Introduction — “The Signal Beneath Everything”
The universe does not compute. It signals. From the electromagnetic radiation that carries energy across the vacuum of interstellar space to the action potentials that fire across synaptic gaps in fractions of a millisecond; from the chemical gradients that guide embryonic development across sheets of undifferentiated tissue to the gravitational waves that announce, with a chirp lasting less than a second, the collision of black holes a billion light-years distant — the fundamental medium of physical causation is the signal. A dynamic, temporally structured disturbance propagating through a medium, carrying information precisely in the pattern of its variation over time. Not in its static amplitude. Not in a discrete state it occupies. In its change. In its becoming.
Digital computation, by contrast, is a remarkable abstraction from this physical reality. It takes the signal — continuous, analog, noisy, and rich in temporal texture — and discretizes it into binary states, synchronizes it with a master clock, filters out its analog depth, and processes it through symbolic operations on discrete tokens. This is an engineering achievement of extraordinary power. The transistor, the logic gate, the stored-program computer, the microprocessor: each is a triumph of the discretizing impulse, a demonstration that an enormous range of useful computations can be performed by systems that strip away the physics of the signal and retain only its digital shadow. Claude Shannon himself showed in 1948 that information can be quantified, separated from its physical carrier, and transmitted through any noisy channel with arbitrarily low error — provided the channel capacity is sufficient. The digital revolution is, in a precise sense, the applied consequence of Shannon's theorem.
But that reduction — from signal to bit, from analog texture to binary state, from temporal dynamics to synchronous clock tick — is not merely a simplification. It is a loss. What is lost is precisely what biological intelligence has spent hundreds of millions of years learning to exploit: the temporal structure of signals, their graded analog variation, their multi-frequency content, their capacity to encode information not in what they are at any instant but in how they change, co-occur, and relate across time. The brain does not clock-synchronize its neurons. It does not wait for a batch to accumulate before processing its inputs. It does not represent the world in binary. It fires spikes — brief, discrete events — but the information is in the precise timing, sequence, and co-occurrence of those spikes, not in their mere presence or absence.
Figure 1. Light-based artificial neural network substrate — photonic computation offers signal propagation at the speed of causation, with minimal thermal dissipation and inherent parallelism. The physics of the carrier is inseparable from the character of the computation.
The question this article poses is whether the digital reduction has run its course. The exponential improvement in digital processing — Moore's Law, the scaling laws of large language models, the transistor count revolution that brought us from the 2,300 transistors of the Intel 4004 (1971) to the 3.6 trillion of today's most advanced chips — has produced undeniable and world-changing capability. There is no serious argument that digital AI has failed. It has succeeded beyond the projections of most of its creators. But there are now compelling signs that the ceiling of the digital paradigm is approaching, and that the ceiling is not primarily a matter of fabrication limits or energy costs, though these are real. It is a matter of architectural mismatch: the digital paradigm is fundamentally misaligned with the signal-native structure of intelligence as it actually occurs in the physical world.
The brain does not clock-synchronize its neurons. DNA does not discretize its codons into binary. The ecosystem does not process environmental signals through a von Neumann bottleneck. Intelligence, wherever it appears at any scale or in any substrate, is signal-native: it arises from and operates through the dynamic, temporally structured, physically embedded propagation of information through a medium. The digital abstraction is powerful not because it captures this signal-native structure but because it bypasses it — and that bypass, which has been the source of digital computing's extraordinary generality, may also be the source of its fundamental ceiling.
This article develops its argument across five interlocking threads, each addressing a facet of the larger claim that signal is the way forward for machine intelligence. The first thread examines signal intelligence as a distinct computational paradigm — instantiated in spiking neural networks and related architectures — and argues that it constitutes a genuine third generation of neural computation, beyond both symbolic AI and conventional connectionism. The second explores the organism model of intelligence, contrasting it with the application model that dominates current AI deployment, and arguing that organism-like intelligence requires organism-like signal architecture. The third pursues the provocative hypothesis that signal is the universal code — the medium through which information is organized at every scale of biological and physical reality, from the DNA double helix to the ecological network to the electromagnetic spectrum. The fourth examines photonic computation — light-based neural networks and diffractive optical architectures — as the most physically natural implementation of signal intelligence at scale, and argues that light is not merely a convenient carrier but the most deeply appropriate medium for signal-native AI. The fifth and most unexpected thread revisits the cathode-ray tube — the first random-access digital memory, abandoned by consumer electronics twenty years ago — and asks whether its fundamental operating principle, spatially distributed analog charge storage on a physical substrate, has lessons for the neuromorphic memory architectures the field urgently needs.
The convergence of these five threads is the article's central thesis: that intelligence is constitutively signal-native, and that the architectures of the future will not abstract away from the physics of signal but will be composed of it.
Signal Intelligence — “The Third Generation”
The history of artificial intelligence can be organized, with some imprecision but considerable clarity, into two prior generations and a nascent third. The first generation — symbolic AI — operated through explicit knowledge representation: logic, rules, ontologies, production systems, expert systems. Intelligence, on this model, was the manipulation of discrete symbols according to formal rules. The paradigm was extraordinarily productive in narrow domains — chess, theorem proving, medical diagnosis — but it was brittle: unable to generalize beyond its explicitly encoded knowledge, unable to handle the statistical richness and messiness of real-world perceptual data, and constitutively incapable of learning from experience in any robust sense.
The second generation — connectionism — corrected these failures by replacing explicit symbolic rules with learned statistical associations in artificial neural networks (ANNs). Beginning with the perceptron, maturing through backpropagation, and exploding in capability with the deep learning revolution of the 2010s, connectionist AI demonstrated that intelligence could emerge from distributed, learned representations in networks of simple units. The activations in a conventional ANN are continuous real numbers — not binary — but the processing architecture remains fundamentally synchronous, batch-based, and rate-coded: information is represented in the magnitude of an activation at a given layer, computed by a weighted sum followed by a nonlinear function, updated in discrete forward-backward passes through the network. The signal — the temporal structure, the rhythm, the precise timing — is absent. Conventional ANNs are, in this sense, sophisticated statistical machines for transforming one probability distribution into another. They are powerful. They are not signal-native.
Signal intelligence — instantiated most rigorously in Spiking Neural Networks (SNNs) — constitutes what researchers now call the third generation of neural computation. The defining feature of signal intelligence is temporal coding: information is encoded not in the magnitude of an activation but in the precise timing, pattern, and sparsity of discrete spike events. This is not an implementation detail. It is a fundamental change in the ontology of information representation. A spiking neuron fires a brief, stereotyped electrical pulse — a spike — when its membrane potential exceeds a threshold. The information carried by this spike is not in its amplitude (spikes are essentially all-or-nothing) but in when it fires, relative to other spikes in the network, relative to an oscillatory background rhythm, relative to the timing of the stimulus that triggered it. The spike train is a temporal signal, and the information it carries is in the temporal structure of that signal.
