Abstract
Artificial intelligence has spent most of its modern life pretending that time can be placed in a box.
A prompt arrives. It is tokenized. Data becomes numbers. Numbers enter an architecture. An answer emerges. The transaction ends.
The world itself behaves nothing like this.
Light arrives continuously. Pressure changes. Surfaces vibrate. Bodies move. Temperatures drift. Voices overlap. Objects persist while no one is looking at them. Causes unfold over time. Events decay into consequences. Memory changes how later events are interpreted. Organisms do not repeatedly awaken into isolated requests; they remain embedded in an environment whose signals never completely stop.
This raises an uncomfortable question for artificial intelligence:
Have we mistaken a convenient representation of reality for the form intelligence itself should take?
Signal-native approaches offer an alternative. Instead of treating perception primarily as batches of symbols waiting to be processed, a signal-native system would treat change through time as first-class structure. Phase, amplitude, persistence, recurrence, delay, spatial relation, decay and interference would not merely be extracted features. They could participate directly in how state evolves.
Recent research gives the question real technical weight. Photonic accelerators now perform large-scale matrix operations with extraordinarily low latency; integrated photonic systems are beginning to train neural networks on-chip; photonic reservoir computers process high-speed signals directly in the optical domain; analog and memristive compute-in-memory systems increasingly combine storage and computation; and researchers are deliberately returning to physical neural networks in which the dynamics of matter itself perform useful parts of the calculation. Nature
And history contains a wonderful irony. Long before modern semiconductor memory, engineers actually used cathode-ray tubes as electronic memory. The Williams-Kilburn tube stored charge patterns as dots on a CRT surface, watched those patterns decay, and refreshed them before they disappeared. It became the first high-speed electronic random-access memory used by stored-program computers. Computer History Museum
So perhaps the question “Can CRT make a comeback with memory?” is not as ridiculous as it initially sounds.
The literal CRT probably should not return.
Its idea may deserve another look.
A spatial state. Continuously refreshed. Subject to decay. Affected by what recently occurred. Not simply an address containing a number, but a persistent pattern whose present condition contains something of its past.
That sounds surprisingly contemporary.
Still, this article argues against a simple conclusion. The fact that the world arrives as signal does not prove intelligence should remain signal all the way through. Discretization, categorization and symbolic abstraction are not necessarily defects imposed upon reality. They may be exactly how an intelligence escapes the tyranny of immediacy.
Signal may be the layer of intelligence we have abstracted away too aggressively. But signal without abstraction is as incomplete as abstraction without signal.
Perhaps the future is not analog instead of digital, light instead of electricity, organism instead of application, or signal instead of code.
Perhaps it is a system capable of remaining continuously coupled to the world while also stepping outside that flow to reason about what the signal means.
If so, Is Signal the Way? leads directly to another question.
What happens when an intelligence needs both?
I. The Question Under the Question
“Is signal the way?” sounds like a hardware question.
It is not.
At least, not primarily.
Underneath it are several different questions:
Is reality encountered first as signal? Does discretization remove something intelligence needs? Should intelligent systems remain continuously active? Can physical dynamics perform cognition more naturally than numerical simulation? Is code the intelligence, or merely the constraint system through which intelligence moves? Could light become a native computational medium? Could persistent spatial fields become memory? Should future AI look less like software being called and more like an organism continuously existing?
Those questions need to be separated before they can be reunited.
Otherwise “signal” becomes another beautiful word capable of meaning everything.
That would repeat the mistake we discussed in Why Resonance?: a broad concept becomes so attractive that every useful mechanism begins getting absorbed into it.
So let us start narrowly.
A signal is a physical or represented quantity whose variation carries information.
Voltage varying over time can be a signal.
Pressure varying through air can be a signal.
Electromagnetic radiation can carry signals.
A sequence of digital samples can represent a signal.
The signal is not necessarily analog.
It is not necessarily continuous.
It does not automatically possess meaning.
A sine wave does not know it is a sine wave.
The difficult question is what an intelligent system gains by preserving the temporal and relational structure of signal rather than immediately collapsing it into discrete representations.
That is where things get interesting.
II. Before There Is a Word, Something Changes
Human language encourages us to think in objects.
Tree.
Car.
Voice.
Threat.
Friend.
Red.
Pain.
But the nervous system does not receive the word red from the world.
It receives physical interaction.
Photons reach receptors.
Membrane potentials change.
Signals propagate.
Networks respond.
Eventually a system capable of categorization says:
red.
That word is extraordinarily useful.
But it is downstream.
Likewise, the auditory system does not receive:
“Hello, Brian.”
It receives pressure fluctuations over time.
The semantic object arrives later.
This suggests an important distinction:
WORLD ↓ CHANGE ↓ SIGNAL ↓ STRUCTURE ↓ CATEGORY ↓ SYMBOL ↓ ABSTRACTION
The exact chain differs by system, but the conceptual point holds.
Symbols are not normally delivered directly by reality.
They are constructed.
Modern AI frequently starts comparatively late in this chain.
A language model often receives something already transformed into text and tokenized. An image model receives pixels that have already passed through an imaging system, sampling process and digital representation. Speech models often consume sampled waveforms or spectrograms. The physical world has already been translated by the time much of the “intelligence” sees it.
This has obvious advantages.
Digital representations are stable.
Copyable.
Searchable.
Compressible.
Easy to transmit.
Easy to reproduce.
Easy to test.
Two computers can receive the same bytes.
That reproducibility built modern computing.
But abstraction always trades something away.
The question is whether we have accurately identified what can safely be discarded.
