Our Live AI Brain Experiment

What We Are Building

Our project is not an attempt to simulate the human brain. That would require resources far beyond what any small team can muster. Instead, we are building a minimal artificial brain, a system with just enough complexity to exhibit emergent cognitive behaviors like pattern recognition, simple learning, basic memory, and adaptive responses to changing inputs. The goal is to demonstrate the principles covered throughout this guide in a working system that anyone can study and learn from.

The system is designed around three core components. First, a spiking neural network that processes information through discrete timing events rather than continuous activations, capturing one of the most fundamental operating principles of biological neural tissue. Second, a plasticity mechanism that allows the network to modify its own connections based on experience, implementing a form of learning that does not require external training signals. Third, a sensory-motor loop that connects the network to a simple simulated environment, giving it inputs to process and outputs that affect its world, providing the grounding that purely abstract neural networks lack.

Why Build It in Public

Most artificial brain research happens behind closed doors in university labs and corporate research departments, with results published in academic papers that are difficult for non-specialists to read and impossible to reproduce without significant expertise and resources. We believe there is value in doing this work transparently, showing not just the polished results but the messy process of building, testing, failing, and iterating that characterizes real scientific and engineering work.

Building in public also creates accountability. When we make design decisions, we document our reasoning so that others can evaluate whether our choices are well-founded. When we encounter problems, we describe them honestly rather than glossing over them in a results section. And when our system does something unexpected, whether impressively emergent or embarrassingly broken, we share it with full context so that the observation is scientifically useful.

This approach serves our educational mission. The articles in this guide explain the science of artificial brains in depth, but there is no substitute for seeing how that science plays out in an actual implementation. The gap between "understanding how spiking neural networks work in theory" and "building a spiking neural network that actually does something interesting" is enormous, and documenting that gap is one of the most valuable things we can offer.

The Architecture So Far

Our current implementation uses a network of Izhikevich neurons, chosen because they can reproduce a wide range of biological firing patterns while remaining computationally efficient enough to run on consumer hardware. We started with a population of 1,000 neurons divided into excitatory (80%) and inhibitory (20%) groups, matching the approximate ratio found in the mammalian cortex.

Connectivity follows a small-world topology: neurons are primarily connected to their spatial neighbors (local connectivity), with a smaller number of long-range connections that span the network. This topology is inspired by the organization of cortical circuits, where local connections within cortical columns are dense and long-range connections between distant regions are sparser but functionally important for integrating information across the network.

Learning is implemented through spike-timing-dependent plasticity (STDP), the biological learning rule that strengthens synapses when pre-synaptic spikes consistently precede post-synaptic spikes and weakens them when the timing is reversed. We also implement a homeostatic plasticity mechanism that adjusts overall synaptic strengths to maintain stable network activity, preventing the runaway excitation or complete quiescence that can occur in networks with purely Hebbian learning.

The sensory-motor loop connects the network to a simple grid-world environment where a simulated agent must navigate to food sources while avoiding obstacles. Sensory input from the environment is encoded as patterns of spikes that drive a subset of neurons in the network. Motor output is decoded from the activity of another subset of neurons, with the most active motor neuron determining the agent direction of movement at each time step.

What We Have Learned So Far

Several findings from our ongoing experiment have deepened our understanding of the challenges and opportunities in artificial brain construction.

Parameter sensitivity is real. The behavior of the network is exquisitely sensitive to parameters that might seem minor: the time constant of synaptic plasticity, the ratio of excitatory to inhibitory connection strengths, the initial distribution of synaptic weights. Small changes to these parameters can shift the network from chaotic, seizure-like activity to complete quiescence, with the narrow band of useful dynamics in between. This mirrors the experience reported by large-scale brain simulation projects and underscores why parameter estimation from biological data is so critical.

Emergence is real but fragile. Under the right conditions, our network develops spatial maps of its environment, with groups of neurons becoming selectively responsive to specific locations in the grid world. This emergent place-cell-like behavior was not programmed; it arises from the interaction between STDP learning, the spatial structure of the sensory input, and the recurrent dynamics of the network. However, these representations are fragile and can be disrupted by small perturbations, suggesting that biological brains have stabilization mechanisms that we have not yet implemented.

The brain is not just a neural network. Our most humbling lesson has been discovering how many biological mechanisms we initially dismissed as "implementation details" turn out to be functionally important. Neuromodulation (global signals that change the behavior of the entire network based on reward or novelty) proved essential for making learning goal-directed rather than random. Homeostatic plasticity proved essential for maintaining stable dynamics over long timescales. And the specific connectivity patterns of excitatory and inhibitory neurons, not just the excitation-to-inhibition ratio but the specific wiring motifs, significantly affect the computational capabilities of the network.

Next Steps

Our current priorities are expanding the network to incorporate multiple interacting populations that serve different functional roles (analogous to different brain regions), implementing a simple reward system based on dopaminergic neuromodulation that allows the network to learn from the consequences of its actions, and developing better tools for visualizing and analyzing the network dynamics so that we can understand what the network is doing when it succeeds and when it fails.

We will continue updating this page as the project progresses. Each major milestone, design decision, and unexpected discovery will be documented here, creating a living record of what it takes to build an artificial brain from first principles. We hope this transparency will be useful to other researchers and students who are interested in this fascinating and challenging field.