NSF bets $2M on training organoids, with one risky hypothesis

What the award commits to

Genre first, because it governs everything. This is a grant abstract. It is evidence about what a review panel found credible and what investigators have committed to attempting. It is not evidence about the world. It is also a public facing document compressed for a lay reader and a mixed panel, so controls I describe as missing may exist in the full project description and simply not have survived the compression. Read what follows as claims about what is on the public record and what a reader should demand of the eventual papers.

The award is a continuing grant from the NSF Directorate for Engineering, through the Office of Emerging Frontiers in Research and Innovation, jointly funded with the Directorate for Mathematical and Physical Sciences. BEGIN OI stands for Biocomputing through EnGINeering Organoid Intelligence, a programme established by solicitation NSF 24-508, posted in November 2023, which anticipated roughly 15 awards against $30,000,000 over two fiscal years. This is one of them. It runs from September 2024 to an estimated August 2028, with an estimated total of $2,000,000, of which $1,775,000 was obligated in fiscal 2024 and $225,000 in fiscal 2025. The principal investigator is Pamela Abshire at the University of Maryland, College Park, with co-investigators Timothy Horiuchi, Erika Taylor and Marc Dandin.¹

The composition matters. Abshire works in mixed signal very large scale integration and integrated bioelectronics; Horiuchi is a neuromorphic engineer known for analog circuit models of sensorimotor and echolocation circuits. This is an engineering led team rather than a stem cell biology led one, and that predicts both its strengths and its blind spot.

Three thrusts are named: biocomputing theory to identify functional neuronal networks; organoid culture plus hardware interfaces using high density microelectrode arrays and optogenetics; and discussion and analysis of ethical concerns. Three hypotheses are stated. First, that 3D cortical organoids can be trained to respond selectively to one pattern out of many distractor stimuli. Second, that this association can be established simultaneously for many distinct patterns. Third, that closed loop training incorporating mechanistic insight is more effective than either open loop training or closed loop training that ignores underlying neural structure. One further framing deserves attention: organoids are described as the nodes of a three dimensional network, not as the computational substrate itself.

How the hypotheses are structured

Read as an engineering specification, these are not three coordinate hypotheses. The second strictly entails the first: establish many pattern specific associations simultaneously and you have trivially established one. The first is therefore not really a hypothesis but a de-risking milestone written in a hypothesis's grammar, and its function in a four year continuing grant is to guarantee a publishable result in year two regardless of what happens to the harder claim. The third is orthogonal to both, because it is a claim about controllers rather than about tissue, and could in principle be evaluated on any trainable substrate.

The structural consequence is that this project cannot fail loudly in its first half.

The specification objection is the unquantified quantifier. Many distinct patterns has no number. Simultaneously has no retention interval. Selectively has no effect size and no baseline. As written, many can retreat to two under pressure, and simultaneously can mean that both were above chance in one session rather than that a joint decoder held both above chance across sessions without retraining. A specification with an unbounded quantifier and no acceptance threshold is difficult to falsify by construction, and that is the most important structural fact about this document.

Where a skeptic should push

The dominant risk to the second hypothesis is not failure. It is spurious success, and the mechanism is worth stating carefully because it is the single most important thing in this analysis.

A fixed, non-plastic, high dimensional nonlinear dynamical system with fading memory is a reservoir. Reservoirs support many simultaneously decodable input discriminations with no learning whatsoever, because the discrimination lives entirely in the readout that the experimenter trains, not in the substrate. A three dimensional neuronal network stimulated with distinct spatiotemporal patterns is close to an ideal physical reservoir. So the second hypothesis can be confirmed by tissue that stored nothing, and the failure mode gets worse as you add patterns, because high dimensional separability is exactly what reservoirs are good at and exactly what plastic storage is bad at. The hardest claim to establish is simultaneously the easiest claim to appear to establish. Any paper reporting multi-pattern training in organoids must therefore demonstrate that something changed in the tissue, not merely that a decoder succeeded.

The most load bearing unstated assumption is that training an organoid is a control problem with a stable plant. Cortical organoids are developmentally non-stationary over the timescale of these experiments: the substrate matures, rewires and drifts while you are writing associations into it, so any learned state is confounded with development. This is worse than the more visible addressability problem, for a specific reason. Drift is a slowly varying, low dimensional signal shared across recording channels, which is precisely the structure that inflates apparent decodability. If training on a pattern precedes testing on it in wall clock time, a decoder can exploit maturation correlated firing changes and report selectivity that is really a clock. Together with the reservoir problem, that gives two independent routes to a positive result with no storage in the tissue. Addressability, by contrast, mostly produces null results and honest complaints about yield. It is the more visible problem and the less dangerous one.

There is a principled reason to expect low capacity and short retention, and it is not merely interference. Cortical organoids have no subcortical neuromodulatory afferents. There is no cholinergic, dopaminergic or noradrenergic third factor to gate plasticity, no consolidation phase, no replay, no sleep like offline period. Every gating signal must arrive exogenously through the stimulation channel, which means the experimenter's protocol is the credit assignment mechanism, while the tissue supplies local Hebbian rules plus homeostatic scaling that actively works against retaining written asymmetries. The honest deliverable for the second hypothesis is therefore a capacity curve, performance against the number of concurrently trained patterns, with a measured knee and a retention time constant. A knee at two or three, reported plainly, would be a genuine and citable result. A binary claim that many was achieved would not.

