Lenia as a Living Environment for Biological Neural Networks

A Closed-Loop Experiment Concept

Author: Peter Müller, in collaboration with Claude (Anthropic)
Date: March 2026
Status: Concept — seeking experimental partner

Abstract

We propose a closed-loop experiment coupling living biological neurons (via Cortical Labs CL1 or FinalSpark Neuroplatform) with Lenia, a continuous, multi-dimensional cellular automaton system. Unlike existing biocomputing experiments (Pong, Doom), we impose no predefined goal function. Instead, Lenia serves as a dynamic environment whose state is influenced by neural output, and whose dynamics in turn stimulate the neural network. We observe whether the system self-organizes toward a dynamically stable regime — and whether neural networks exhibit intrinsic drives toward complexity. This experiment has no direct precedent in published literature.

1. Background

1.1 Biological Neural Networks as Computational Substrates

Recent work by Cortical Labs (DishBrain, CL1) and FinalSpark (Neuroplatform) has demonstrated that living neurons cultured on multi-electrode arrays (MEAs) are capable of adaptive, goal-directed behavior when embedded in closed-loop feedback systems. Notably:

• DishBrain neurons learned to play Pong within minutes using Karl Friston's Free Energy Principle as the feedback mechanism — destabilizing stimulation for misses, stabilizing stimulation for hits.
• Cortical Labs CL1 subsequently demonstrated neurons playing Doom, extending the complexity of controllable behavior.
• FinalSpark Neuroplatform provides remote access to 16 human brain organoids with both electrical and chemical stimulation (dopamine, glutamate, serotonin via photocaged release), enabling reward-signal experiments.

The common structure of all these experiments: neurons receive sensory input, generate motor output, receive feedback. The feedback is always predefined by the experimenter.

Our experiment removes this predefinition entirely.

1.2 Lenia as a Dynamical Environment

Lenia (Chan, 2019) is a continuous generalization of Conway's Game of Life. It operates on a real-valued field A(x,y,t) updated by a convolution-based growth function:

A(t+dt) = clip(A(t) + dt · G(K * A(t)))

Where:

• K = kernel (neighborhood structure)
• G = growth function with mean μ and width σ
• μ and σ determine the dynamical regime

Lenia exhibits three qualitatively distinct regimes determined primarily by the μ/σ ratio:

This regime structure creates a natural analog to ecological stability: collapse is "bad" for any embedded entity, stasis is "safe but sterile," and the complex regime at the edge of chaos is maximally rich. Importantly, hysteresis is observed: returning from collapse costs more than preventing it.

2. Experimental Design

2.1 Core Hypothesis

Biological neural networks embedded in a Lenia environment — with bidirectional coupling and no imposed goal — will through synaptic plasticity mechanisms self-organize toward the dynamically stable, complex regime at the edge of chaos.

Secondary hypothesis: Neural networks in a static (stasis) environment will exhibit homeostatic plasticity responses, increasing intrinsic excitability until they perturb the environment — demonstrating an intrinsic drive toward complexity.

2.2 The Closed Loop

┌─────────────────────────────────────────┐
│                                         │
│   LENIA FIELD STATE                     │
│   A(x,y,t) → spatial activation map    │
│         │                               │
│         ▼                               │
│   field_to_stimulus(A)                  │
│   Maps field values to electrode        │
│   stimulation amplitudes                │
│         │                               │
│         ▼                               │
│   BIOLOGICAL NEURONS                    │
│   800k neurons (CL1) or                 │
│   16 organoids (FinalSpark)             │
│         │                               │
│         ▼                               │
│   spikes_to_params(spike_train)         │
│   Maps neural output to Lenia           │
│   parameter deltas (Δμ, Δσ, ...)        │
│         │                               │
│         ▼                               │
│   LENIA FIELD UPDATE                    │
│   New state feeds back to neurons       │
│                                         │
└─────────────────────────────────────────┘

2.3 Experimental Phases

Phase 0 — Baseline (Coexistence, no coupling)
Lenia runs in stable complex regime. Neural network receives field-derived stimulation but cannot influence Lenia parameters. Baseline firing rates recorded. Duration: several hours.