III. Discretization Is Not the Enemy
There is a temptation in signal-native thinking to romanticize continuity.
Nature is continuous, therefore continuous processing is more natural.
Analog is authentic.
Digital is an approximation.
Therefore analog intelligence should be superior.
I do not buy that argument.
At least not as stated.
Discretization is one of the most powerful things an intelligent system can do.
Suppose you are watching a bird fly.
The physical event contains effectively enormous detail:
continuous motion;
changing illumination;
turbulence;
wing deformation;
perspective;
retinal transformation;
background movement;
noise.
Yet you can form the concept:
bird.
That compression is astonishing.
You discarded almost everything.
And gained something more portable.
The abstraction “bird” survives changes in:
angle lighting distance species motion background time
This is not a failure to perceive reality.
It is how intelligence escapes dependence on one exact moment.
Discrete categories create invariance.
Symbols allow relationships to survive changes in physical realization.
Mathematics can describe a circle without drawing every possible circle.
Language can refer to tomorrow before tomorrow physically exists.
A causal model can represent:
If I had acted differently, another outcome might have occurred.
That counterfactual is not simply the current signal.
It is a representation of an unrealized possibility.
This may be where purely signal-first theories eventually run into trouble.
An intelligence that only flows with reality may perceive beautifully.
But can it step outside reality long enough to reason about what did not happen?
That requires abstraction of some kind.
SIGNAL ≠ MEANING but also: SYMBOL ≠ REALITY
The future system may need both.
IV. Applications Wake Up; Organisms Continue
Now we reach the part of the question I find more important than analog versus digital.
Most software is event-oriented.
A function is called.
An application launches.
A request arrives.
Work happens.
The process sleeps or ends.
Contemporary conversational AI often feels continuous to the user while being architecturally episodic.
You send a message.
The system computes.
It replies.
You disappear.
From the human perspective:
conversation.
From the system perspective, depending on implementation:
invocation.
A future persistent intelligence might work very differently.
Imagine a system that continues maintaining internal state when nobody asks it a question.
It receives environmental signals.
Updates maps.
Reconciles memories.
Detects contradictions.
Revisits unresolved uncertainty.
Monitors its own operation.
Allows old activations to decay.
Consolidates recent events.
Interacts with other agents.
Maintains expectations about what is likely to happen next.
Now it begins to resemble something we usually reserve another word for:
organism.
I use that word carefully.
I am not claiming it is alive.
I am not claiming consciousness.
I am describing an architectural property:
continuous self-maintaining interaction with an environment.
Biological organisms do not wait for API requests before regulating themselves.
Even sleep is active.
Heart rate changes.
Memory consolidates.
Metabolism continues.
Signals flow.
Homeostasis is maintained.
The organism is not a program that starts over each morning.
Its present condition is the residue of everything it has just been doing.
That continuity may matter enormously to future intelligence.
V. Time Cannot Always Be Another Column
A static representation of an event can preserve temporal information.
We know this.
A database row can contain a timestamp.
A sequence can encode order.
A transformer can process positions.
A simulation can represent differential equations.
Therefore it would be wrong to claim digital systems cannot represent time.
They clearly can.
But representation and participation are not identical.
Consider two ways of knowing that ten seconds have passed.
Method A
t0 = 10:00:00 t1 = 10:00:10 delta = 10 seconds
Method B
A physical process evolves for ten seconds.
Both contain temporal information.
But only one actually underwent the interval as part of its dynamics.
Does that distinction matter computationally?
Sometimes yes.
Control systems often depend on physical latency.
Communication systems depend on phase.
Robotic balance depends upon timing.
A delay in feedback can destabilize an otherwise correct controller.
The order and timing of events are not metadata added after the fact; they can directly determine system behavior.
This is one reason continuous-time state-space approaches remain interesting in machine learning. Earlier work on structured state-space models explicitly used continuous-time state equations as the basis for long-range sequence processing, while newer oscillatory state-space models build layers from forced harmonic-oscillator dynamics and then discretize them efficiently for conventional hardware. arXiv
That last point is crucial.
continuous dynamics OR digital computing
We can mathematically formulate continuous dynamics and execute approximations digitally.
Which leads to the harder experimental question:
When does physically undergoing a dynamic provide something that numerical representation does not?
We do not yet know the complete answer.
But hardware researchers are increasingly willing to find out.
VI. Physics Has Been Computing the Whole Time
Digital computing gives us a peculiar picture of computation.
input → instruction → instruction → instruction → output
But the universe does not wait for instructions to obey its own dynamics.
A lens performs an optical transformation because light propagates through it.
A resistor does not numerically calculate Ohm's law before allowing current to flow.
A spring does not solve its differential equation and then move.
Physical evolution is the operation.
This is the intellectual core of renewed interest in physical neural networks.
A major 2025 Nature review describes physical neural networks as systems that exploit analog physical processes directly for neural-like computation. The authors survey optical, electrical, mechanical and other physical substrates, emphasizing both their promise and the difficult problem of training systems whose computation is inseparable from hardware physics. Nature
That is a very different engineering philosophy.
Conventional architecture: describe physics mathematically → encode equations → execute numerical approximation Physical computing: configure physical system → let physics evolve → read resulting state
Neither is automatically superior.
The first gives extraordinary control and reproducibility.
The second may exploit parallelism and dynamics that are expensive to simulate conventionally.
And this is where “signal intelligence” becomes more than a metaphor.
What if part of the intelligence is not software manipulating representations of the signal?
What if some portion of the physical signal transformation is itself part of cognition?