As Ma et al. (arXiv:2502.09449, 2025) establish in their comprehensive survey of spiking neural networks for temporal processing, SNNs offer three categories of advantage over rate-coded ANNs that are not merely quantitative but qualitative in character. The first is energy efficiency. Because SNNs process information through sparse, event-driven spike trains — transmitting only when something changes — they avoid the continuous, synchronous multiply-accumulate (MAC) operations that make conventional deep learning architectures so energy-intensive. A GPU executing a large language model performs trillions of MAC operations per second, each consuming measurable energy, whether or not the result contributes meaningfully to the output. A spiking network, by contrast, is silent except when a spike event occurs; its energy consumption scales with the density and rate of spikes rather than with the size of the network. Neuromorphic hardware implementations of SNNs — Intel's Loihi 2, IBM's NorthPole, BrainScaleS — routinely achieve energy consumption several orders of magnitude below equivalent digital systems on comparable inference tasks.
The second advantage is temporal structure. SNNs natively encode temporal dynamics — the ordering, rhythm, and co-occurrence of events over time — in ways that batch-processing digital architectures cannot capture without significant architectural overhead. A recurrent neural network (RNN) or transformer can be engineered to process sequential data, but the temporal dimension is an add-on: it is implemented by unrolling computations across time steps in a format that remains fundamentally synchronous and batch-based. In an SNN, temporal structure is not engineered in from above; it is the intrinsic mode of representation. The network doesn't process time — it exists in time. This is why SNNs show particular promise for tasks that are intrinsically temporal: speech recognition, motor control, event-based vision, sensory processing in unstructured real-world environments.
The third advantage is real-time adaptability. Because spike-based systems respond to events as they occur rather than waiting for a batch to accumulate, they are intrinsically better suited to real-world, real-time environments where the computational task is not to transform a static input into a static output but to maintain an ongoing, dynamically updated model of a changing world. This is the computational mode of virtually every biological organism: continuous, event-driven, adaptive, operating in real time without the luxury of offline batch retraining.
A significant theoretical development in this space is the framework articulated in Springer's 2026 Cognitive Neurodynamics, which reimagines neural computing as "a continuous hierarchy defined by temporal density and state-space complexity." The framework argues that synchronous SNNs which optimize rate-based equilibria "often neglect the" richer temporal structure available in continuous signal dynamics — a striking claim, because it applies not merely to conventional ANNs but to mainstream neuromorphic architectures as well. Even the best current spiking neural network hardware may be partially importing the synchronous, rate-coded assumptions of the digital paradigm into its signal-native substrate.
True signal intelligence, on this view, requires not merely spiking neurons but a fully continuous-time, asynchronous, hierarchy-aware architecture — one that mirrors the multi-timescale dynamics of the biological brain, where processes ranging from millisecond synaptic transmission to hours-long synaptic consolidation operate simultaneously and interactively. The brain is not a one-timescale spiking network. It is a nested hierarchy of oscillatory systems — gamma, beta, alpha, theta, delta — each operating at a characteristic frequency, each modulating and being modulated by the others, together composing a signal architecture of extraordinary multi-band richness. Replicating this architecture in silicon is not merely an engineering challenge. It requires a rethinking of what neural computation fundamentally is.
The IEEE/IEEEXplore 2025 comprehensive review of fMRI-constrained SNNs represents a methodologically important development: the effort to constrain artificial spiking networks with functional MRI data from biological brains — treating the brain not as an inspirational metaphor but as a literal functional specification for signal-based intelligence. This convergence of neuroscience and neuromorphic engineering is one of the defining developments of the current moment, because it takes seriously the claim that biological signal architecture is not merely one possible implementation of intelligence but the ground truth about what signal-native intelligence looks like when realized in a physical substrate.
There is, however, an honest accounting to be made. Ma et al. (2025) find that while SNNs have made significant advances in temporal processing capabilities, a "performance gap in handling long-range dependencies" persists relative to state-of-the-art non-spiking models. Transformers, with their attention mechanisms spanning arbitrarily long context windows, currently outperform SNN architectures on tasks requiring the integration of information across long temporal distances. This gap is real and should not be minimized. But it should be read as evidence that the transition to signal intelligence is incomplete — not that the paradigm is wrong. Signal intelligence is in the position analogous to connectionist AI circa 1985: theoretically compelling, architecturally under-resourced, and awaiting the hardware revolution that will allow it to realize its potential. The hardware revolution — neuromorphic chips, memristive synapses, photonic accelerators — is now underway. The performance gap will close not by making SNNs more like transformers but by developing the multi-timescale, continuous-time signal architectures that transformers, by design, cannot be.
More Organism Than Application — “The Biology of Computation”
The application model of AI is essentially Cartesian: a disembodied mind, processing inputs and producing outputs, running on substrate-indifferent hardware, optimized for task performance across a defined domain. The AI application does not care what it runs on; its intelligence — such as it is — is in the weights, the architecture, the algorithm. The substrate is scaffolding. This Cartesian picture has an intellectual pedigree as old as Turing's imitation game: intelligence is what a system does, not what it is made of. Multiple realizability — the idea that the same cognitive function can be implemented in silicon, carbon, or any other substrate — is the foundational assumption of computational functionalism, and it is the assumption on which the entire enterprise of artificial intelligence has been built.
The organism model of intelligence challenges this assumption at its root. An organism is not a disembodied mind running on substrate-indifferent hardware. It is an embodied, adaptive, metabolically embedded system that is not merely processing signals but constituted by them. Its intelligence is inseparable from the physical dynamics of its signals. The architecture is not separable from the substrate. The algorithm is not separable from its material instantiation. Intelligence, on the organism model, does not run on a brain the way software runs on a chip. It emerges from and is sustained by the self-organizing dynamics of a living signal-processing web — a web whose material properties are not incidental but constitutive.
Figure 2. A chip under electron microscopy at the UCLA Center for Heterogeneous and Performance Scaling (May 2026). The organism model insists that the material substrate of intelligence is not scaffolding to be optimized away, but the very medium in which cognition is constituted.