Expected outcome: Stable, non-adapted firing. Neurons as passive observers.

Phase 1 — Parameter Release (Single parameter)
Neural spike rate is mapped to a single Lenia parameter (proposed: μ). Neurons can now shift Lenia between regimes.

Mapping function (initial):

Δμ = α · (mean_spike_rate - baseline_rate)

Expected observations:

• Random perturbations push Lenia toward stasis or collapse
• In stasis: reduced stimulation → homeostatic upregulation → increased firing → Lenia pushed back toward complex regime
• In collapse: zero stimulation → homeostatic crisis → strong perturbation attempts

Key question: Does the network accidentally discover that certain firing patterns prevent collapse?

Phase 2 — Multi-parameter Release
Multiple Lenia parameters become available (μ, σ, kernel parameters). The neural network now controls a genuinely multi-dimensional dynamical system — comparable in complexity to the original Game of Life, but continuous.

This is the heart of the experiment. No single parameter controls system stability. Combinations matter. The search space is large.

Expected observation: Spike-Timing Dependent Plasticity (STDP) selectively reinforces synaptic connections whose output correlates with beneficial return stimulation. Over hours, the network develops structured response patterns — not through explicit learning, but through selection of plasticity states that produce stable environments.

Phase 3 — Long-term observation
After multi-parameter coupling is established, we observe the long-term trajectory of the joint system:

• Does it stabilize at a fixed parameter set? (attractor)
• Does it exhibit ongoing variation within the complex regime? (limit cycle or strange attractor)
• Does it drift? (slow biological timescale parameter shift due to synaptic remodeling over hours)

2.4 The Stasis Problem — Core Experiment

A specific sub-experiment of high theoretical interest:

1. Set Lenia parameters to deep stasis (no dynamics, flat field)
2. Release parameter control to neurons
3. Observe:
   • Does the network remain quiescent indefinitely?
   • Or does homeostatic plasticity generate sufficient spontaneous activity to perturb Lenia out of stasis?

If the network spontaneously breaks stasis: This constitutes empirical evidence for an intrinsic drive toward environmental complexity in biological neural networks — not as a learned behavior, but as an emergent consequence of homeostatic physiology.

This would be a significant finding independent of the Lenia framework.

3. Theoretical Framework

3.1 Free Energy Principle (Friston)

The original DishBrain experiment used externally imposed free energy (destabilizing stimulation) as a feedback signal. Our experiment removes this: Lenia collapse IS the destabilizing signal. No artificial reward is needed. The physics of Lenia provides the feedback structure naturally.

This is not a minor technical difference — it means we are testing whether biological neural networks can exploit environmental physics as an implicit reward signal, without any experimenter-imposed fitness function.

3.2 STDP as Causality Detection

Spike-Timing Dependent Plasticity strengthens synaptic connections when presynaptic firing precedes postsynaptic firing within a ~20ms window. In our closed loop:

• Neuron fires → Lenia parameter shifts → Lenia dynamics change → Neuron receives altered stimulation

If the loop latency is consistent, STDP will detect this causal chain and reinforce the connections responsible for beneficial perturbations. The network learns the causal structure of its environment without explicit instruction.

3.3 Coupled Dynamical Systems — Two Order Parameters

Lenia has its μ/σ order parameter determining its regime. Biological neural networks have an analogous parameter: baseline firing rate vs. variability. When coupled, these form two interacting order parameters. The bifurcation structure of this coupled system is unknown. Possible outcomes:

• Mutual entrainment at a shared attractor
• Oscillatory dynamics (each system alternately destabilizes the other)
• Slow neural drift continuously shifting the Lenia attractor

This represents a genuine open problem in dynamical systems theory.

3.4 Edge of Chaos as Expected Attractor

The "Edge of Chaos" — the boundary between order and collapse — is the regime where complex systems maximize information processing capacity. Living systems generally operate near this boundary. If our experiment confirms that neurons + Lenia gravitates toward this regime, it provides:

1. A mechanistic explanation (stasis starves the network; collapse destroys input)
2. An experimental demonstration that this attractor can emerge from pure closed-loop physics without design

4. Technical Implementation

4.1 Technology Stack

Both platforms are Jupyter Notebook compatible. The shared language is Python + NumPy, making integration technically straightforward.