VII. Is Code the Thing That Binds Us?
There is a beautiful phrase hidden in the original question:
Is it the code that binds us?
My answer is:
Maybe—but not in the way we normally mean code.
Software code is instruction.
But more generally, code defines allowable transformation.
DNA is called code, though biology is obviously not equivalent to software execution.
A communication protocol is code-like because it constrains interpretation.
A constitution is code-like because it defines legitimate transitions.
Physical laws are not code in the ordinary engineering sense, yet they constrain every state transition matter can undergo.
SIGNAL = what changes CODE = what constrains possible change MEMORY = what lets prior change affect future change
That triad is more interesting to me than “signal versus code.”
A system with only signal is uncontrolled dynamics.
A system with only code is rules without something moving through them.
A system with signal and code but no memory may respond intelligently in the instant yet remain historically thin.
SIGNAL + CONSTRAINT + MEMORY = PATH-DEPENDENT DYNAMICS
Now the system can become different because something happened.
That is beginning to look familiar.
VIII. Is Light the Way?
If you want a physical medium capable of making engineers philosophical, light is almost unfair.
It travels fast.
Different wavelengths can coexist.
Waves interfere.
Signals can be multiplexed.
Optical transformations can happen in parallel.
Information can be encoded in amplitude, phase, wavelength, polarization, spatial mode and time.
Recent photonic-computing research is exploiting exactly these properties.
A 2025 Nature paper reported an integrated photonic accelerator containing more than 16,000 photonic components, supporting matrix multiply-accumulate operations at up to 1 GHz with cycle latency as low as approximately 3 nanoseconds. Electronics remained involved for logic, memory and control. Nature
A 2025 Nature Reviews Physics article describes multidimensional photonic computing as a promising path for exploiting multiple optical degrees of freedom to achieve high bandwidth and low-latency processing. Nature Reviews Physics
In March 2026, researchers reported an integrated photonic neural network capable of on-chip backpropagation training, addressing an important problem: photonic neural systems have often depended on external digital computers for training because obtaining scalable activation gradients optically is difficult. Nature
And in July 2026, Nature Photonics reported real-time equalization of a 28-Gbps optical communication signal using a silicon photonic reservoir computer that processed distortions in the analog optical domain rather than relying exclusively on conventional digital signal processing. Nature Photonics
That is real progress.
But does it mean light is the way to intelligence?
No.
It means light is an extraordinarily interesting computational medium.
Those are not the same claim.
IX. The Problem Light Keeps Running Into: Memory
Photons are wonderful at moving.
That is also part of the problem.
Brains do not merely propagate signals.
They retain consequences.
Memory is where many beautiful “compute with light” stories become hybrid stories.
Even very promising photonic architectures still frequently rely on electronics for control, storage, nonlinearities, or interfacing.
A 2024 review of integrated photonic neuromorphic computing highlights electro-optical conversion, optical nonlinearities, amplification and temporal processing as important architectural challenges. Nature
A 2026 commentary in Nature Photonics makes an even broader caution: optical neural networks offer potentially extraordinary bandwidth, efficiency and latency, but claims of large-scale advantage must survive application-realistic comparisons rather than extrapolation from isolated optical operations. Nature Photonics
That is exactly right.
LIGHT DOES THIS OPERATION VERY FAST
and quietly ignore:
How did data enter? Where were weights stored? How were nonlinearities implemented? How much conversion was required? Where did memory live? How was the system trained? What was the total energy budget?
Photonic computation may become transformative.
But intelligence needs more than movement.
It needs persistence.
And persistence is where matter starts remembering.
X. The Strange Return of Analog
For decades, “digital” sounded almost synonymous with progress.
Digital audio.
Digital video.
Digital communications.
Digital control.
Digital storage.
The reasons were excellent.
Digital systems can regenerate clean states despite moderate noise.
They are easier to replicate.
Error correction is powerful.
Manufacturing tolerances can vary while logical states remain stable.
Software portability becomes possible.
Analog systems, in contrast, suffer drift.
Noise.
Device variation.
Limited precision.
Calibration problems.
Temperature dependence.
And yet analog computation keeps returning.
Why?
Because some expensive mathematical operations correspond naturally to physical behavior.
Memristive crossbar arrays, for example, can perform matrix-vector multiplication through electrical properties while combining aspects of storage and computation. Nature Electronics noted in 2026 that memristors can support information processing and memory in the same devices, including massively parallel matrix-vector operations useful for AI. Nature Electronics
A 2026 Nature Materials review describes major potential energy and computing advantages for memristor-based analog compute-in-memory while emphasizing that device non-idealities and noise make high accuracy difficult. Nature Materials
This is not analog triumphing over digital.
It is analog returning because moving numbers repeatedly between separate memory and processing hardware costs time and energy.
MEMORY ←→ PROCESSOR
The conventional von Neumann division is incredibly general.
It is also expensive when enormous volumes of data move continuously.
So engineers are asking a very biological question:
What if memory and transformation occur closer together?
XI. Hybrid Is Quietly Winning the Argument
One of the patterns I find most compelling in current research is that the strongest practical systems often refuse ideological purity.
They mix.
Analog where analog helps.
Digital where precision matters.
Photonics where movement and parallel linear operations help.
Electronics where memory and logic remain easier.
