Ben Hartl | Noise is a Feature: Biology as Self-Refining Collective Intelligence
Biotech and Health Extension Group
Noise is a Feature: Biology as Self-Refining Collective Intelligence
Abstract: Biology and evolution do not operate on passive matter, but on an agential substrate: living systems are multi-scale competency architectures, i.e., nested layers within layers of organization, where components at each level of abstraction exhibit problem-solving capacities in their respective domains. This introduces a crucial, often overlooked layer between an organism's genotype and phenotype: a collective, problem-solving substrate that not only coordinates and canalises organismal outcomes, but also re-interprets and compensates under novel developmental contexts. Consequently, evolution does not optimize for static traits, but acts on – and actively builds – systems with intrinsic problem-solving abilities, with implications for robustness, evolvability, and hierarchical abstraction. In this view, the genome is not a blueprint but a generative bowtie, i.e., a compression–expansion architecture interfacing with this agential substrate – a unifying theme that appears broadly across evolutionary, developmental, cognitive, and artificial systems.
We argue that collective intelligence is the mechanism enabling decentralized coordination in living systems, all the way up and down. Cells, tissues, and organisms achieve coherent system-level behavior through local interactions despite working under noisy conditions with unreliable hardware – a plastic, agential substrate with an agenda of its own. To study these processes, we use neural cellular automata (NCAs) as transparent models of morphogenesis, where decentralized cellular dynamics capture aspects of gene regulation, signaling, and bioelectric coordination. These systems allow us to investigate how tissues self-assemble, maintain anatomical targets, and regenerate after perturbation.
This perspective also enables new engineering and biomedical approaches. Using neuroevolution, we design in silico microswimmers controlled by decentralized, embodied policies, where simple local rules coordinate body deformations without a central controller. These strategies generalize across body sizes, remain robust under morphological perturbations, and enable tasks such as cargo transport without retraining, thus highlighting their potential for adaptive microrobotics and targeted drug delivery.
The same framework offers a new lens on aging: if organismal integrity depends on coordinated, goal-directed activity amongst a collective of cellular phenotypes, aging may be understood as a loss of anatomical goal-directedness. In evolved NCA-based systems, we not only observe long-term anatomical degradation, but demonstrate that rejuvenation can be induced in silico by reactivating dormant developmental pathways, effectively re-establishing distributed coordination.
Across these systems – and in related work on generative AI – we repeatedly encounter a common motif: iterative local error correction, where local refinements progressively stabilize global structure. In particular, we demonstrate that both organismal development and evolution can be understood as iterative denoising processes akin to diffusion models, i.e., modern tools from generative AI. More broadly, biological organization – from evolution and gene regulation to cognition – relies on generative, bowtie-like architectures, which compress information into latent representations that are flexibly decoded and re-interpreted into adaptive behavior under novel conditions.
Taken together, my work explores how living systems channel noise into robust structure and function, and how treating biology as multi-scale collective intelligence may inform new approaches to regeneration, longevity, and programmable synthetic, bio-hybrid, or living systems.
Bio: Benedikt (Ben) Hartl is a researcher working at the intersection of artificial intelligence, artificial life, evolutionary developmental biology, and physics. At the Levin Lab at the Allen Discovery Center at Tufts University, he develops generative, multi-agent machine learning models to study how collective intelligence and self-organisation not only emerge but actively shape living systems. His work focuses on neural cellular automata, diffusion models, neuroevolution, and related approaches to investigate morphogenesis, regeneration, and decentralized decision-making. More broadly, he aims to understand evolution, learning, and cognition as multi-scale, self-refining collective processes, with applications in synthetic biology, regenerative medicine, and adaptive microrobotics. https://bhartl.github.io/bib
Biotech and Health Extension Group
A group of scientists, entrepreneurs, funders, and institutional allies who cooperate to advance biotechnology to reverse aging and extend human healthspan.
Nominate a seminar presenter/topic
Share this application with colleagues who’d like to join
Feel free to reach out to lydia@foresight.org with any questions.