4.2 Timing Considerations

CL1 operates at up to 25kHz. Lenia at 30–60 fps. These timescales are mismatched by ~3 orders of magnitude.

Solution: Integrate neural spikes over time windows (e.g., 100ms) to produce smooth parameter estimates. Lenia updates on its own frame rate. This is a standard rate-coding approach and does not require special hardware.

# Core loop pseudocode
WINDOW_MS = 100
with cl.open() as neurons:
    for tick in neurons.loop(ticks_per_second=1000, stop_after_seconds=DURATION):
        spike_buffer.append(tick.analysis.spikes)
        
        if time_since_last_update > WINDOW_MS:
            # Compute firing rates per electrode
            rates = compute_rates(spike_buffer)
            
            # Map to Lenia parameters
            lenia.μ += α * (rates.mean() - baseline)
            
            # Update Lenia one step
            lenia.step()
            
            # Convert Lenia field to stimulation
            stim = field_to_stim(lenia.field)
            neurons.stimulate(stim)
            
            spike_buffer.clear()

4.3 Mapping Functions (Experimental Variables)

The critical design decisions — which we treat as experimental variables, not fixed parameters:

field_to_stimulus(A): How does the spatial Lenia field become electrode stimulation?

• Option 1: Spatial downsampling to electrode grid (preserves spatial structure)
• Option 2: Global statistics (mean, variance of field → uniform stimulation intensity)
• Option 3: Regime detection (stasis/complex/collapse → qualitatively different stimulation patterns)

spikes_to_params(spikes): How does neural output shift Lenia parameters?

• Start: mean firing rate → Δμ only (single degree of freedom)
• Later: spatial spike patterns → multiple parameter deltas
• Advanced: learned mapping via STDP readout

5. Platform Comparison and Access

Preferred starting point: FinalSpark. Geographic proximity (Switzerland), lower cost, and critically: dopamine-based reward stimulation allows us to optionally reinforce specific Lenia states — a clean experimental handle not available via electrical stimulation alone.

6. What This Experiment Could Show

Minimum result (if coupling works at all):

Neural activity measurably influences Lenia parameters, and Lenia state measurably influences neural activity. Closed-loop confirmed in wetware.

Medium result:

The system avoids collapse significantly more often than random parameter perturbation would predict. Neural network develops non-random response structure.

Strong result:

System gravitates to the complex Lenia regime without being pushed there. Stasis is spontaneously broken by homeostatic neural activity.

Maximum result:

Long-term observation reveals dynamic stability with continuous novelty generation at the edge of chaos — demonstrating that biological neural networks embedded in a physics-rich environment self-organize toward complexity as an emergent property of the coupling, with no designed fitness function.

7. Why This Has Not Been Done

Existing biocomputing experiments all share one structure: the experimenter defines what success looks like. Pong has a score. Doom has a survival criterion. These are human-imposed fitness functions.

Our experiment removes this entirely. The only "success" is determined by the physics of Lenia itself — and we observe what the neurons do with that.

This is not a minor variation. It is a different class of experiment: from supervised biological learning to emergent biological self-organization.

8. Next Steps

1. Lenia baseline notebook — establish stable Lenia regime, document parameter space
2. Mock neural interface — replace neural input with random noise, verify loop architecture
3. Contact FinalSpark — present concept, apply for research access
4. First coupling experiment — single parameter (μ), single mapping function, document trajectories
5. Publication — results in any of the above outcome categories are publishable

References

• Chan, B. (2019). Lenia: Biology of Artificial Life. Complex Systems, 28(3).
• Friston, K. (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience.
• Kagan, B. et al. (2022). In vitro neurons learn and exhibit sentience when embodied in a simulated game-world. Neuron, 115(6).
• Sims, K. (1994). Evolving Virtual Creatures. SIGGRAPH Proceedings.
• FinalSpark (2024). Open and remotely accessible Neuroplatform for research in wetware computing. Frontiers in Artificial Intelligence.

This document was developed in a conversation between Peter Müller and Claude (Anthropic), March 2026. It represents a genuine research proposal, not a thought experiment.