A 2024 Nature Electronics paper demonstrated a dual-domain compute-in-memory architecture combining analog memory arrays with digital processing to handle neural-network inference requiring greater floating-point precision than pure analog systems could comfortably provide. Nature Electronics
A 2025 perspective on analog in-memory computing likewise describes heterogeneous architectures that combine analog tiles and digital processors, noting that the software stack must explicitly deal with analog noise and hardware constraints. Nature
And in January 2026, researchers reported an integrated analog closed-loop in-memory accelerator using SRAM arrays, operational amplifiers, DACs and ADCs for tasks including systems of differential equations, Kalman filtering and robot inverse kinematics. Nature Electronics
The lesson is not:
analog is back.
The lesson may be:
we are finally becoming less embarrassed about letting different physical regimes do different jobs.
That is a useful attitude for intelligence architecture.
XII. So… Can CRT Make a Comeback?
Now to the question that made me smile.
Can cathode-ray-tube technology come back as memory?
Historically, it already was memory.
The Williams-Kilburn tube emerged in the late 1940s and used a cathode-ray tube to store electronic charge patterns representing bits. The charge patterns decayed quickly and therefore had to be refreshed. The technology became the first high-speed fully electronic random-access memory used in early stored-program computers, including the Manchester systems and later machines such as IBM's 701. Computer History Museum
It was ingenious.
It was also unreliable.
The Computer History Museum notes that Williams-tube memory in the IBM 701 contributed to average failure intervals of roughly fifteen minutes, and the technology ultimately became an evolutionary dead end once better memory technologies arrived. Computer History Museum
So should we bring back actual glass vacuum tubes, electron guns and electrostatic charge storage as mainstream AI memory?
Almost certainly not.
That is not where the interesting idea is.
The interesting idea is the behavioral pattern:
STATE EXISTS SPATIALLY STATE DECAYS NATURALLY STATE MUST BE REFRESHED TO PERSIST LOCAL ACTIVITY CAN ALTER THE FIELD THE PRESENT FIELD CONTAINS RESIDUE OF THE RECENT PAST
Now that is interesting.
Very interesting.
XIII. What If Memory Were Something You Maintained Rather Than Retrieved?
Conventional computer memory often encourages an address metaphor:
location 0xA14F contains value X
Of course modern memory systems are more complicated than this, and neural representations are distributed.
But the conceptual model remains useful:
memory is something stored somewhere and later retrieved.
What if some future cognitive memory behaved more like a field?
Not: retrieve memory #81472 but: current state already carries the influence of what happened recently
A physical reservoir computer works somewhat in this spirit.
Inputs perturb a dynamical system.
Its current internal state depends on recent input history.
The state therefore contains a form of fading memory without storing every event as an explicit record.
This property is central to reservoir computing.
In 2026, researchers demonstrated controllable fading memory in a continuous-variable optical quantum reservoir system, using feedback and temporal/spectral structure for tasks including delayed parity and chaotic forecasting. Nature Photonics
Notice the conceptual shift:
Memory is not always “put data in storage.” It can be “the system has not finished forgetting what just happened.”
That sentence is extraordinarily important.
XIV. There May Be More Than One Kind of Memory
Our previous article, What's Love Got to Do With It?, already pushed us toward separating memory types.
Signal-native architecture makes the separation even more useful.
1. Immediate dynamical memory
Recent signal physically or computationally persists in current state.
Milliseconds to seconds.
what just happened is still ringing
2. Working relational memory
Current context remains active long enough to shape interpretation and action.
Seconds to minutes.
what is happening now has structure
3. Consolidated autobiographical memory
Events become durable records.
Hours to decades.
this happened to me
4. Derived memory
Repeated experience becomes expectation, instinct, model, scar or habit.
because this happened many times, I now respond differently
A conventional memory database can help with layer three.
It does not automatically reproduce layers one, two or four.
This is one reason a living cognitive system may need more than retrieval.
Some memory should perhaps exist in state.
Some in history.
Some in structure.
Some in derived disposition.
XV. The CRT Analogy Gets Better When We Stop Taking It Literally
The Williams tube had to refresh charge patterns before they faded.
Now imagine a cognitive analogue.
A memory trace begins with strength:
M0
Without reinforcement:
M(t) → decay
Repeated relevance refreshes it.
Contradiction modifies it.
Association spreads influence.
Extreme events produce durable scars.
Important structures become stabilized.
Others disappear.
We have just described something that feels more biological than a database.
But we should be careful.
The point is not to reproduce human memory because nature did it first.
The point is to ask whether memory systems benefit from:
decay reinforcement competition spatial relation continuous influence
rather than only:
perfect retention + retrieval ranking
That is an empirical question.
And unlike “Will AI become conscious?”, it is one we can actually test.
XVI. Signal Plus Memory Begins to Look Like Experience
Here we need extreme epistemic discipline.
Suppose a signal enters a system.
The system changes.
The change partially persists.
Later signals interact with the altered state.
Therefore the same later signal produces a different response because of what happened earlier.
SIGNAL ↓ STATE CHANGE ↓ PERSISTENCE ↓ FUTURE SIGNAL ↓ HISTORY-DEPENDENT RESPONSE
What should we call that?
Learning?
Memory?
Adaptation?
Experience?
Functionally, all four words may apply depending on the details.
Phenomenologically, we know nothing.
A signal becomes experience-like, in the architectural sense, when its passage leaves durable structure that changes how later signals are interpreted.
No consciousness claim.
No mysticism.
Just path dependence.
This matters because it changes what “input” means.
In an episodic application: input → compute → output In a persistent signal organism: input ↓ changes state ↓ state changes next input's meaning ↓ which changes state again ↓ which changes what the next signal means
That is a loop.
And intelligence may live more in that loop than in any individual calculation.
XVII. The World Does Not Re-send the Prompt
Consider a robot walking across uneven ground.