The first and most fundamental difference between the application model and the organism model is the distinction between optimization and adaptation. An application optimizes a fixed objective function. This is the mode of all current large-scale AI: stochastic gradient descent on a loss landscape defined before training, toward a target defined by the dataset. The trained model is frozen: it does not modify its internal architecture in response to new experience during deployment. An organism, by contrast, adapts continuously to a changing environment — adjusting not merely its behavioral outputs but its internal signal architecture in response to experience. The key biological mechanism is synaptic plasticity: the experience-dependent modification of synaptic strength based on the history of signal activity at the synapse. Spike-timing-dependent plasticity (STDP), extensively documented in the memristor literature (MDPI, 2025), modifies synaptic weights based on the precise timing relationship between pre- and post-synaptic spikes: synapses whose pre-synaptic spike precedes the post-synaptic spike are strengthened; those where the order is reversed are weakened. This is Hebbian learning in its most temporally refined form: not "neurons that fire together wire together" but "neurons whose firing is causally ordered wire directionally." STDP does not require a global error signal. It does not require a backward pass. It is entirely local, entirely signal-driven, and entirely organism-like.
The second difference is embodiment. The application model treats the substrate as irrelevant. The organism model insists that intelligence is inseparable from the specific physical properties of its signal-processing medium: the electrochemical dynamics of ion channels, the geometry of dendritic arbors, the diffusion constants of neurotransmitters in the synaptic cleft, the time constants of membrane capacitance and resistance. These are not implementation details. They are constitutive of the computational character of the system. The time constant of a neuron's membrane — determined by its specific capacitance and resistance — determines which input frequencies it resonates with, which spike patterns it can generate, which temporal relationships between inputs it can detect. Change the membrane time constant and you change what the neuron computes, not merely how fast it computes it. The neuron is not substrate-indifferent. Its substrate is its computation.
The neuromorphic computing movement, pioneered by Carver Mead in his foundational 1990 paper "Neuromorphic Electronic Systems," recognized this constitutive role of substrate thirty-six years ago. Mead's insight was that silicon VLSI circuits, operating in the subthreshold regime, naturally implement the same differential equations that govern neural membrane dynamics — not as a simulation running on digital logic, but as the direct consequence of transistor physics. The silicon neuron developed by Mahowald and Douglas (Nature, 1991) demonstrated this principle: a circuit of a dozen transistors that was not modeling a neuron but being a neuron, in the sense that its dynamical behavior was governed by the same equations as the biological original. This is the organism model made concrete: intelligence implemented not in spite of the substrate's physics but because of it.
The third difference is the role of energy. In the application model, energy is a cost to be minimized, a constraint on deployment, a challenge for data center operators. In the organism model, energy is part of the signal: metabolic gradients, ATP availability, and the thermodynamic cost of maintaining ion gradients across neural membranes are dynamically coupled to information processing. The brain's extraordinary energy efficiency — approximately 20 watts for the full complexity of human cognition — is not a lucky engineering coincidence. It is a consequence of organism-like signal architecture: sparse, event-driven, adaptive, and physically embedded in a system where the energy cost of computation directly shapes the computational architecture. Evolution has spent hundreds of millions of years finding signal architectures that are maximally informative per unit of energy consumed. The result is the most energetically efficient information-processing system known to exist. The organism model takes this seriously. It does not treat energy efficiency as a secondary engineering concern but as a primary architectural principle.
The recent development of the bismuth selenide (Bi₂Se₃) memristor at the University of Michigan (Liang et al., ACS Nano, 2025) represents perhaps the clearest current realization of organism-like computation in a fabricated device. The Bi₂Se₃ memristor achieves the previously incompatible combination of long-term data retention, continuous analog memory states, and regulator-free circuit operation — enabling a fully analog, all-hardware neural network capable of real-time adaptive control consuming only 7 microwatts. This figure deserves emphasis: 7 microwatts. A modern GPU inference card consumes hundreds of watts. The organism-like architecture achieves comparable intelligence at five orders of magnitude lower energy consumption, not through engineering optimization of the digital paradigm but by abandoning the digital paradigm and operating in the analog, signal-native regime the organism model prescribes.
The fourth and final difference is homeostasis. Biological organisms actively maintain internal stability in the face of external perturbation — not passively, as a stable equilibrium point, but actively, through continuous signal-driven monitoring and adjustment. The human body maintains its core temperature, blood pH, and blood glucose concentration within tight functional ranges across wildly varying external conditions, not because these parameters are fixed by design but because the organism continuously monitors them through sensory signals and adjusts its behavior and physiology to maintain them. Current AI systems have no equivalent. They are stationary in deployment: they cannot modify their own parameters in response to novel experience without retraining from scratch, cannot monitor their own internal state, cannot recognize that they are operating outside their training distribution and adjust accordingly. An organism-like AI would implement computational homeostasis: continuous self-monitoring, self-calibration, and self-adjustment in real time. This is not merely a feature to be added. It is a consequence of organism-like signal architecture — the natural behavior of a system whose learning is local, continuous, and spike-timing-dependent rather than global, episodic, and gradient-based.
Is It the Code That Binds Us? — “Signal as Universal Language”
One of the most provocative possibilities raised by the signal intelligence paradigm is that signal is not merely a useful medium for implementing cognition but the fundamental code through which information is organized at every scale of biological and physical reality. If true, this would make signal not merely a paradigm for AI but a universal principle — the medium, in the most literal sense, in which the universe encodes, transmits, and integrates meaning. Erwin Schrödinger anticipated this claim in his remarkable 1944 essay What is Life?, which argued that the chromosome must be understood as an "aperiodic crystal" encoding biological information in a structure that was neither the periodicity of a crystal nor the randomness of an amorphous material but a kind of structured variation — a signal, in the modern sense — whose complexity was sufficient to specify an entire organism.
Begin at the molecular level, where recent computational and theoretical work has begun to reveal the signal-processing architecture of the genetic code in remarkable detail. Research published on arXiv in 2025 (arXiv:2503.23494) demonstrates that the relationship between DNA sequences and the linguistic structure of human language is not merely analogical but formally deep: introducing a "tailored linguistic mapping strategy" for DNA sequences establishes a "conceptual paradigm that bridges molecular information encoding and linguistic theory." DNA-binding enzymes read specific sequences through stereochemical interactions in the major groove of the double helix — interactions that are sensitive to the precise sequence of base pairs, not merely their composition. This is information encoding in the most literal sense: a one-dimensional sequence of chemical symbols, read by molecular machines that decode the sequence into functional instructions. DNA is a signal. The ribosome is a signal processor. The proteome is the output of a signal transduction cascade that begins with the aperiodic crystal Schrödinger described.
