The value evolution wrote into a brain is in the weights, not the wiring diagram

What the work claims

This is a benchmarking and optimisation study, a methods paper with a biological substrate rather than a neuroscience result. The setup is classic reservoir computing: a fixed recurrent network, here an echo-state network with tanh neurons and spectral radius pinned to 0.97, transforms a temporal input into a high-dimensional state trajectory, and only a linear ridge-regression readout is trained. What varies is the recurrent weight matrix, seeded from one of six real connectomes spanning six orders of magnitude in neural complexity: C. elegans at 279 neurons, Drosophila at 49 nodes, mouse at 112, rat at 73, macaque cortex at 29 regions with continuous fractional-labelled-neuron densities, and human at 83 parcels from diffusion tractography. Four gradient-free, biologically inspired optimisers, particle swarm, differential evolution, grey wolf, and whale optimisation, perturb the non-zero edge weights while the connectome's sparse topology stays frozen.¹

The headline claims are three. First, all four optimisers, initialised as Gaussian perturbations of the biological weights, significantly outperform the unoptimised connectome on all six species across all four benchmarks, which are memory capacity, Lorenz-attractor prediction, NARMA-10 system identification, and Mackey-Glass chaotic time-series prediction, with evaluation on held-out data using a separate random seed from the optimisation objective. Second, the whale optimisation algorithm leads everywhere: up to a 17-fold memory-capacity gain in the C. elegans reservoir, from 1.39 to 23.91, an 89 percent error reduction on Mackey-Glass for the human connectome, and an average improvement of 214 percent across species and tasks. Third, and the claim the authors themselves treat as crucial, random initialisation on the same connectome topology falls at or below the biological baseline, so the evolved advantage is carried by the weight values, not recoverable from the wiring diagram alone.

How it works

An echo-state network needs its recurrent matrix to satisfy the echo-state property, roughly that old inputs wash out rather than echo forever, which the authors enforce by rescaling every matrix to a spectral radius of 0.97, a value prior work associates with maximal linear memory. Each connectome contributes topology and starting weights: synapse counts from electron microscopy for the worm, confocal fluorescence for the fly, tracer injection densities for rodent, retrograde fractional-labelled-neuron densities for macaque, and streamline counts for human. The optimisers then search the space of non-zero edge weights. The initialisation is the experimentally important detail: populations start as Gaussian perturbations of the biological weights, with standard deviation 0.3 times the biological spread and values clipped to between zero and three times the maximum biological weight, and a particle-swarm run initialised uniformly at random on the same topology serves as the negative control. Memory capacity is measured as the sum of squared correlations between the reservoir state and the last 50 input steps, with theoretical maximum equal to the neuron count, and the predictive tasks are scored by normalised root-mean-squared error. Every condition is run ten times, and significance is claimed at p below 0.001 by paired t-test.¹

Two results give the mechanism its shape. The worm connectome starts with the worst memory capacity of any species, 1.39 plus or minus 0.20 against a theoretical maximum near 279, and therefore has the most headroom, ending at 23.91 after whale optimisation. The macaque connectome starts with the best memory capacity, 4.75 plus or minus 0.24 despite having only 29 nodes, because its 72.7 percent density and continuous weight distribution give the dynamics more to work with; it gains least in relative terms, plus 226 percent. Gains track starting weakness, not phylogenetic proximity to anything in particular, which is what you would expect if the optimisers are harvesting unused capacity rather than discovering species-specific secrets.

