A thousand electrodes, folded into a living sphere
A four year, two million dollar federal program proposes to build soft bioelectronics carrying over a thousand electrodes, integrate them in three dimensions through a brain organoid, and close a loop that reads the tissue and stimulates it back. The headline is biological computing. The load bearing engineering is an acquisition problem: getting a thousand quiet channels out of a growing millimetre of tissue.
Source: Neuron-Soft Organoid-Computer Interfaces for Long-Term Three-Dimensional Neural Network Computing, NSF award 2422348 (EFRI BEGIN OI), 2024 to 2028. Primary source. Read: the full award abstract and metadata via the NSF award API. This is a funded proposal, not a result; every capability below is a stated plan unless attributed to the group's separately published work.
What the work claims
The award funds a research program, not a finished device, and it should be read as an intention with money behind it rather than as data. Its central bet is that two limits of organoid recording can be attacked together. First, that soft electronics with over one thousand electrodes can be integrated seamlessly in three dimensions through a brain organoid for long term, stable recording and stimulation.1 Second, that an embedded sensing to computation to actuation loop, driven by reinforcement learning, can generate evolving stimulation patterns that steer the tissue's maturation toward chosen sub-regional and functional specificity.1
What makes this bold is the word seamless and the number three. Almost every organoid recording to date happens on a two dimensional floor: the organoid sits on a planar microelectrode array and the array sees the units nearest the glass. A volumetric interface that samples the interior at a thousand sites, and does so while the tissue is still developing, would be a different instrument class. The program is led from Harvard with a University of Texas at Austin partner, and it descends from a real published lineage of soft electronics that become embedded as tissue grows around them.2
How it is meant to work
The enabling idea is mechanical, not electrical. Conventional probes are stiff relative to neural tissue by roughly seven to eight orders of magnitude, a kilopascal-scale tissue against silicon near 150 gigapascals, so they wound the tissue and provoke a glial response. The proposed approach instead engineers the bending stiffness of an electrode mesh so low that the developing tissue folds and grows around it, distributing the electronics through the volume the way scaffolding is enveloped by a building rather than driven into a finished one. The award calls this mechanics-driven soft bioelectronics design combined with three dimensional microfabrication.1 The group's prior Nature report, listed on the award record and not independently reviewed here, describes exactly this manoeuvre for cortical tissue: seed the electronics early and let embryonic-style development carry them inward.2
On top of that substrate sits the loop. Embedded sensors read spiking activity; a reinforcement learning policy chooses a spatiotemporal stimulation pattern; embedded actuators deliver it; the tissue's response is read back and scored. Run over developmental time, the loop is meant to push the organoid toward a target internal organization rather than merely watch it drift. For the acquisition chain, the important consequence is that the same distributed array must record microvolt signals and deliver stimulation, often close together in time and space.
Where a skeptic should push
The single most load bearing assumption is that over a thousand electrodes can be turned into over a thousand usable channels out of a millimetre-scale sphere, and kept usable for months. That is an interconnect problem before it is a recording problem. Every electrode needs a conductive lead, and leads do not shrink as fast as electrode sites. Routing a thousand traces out of a three dimensional volume, through a soft substrate that must also flex, is the wall that has historically separated electrode count on a datasheet from channel count in a working experiment. The standard escape, multiplexing many electrodes onto a few lines or digitising inside the substrate, relaxes the wire count but spends temporal fidelity or injects heat and noise into the tissue, so the constraint moves rather than vanishes. The award abstract states the ambition but, being a plan, reports no yield, no noise floor, and no channel count actually achieved.
Two more pressures deserve naming, and an adversarial read of this piece sharpened both. The tissue grows, and electronics fixed at assembly are strained by that growth: a contact placed correctly on day one can drift, delaminate, or change its interface impedance by week eight. The program's own answer is to embed the electronics during development, so the tissue grows around a compliant mesh instead of around a probe forced into finished tissue, and the group's Nature report is the in vivo precedent for exactly that manoeuvre. That mitigates insertion trauma and lets growth accommodate the device, but it does not abolish chronic strain, and it has not been shown at anything close to a thousand-channel volumetric density. The second pressure is that the loop demands concurrent stimulation and recording on one array. A stimulus can dwarf a neural signal by three to four orders of magnitude on a nearby contact, and by far more at the stimulating site itself; the usual defence, blanking the amplifier for a few hundred microseconds, collides with the sub-millisecond onset of an evoked response. The loop that steers maturation is itself biologically slow, unfolding over days, so blanking does not throttle its rate; what it degrades is the quality of the signal the policy is scored on, by discarding the earliest and most informative part of each response. Finally, the biological claim, that patterned stimulation can direct functional specificity, is the least demonstrated part and should be read as a hypothesis the program intends to test, not a capability it possesses.