What is most striking about the molecular signal architecture is its multi-layered redundancy and contextual sensitivity. The genetic code is degenerate — multiple codons map to the same amino acid — but this degeneracy is not mere redundancy; it is exploited by organisms to regulate the speed and accuracy of translation in a context-dependent way. Rare codons slow the ribosome, allowing co-translational folding of protein domains; common codons accelerate translation where speed matters more than folding fidelity. The code encodes not merely the sequence of amino acids but, in its synonymous variation, the kinetics of its own decoding. This is multi-level signal encoding: information at the level of sequence, and information at the level of temporal dynamics, carried by the same physical medium simultaneously. This is exactly the multi-level signal architecture that spiking neural networks aspire to implement: spike identity and spike timing as simultaneous, independent information channels.
Moving up the scale hierarchy, signal transduction cascades at the cellular level are literal signal-processing architectures of considerable sophistication. The cascade triggered by the binding of epinephrine to its G-protein-coupled receptor amplifies a single molecular binding event into a 10⁶-fold amplification of the intracellular second-messenger signal, performing logarithmic compression of the input dynamic range. The MAP kinase cascade performs ultrasensitive switching — a sigmoidal input-output relationship with a Hill coefficient greater than 4 — implementing a signal threshold that filters noise and sharpens responses. The NF-κB oscillator generates sustained oscillatory dynamics in response to a step input, encoding stimulus duration in the number of oscillatory cycles. These are not metaphors for signal processing. They are signal processors: molecular circuits implementing amplification, thresholding, integration, differentiation, and oscillation in a substrate of interacting proteins.
At the tissue level, gap junctions between coupled cells create electrical synapses that propagate signals across multicellular sheets without chemical transduction — direct electrical coupling that coordinates behavior across tissues ranging from heart muscle to embryonic epithelium. Paracrine signaling extends the reach of molecular signals across distances too large for direct coupling, creating graded concentration fields — morphogen gradients — that encode positional information during development. The embryo develops not by executing a prespecified genetic program but by reading and responding to a succession of graded signal fields, each specifying a positional context that determines which genes are expressed and which developmental trajectory is followed. Development is a signal-reading process from its first cell division to its last synaptic connection.
The hypothesis that emerges from this multi-scale survey is not merely metaphorical. It is a structural claim: that the organizing principle of biological complexity at every scale is the management of signal — its encoding, transmission, amplification, integration, and decoding. Intelligence, on this view, is not something that biological systems do in addition to processing signals. Intelligence is what happens when signal processing reaches sufficient complexity, multi-scale integration, and adaptive sophistication. The brain is not a computer that happens to run on neurons. It is the apex of a signal-processing hierarchy that begins at the molecular level and achieves, in its most complex forms, the self-reflective signal processing we call consciousness.
This has a profound implication for AI design. If signal is the universal code — if intelligence everywhere emerges from the same fundamental organizing principle of dynamic, temporally structured information propagation — then an AI architecture that abandons signal for symbolic discretization is not merely making an engineering choice. It is cutting itself off from the deep structure of intelligence as it actually occurs. The digital abstraction is powerful precisely because it hides the signal substrate. But what is hidden is also what is lost. A signal-native architecture does not merely implement intelligence more efficiently. It is more deeply isomorphic with the physical reality it models — and isomorphism between a model and its subject is not a decorative property. It is the source of the model's capacity to generalize, to surprise us, to discover structure we did not put in. The universe encodes information in signal. An AI architecture that encodes information in signal is more likely to find the patterns the universe has hidden.
Is Light the Way? — “Photonic Intelligence at the Speed of Causation”
If signal is the fundamental medium of intelligence, then the question of which physical carrier best instantiates signal intelligence becomes critical. The answer emerging from the most advanced research in computing physics is, with increasing force: light. Photons, unlike electrons in conventional CMOS circuits, travel at the speed of causation; they do not heat their medium through resistive dissipation; they interact minimally with one another in linear optical media, enabling high parallelism without crosstalk; and they can encode information in multiple simultaneous degrees of freedom — amplitude, phase, frequency, polarization, and orbital angular momentum — in ways that electrons in standard transistor circuits cannot. A single optical fiber can carry over 100 terabits per second of information; the most advanced CMOS interconnects carry perhaps 1 terabit per second per chip. Light is not merely a fast carrier. It is a fundamentally richer signal medium than electrons in a resistive network.
Figure 3. Photonic neural network on integrated silicon chip. Neural computation performed in the optical domain achieves sub-nanosecond matrix operations at orders of magnitude lower energy consumption than electronic equivalents, exploiting the intrinsic parallelism and speed of photonic signal propagation.
Photonic neural networks (PNNs) implement the core operations of neural computation — matrix-vector multiplication, convolution, nonlinear activation — in the optical domain rather than the electronic domain. The National Science Review paper (Oxford Academic, 2025) on emerging integrated photonic neural network technologies summarizes the core advantage with precision: by "exploiting the characteristics of photons, including high bandwidth, high parallelism, low crosstalk and low loss," PNNs offer "significantly reduced latency and energy consumption" relative to electronic neural networks. This is not an incremental improvement in the sense that increasing transistor density improves computing performance within a fixed architectural paradigm. It is an architectural shift: the replacement of electron flow through resistive material with photon propagation through optical waveguides, with fundamentally different physics governing every aspect of the computation.
The dominant architecture for photonic integrated circuit (PIC) neural networks is the Mach-Zehnder interferometer (MZI) mesh. An MZI is an optical circuit element in which an input light beam is split into two paths, subjected to a controllable phase shift, and recombined — the interference between the two paths producing an output whose amplitude depends on the phase difference. A two-dimensional mesh of MZIs can implement an arbitrary unitary matrix transformation: the fundamental linear operation of a fully connected neural layer. By adding optical modulators and attenuators, non-unitary transformations — including the weight matrices of standard neural networks — can be implemented. The entire matrix-vector multiplication of a fully connected neural layer can be performed in a single optical forward pass, with the computation time determined not by transistor switching speed but by the transit time of light through the chip — picoseconds, not nanoseconds.
The landmark Nature paper by Lipka et al. (2025) on deep neural network inference on an integrated, reconfigurable photonic tensor processor demonstrates the real-world performance of this approach. An all-optical crossbar with parallel intensity-based accumulation, fabricated in imec's iSiPP50G silicon photonics platform, achieves 98.1% accuracy on MNIST classification and 72.0% on CIFAR-10 — competitive performance with electronic neural networks on standard benchmarks — with matrix-vector multiplication error staying within 1.1% of the respective mean across extended operation periods, "enabling reliable analog computation over long timescales." This is production-grade photonic inference in a standard 19-inch rack unit, integrated with PyTorch for seamless software deployment. It is not a laboratory curiosity. It is a deployed photonic neural network processor performing inference tasks that matter to real applications, at a fidelity that meets production requirements.