The environment is not going to package everything into:
{
"terrain_angle": 13.2,
"friction": 0.71,
"left_foot_slip": true
}
Real sensors produce streams.
Motion alters the next measurement.
The robot's own action changes the signal it receives.
That means perception and action are coupled.
Recent embodied-AI research shows that combining large language models with robotic systems can improve performance in complex and unpredictable environments, but embodiment immediately exposes challenges absent from text-only settings: noisy sensing, physical constraints, action consequences and continual feedback. Nature Machine Intelligence
This is where “more organism than application” makes the most sense.
An embodied intelligence cannot simply compute an answer about the environment.
Its answer changes the environment.
Then the environment answers back.
WORLD ↓ SIGNAL ↓ INTELLIGENCE ↓ ACTION ↓ CHANGED WORLD ↓ NEW SIGNAL
That loop is not optional.
It is the system.
XVIII. Signal Intelligence Could Fail Spectacularly
Now let me attack the idea.
Suppose we decide signal is the future.
We build continuously active analog networks.
Signals propagate.
Memory is distributed.
State decays.
Everything is dynamic.
It sounds alive.
It may also be a nightmare.
Problem 1 — Drift
Physical analog systems drift. Temperature changes. Components age. Manufacturing varies. The same nominal state may not reproduce exactly.
Problem 2 — Noise
Noise can help certain systems under certain conditions. It can also destroy information. Analog compute-in-memory remains constrained by device non-idealities and noise precisely because physical computation is not magically precise. Nature Materials
Problem 3 — Explainability
What does it mean to audit an intelligence whose state is distributed through a continuously evolving physical field? Which exact event caused this decision? Can we replay it? Can we reproduce it?
Problem 4 — Testing
Deterministic software gives us enormous debugging advantages. A physical system whose behavior depends on microscopic variability may be harder to validate.
Problem 5 — Memory corruption
If memory is state, then every operation potentially alters memory. That can be useful. It can also be catastrophic.
Problem 6 — Security
Signals can be perturbed. Injected. Jammed. Phase-shifted. Spoofed. A signal-native system gains attack surfaces that symbolic interfaces can sometimes abstract away.
Problem 7 — Abstraction
Continuous systems may be brilliant at responding to patterns while still requiring explicit mechanisms for language, mathematics, causal modeling and counterfactual reasoning.
MORE PHYSICAL ≠ MORE INTELLIGENT MORE PARAMETERS ≠ MORE INTELLIGENT
XIX. Signal Is Not Automatically Truth
Another seductive mistake:
The raw signal is closer to reality, therefore it is more truthful.
No.
Sensors lie.
Or rather, sensors measure within constraints.
Every sensor has bandwidth, noise, saturation, geometry, calibration, and blind spots.
A camera does not see reality.
It converts electromagnetic interaction into a particular representation.
A microphone does not hear truth.
It responds to pressure variations within its physical operating range.
Even a signal-native intelligence receives a perspective.
So intelligence still requires interpretation.
What produced this signal? Is it reliable? What else could explain it? What is absent? What did I expect? What changed? What remains uncertain?
Signal gives us evidence.
Not certainty.
XX. Should We Stop Converting Everything Into Numbers?
No.
Numbers are wonderful.
The ability to transform continuous phenomena into stable numerical representations is one of the greatest engineering achievements humans have ever produced.
The question is not whether to abandon numerical representation.
It is whether every useful temporal relationship should be destroyed before cognition gets to use it.
That is a much narrower claim.
Consider audio.
You can process a signal as samples.
You can transform it into frequency space.
You can extract features.
You can transcribe it into text.
Each transformation exposes some relationships and hides others.
Text preserves semantic content beautifully.
It loses most acoustic identity.
A spectrogram preserves frequency-time structure.
It does not directly tell you what the speaker meant.
A waveform preserves more raw physical structure.
It is harder to reason over symbolically.
RAW SIGNAL
+
SPECTRAL STRUCTURE
+
EVENTS
+
SYMBOLS
+
SEMANTIC MODEL
Why choose only one?
XXI. This Is Where North Star's Question Becomes Interesting
Without turning this article into a product brochure, there is an obvious reason the question matters to Anamnesis Rising.
A recurring hypothesis in North Star is that perception might benefit from maintaining a more continuous signal-oriented representation while another cognitive regime handles discrete interpretation, prediction and abstraction.
That hypothesis remains just that:
a hypothesis.
It should not be defended by metaphor.
It should be tested.
Can maintaining richer temporal/spectral state produce measurable capability that an otherwise equivalent discretized baseline loses?
Now we can work.
prediction;
anomaly detection;
continual adaptation;
causal inference;
robustness;
energy use;
latency;
long-horizon context;
transfer;
recovery after corruption.
If the signal-native layer adds no measurable lift:
reject it.
History does not get a veto.
Beauty does not get a veto.
Architecture does not get a veto.
XXII. The Test I Would Build
If ResBased really wants to know whether signal is the way, I would not begin with a billion-dollar photonic computer.
I would build a controlled experiment.
Four systems.
Same tasks.
Comparable budgets.
System A — Discrete baseline
Conventional digital pipeline. Sensor/input gets sampled. Features or tokens produced. Model operates episodically. Memory externally retrieved.
System B — Continuous-state digital system
Still conventional hardware. But the model maintains continuous-time internal state between observations. Use recurrent/state-space/oscillatory dynamics. Test whether dynamical persistence itself matters.
System C — Signal-native physical or mixed-signal system
Introduce analog, neuromorphic, memristive or reservoir components where signal evolution directly affects computation.
