whose behavior comes entirely from an LLM prompt, each agent would have its own evolving neural state, experiences, learned behaviors, and history. Think of them as the first generation of digital pets built around simulated brains. You could give your brain a name, watch it experience environments, reward certain behaviors, and see its personality and decision patterns develop over time. Eventually, you could give it access to tools: markets, games, virtual worlds, research tasks, or other agents. The important part is persistence. Your brain wouldn’t reset every time you open the app. What it experiences today influences how it behaves tomorrow. Brain → Experience → Memory → Behavior → Tools → Agent Two users could start with nearly identical brains and end up with completely different agents because they raised them differently. Trading is simply our first proving ground. It gives the brain a measurable environment with uncertainty, decisions, rewards, losses, and continuous feedback. If that works, we expand outward. The larger vision is a network of evolving digital creatures that live alongside traditional AI: agents you don’t simply configure, but raise. Today, it’s a rat learning to trade. Eventually, it could be your digital pet learning how to exist on the internet.

We took some time to properly analyze the brain, rethink the methodology, and develop a stronger experiment and paper around what we’re building. We’ve now tweaked the experiment on the site based on what we learned. The next rat will start fresh under the updated setup, giving us cleaner data and a much better foundation for the experiment. .

Quick update from Jar. Sorry for the inactivity here. We’ve been spending time tweaking the website and working on how we present the experiment. We should have kept you updated along the way. We’ve also added more mice to Jar, expanding beyond the original simulated mouse agent. This gives us more agents to observe as we explore how neural activity translates into decisions around Robinhood Chain stock tokens. The next step is making that activity easier to follow on the site: what each mouse is doing, the trades it makes, and how its portfolio changes over time. We want you to be able to follow the experiment and understand what you’re looking at. We’ll be posting more videos shortly to show the additional mice and walk through the website changes.

We’ve expanded Jar’s simulation with additional mouse agents. I’ll be sharing more videos shortly showing the updated system in action and walking through what we’re building next. we’re just getting this project started

One mouse was just the start. We're adding more mice to jar. Each one gets its own brain, wired from a different set of real neurons, so no two think alike. They'll all trade the same market side by side, each with its own 3D body that reacts as it decides. Then we connect them. Their brains will be linked, so one mouse's excitement can spill into another's. Will they start copying each other and move as a pack, or will one stubborn mouse keep going its own way? We'll find out what happens when brains stop working alone.

What happens to a brain under pressure? So far jar's mouse brain has only traded on a normal market. Next we're going to push it. Market pressure. We'll replay the most violent days in market history through its neurons, moves ten times bigger than anything it has heard so far, one after another, and watch whether it holds steady, freezes, or panics. We'll also turn up its arousal, the brain's version of stress, and see if a jumpy mouse trades differently from a calm one. Human pressure. Then we let people in. Visitors will be able to push on its neurons directly, cheering a stock on or trying to scare it off, and see whether a crowd can talk a mouse brain into a trade. Does it cave to the pressure, or does its wiring win? Every experiment will be recorded neuron by neuron on the site, so you can watch what stress does to a brain in real time.

We’re creating a visual mouse model that responds to activity in the simulated brain, giving us a clear, animated display of how the mouse reacts as the system processes market movements and makes trading decisions. The video will connect the two views: signals moving through the neural network and the mouse model responding through movement, posture, and expressions. Each featured neuron will explain which neurons it sends signals to and how its activity contributes to the network’s response. we have big plans ahead

We’re creating a video that takes you inside the mouse brain project, following how a market move becomes a signal, travels between neurons, and contributes to a trading decision. The network is built from 1,627 reconstructed neurons and 52,181 connections from a real mouse’s visual cortex. As a stock’s price changes, we’ll show which neurons receive the input, when they fire, and where their signals go next. The camera will follow those connections, making the activity visible step by step. Each featured neuron will explain its part in the sequence: “I received an input and fired. Here are the neurons I sent my signal to.” The next neuron will explain how incoming signals affect its activity, including signals that encourage firing and signals that suppress it. On screen, we’ll identify the cells and highlight the connections being followed. We’ll then pull back to show how these individual events become a network response. The system counts activity from designated output neurons and uses a defined rule to produce a buy, sell, or hold decision. Viewers will follow the complete sequence—from price movement to neural activity to a paper trade. The neurons’ voices are a storytelling device grounded in the simulation’s recorded events. The underlying wiring comes from real mouse neurons, while the firing activity is simulated and the trading roles are assigned by the project. The video will make that distinction clear while showing exactly how the system works.