The third hypothesis is the scientifically serious one, because it is the only claim that would generalise beyond this substrate, and it is nearly unfalsifiable as stated because mechanism informed has no operational definition. A rigorous test needs a pre-registered mechanistic model making a prediction that differs from the blind controller before any data are seen, resource matched arms with equal parameter count and stimulation dose and tuning budget, a strong blind baseline such as Bayesian optimisation rather than random search, counterbalanced order with time as an explicit covariate, and organoid level sample size with differentiation batch as a nesting factor. Above all it needs a mechanism scrambled arm: a deliberately wrong mechanistic model of matched complexity. If the scrambled model performs as well, the benefit came from having any structured prior rather than from the mechanism being true. That control is what separates the hypothesis from a tautology, and it is the one least likely to be run.

The predictable degenerate version is three arms in the aims and two in the paper: mechanism informed closed loop against open loop fixed interval stimulation, which is a strawman, with the blind closed loop arm absent or underpowered. Mechanism informed will be operationalised after the fact as targeting electrodes with high spontaneous pairwise correlation, which is a data driven heuristic rather than a mechanism and would survive being wrong about every biophysical detail. Effect sizes will be computed per trial or per electrode, inflating the sample by orders of magnitude over the four to six organoids underneath. Watch for exactly that.

What this means for organoid intelligence

The non-obvious implication is architectural and it is hiding in one phrase. The abstract describes organoids as the nodes of a three dimensional network. That is a real departure from the single culture approach that has defined the field's best known demonstrations, and it quietly concedes that a single organoid is not the computational unit. Modularity is a reasonable response to the two properties that make monolithic organoids hard to compute with, namely diffusion limited size and the necrotic core that follows from it. Keeping each node small enough to stay oxygenated, optically accessible and electrically addressable, then composing them, is a sensible engineering answer.

It also creates a harder problem than it solves. A network needs its nodes to remain differentiated in order to carry distinct state, needs defined and durable inter-node connectivity, and needs latency it can reason about. None of those are solved. The field has demonstrated that fused neural tissues develop connections; it has not demonstrated connections with specified strength, direction and stability over the months such a machine would need to run. So this award is making a bet on composition before the composition primitives exist, which is either prescient or premature and will not be distinguishable for several years.

The ethics thrust deserves a plain reading rather than a diplomatic one. Its stated purpose is discussion and analysis to build awareness, literacy and reasoning capacity among researchers and the wider community. That is ethics as education and capacity building. There is no deliverable in it that can halt or modify an experiment: no oversight body, no pre-specified moral status threshold, no stopping criterion, no independent review trigger. It is deliberative, not governing. That is defensible, because cortical organoids at accessible scales, without vascularisation, thalamic input or sensory afferents, are widely taken to be far from anything with morally relevant experience, so there is little live moral hazard for a governance mechanism to govern. The value is anticipatory: building the vocabulary and the trained personnel before the capability arrives is the only time it is cheap. The structural signal is the more interesting one. Once a panel funds embedded ethics in an engineering award, subsequent proposals in that programme line carry one too, and a cohort acquires a shared normative apparatus by construction rather than after an incident. The test of substance is observable: whether the ethics work produces independently authored output or only a paragraph in the final report.

The genuine threat is a hype correction, and its mechanism is the wording. Because the second hypothesis has no threshold and the first is entailed by it, a weak result can be reported as a positive one without anybody lying. Because the reservoir and drift confounds both produce positive looking data, a null result may not be cleanly detectable even in principle. The risk to the field is therefore not a public failure but a slow accumulation of unfalsifiable positives that later fail to replicate, which is the pattern that has damaged other fields far more than honest negative results ever did. A capacity curve with a knee at two would be worth more to organoid intelligence than another headline about tissue that learned.

The bottom line

Nothing here is a result. What is established is institutional: the National Science Foundation has committed $2,000,000 through 2028, with the large majority already obligated, to an engineering led programme that treats training neural tissue as a control problem and writes ethics in as a funded thrust. That is a real signal about the field's transition from curiosity to programme, and it is independent of whether the science works.

What would count as success: a published capacity curve showing performance against the number of concurrently trained patterns with a measured knee and a retention time constant; evidence that something changed in the tissue rather than in the decoder, ideally by showing the trained state persists through a period without stimulation and survives decoder retraining; and a mechanism scrambled control arm in the third hypothesis. What would count as failure, and should be reported rather than buried: a knee at one, retention shorter than the training session, or a scrambled mechanism performing as well as the real one.

What a reader should watch for between now and 2028 is narrower than the science. Watch whether the sample size is reported per organoid or per electrode. Watch whether time is counterbalanced or merely mentioned. Watch whether the mechanism blind closed loop arm survives from the aims into the paper. Those three details will tell you more about the result's reliability than any effect size in the abstract.