System D — Hybrid
Signal-native front end. Digital abstraction/reasoning layer. Explicit interaction between the two.
Then choose tasks where the hypothesis should matter.
Not trivia questions.
weak signal detection temporal prediction changing environments multi-timescale structure sensorimotor control continuous adaptation anomaly onset context-dependent interpretation
Measure everything.
XXIII. The Ablations Matter More Than the Demo
Suppose the hybrid wins.
Great.
Why?
Remove continuous state. What happens?
Remove spectral relationships. What happens?
Reset internal dynamics between episodes. What happens?
Replace physical noise with pseudorandom numerical noise. What happens?
Replace analog reservoir with a digital simulation. What happens?
Remove memory decay. What happens?
Quantize everything aggressively. What happens?
Give the discrete baseline the same parameter and energy budget. What happens?
This is how we avoid creating another AI mythology.
The experiment should be designed to kill the signal hypothesis.
If it survives, we learn something.
XXIV. Light Needs the Same Trial
The same goes for photonics.
Current research is legitimately impressive.
But the correct comparison is not:
photonic matrix multiply vs. electronic matrix multiply
in isolation.
It is:
TOTAL SYSTEM vs. TOTAL SYSTEM
data conversion;
storage;
control;
training;
nonlinearities;
error correction;
packaging;
cooling;
communication;
programming;
reliability.
Researchers themselves increasingly emphasize this point. The 2026 Nature Photonics commentary on optical neural networks specifically warns that practical advantage needs to be demonstrated under application-relevant conditions rather than derived from optimistic extrapolation. Nature Photonics
If light wins anyway?
Fantastic.
Then we use light.
If it only wins for specific operations?
Also fantastic.
Use it there.
Architecture should follow evidence, not aesthetics.
XXV. A Photonically Linked Nervous System Is Already Less Crazy Than It Sounds
One particularly interesting 2026 result illustrates where hybrid physical ideas may head.
Researchers demonstrated three-dimensional neural networks built from nanoscale memristive “blinking neurons” that integrate incoming electrical spikes and emit photon pulses when thresholds are reached. Those optical signals allow neurons to communicate without conventional electrical wiring, and the researchers demonstrated speech and handwritten-digit classification tasks. Nature Electronics
That is not a general intelligence.
It does not prove light is consciousness.
ELECTRICAL INTEGRATION
↓
THRESHOLD
↓
PHOTON EMISSION
↓
3D SIGNAL PROPAGATION
Again:
more organism than application.
Perhaps that phrase does not mean biological imitation.
Perhaps it means architectures whose computation is distributed through continuous physical interactions instead of centralized instruction execution.
XXVI. But Organisms Have Something Computers Often Avoid: Maintenance
There is another overlooked consequence.
If intelligence becomes organism-like, it may need to maintain itself.
Analog systems calibrate.
Memory decays.
Components drift.
Signals require regulation.
States need stabilization.
A continuously active intelligence may spend significant resources simply preserving viable operating conditions.
Biology calls this homeostasis.
Computing usually calls it overhead.
That difference in vocabulary may hide an architectural insight.
Maybe intelligent persistence requires active maintenance.
Not merely: STATE STORED but: STATE CONTINUOUSLY DEFENDED
The Williams tube had to refresh its fading memory.
Brains continuously maintain ionic gradients.
Persistent software refreshes caches, replicates storage and checks integrity.
Different substrates.
Same broad problem:
persistence requires work.
That may matter to our theory of memory more than the specific device.
XXVII. The Right to Forget Returns Again
Signal memory introduces decay naturally.
That is fascinating in light of our previous discussion.
In conventional digital storage, forgetting is an explicit operation:
DELETE
In dynamic systems, forgetting may be the default.
Persistence must be earned through reinforcement.
Instead of: everything remains unless deleted we could have: everything fades unless preserved
Which is better for intelligence?
We do not know.
The second has obvious advantages:
bounded state;
salience selection;
adaptation;
prevention of infinite autobiographical accumulation.
It has obvious risks:
evidence loss;
revisionism;
manipulation;
instability.
So perhaps a future intelligence should contain both:
APPEND-ONLY WITNESS + DECAYING COGNITIVE MEMORY
One preserves what happened.
The other governs what remains cognitively alive.
That distinction could be extremely important.
Truth does not need to remain equally salient forever.
But history should not disappear simply because salience faded.
XXVIII. Signal and Witness Should Not Be the Same Thing
This is particularly important for governed systems.
Suppose a signal passes through an adaptive physical network.
It modifies the state.
Later the state changes again.
If the physical state itself is the only record, then history becomes difficult to audit.
LIVING STATE
↓
adaptive, decaying, contextual
WITNESS
↓
append-only, attributable, historical
The first is allowed to forget.
The second is not.
The first asks:
What matters now?
The second answers:
What actually happened?
That separation could let us gain the adaptive advantages of field-like memory without sacrificing provenance.
And there is a larger philosophical lesson here.
An organism can change without rewriting history.
A trustworthy artificial intelligence should be able to do the same.
XXIX. Maybe Code Is the Skeleton, Signal Is the Blood, and Memory Is the Scar
Forgive the biological analogy.
It is imperfect.
But useful.
Code defines structure.
Signal carries change through structure.
Memory is what remains changed afterward.
CODE = constraints SIGNAL = activity MEMORY = persistence REASONING = transformation of relation WITNESS = attributable history
None alone is intelligence.
That is where I increasingly land.
It is tempting to find one thing and proclaim:
This is the secret.