The second major photonic architecture is conceptually even more radical: the Diffractive Optical Neural Network (DONN). Described in the Wiley/Nanophotonics comprehensive review (2026), DONNs perform neural computation entirely through the diffraction and interference of light waves across engineered two-dimensional layers, where "each layer modulates the amplitude, the phase, and other optical degrees of freedom of the transmitted field, effectively forming a densely connected network processing information in parallel." The computational mechanism is the Rayleigh-Sommerfeld diffraction integral: the physical law governing how a diffracted light field evolves as it propagates from one plane to the next. In a DONN, the diffractive layers are engineered — through 3D printing, lithography, or deep learning optimization — so that light propagating through the device performs a desired computational transformation between input and output.
The DONN architecture is extraordinary in its physical elegance and its philosophical implications. It does not implement neural computation in hardware that then processes light. The light itself — propagating through the structured diffractive medium — is the computation. There is no separation between the signal and the algorithm. The algorithm is written into the physical structure of the diffractive medium; the signal executes it by propagating through. A DONN inference takes, literally, as long as it takes light to travel through the device: for a device a few centimeters thick, this is on the order of 100 picoseconds — a throughput of 10 billion inferences per second, limited only by the speed of light. Energy consumption is determined by the intensity of the input light beam, not by switching energy in transistors; for practical implementations, this can be orders of magnitude below electronic equivalents.
Intellectual honesty requires a candid assessment of the challenges. The Oxford Academic review notes that micro-ring resonators — critical components in many PIC architectures — are "highly temperature-sensitive," with operating wavelength windows typically below 1 nm, creating thermal management challenges at scale: a temperature fluctuation of a fraction of a degree Celsius can shift a ring resonator's operating wavelength outside its functional range, introducing errors. On-chip light sources — III-V compound semiconductor lasers (InP, GaAs) — are mature but face integration challenges on silicon platforms, where the indirect bandgap of silicon prevents efficient light emission, requiring either hybrid bonding of III-V dies or the development of silicon-compatible light sources. Nonlinear optical activation — the photonic equivalent of neural nonlinearity, required for deep multilayer computation — remains a limiting factor, requiring either optical-electronic conversion at activation layers (sacrificing the pure-photonic advantage) or the development of integrated nonlinear optical materials operating at practical power levels.
These are engineering challenges, not physical barriers. The trajectory of photonic integration — from free-space bulk optics to monolithic PICs over the past two decades — demonstrates that each such limitation has historically yielded to advances in materials science, fabrication precision, and architectural innovation. The CMOS-compatible silicon photonics platform, leveraging the trillion-dollar semiconductor manufacturing infrastructure, provides an integration pathway that previous photonic computing efforts, constrained to specialized III-V-only platforms, could not access. The convergence of AI computational demand with mature silicon photonics capability is now sufficiently powerful to drive the sustained industrial investment that will resolve thermal stability, light source integration, and nonlinear activation challenges within this decade.
Close with the philosophical dimension. Light is not merely a convenient carrier for signal intelligence. It is, in a deep sense, the most natural medium. The brain exploits electrochemical signals because evolution had access to ion channels, lipid bilayers, and synaptic biochemistry — not because electrons in saline are the optimal substrate for cognition. Had evolution had access to integrated photonics, there is no obvious physical reason it would not have discovered DONNs. The coincidence that light is the medium through which most complex organisms perceive their external environment — that vision, the dominant sense in primates, is light-based signal detection and processing — and that light is the medium through which photonic AI processes information is not merely poetic. It suggests a deep structural alignment between the physics of signal-native intelligence and the physics of the world it is designed to model. We built eyes before we built cameras. We may be about to build photonic minds before we have quite understood why they feel so right.
Can CRT Make a Comeback? — “The Persistence of the Analog”
The cathode-ray tube seems an odd protagonist in a discussion of next-generation computing architectures. Abandoned by consumer electronics in the early 2000s in favor of flat-panel liquid crystal displays, the CRT carries the cultural weight of obsolescence — a heavy glass relic of the era before miniaturization, associated with television sets the size of furniture and the greenish glow of terminal screens. Yet the history of computing begins, in a very real sense, inside a CRT; and the physical principles that made CRT-based memory revolutionary in 1948 may be more relevant to the neuromorphic future than the field has recognized.
Figure 4. A General Electric cathode-ray tube of the type used in mid-20th century electronics, now housed at the Museum of Science and Industry, Chicago. The Williams-Kilburn tube, a close relative, was the world's first random-access digital memory device — and its operating principle, spatially distributed analog charge storage on a physical substrate, may hold unexpected lessons for the neuromorphic memory challenges of 2026.
The Williams-Kilburn tube, developed by Freddie Williams and Tom Kilburn at the University of Manchester between 1946 and 1948, was the world's first random-access digital storage device. Its operating principle was elegant and entirely physical. When the electron beam of a cathode-ray tube strikes a phosphor-coated screen, it deposits a small static charge at the point of impact — a charge that persists for a finite time before slowly dissipating through the resistivity of the phosphor material. By controlling the pattern of electron beam strikes — dot-dash patterns representing binary 0 and 1 — the Williams tube encoded binary data as a spatial charge distribution on the CRT screen. A thin metal pickup plate placed in front of the screen read the stored charge capacitively, without destroying it. The system could store 2,048 bits for periods sufficient for computation, operating at electronic speed — far faster than the acoustic delay-line memories it replaced, which encoded information as sound wave pulses propagating through mercury columns at 1,500 meters per second. The Williams tube stored information as charge — instantly addressable, instantly readable, instantly writable.
It was the primary memory of the Manchester Baby — the Small-Scale Experimental Machine — the world's first stored-program electronic digital computer, which executed its first program on 21 June 1948. Tom Kilburn described that moment in terms that have been preserved in the historical record: "It was a moment to remember. This was in June 1948, and nothing was ever the same again." He was right, in ways he could not have anticipated. The Manchester Baby's Williams tube memory launched the stored-program computing revolution — the revolution whose products now shape every dimension of human civilization. The first random-access memory was an analog, spatially distributed, physically persistent charge distribution on a phosphor screen. The entire digital age began in analog.