What it demands of the acquisition chain
Read strictly as hardware, this program relocates the hard problem from the amplifier to the wire and the boundary. If the field can already build low noise amplifiers at thousands of channels, the binding constraints become interconnect density, multiplexing, and stimulate-record isolation. A thousand volumetric channels almost certainly cannot each own a dedicated off-array wire, which forces on-array multiplexing or active in-substrate electronics; both trade temporal fidelity or add heat and noise inside living tissue. That is the non-obvious implication: a volumetric organoid interface is less an electrode advance than a packaging and routing advance, and it will be judged on usable, low-drift channel yield rather than on nominal electrode count.
The genuine opportunity is a new instrument category. A distributed, embedded, closed loop array would let the acquisition chain do things a planar passive MEA cannot: sample the interior, hold contacts through development, and act as well as observe. The genuine threat is twofold. Technically, the interconnect and strain problems could make the delivered system a few dozen good channels wrapped around a thousand-electrode press release, the classic gap between specification and measured result. Ethically, an array whose explicit purpose is to direct the development of human-derived neural tissue is dual-use by construction, which is why the program carries a dedicated ethics thrust rather than an afterthought. For anyone building organoid instrumentation, the lesson is to specify closed-loop arrays by their artifact-limited effective channel count under simultaneous stimulation, not by their electrode census.
The bottom line
This is a well funded hypothesis, not an established result. Its strongest, most concrete element is the mechanics-driven soft interface, which rests on a published lineage; its weakest is the leap from a thousand electrodes to a thousand stable, artifact-clean channels closing a developmental loop. What would confirm the direction is a demonstrated volumetric recording with a reported per-channel noise floor, a usable-channel yield, and stability quantified across weeks of tissue growth, together with a loop that measurably changes maturation. What would break it is an interconnect or strain ceiling that leaves the effective channel count far below the electrode count, or a stimulation artifact that the loop cannot see past. Until those numbers exist, the honest reading is that the interface is the science, and the science is unproven.
Frequently asked questions
Is this a working device or a proposal?
A proposal. NSF award 2422348 is an active four year program funded in 2024. The abstract states goals and an approach; it reports no measured channel count, noise floor, or yield. We treat every capability as a plan, and label the group's separately published Nature work as prior art we did not independently review.
Why call interconnect the real bottleneck rather than the amplifier?
Because thousands of low noise amplifier channels already exist in commercial CMOS arrays. What does not scale as easily is getting a thousand conductive leads out of a soft, three dimensional, growing volume without adding stiffness, noise, or heat. Electrode count on a datasheet and usable channel count in an experiment diverge precisely at that wire.
What is the stimulate-and-record problem?
The same array must both read microvolt spikes and deliver stimulation. A stimulus can exceed the neural signal by several orders of magnitude on nearby contacts, and the standard fix, briefly blanking the amplifier, can hide the fast responses a closed loop needs. Effective performance is therefore set by artifact rejection, not by nominal channel count.
Does growth of the tissue threaten the electronics?
Yes. Electronics integrated during development are then subjected to the tissue's continued growth. Fixed contacts can drift, delaminate, or change interface impedance over weeks. Long term stability at high electrode density through active growth is unproven and is one of the program's central risks.
Why does a hardware analysis raise ethics?
Because the array's stated purpose is not only to observe but to steer the maturation of human-derived neural tissue toward chosen functions. That is dual-use by construction, which is why the award includes a dedicated ethics thrust. Instrumentation that acts on living tissue cannot be evaluated on signal quality alone.
References
- National Science Foundation. EFRI BEGIN OI: Neuron-Soft Organoid-Computer Interfaces for Long-Term Three-Dimensional Neural Network Computing. Award 2422348, Harvard University (PI J. Liu) with University of Texas at Austin. 2024 to 2028. nsf.gov/awardsearch/showAward?AWD_ID=2422348. Accessed 2026-07-23.
- Sheng H, Liu R, Li Q, et al. Brain implantation of soft bioelectronics via embryonic development. Nature. 2025; 642. doi:10.1038/s41586-025-09106-8. Among the publications listed on the award record; cited here as prior art and not independently reviewed. Accessed 2026-07-23.