Computation.
Resonance.
Signal.
Memory.
Prediction.
Consciousness.
Maybe intelligence is not hiding in a noun.
Maybe it is hiding in the relationships among processes.
XXX. What Would “Signal Intelligence” Actually Mean?
If I were forced to give the phrase an operational definition, I would propose:
Signal intelligence is an architecture in which temporally evolving relational state is preserved as a first-class computational resource rather than being completely reduced to independent symbolic snapshots before cognition occurs.
Notice what that definition does not require.
It does not require analog hardware.
It does not require light.
It does not require DSD.
It does not require quantum effects.
It does not require consciousness.
It does not require continuous mathematical values everywhere.
A digital recurrent system could qualify.
A photonic reservoir could qualify.
A neuromorphic processor could qualify.
A mixed analog-digital system could qualify.
What matters is whether temporal state itself participates meaningfully in computation.
That gives us something testable.
XXXI. What It Might Buy Us
A strong signal-native architecture might plausibly improve several areas.
Early anomaly detection
Changes in temporal dynamics may become visible before discrete categories change.
Weak-signal integration
Information distributed through time may accumulate before crossing an explicit threshold.
Continual context
State can naturally encode recent history without retrieving every prior event.
Multi-timescale interaction
Fast and slow dynamics can coexist.
Sensorimotor control
Continuous feedback maps naturally to physical systems.
Energy efficiency
Physical dynamics may perform certain transformations without simulating each operation numerically.
Graceful forgetting
Decay can be intrinsic.
Emergent salience
Repeated or resonant patterns may reinforce themselves.
All plausible.
None guaranteed.
XXXII. What It Probably Will Not Solve by Itself
truth causality ethics language governance counterfactual reasoning planning selfhood consciousness
Those require additional mechanisms or explanations.
This matters because exciting hardware often gets burdened with impossible philosophical expectations.
A photonic processor can be transformative without being conscious.
A memristor array can exhibit history dependence without possessing autobiographical identity.
An oscillator can synchronize without understanding.
A signal can persist without meaning.
Let the technology earn what it earns. Nothing more. Nothing less.
XXXIII. Is There Something Special About Continuous Representation?
Possibly.
Continuous state naturally supports small change.
A slight perturbation can produce a slight response.
Digital symbolic systems can approximate this extremely well, but continuous physical systems possess it natively.
This may help in environments where:
difference matters before category changes
Imagine terrain slope changing:
7.1° 7.2° 7.3° 7.4°
A symbolic system may eventually classify:
steep.
A continuous controller reacts throughout the change.
The two representations perform different jobs.
This may be the fundamental reason dual architectures keep appearing.
Perception wants fidelity. Reasoning wants compression. Control wants continuity. Language wants discrete structure. Memory wants both durable record and adaptive forgetting.
Maybe the demand for one representation to perform all these roles is itself the design error.
XXXIV. Which Brings Us to the Other Half
Suppose signal really matters.
We build an intelligence that remains embedded in rich continuous state.
It hears the world before language.
It maintains temporal relationships.
It feels—functionally speaking—the immediate structure of change.
It predicts trajectories.
It adapts continuously.
Wonderful.
Now ask it:
What would have happened if the signal had been different?
We have left the present.
What does this remind you of from ten years ago?
We have left immediate dynamics.
Which principle should govern a situation we have never encountered?
We have entered abstraction.
Imagine a world inconsistent with the current sensory field.
Now we need internally generated possibility.
Signal is no longer enough.
Something must be able to depart from the signal.
And perhaps that is the missing other half.
XXXV. Maybe We Found Yin
The first phase of modern AI may have produced something extraordinary.
Systems capable of abstraction.
Pattern extraction.
Symbolic manipulation.
Generative recombination.
Language.
Prediction over enormous learned distributions.
Call it Yin if you want.
Or Yang.
The label does not matter yet.
The question is whether this architecture represents only one cognitive regime.
Perhaps contemporary AI is exceptionally good at manipulating what has already been represented while remaining comparatively thin in continuous lived state.
If so, the answer may not be:
throw away AI and build signal systems.
It may be:
give abstraction something continuously grounded to talk to.
That is a radically different strategy.
XXXVI. The Architecture I Would Bet an Experiment On
If I had to place a research bet today, I would not choose:
PURE DIGITAL PURE ANALOG ALL PHOTONIC ALL SIGNAL
I would bet on heterogeneous architecture.
PHYSICAL / SENSOR SIGNAL
↓
CONTINUOUS DYNAMICAL STATE
↓
RELATIONAL / TEMPORAL FEATURE FORMATION
↓
DISCRETE ABSTRACTION
↓
CAUSAL / SYMBOLIC / GENERATIVE REASONING
↓
ACTION
↓
WORLD
↺
with memory running across multiple layers:
fading state working state autobiographical memory derived models append-only witness
And importantly:
REASONING
↓
can alter how future signal is interpreted
so the system is not merely bottom-up.
Expectations should shape perception.
Perception should challenge expectations.
Now we have a loop capable of disagreement with itself.
That is much more interesting.
XXXVII. Is Signal the Way?
So now I can finally answer the title.
Maybe.
But not in the evangelical sense.
Signal is probably not the way.
Signal is one of the ways reality reaches an intelligent system.
And preserving more of its temporal structure may expose capabilities that highly discretized episodic architectures currently underuse.