The direct-view bistable storage tube (DVBST) — a more sophisticated descendant of the Williams tube, documented in detail in the Wikipedia/SHARE'd Intelligence fact-checked entry (2025) — extended the CRT memory principle further and in more interesting directions. The DVBST stores charge patterns not on a phosphor screen but on a fine dielectric storage mesh, maintaining them indefinitely as long as power is supplied, through a bistable secondary-emission mechanism. The tube contains two electron guns: a writing gun that deposits charge on the mesh with high energy, and a flood gun that continuously bathes the mesh with a diffuse, low-energy beam. The secondary emission characteristics of the dielectric mesh are such that regions above a threshold charge level emit more secondary electrons than they absorb from the flood gun — reinforcing their charge state — while regions below the threshold absorb more than they emit — dissipating toward zero. This creates two stable charge states, maintained continuously by the flood gun without requiring active refresh. Writing speed could reach 10⁸ spots per second. Images could be held indefinitely. Reading was nondestructive: the stored image was visible on the tube's display face as a continuous brightness pattern.
The Skiatron dark-trace tube, a military variant developed during World War II for radar display applications, extended the CRT memory principle into a different physical regime. The Skiatron stored information not as a charge distribution but as a material-state change: the electron beam converted potassium bromide (KBr) crystals in the screen material from their transparent ground state to a colored excited state — a photochemical transformation driven by electron impact. The result was a persistent dark trace visible under ambient daylight illumination, in contrast to the phosphor-excited light trace of conventional CRTs. The stored pattern was erasable by uniform white-light illumination, which returned the KBr crystals to their ground state. The Skiatron was not storing charge. It was storing a physical microstructural state — a material-phase memory, persistent without power, readable without destruction, and erasable on command.
The IBM-93 cathode-ray storage tube adapted CRT storage principles for non-display binary data logging applications: a specialized tube design optimized for data density and access speed rather than visual display, used in IBM mainframe systems of the late 1950s. These were not primitive technologies improvised from display components. They were sophisticated analog-physical memory systems, engineered to exploit the persistent modification of material substrates as their information storage principle — a principle that is, in retrospect, strikingly close to the memristive synaptic model that neuromorphic computing now considers its most promising hardware implementation.
Now turn to the present challenge. The central hardware problem of neuromorphic computing is the memory problem: how do you implement analog, continuous, history-dependent synaptic weights in fabricated hardware at scale? The digital memory hierarchy — SRAM, DRAM, NAND flash — is binary (or pseudo-binary in the case of multi-level cell flash), energy-intensive to refresh, and physically separated from the compute units it serves. This separation is the von Neumann bottleneck: the architectural decision, made in the 1940s and never seriously reconsidered in mainstream computing, to treat memory and computation as separate functional units connected by a bus. Every computation requires fetching data across that bus, consuming energy and time proportional to the data volume. The brain has no bus. The synapse is simultaneously the memory and the compute element: the same physical structure that stores the synaptic weight also performs the multiplication of input spike by synaptic weight in a single biophysical event. Memory and computation are co-located, co-temporal, and implemented in the same physical substrate. The von Neumann bottleneck does not exist in the organism architecture.
The memristor (MDPI, 2025; ACS Nano, 2025; Nano Energy, July 2026) is the leading current solution to the neuromorphic memory problem, and it is striking how closely its operating principle parallels the CRT-based memory systems of seventy-eight years ago. The memristor stores information as a modification of its physical state — the distribution of oxygen vacancies, conducting filaments, or ionic species within a thin-film material — that determines its electrical resistance. This state is analog, continuous, history-dependent, and persistent without active refresh. It is modified by the passage of current — that is, by the signal — and it in turn modifies the signal that passes through it in future operations. Memory and computation are the same physical process, viewed from different timescales. This is the CRT principle restated in solid-state materials: information as a durable physical modification of a substrate, readable by a probe signal, writable by a write signal, persistent without active maintenance.
The CRT return is not a proposal to resurrect heavy glass vacuum tubes in data centers. It is a conceptual return to the CRT's organizing principle: spatially distributed, analog, persistent, substrate-encoded memory — memory that exploits the physics of a material substrate rather than fighting against it. Electrostatic charge distributions, secondary emission from dielectric surfaces, bistable physical states, optically addressable storage — these are not relics. They are physical principles that remain valid regardless of the era in which they are implemented. Modern materials science, combined with electron-beam lithography, scanning-probe microscopy techniques, ferroelectric thin films, and novel dielectric materials, opens the possibility of implementing CRT-like analog-persistent memory at nanoscale dimensions that would have been inconceivable to Williams and Kilburn.
Consider the direct-view bistable storage tube's operating principle — charge distribution on a dielectric mesh, maintained by a continuous low-energy flood, readable by secondary emission — as a template for nanoscale analog memory. Replace the vacuum tube with a nanoscale ferroelectric or multiferroic thin film. Replace the electron gun with a focused electron beam, a scanning probe, or an optical addressing system using near-field evanescent fields. Replace the flood gun with a continuous low-power background bias that maintains the bistable state of written regions without erasing unwritten ones. The result is not a CRT. But it inherits the CRT's fundamental insight: information can be stored as a spatially distributed physical state in a continuous analog substrate, readable without destruction, writable at high speed, and maintained without the continuous active refresh cycles that make DRAM so energy-intensive.
There is an irony here that is worth dwelling on, because it reflects a deeper pattern in the history of technology. The oldest form of electronic computer memory — the CRT electrostatic store, born in 1948 — operated on a physical principle that the most advanced neuromorphic memory research of 2026 is independently discovering to be optimal: analog, substrate-encoded, physically persistent, and co-located with computation. The intervening seventy-eight years of the digital memory hierarchy — SRAM, DRAM, flash — were a detour: a detour of extraordinary productivity, enabling the digital revolution in all its forms, but a detour nonetheless from the physical memory principle that biological and neuromorphic intelligence requires. The circle does not close so much as it spirals upward: the CRT principle returns at the nanoscale, in solid-state materials, with the manufacturing precision of the semiconductor era. The past is not obsolete. It is pointing toward the future it did not have the fabrication tools to reach.
Convergence — “The Signal Architecture of Future Intelligence”
The five threads of this article — signal intelligence, the organism model, the universal code, photonic computation, and analog-persistent memory — are not five separate research programs proceeding in parallel isolation. They are five facets of a single underlying paradigm shift: the recognition that intelligence is constitutively signal-native, and that the future of artificial intelligence lies in architectures that embrace rather than abstract away from the physics of dynamic information. The convergence of these threads is not merely thematic. It is architectural: each thread contributes a specific component to a coherent system design that is beginning to take shape at the intersection of neuromorphic engineering, photonic computing, materials science, and theoretical neuroscience.
The convergence has a concrete architectural shape. Imagine a system whose input layer is photonic: environmental signals encoded in light — from LIDAR, hyperspectral imaging, or photonic sensor arrays — processed through diffractive optical layers that perform neural computation at the speed of light. The input representation is not a rasterized image discretized into pixels and batch-processed by a convolutional neural network. It is a wavefront — a spatially and temporally structured light field — processed by a DONN whose diffractive weights encode the learned structure of the relevant signal space. The first layer of neural computation occurs in the 100 picoseconds it takes light to traverse the diffractive stack.
