A 7 pA current stands between baseline and collapse

What the work claims

Be clear about the kind of work this is: a single-author simulation study, no experiments, no validation against biological recording. The model is a 10,000-neuron network of Izhikevich two-variable spiking neurons, split into a sensory processing layer of 5,000 regular-spiking neurons and a default-mode-style subsystem of 5,000 intrinsically bursting neurons, of which a 1,000-neuron core receives a continuous tonic current of 7.0 pA, standing in for ascending brainstem neuromodulation.¹ Connectivity is random and sparse, 1,000 outgoing projections per neuron for 10 million synapses, with a top-down weight asymmetry: projections from the pacemaker subsystem carry weight 0.6, sensory projections 0.4, encoding the intended functional dominance of the endogenous system. A stimulus of 16.0 pA is applied to 1,500 sensory neurons for 300 ms in the middle of a 900 ms simulation, on a background of Gaussian noise with standard deviation 3.

The paper's quantitative instrument is a proxy for integrated information, computed from the determinant ratio between the full covariance matrix of the most active neurons and a block-diagonal version that treats sensory and default-mode populations as independent. The claim is that this quantity, denoted phi, tracks three dynamical regimes: near zero during quiescence, 1 to 2.5 bits during pre-burst priming, and 10 to 15 bits during large synchronized population bursts, with peaks of 12.5 to 15.0 bits when sensory and pacemaker populations co-activate. From this the paper argues for the emergence of a persistent, self-sustaining neural representation of self, and situates the model in the integrated information theory literature on consciousness.

How it works

The mechanism worth keeping is the ablation, because it is the one part of the paper with a control. Across 30 random network realizations, the author compares the intact condition, tonic current 7.0 pA to the pacemaker core, against an ablated condition with zero tonic drive. In the baseline window before stimulation, the intact network holds an integrated-information baseline of 7.87 plus or minus 0.12 bits versus 4.67 plus or minus 0.57 bits ablated, a paired difference reported as t = 30.06 with p = 2.08 x 10^-23. During stimulus onset the intact network jumps to 14.27 plus or minus 0.76 bits while the ablated network responds late and weakly at 6.42 plus or minus 0.10 bits, t = 55.98 with p = 4.20 x 10^-31. After stimulus offset, the ablated network collapses toward zero integration, averaging 5.23 plus or minus 0.40 bits across the window, while the intact network sustains 13.13 plus or minus 1.35 bits, t = 32.34 with p = 2.65 x 10^-24.

The dynamical picture behind those numbers is plausible and, on its own terms, coherent. The pacemaker core fires persistently and asynchronously on tonic drive, keeping the recurrent network near a threshold where modest input can recruit large assemblies. The top-down weight bias lets this endogenous activity shape sensory responses. Synchronized bursts then sweep the whole network to a saturated global firing rate around 1,000 Hz, and the covariance proxy peaks because every neuron is co-active with every other. Remove the tonic current and the pacemaker core falls silent, the network stops sitting near threshold, and after the stimulus ends it cannot re-establish its prior state. Everything else in the paper, the language of qualia and self, is interpretation layered on this mechanism.

Where a skeptic should push

Start with the instrument. The phi used here is not the integrated information of integrated information theory, which is defined over causal mechanisms and is intractable to compute exactly for systems of this size. It is a covariance-based approximation taken from earlier work, evaluated over the 200 most active neurons in sliding windows of 5 to 80 milliseconds. Covariance has a well-known failure mode that matters enormously here: common input inflates it. During a global burst, shared drive makes every neuron co-vary whether or not any information is being integrated, so a high phi reading is precisely what an over-synchronized, information-poor network should produce. The paper's strongest phi peaks occur during saturating population events, which is evidence for the confound, not against it. A control that would separate integration from common drive, for instance comparing phi against phase-shuffled surrogates, is not reported.

