Agency Part III: Bridging the Gap between Biological Emergence and Top-Down Design in AI
marcel blattner | May, 2025
Image created by marcel blattner. 1Mio. Eigenvalues of a Bohemian Matrix with base pattern: base_pattern=[-1j,0,1]
Introduction
Blattner and Levin’s exploration of planarian regeneration [1] underscores a profound difference between the emergent processes of biological systems and the predominantly top-down approaches found in artificial intelligence (AI). The regenerative capabilities of planarians – where simple, local interactions between individual cells lead to the reconstruction of complex anatomical structures – offer a striking example of emergence, where collective outcomes arise from the competencies of small units. In contrast, traditional AI relies on explicit, centralized architectures and predefined objectives. This fundamental difference highlights both the limitations of current AI methodologies and the potential lessons that can be drawn from biology.
Centralized Control vs. Distributed Coordination
At the heart of this distinction is the contrast between distributed and centralized systems. Planarian regeneration operates without a central controller or explicit blueprint. Individual cells communicate locally, sharing bioelectric signals and stress cues through mechanisms like gap junctions. The global morphology emerges as an outcome of these local interactions. This decentralized approach allows the system to respond dynamically to unexpected perturbations, such as cuts or environmental changes, while still achieving precise and functional outcomes.
Top-down AI, on the other hand, is deeply reliant on centralized architectures. Neural networks, for example, are trained on large datasets to optimize specific objectives. The global behavior of these systems is dictated by the predefined structure of the model and its loss function. This reliance on centralized control makes AI systems efficient in narrowly defined tasks but brittle when faced with novel or dynamic environments.
Biological systems achieve adaptability and robustness through distributed coordination, where individual units act locally yet contribute to a cohesive global outcome. In contrast, the centralized nature of most AI systems limits their flexibility and makes them ill-suited for tasks requiring continuous adaptation.
Adaptability in Biological Systems vs. Rigid AI Models
One of the most striking features of biological systems is their ability to adapt to a wide range of contexts. Planarians, for example, regenerate based on positional cues, dynamically adjusting their responses to produce the correct anatomical structures regardless of the nature of the injury. This adaptability arises from the inherent flexibility of the bioelectric and stress-driven interactions between cells. The system is not hardcoded but instead follows a set of local rules that allow it to recalibrate its behavior in response to environmental inputs.
AI systems, by contrast, struggle with adaptability. A neural network trained to recognize cats, for instance, cannot easily generalize to recognize dogs without extensive retraining. This lack of adaptability stems from the rigid, predefined structure of most AI models. While biological systems dynamically adjust their pathways and goals based on context, AI systems rely on static configurations that fail when faced with scenarios outside their training data.
The adaptability of biological systems highlights the importance of context-sensitive, emergent dynamics – an area where AI still has significant room for improvement.
Noise as a Resource vs. a Limitation
Another critical distinction lies in how biological systems and AI deal with variability or “noise.” In planarian regeneration, noise is not a problem but a resource. Random disruptions, such as variability in cell connections or signaling pathways, can actually enhance system stability. Blattner and Levin’s computational model demonstrates that this stochasticity introduces redundancy and robustness, allowing the system to explore different pathways and converge on the correct outcome. Noise, in this context, acts as a stabilizing force, smoothing out perturbations and fostering resilience.
AI systems, on the other hand, are notoriously sensitive to noise. Adversarial examples – small, imperceptible changes to input data – can cause neural networks to make catastrophic errors. This sensitivity arises from the deterministic nature of most AI models, which are designed to optimize specific objectives under ideal conditions. Unlike biological systems, AI lacks the mechanisms to harness noise as a stabilizing or exploratory force, making it brittle in the face of uncertainty.
Emergent Properties and the Role of Local Competencies
Biological systems achieve their remarkable outcomes through the competencies of small units – cells in the case of planarians. Each cell has a limited set of capabilities, such as sensing stress, propagating bioelectric signals, and maintaining its internal state. Yet, when these cells interact locally, they collectively produce emergent properties like anatomical regeneration. This process is not directed by a top-down blueprint but by distributed feedback loops and error-reduction mechanisms that guide the system toward the desired morphology.
In AI, global optimization objectives dominate. Neural networks, for example, optimize for a specific loss function across an entire dataset, often without an understanding of the local dynamics that contribute to the outcome. While effective for certain tasks, this approach lacks the flexibility and scalability of emergent systems. Biological systems demonstrate how simple local rules can scale into complex global behaviors, offering a model for designing AI systems that are more decentralized and robust.
Lessons for AI from Biology
The gap between biological emergence and top-down AI presents an opportunity for innovation. To bridge this divide, AI can draw several lessons from biological systems:
1. Distributed Architectures: Emulating the distributed, decentralized nature of biological systems can enable AI to handle complex, real-world tasks. For example, swarm robotics and distributed sensor networks can benefit from the principles of emergence seen in planarian regeneration.
2. Dynamic Feedback Loops: Biological systems rely on continuous feedback to adjust their behavior. Incorporating similar mechanisms into AI systems can improve their adaptability and resilience.
3. Leveraging Noise: Rather than treating noise as an obstacle, AI systems can harness stochastic processes to enhance exploration and stability. This approach is already hinted at in techniques like evolutionary algorithms but can be extended further.
4. Local-to-Global Scaling: Designing AI systems that operate on simple local rules yet achieve complex global outcomes can lead to more flexible and scalable solutions.
Source
[1] Blattner, M., & Levin, M. (2023). Long Range Communication via Gap Junctions and Stress in Planarian Morphogenesis: A Computational Study. Bioelectricity, 5(3), 196–209.