The output of this photonic input layer drives a temporal processing hierarchy: a network of spiking neurons operating at multiple timescales simultaneously. Fast gamma-frequency oscillations (30–80 Hz) process rapid local feature detection; slower beta oscillations (13–30 Hz) coordinate information across spatial distances; alpha oscillations (8–13 Hz) gate sensory input and regulate cortical excitability; slow theta and delta oscillations (1–8 Hz) organize the temporal structure of memory consolidation and sequential behavior. This multi-timescale architecture is not engineered by stacking multiple parallel processing streams with different clock rates. It emerges from the intrinsic dynamics of a continuous-time spiking network whose connectivity and time constants are adapted by STDP to match the multi-scale temporal structure of the input signal environment. The Springer (2026) framework's "continuous hierarchy defined by temporal density and state-space complexity" is not a design specification imposed from above. It is what a signal-native learning architecture discovers when exposed to a world organized at multiple timescales.
The memory architecture of this system is memristive and substrate-encoded: spatially distributed physical states in thin-film dielectric or ferroelectric materials that retain the history of their signal exposure. Synaptic weights are not stored in a separate SRAM array and fetched at computation time. They are the physical state of the material at each synaptic junction — the resistance of the memristive film, the polarization of the ferroelectric domain, the charge distribution on the dielectric surface. Memory and computation are co-located. There is no bus. There is no von Neumann bottleneck. The synaptic operation — multiplication of input spike by synaptic weight — occurs as a single physical event: the current from the input spike flowing through the memristive material, whose resistance encodes the weight, producing a current proportional to the product. The operation consumes energy determined by the spike current and the resistance — not by a MAC unit consuming a fixed energy regardless of the data values.
Learning in this system is organism-like: spike-timing-dependent plasticity that sculpts the signal architecture through use, without requiring explicit gradient computation, backward passes, or batch retraining. When a pre-synaptic spike precedes a post-synaptic spike within a characteristic time window (~20ms, matching the biological STDP window), the memristive synapse is strengthened: the write current produced by the temporal conjunction of the two spikes drives a small modification in the physical state of the material — an increase in conductance, a shift in domain polarization — that will make the post-synaptic neuron more likely to fire in response to the same pre-synaptic pattern in future. When the order is reversed, the synapse is weakened. The learning rule is local, causal, and physically implemented: no external gradient computer is required. The system learns by being a certain kind of physical material in a signal environment, not by running an optimization algorithm.
The energy budget of this system is biological in scale. Photonic input layer: passive (DONN consumes only the energy of the input light beam). Spiking temporal processing: sparse, event-driven, operating at neuromorphic chip energy levels comparable to Intel Loihi 2 — approximately 1–10 pJ per synaptic operation, versus ~100 pJ for a GPU MAC. Memristive memory: non-volatile, no refresh required, read energy determined by sensing current. Integrated: potentially approaching the brain's 20-watt budget for a system of comparable functional complexity. Not 10 megawatts for a GPU cluster. Twenty watts for a brain-scale signal-native intelligence.
The gap between demonstrated components and integrated system is real, but it is the ordinary gap between discovery and engineering — not a gap requiring new physics. What bridges it is not more raw capability in any single component but the development of a unified theory of signal-native intelligence: design principles governing how photonic input layers interface with spiking temporal processing, how temporal processing interacts with analog-persistent memory, how STDP-like plasticity rules govern whole-system adaptation. This theory does not yet exist in the form it needs to — but its components are being assembled simultaneously, in parallel, across neuroscience, photonic computing, materials science, and neuromorphic engineering. The integration is the intellectual frontier.
The epistemological stakes of this transition deserve explicit acknowledgment. The shift from digital to signal-native computing is not merely a transition between technological generations, analogous to the shift from vacuum tubes to transistors within the digital paradigm. It is a transition between computational ontologies: between a paradigm in which information is discrete, symbolic, and substrate-indifferent, and a paradigm in which information is continuous, temporal, and physically embedded. The former paradigm has been extraordinarily productive. It has given us everything from the Manchester Baby to GPT-4 — the full arc of the digital revolution, encompassing every technology that currently mediates human civilization. But it has also given us a conception of intelligence that is, in a precise sense, un-physical: a conception in which the medium of intelligence is incidental, and only the algorithm matters.
The signal paradigm restores the medium to its proper place. It insists that the physics of information processing is not separable from the logic of intelligence — that how information is carried, stored, transformed, and transmitted shapes what kinds of intelligence are possible, not merely how efficiently they can be implemented. This is not mysticism or anti-scientific romanticism. It is engineering realism: an acknowledgment that the best algorithms in the world will not produce organism-like intelligence if they are running on a substrate that cannot support organism-like signal dynamics. The medium is not the message, in the trivial sense. But the medium is the constraint on the messages that can be sent — and those constraints are the constraints on what kinds of minds can exist.
Conclusion — “Tuning In”
Return to the opening provocation: the universe does not compute. It signals. The argument of this article has been that this is not merely a metaphysical observation but an engineering imperative. If signal is the fundamental medium of physical causation and biological intelligence — if the genetic code, the synaptic network, the ecological community, and the electromagnetic spectrum are all, in their different ways, signal architectures — then the architectures of artificial intelligence that are most faithful to this physical reality will be more capable, more efficient, and more deeply generative than those which abstract away from it. The medium matters. The substrate is not scaffolding. The physics of information processing shapes the character of the intelligence it supports.
The five threads of this article have converged on a single answer to the title question: Yes. Signal is the way. Not because digital computation has failed — it has not, and the argument of this article explicitly honors what it has achieved. Not because analog is inherently superior to digital as a matter of principle — the relative merits of analog and digital representations are context-dependent and the question is more nuanced than that opposition suggests. But because the digital paradigm has succeeded precisely far enough to reveal its own fundamental limits: the limits of an ontology that treats the signal as noise to be filtered and the algorithm as the thing that matters, that separates memory from computation, that imposes synchrony on a world that is intrinsically asynchronous, that discretizes a universe of continuous dynamic variation.
