Oscillator reservoirs and the durable claim of the wet substrate

What the work claims

This is a grant, a funded plan rather than a finding, and it must be weighted as intention, not evidence.¹ It is a Small Business Innovation Research Phase I award of 272,615 dollars, obligated in fiscal year 2024, to build what the awardee calls an Oscillator Processing Unit. The thesis is that two separate approaches to analog computing, physical reservoir computing and adaptive oscillators, can be converged into a single device, and that because an oscillator core does not separate where information is stored from where it is processed, such a device could sidestep the von Neumann bottleneck and act as both an inference processor and a general-purpose computer. The award record also lists four associated journal publications by the principal investigator on Duffing adaptive oscillators and physical reservoir computing, which establishes that this is an active and published line of work rather than a paper proposal.

Because there are no benchmark results in an award abstract, the honest read is limited: this tells us what a credible group intends to build and why, and it lets us examine the paradigm carefully, but it cannot tell us whether the device works or how well.

How it works

Reservoir computing is the key idea to understand, because it is also the theory beneath most organoid computing. You take a fixed, high-dimensional nonlinear dynamical system, the reservoir, and drive it with your input. The reservoir's rich, fading-memory response projects the input into a space where a task becomes easy, and crucially you train only a simple linear readout of the reservoir state. Nothing inside the reservoir is trained, which is what makes the approach tolerant of messy, un-rewireable media. Physical reservoir computing uses a physical system, here a network of oscillators, as that medium. Adaptive oscillators are the second strand: oscillators that adjust their own natural frequency to lock onto features of the input, a limited form of self-organization. Fusing the two is the award's central bet.

The phrase that carries the most weight is that in an oscillator core, memory and processing are not independent. That is a precise statement of why analog physical computing appeals: a conventional computer spends most of its energy shuttling data between separated memory and logic, the von Neumann bottleneck, whereas a physical dynamical system holds its recent history in its own state, its fading memory, and transforms that history through the same dynamics. Computation and storage are the same physical event. The award frames this as a route to smaller, faster, lower-power edge processors that could also be manufactured domestically.

Where a skeptic should push

The first caution is category: this is a small Phase I award, and Phase I exists to test feasibility, not to ship a processor. The market framing, an edge-computing market projected at 157 billion dollars by 2032, is standard proposal scaffolding and carries no technical weight. The strongest claim to interrogate is the last one, that an oscillator reservoir could serve as both an AI inference processor and a generalized computing processor. Reservoir computers are well suited to temporal pattern tasks, classification and prediction on time series, and much less obviously suited to general computation, which needs addressable memory and controllable, composable logic. Calling a fixed nonlinear reservoir a general-purpose processor is the kind of scope that a Phase I abstract can assert and a Phase I budget cannot demonstrate. And because there are no results in the record I have read, everything here is potential, which is exactly what a grant is meant to describe.

The oscillator reservoir and wetware's real claim

Reservoir computing is the framework that makes organoid intelligence coherent, so a serious effort to build a better physical reservoir is aimed, unintentionally, at the heart of the field. Most demonstrations of computing with living neurons follow the reservoir recipe: the tissue is the untrained nonlinear medium, a stimulus is the input, and only a readout is trained. That recipe is what lets anyone compute with a substrate they cannot design or rewire, and it is genuinely the right posture for organoids, because it tolerates drift by re-fitting the readout as the tissue changes. The non-obvious implication is that this strength is also a trap. If the reservoir contract asks nothing of the medium beyond being a rich nonlinear dynamical system, then the medium is interchangeable, and an oscillator network, a nonlinear optical cavity, and a cortical organoid are, from the readout's point of view, competing suppliers of the same service. The living tissue has no privileged claim unless it delivers a kernel that silicon cannot cheaply match.

That is where this award becomes a substrate-obsolescence argument written in someone else's grant. The property it advertises, memory and processing fused in the same physics, is precisely the property organoid advocates cite for neurons. A manufacturable oscillator core that delivered that property would compete directly with the wet substrate, and on the axes that matter for deployment it would hold structural advantages: it runs at room temperature without a wet lab, it is deterministic and reproducible, it can be copied, and it does not die in weeks. Grounding this in the award's own mechanism, if the fading-memory nonlinear kernel is what organoids are prized for, and an oscillator unit supplies that kernel in silicon, then the biological reservoir is left competing on biological fidelity, which is a reason to build organs for disease modeling, not a reason to compute on them.

The genuine opportunity, and the clearest durable claim for the wet substrate, is the one thing a fixed oscillator reservoir cannot do: change its own internal weights. A reservoir is by definition untrained inside, and adaptive oscillators bend only their frequencies. A living network has real synaptic plasticity, so in principle it is not merely a reservoir but a reservoir that can partly rewrite itself in response to use. That is the property most likely to justify computing on tissue rather than on an oscillator core, alongside a weaker candidate: that the dense three-dimensional recurrence and multi-timescale dynamics of living tissue could furnish a richer kernel than an oscillator network, though that is an empirical race silicon might win rather than a categorical difference. Plasticity is also, soberingly, the property that the learning-rule ceiling makes hardest to exploit, because in-tissue learning is local and noisy. So the balanced reading is that this grant strengthens the case for silicon reservoirs across the board and narrows the wet-substrate case toward what a fixed reservoir cannot copy, chiefly controllable plasticity, which no fixed physical reservoir, however well engineered, can reproduce.

The boundary here is unusually important because the source is thin. I read the award record and confirmed the associated publications exist; I did not read those papers, and I make no claim about their specific results. Everything above analyzes the paradigm the award commits to, not a demonstrated device.

The bottom line

Established, as facts of the award: NSF funded a Phase I effort to unify physical reservoir computing and adaptive oscillators into an oscillator processing unit that exploits fused memory and processing to attack the von Neumann bottleneck, and the principal investigator has a published record in this area. Everything about the device's performance is aspirational, and the general-purpose-processor framing is the least supported part. For organoid intelligence the calibrated conclusion is that reservoir computing both enables the field and endangers it: it is the right way to compute with an un-rewireable living medium, and it makes that medium interchangeable with any good physical reservoir, so the durable case for wetware narrows toward what a fixed reservoir cannot copy, chiefly controllable plasticity, with a richer dynamical kernel a weaker second possibility. The threat would be confirmed if an engineered oscillator reservoir matches a living reservoir on a temporal task where the tissue's plasticity was supposed to matter; the wet substrate's case would be vindicated only where controllable in-tissue plasticity delivers performance a fixed reservoir provably cannot.