Where a skeptic should push

The load-bearing assumption is that these reservoirs still meaningfully represent biology after the processing they undergo, and here the paper makes several quiet but consequential simplifications. The spectral-radius rescaling to 0.97 erases the biological timescales and gain structure that presumably matter to the real nervous systems. The weight clipping to non-negative values discards the sign of every connection, so excitatory-inhibitory structure, the single most important organising principle of real neural networks, is flattened; the reservoirs are recurrent excitatory webs wearing connectome topology as a costume. The benchmarks are low-dimensional scalar tasks, memory and chaotic prediction, not the perceptual, motor, or relational computations the connectomes actually evolved for, so showing that optimised connectomes do well on them is as much a statement about the tasks as about the brains. And the evidence is ten runs per condition with a paired t-test, which the authors themselves name as a statistical-power limit; they also note that input-node assignment is random for all species except the fly, that the human connectome is an indirect diffusion-MRI proxy for axonal connectivity, and that algorithm rankings may shift with larger optimisation budgets. None of this overturns the core comparative finding, random weights fail where biological weights succeed, but it bounds the interpretation: the study demonstrates that evolved weight distributions are a useful inductive bias for echo-state networks, not that it has located where computation lives in a real brain.

Evolved wiring as a benchmark for living tissue

For organoid intelligence, the substrate-obsolescence reading comes first, because it is the uncomfortable one. A standard argument for computing on living tissue is that biological neural structure carries computational value that random or engineered substrates lack. This paper puts that claim in its sharpest form yet, and then undercuts the necessity of wetware to obtain it. Everything evolution deposited, the topology and the weight distribution, can be lifted out of an atlas, dropped into a conventional echo-state network, and improved upon by a generic metaheuristic in silico, with no cells to feed, no ethics board, no batch variability, and a held-out evaluation protocol. The connectome reservoir does not just match the living-substrate pitch, it beats it on convenience and, after optimisation, on the benchmarks used here. Anyone claiming that organoid or tissue reservoirs are valuable because biological structure is valuable now has a specific, published baseline to clear: a swarm-optimised connectome echo-state network, which is embarrassingly cheap to run.

The opportunity, though, is the subtler and more durable point. The negative control is the result that matters for the wetware field. Topology alone carried nothing; random weights on the worm's wiring failed to reach even the unoptimised biological baseline. The information evolution wrote is in the realised synaptic weights, the full history of development and experience compressed into a number per edge. If that is true of evolved nervous systems, the analogous claim for self-organised neural tissue is direct: an organoid's computational value, if it has any beyond a generic dynamical system, lives in its realised weights and therefore in its developmental history, its morphogen exposure, its maturation time, its activity. That reframes the field's reproducibility problem in a useful way. It is not enough to grow tissue with the right cell types and the right rough structure; the weight distribution is the product, and it is exquisitely sensitive to history. The practical consequence is that protocol variables the field treats as nuisance, plating density, induction timing, oxygen, feeding, may be exactly the variables that write the inductive bias, and they belong in the reportable specification of any organoid claimed as a computing substrate.

There is also a concrete benchmarking gift here. The paper supplies a protocol that organoid groups could adopt verbatim: four canonical tasks, a held-out seed distinct from the optimisation seed, ten runs, memory capacity and NRMSE. Running living tissue reservoirs on exactly these tasks, against exactly these connectome-optimised baselines, would give the field its first honest cross-substrate comparison, with the in silico evolved connectomes as the control arm that currently sets the bar. And one mechanistic hint deserves testing in tissue: the optimisers gain most where the starting dynamics are weakest, which suggests closed-loop training of living reservoirs should expect its largest payoffs early, on under-performing cultures, rather than on ones already sitting near a good regime. That is a hypothesis, not a result, but it is the kind this paper equips the field to actually run.

The bottom line

Established within the paper's own design: biologically initialised echo-state networks consistently outperform their unoptimised connectomes across six species and four tasks under held-out evaluation, random initialisation on the same topologies does not, and whale optimisation leads throughout, with a 17-fold memory-capacity gain in C. elegans and an average improvement of 214 percent. Not established: that the optimised reservoirs preserve anything biologically meaningful, that the results say where computation lives in real brains, or that the rankings are stable under larger budgets or richer tasks; sign structure and biological timescales were discarded by construction. For organoid intelligence the deflationary lesson is that the value of biological structure can currently be captured without biology, so living substrates must justify themselves on axes a connectome file cannot, such as self-repair, continual adaptation, or energy cost at scale. The constructive lesson is that if realised weights are where the value sits, then developmental history is the spec, and the field should start reporting it and benchmarking tissue against these exact baselines.