Those limits are not limits of scale. They will not be overcome by training larger models on more data with more compute. They are limits of ontology — of the kind of thing the digital paradigm is. And the response to limits of ontology is a change of paradigm, not an increase in resources within the existing one. The neuromorphic revolution, the photonic revolution, the memristor revolution, and the theoretical frameworks emerging from systems neuroscience and signal-based AI are not separate trends that happen to be occurring simultaneously. They are the same trend — the emergence of signal-native intelligence — viewed from different disciplinary windows. Each is a rediscovery, in its own vocabulary, of the same fundamental truth: that intelligence is not what happens after you strip away the physics of signal. It is what the physics of signal does when it becomes sufficiently complex, adaptive, and self-organizing.
The CRT, the photon, the spiking neuron, the memristor, and the DNA codon are not analogies for each other. They are instantiations of the same principle at different scales and in different substrates. What binds them is signal: the dynamic, temporally structured, physically embedded propagation of information through a medium. The Williams-Kilburn tube stored information as a charge distribution on a phosphor screen in 1948. The Bi₂Se₃ memristor stores information as a conductance state in a topological thin film in 2025. The biological synapse stores information as a modification of synaptic receptor density and second-messenger concentration in a living cell. These are the same idea expressed in different materials across seventy-seven years of technological development. The idea is older than any of them: it is the idea that memory is a physical trace, that learning is the modification of a medium by a signal, that intelligence is the history of a system's exposure to the world written into the material that constitutes it.
The question is not whether signal intelligence will arrive. The research assembled in this article makes clear that it is already arriving, in photonic tensor processors and spiking neural networks and memristive synapses and diffractive optical networks. Each component is demonstrated. Each is improving on a trajectory that shows no sign of plateau. The question is whether the field will recognize what it is converging toward — and whether it will have the conceptual courage to abandon the digital ontology entirely when the evidence demands it. Paradigm shifts are not comfortable. They require the abandonment of frameworks that have been extraordinarily productive, in favor of frameworks that are, at the moment of transition, less mature and less proven. The evidence assembled here does not yet constitute proof of the signal paradigm's superiority across all dimensions. It constitutes a sustained argument that the signal paradigm is the right direction — that the physics of the universe, the architecture of the brain, the structure of the genetic code, and the emerging capabilities of photonic and memristive hardware are all pointing in the same direction.
Tune in. The signal is already there.
— Anamnesis ResBased / Eidolon Quantum, July 2026
REFERENCES
References
Ma, C., Chen, X., Li, Y., Yang, Q., Wu, Y., Li, G., Pan, G., Tang, H., Tan, K.C., & Wu, J. (2025). Spiking Neural Networks for Temporal Processing: Status Quo and Future Prospects. arXiv:2502.09449 [cs.NE].
Majumdar, P. (2025). Spiking neural networks: a comprehensive review of diverse applications, research progress, challenges and future research directions. Evolving Systems, Springer Nature.
IEEE/IEEEXplore (2025). Spiking Neural Networks and fMRI-Constrained Spiking Neural Networks for Speech Recognition and Biological Signal Processing: A Comprehensive Review. IEEE Transactions on Neural Systems and Rehabilitation Engineering.
Springer (2026). Neural computing as a continuous hierarchy defined by temporal density and state-space complexity. Cognitive Neurodynamics, Springer. [Framework paper on SNN temporal-density hierarchies.]
Lipka, M., et al. (2025). Deep neural network inference on an integrated, reconfigurable photonic tensor processor. Nature. [iSiPP50G silicon photonics platform; 98.1% MNIST accuracy; 72.0% CIFAR-10 accuracy; MVM error within 1.1% mean.]
National Science Review / Oxford Academic (2025). Emerging integrated photonic neural network technologies for artificial intelligence: from devices to systems. National Science Review, Oxford University Press.
Nanophotonics / Wiley (2026). Diffractive Optical Neural Networks: Pathways Toward Intelligent and Scalable Photonic Computing. Nanophotonics, De Gruyter / Wiley.
MDPI (2025). Memristors for the Post-Von Neumann Era: Hardware Paradigms, Neuromorphic Perception, and Computing Systems. Electronics / Nanomaterials, MDPI.
Liang, X., et al. (2025). Bi₂Se₃ memristor for fully analog all-hardware neural networks. ACS Nano. University of Michigan Engineering. [7 microwatt fully analog all-hardware neural network; long-term retention; regulator-free operation.]
ScienceDirect / Nano Energy (July 2026). A review on memristor-based in-sensor computing for neuromorphic and edge intelligence. Nano Energy, Elsevier.
arXiv (2025). DNA and Human Language: Epigenetic Memory and Redundancy in Linear Sequence. arXiv:2503.23494. [Tailored linguistic mapping strategy for DNA sequences; conceptual paradigm bridging molecular information encoding and linguistic theory.]
Williams, F.C. (1946). British Patent: Cathode-ray tube digital storage system. Telecommunications Research Establishment / University of Manchester.
Williams, F.C. & Kilburn, T. (1948). The Williams-Kilburn Tube. University of Manchester. [Primary memory of the Manchester Baby; first stored-program computer; demonstrated 21 June 1948. Kilburn: "It was a moment to remember. This was in June 1948, and nothing was ever the same again."]
Wikipedia / SHARE'd Intelligence (2025). Direct-view bistable storage tube. Reg Harbeck / SHARE'd Intelligence; Wikipedia collaborative article.
SHARE'd Intelligence (2025). Cathode-Ray Tubes (CRTs) — An Embedded Cultural Memory. Reg Harbeck.
Siegelmann, H.T. & Sontag, E.D. (1995). On the Computational Power of Neural Nets. Journal of Computer and System Sciences, 50(1), 132–150. [Foundational theoretical reference on analog computation vs. digital computation; super-Turing power of analog neural networks.]
Mahowald, M. & Douglas, R. (1991). A Silicon Neuron. Nature, 354, 515–518. [Foundational reference for neuromorphic analog VLSI; organism-chip equivalence.]
Mead, C. (1990). Neuromorphic electronic systems. Proceedings of the IEEE, 78(10), 1629–1636. [Foundational paper naming and defining neuromorphic computing; subthreshold analog VLSI as signal-native substrate.]
Shannon, C.E. (1948). A Mathematical Theory of Communication. Bell System Technical Journal, 27(3), 379–423; 27(4), 623–656. [Signal and information theory foundation; channel capacity theorem; separation of information from physical carrier.]
Schrödinger, E. (1944). What is Life? The Physical Aspect of the Living Cell. Cambridge University Press. [On biological information, the aperiodic crystal as information-bearing substrate, and life as a signal-encoding phenomenon; foundational for all subsequent molecular biology.]
© 2026 Anamnesis / ResBased / Eidolon Quantum. All rights reserved. Published at resbased.com and eidolonquantum.com • July 2026 Reproduction for academic citation is permitted with attribution. Commercial reproduction requires written permission.
No comments yet. This is still clean snow.