Second, the ablation, while the best part of the paper, is close to trivial in its design. The 7.0 pA tonic current is the only thing making the designated pacemaker neurons fire; setting it to zero removes their activity by construction, so the observation that the network then loses its baseline and fails to recover is close to baked in. The interesting question, whether some other pattern of endogenous activity could support the same function, is not asked. Third, the statistics are a hybrid that should raise eyebrows: paired t-tests with 30 seeds of a fully deterministic simulation produce p-values below 10^-23, but the randomness being averaged over is only initial connectivity and parameter draws, not sampling uncertainty of any physical kind. The intervals are real evidence of robustness to initialization, nothing more.

Fourth, the framing. A 10,000-neuron model with one parameter setting, one stimulus protocol, one 900 ms window, random unstructured connectivity, and a second affiliation listed as a trademarked lab name is not a foundation for claims about self-consciousness, and to the author's credit, the discussion partially concedes this, explicitly calling the simulation a functional philosophical zombie with access consciousness at most. Readers should hold the paper to that narrower standard, because the narrower version, tonic drive maintains a recoverable baseline state, is genuinely interesting and is the version that survives contact with real neuroscience.

The baseline problem in organoid computing

Here is why a skeptical reading still lands hard on organoid intelligence. The demonstrated mechanism is: a recurrent spiking network with continuous endogenous drive sits in a primed baseline, integrates perturbations as they arrive, and returns to baseline after they stop; remove the drive and the same network responds late, weakly, and cannot recover. Now consider what a cortical organoid on a multielectrode array is. It is a recurrent spiking network with abundant spontaneous burst activity, no sensory afferents, no ascending neuromodulatory tone, no brainstem, no thalamus, and a baseline that wanders over days as the culture matures. Every serious organoid-computing effort currently spends its ingenuity on readout and stimulation while treating the culture's internal state as something to suppress or route around. This paper is a formalized argument that the internal state is the product, and that its maintenance requires exactly the input, continuous modulatory drive, that a dish does not provide.

The opportunity is concrete. If primed baseline states in recurrent tissue are tonic-drive dependent, then neuromodulation is not an optional enhancement for organoid computing but infrastructure: perfused dopaminergic, cholinergic or noradrenergic tone, or electrical or optogenetic substitutes for it, become part of the definition of a working system, alongside media and temperature. It also reframes the spontaneous activity that organoid electrophysiology already measures everywhere. The slow oscillations and burst sequences of dish cultures are usually logged as signatures of maturation or noise to be averaged away. On the paper's reading, that activity is the baseline dynamics doing the thing that makes computation possible, and the right engineering question is whether a given culture's endogenous regime sits in a primed, recoverable state or in one that collapses after every perturbation, which is a measurable question with existing tools.

The threat is hype, and it is specific. Papers like this one show how cheaply a phi-like number can be manufactured: one afternoon of simulation on a desktop machine, a determinant ratio over the most active 200 neurons, and fifteen bits of integrated consciousness emerge during what is functionally a seizure-like saturated burst. If that statistic is ever waved at an organoid recording as evidence of sentience or suffering in a dish, the correct response is now quantitative, not rhetorical. The same confound that inflates phi here, common drive during synchronization, is guaranteed to be present in organoid cultures, which are famous for spontaneous global bursts. A field that cannot distinguish integrated computation from synchronized noise in a 10,000-neuron model it built itself is not ready to certify consciousness in tissue it did not build. Any governance framework for biological computing should treat covariance-based consciousness metrics as unfit for evidentiary use, and this paper is the case study for why.

The bottom line

Established by the paper: in this specific 10,000-neuron random network, continuous tonic drive to an intrinsically bursting core is necessary to maintain a baseline state that responds to and recovers from sensory perturbation, robustly across 30 initializations; without it, post-stimulus integration collapses. Not established: that this has anything to do with consciousness, that the covariance proxy measures integrated information rather than common drive, or that the result generalizes beyond one architecture and one stimulus protocol. The mechanism is confirmed for biological computing the moment someone demonstrates, in cultured tissue, that sustained modulatory drive shifts cultures from a collapse-after-stimulus regime to a recover-baseline regime; it is weakened if phase-shuffled controls show the phi proxy tracks synchrony rather than integration. For organoid intelligence the lasting value of this paper is unintentional: it formalizes why a living computer without neuromodulation is a computer whose operating system is being switched off, and it demonstrates why nobody should ever accept a covariance statistic as proof of inner experience.