braincellsinadish,computingbadly

Problem statement

Current AI hardware (silicon chips) consumes enormous and growing amounts of energy — training GPT-4 required an estimated 50 GWh, roughly the annual electricity consumption of a small city — yet biological neural networks process equivalent or superior information tasks using a fraction of the energy. Despite this, we cannot design, engineer, or fabricate organoid systems capable of processing information dynamically while interfacing with non-living systems. Harnessing complex biological behavior for computing — creating 3D in vitro biological constructs (brain organoids, plant cell constructs, biofilm-based systems) that can receive diverse inputs, process them, and generate outputs that drive engineered devices — remains technically unprecedented. Even the definition of "intelligence" and "learning" in biological computing constructs is unresolved.

Why this matters

Global data center energy consumption is projected to double by 2030. The human brain performs complex pattern recognition, sensory integration, and decision-making on approximately 20 watts — roughly 10,000x more energy-efficient than silicon-based AI for equivalent tasks. If organoid computing could achieve even a fraction of biological neural efficiency, it would fundamentally alter the energy trajectory of AI development. Beyond efficiency, biological computing could enable capabilities that silicon struggles with: continuous learning without catastrophic forgetting, graceful degradation under damage, and processing of biochemical signals that electronic systems cannot detect.

What’s been tried and why it hasn’t worked

"Intelligence" and "learning" have fundamentally different meanings across biology, cognitive science, computer science, and engineering — there is no unified conceptual framework, making it impossible to set clear engineering targets or benchmarks. Creating organoid systems that receive diverse and unexpected inputs and dynamically respond through communications spanning multiple spatiotemporal scales (chemical, optical, mechanical, electrical) is technically unprecedented — no proof of concept exists for bidirectional biological-electronic information exchange at the organoid level. The ethical, legal, and social implications of using living neural tissue as computing substrate are unresolved and may constrain development pathways. Convergent research spanning engineering, biology, computer science, social science, and ethics has been insufficient — each community approaches the problem with fundamentally different assumptions and goals.

What would unlock progress

Organoid systems — brain organoids, plant cell constructs, or biofilm-based constructs — that demonstrably capture real-world input, autonomously process it, and generate outputs driving engineered systems. Interface technologies connecting biological constructs with engineered sensors and devices for sustained bidirectional communication. Defined bounds of "intelligence" and "learning" achievable in engineered biological constructs. Ethical frameworks developed in parallel with technical capabilities rather than retroactively. Neuromorphic computing approaches that bridge biological and silicon paradigms.

Entry points for student teams

A student team can close the loop in simulation before touching tissue: pull an open organoid multi-electrode-array recording from the DANDI archive (dandiset 001872, "The Statistical Analysis of Brain Organoid MEA Data," https://dandiarchive.org/dandiset/001872), read it with SpikeInterface (https://spikeinterface.readthedocs.io/en/stable/), drive a simulated device from the recorded activity, and use a spiking-network model in Brian2 (https://briansimulator.org) or synthetic ground-truth recordings from MEArec (https://mearec.readthedocs.io/en/latest/) to test whether the stimulation fed back changes anything the team is willing to call learning — the deliverable being an explicit, measurable definition of "learning" a wet lab could later apply, which is exactly what the field says it lacks. Doing the same with living cells takes a cell-culture facility with a commercial MEA system and, for human-derived neural tissue, institutional stem-cell and ethics oversight — access held by neuroscience and bioengineering departments, not by an unaffiliated team; a non-neuronal electrogenic substrate (a cardiomyocyte line, an electroactive biofilm) on a low-cost open electrode array lowers that bar but does not remove the culture facility. A second, documents-only door: conduct a systematic review of the ethical frameworks proposed for biological computing and organoid research, identifying gaps relative to the technical capabilities being developed. Relevant skills: bioengineering, electrical engineering, cell biology, neuroscience, ethics, microelectronics, computational neuroscience.