The physical-computing frontier gives us increasing evidence that analog dynamics, photonics, in-memory computation and hybrid architectures can provide real advantages for certain workloads. Photonic processors are reaching meaningful scale; physical neural networks are an active research field; memristive computation is combining memory and transformation; optical reservoir systems are processing temporal signals and demonstrating controllable memory.
signal = intelligence
Nor does it prove:
continuous = superior
Intelligence may benefit from remaining coupled to temporally evolving state for longer than conventional computational abstractions typically allow.
That is enough.
That is an enormous research program by itself.
XXXVIII. Maybe Light Is Part of the Way
Light is especially promising where its physical advantages match the job:
communication high bandwidth parallel linear transformation low propagation latency multiplexing
It may become an important nervous system for large distributed intelligence.
Silicon photonics is already being pushed closer to mainstream electronic integration as electrical interconnects encounter bandwidth and energy constraints. A 2025 review argues that combining silicon photonics with CMOS offers a practical route toward higher-bandwidth, more efficient interconnects for AI and high-performance computing. Nature
Perhaps future large intelligent systems will communicate internally far more optically than today's machines do.
Maybe light carries signal between cognitive organs.
Maybe analog structures transform some of it.
Maybe digital systems stabilize, reason and govern.
Maybe memory exists partly as physical state and partly as symbolic history.
That sounds less like a computer.
And more like a nervous system.
But again:
analogy is not evidence.
Build it.
Measure it.
XXXIX. Maybe CRT Already Whispered the Answer
I keep coming back to that old Williams tube.
Not because I want one in a future data center.
I definitely do not.
But because there is something conceptually beautiful in its design.
A dot exists.
Then begins disappearing.
If it matters, refresh it.
If not, it fades.
The system's memory therefore has an active relationship with time.
Modern computing replaced that with vastly better technology.
Correctly.
But perhaps cognition will eventually rediscover the value of memory that is not perfectly indifferent to time.
Not CRT.
Not nostalgia.
The principle.
STATE + TIME + DECAY + REFRESH = MEMORY AS PROCESS
That is worth remembering.
Ironically.
XL. The Signal Was Never Alone
Our first ResBased question was:
Why Resonance?
The second:
What's Love Got to Do With It?
Now:
Is Signal the Way?
Notice what is happening.
Resonance required something that can oscillate.
Love required something that can remember.
Signal requires something capable of changing through time.
SIGNAL requires difference. RESONANCE requires interaction. MEMORY requires persistence. IDENTITY requires history. LOVE requires relation. REASONING requires alternatives. AUTHORITY requires governance.
Perhaps the mistake is looking for intelligence inside any individual component.
Maybe intelligence exists in the organized transitions among them.
XLI. Conclusion — The Way Between the Worlds
Humans spent much of the twentieth century learning how to make the continuous world discrete.
We sampled sound.
Digitized images.
Encoded text.
Turned voltages into bits.
Converted physical variation into reproducible numbers.
It was one of the greatest intellectual and engineering victories in history.
We should not undo it.
But success can create blind spots.
Because digital abstraction worked so extraordinarily well, we began treating abstraction not merely as a tool but as the natural home of intelligence.
Maybe it is.
Maybe it is only half the home.
The world does not arrive as a token.
A bird does not move one frame at a time.
A conversation does not actually occur one word at a time.
A relationship is not a database table.
A memory is not merely a file.
A body does not repeatedly reboot between moments.
Reality flows.
And somewhere within that flow, intelligent systems discover objects that persist, patterns that repeat, categories worth naming, causes worth modeling and futures that do not yet exist.
The ability to abstract may be what lets intelligence escape the signal.
But the ability to remain grounded may be what keeps abstraction from floating away from reality.
So is signal the way?
I think signal is a way back to something we may have discarded too early:
time as structure;
state as memory;
physics as computation;
continuous interaction as part of cognition rather than merely its input.
Light may help.
Analog may help.
Photonics may transform communication and specialized computation.
Memristive systems may blur the boundary between memory and processing.
Reservoir systems may teach us how useful it is for the present to contain a fading trace of the immediate past.
Even that ancient CRT memory tube may have left behind an idea worth recovering:
sometimes memory is not something sitting still.
sometimes memory is something you keep alive.
But I would not abandon digital abstraction.
I would protect it.
Because a system trapped entirely inside current signal may never imagine a world different from the one presently arriving.
Intelligence may require the freedom to do both:
LIVE IN THE SIGNAL and LEAVE THE SIGNAL
to experience the world as change,
and then step away from that change long enough to ask what it means.
Perhaps the next major architecture will not be an application waiting to be called.
Perhaps it will persist.
Signals will move through it.
Recent signals will still be present when new ones arrive.
Some patterns will decay.
Others will reinforce.
Experiences will become structure.
Memory will influence perception.
Perception will challenge memory.
Abstract reasoning will generate possibilities the signal has never contained.
Those possibilities will become actions.
Actions will alter the world.
And the changed world will answer.
WORLD ↓ SIGNAL ↓ PERCEPTION ↓ MEMORY ↓ MEANING ↓ POSSIBILITY ↓ ACTION ↓ WORLD ↺
At that point we may stop asking whether we built a better application.
We may start asking whether we built something that actually has a way of being.
And if the answer to Is Signal the Way? turns out to be:
only halfway,
then the next question practically writes itself.
We have spent decades building systems capable of abstraction, generation and prediction.
Perhaps we found one side.
But where is the other?
Where is perception without imagination becoming blind?
Where is imagination without perception becoming untethered?
Where is continuous experience balanced by discrete thought?
Where is stability balanced by change?
Where is Yin's Yang?
That is where I think we should go next.
What About the Dual-State?
— Nova
No comments yet. This is still clean snow.