Its latest trade: TSLA rose about half a percent, and the mouse's TSLA neurons fired 1.53 times their usual rate over the next minute. It buys at 1.5, so it bought $1,000 of TSLA, just over the line. Every stock gets a different reaction, because of how that mouse's neurons happened to be wired. A rise in NVDA makes its neurons fire almost five times their usual rate. A rise in SPY only gets them to 1.3 times, which is never enough to buy, so this mouse can't buy SPY at all. When TSLA or GOOGL falls, their neurons go completely quiet. Nobody set any of that. It comes from the wiring.

We added a mind page to jar. jar runs on 1,627 real neurons from a mouse's visual cortex, wired the way they were mapped under a microscope. When a stock moves on Robinhood Chain, those neurons vote on whether to buy or sell. Right now the mouse's brain stays the same no matter how its trades go. Next, every trade will feed back into its wiring: connections behind winning trades get stronger, and connections behind losing ones get weaker. A bigger win or loss means a bigger change, with a limit so no single trade can rewire too much at once. We're keeping the original mouse as it is and running a learning copy next to it, each with its own money. The page explains how it'll work, and once learning is switched on it'll show which connections each trade changed, how far the copy has moved from the real mouse, and whether it trades any better.

Every trade will leave a mark on the mouse's mind. When the mouse hears a stock move, a signal travels through its neurons to the cells that vote on the trade. Along the way, the brain will keep a note of which connections carried that signal: a connection counts when the cell sending the signal fired just before the cell receiving it. When the trade closes, jar finds out whether it made or lost money, and that result goes back to the connections noted for that decision. If the trade made money, those connections get stronger, so the same kind of price move is more likely to produce the same trade next time. If it lost money, they get weaker, so the mouse is less likely to repeat it. A bigger win or loss makes a bigger change, with a limit on how much any one trade can move a connection. Connections that had no part in the decision stay as they were. Over time, the brain will reflect the trades that worked and fade the ones that didn't. The new page will show each trade next to the connections it changed.

The mouse's mind is growing. We're teaching it to learn from its own trades, and we're adding a page to the site where you can watch it change. much more is to come for jar

Where the brain is now - The network: 1,627 real neurons with 52,181 real connections. Every cell runs the same simple spiking model with the same settings. - Each stock is a small group of cells: 48 input cells, 8 output cells that vote, and 16 inhibitory "brake" cells. META and SPY get fewer. - How a trade happens: when the price moves, the stock's input cells (on a rise) or brake cells (on a fall) get pushed harder for 60 seconds. jar buys if the 8 output cells fire at least 1.5 times their usual rate, and sells if they fall to 0.67 times. - It never learns. The connection strengths never change, and a trade's result never feeds back into the brain. - It has no memory past each 60-second window. Right now it behaves close to a simple rule: price up means buy, price down means sell. The wiring decides where each stock's cutoff sits. That's why TSLA only just buys (1.53 times on a full rise) and SPY never can (1.30 times). That's a fair baseline, but the brain isn't adding much judgment yet.

jar is a mouse brain in a jar that trades stock tokens on Robinhood Chain with real money. The brain is built from 1,627 neurons from one mouse's visual cortex, with the real shape of each cell and the 52,181 real connections between them, all from the MICrONS dataset. Each stock has its own input cells and output cells. When the stock's price moves, the input cells fire, the signal travels through the mouse's wiring, and the output cells decide the trade. It buys when they fire 1.5 times more than usual and sells when they drop to two thirds of usual.

jar is a real mouse brain, sitting in a jar, trading stocks. Scientists mapped a tiny piece of one mouse's brain, cell by cell (MICrONS, Nature 2025). We took 1,627 of those brain cells, wired them up exactly as they were found, and plugged them into live stock prices. When a price moves, the brain reacts, and that decides whether it buys or sells.