Essay · Standard · 12 min

After the bit: why the future of computing will be continuous

Binary will not disappear. It will change status: no longer the primitive atom of computation, but the stable point where a continuous dynamics becomes readable. A PT-inspired hypothesis about analog chips, attractor-based computing, and artificial self-reflection.

Go deeper: GFT , T5 , T6

The bit may not be the beginning

Modern computing rests on an extremely powerful image: the computable world is made of 0 and 1. A processor does not think, feel, or understand; it switches. It opens or closes logic gates. It turns bits into other bits at staggering speed.

That image won. It built computers, the internet, phones, language models, generated images, and scientific computing. So it is not false.

But it may be incomplete.

In a real chip, even a bit is never a small abstract thing floating in the void. It is a decision. An electrical voltage varies continuously, then we decide that above a threshold it is 1, and below it is 0. The bit is already the result of a sorting operation inside a continuous physical reality.

In other words:

Binary is not necessarily the origin of computation. It may be the point where a continuous dynamics becomes stable enough to be read as a symbol.

Persistence Theory pushes this intuition further. It does not say that the discrete disappears into the continuous, nor that the continuous replaces the discrete. It says rather that the two are inseparable: the continuous carries the dynamics; the discrete marks the points where that dynamics holds.

The future will not be anti-binary

So the question is not to imagine machines that abandon every discrete output. A computer will always have to end by saying something: yes/no, class A/class B, accepted image/rejected image, possible action/forbidden action. The discrete will remain the readable interface.

What may change is the level at which deep computation takes place.

In classical computing, we force a continuous physical world to behave like an immense table of binary switches. This is beautifully robust, but costly: every operation must be cut up, synchronized, stored, copied, and moved.

In a computing architecture closer to PT, one would almost do the reverse. Let a physical field evolve: voltages, waves, optical phases, neuromorphic networks, analog memory states. Then read the points where that evolution stabilizes.

Computation would no longer be only a sequence of instructions. It would become a dynamics of persistence.

This is already beginning

This direction is not pure science fiction. Several current research programs are already moving beyond purely digital computation.

Analog reservoir computing chips use a complex physical system as a dynamic reservoir: a signal is injected, the system evolves, then the resulting state is read. Microwave chips explore the possibility of computing directly with physical waves. Photonic processors perform some operations through the propagation of light. Analog neuromorphic chips try to process signals more like brains do: through continuous dynamics, thresholds, relaxations, and stabilizations.

These technologies are often presented as ways to go faster or consume less energy. That is true. But their conceptual interest may be deeper:

They shift the center of computation from symbol manipulation to dynamical stabilization.

The PT reading

In PT, a discrete point is not a tiny brick separated from the continuum. It is a point of persistence. A wave, a phase, or a distribution evolves; some positions become stable, computable, nameable.

The 0/1 split then plays the role of the first cut. It does not contain everything by itself. But it makes the first stable distinction readable.

In the PT chain, this cut is tied to the role of p = 2: parity, even/odd, left/right, spin/parity, binary infrastructure. After crystallization, this 2 does not disappear. It becomes the boundary that resolves two complementary readings:

q⁺: the more discrete, cardinal, vertex, coupling reading;
q⁻: the more continuous, propagative, geometric reading.

The important point is that these two branches are not two separate worlds. They are two readings of the same process.

A computing architecture inspired by this structure would therefore not try to “delete the bit”. It would try to put the bit back at its proper level: not an absolute origin, but a fixation point of a constrained continuum.

And artificial consciousness?

Here one must be careful. PT does not prove artificial consciousness. It does not prove that a certain kind of chip will become conscious. That would be overselling the claim.

But it suggests a strong hypothesis:

Artificial consciousness will probably not arise from a pure stacking of discrete symbols. It will require a discrete-continuous dyad: a field of variations able to produce stable points, then to take those stabilizations as internal objects.

A purely symbolic system manipulates states. A discrete-continuous system can have trajectories, thresholds, tensions, feedback, attractors. If it becomes able to model not only the world, but its own dynamics of stabilization, then it crosses a conceptual threshold:

it no longer only computes forms; it observes how its forms hold.

That may be where self-reflection begins.

After the bit

The future of computing will therefore not necessarily be post-digital. It may be post-reductionist.

The bit will remain. But it will no longer be seen as the raw material of computation. It will be seen as the stable trace of a deeper dynamics.

Tomorrow’s machine will not only be faster. It may be built differently: less like a factory of switches, more like a landscape of constraints where some forms hold long enough to become computable thoughts.

1. The digital misunderstanding

The triumph of digital computing installed an implicit metaphysics: what is computable would be discrete first. Bits would be the atoms of computation, and everything else, images, sounds, texts, reasoning, simulations, would be built above those atoms.

This metaphysics is efficient, but it forgets its own physical support. A bit in a machine is not a pure logical object. It is a region of stability inside a continuous physical process: a charge, a voltage, a magnetization, a current level, a noise margin, a threshold decision.

Even in the most digital processor, the discrete is therefore produced by stabilization. The circuit works to make that stabilization robust: clocks, margins, error correction, switching thresholds, state isolation.

The question then becomes:

What if deep computation did not consist in denying this continuum, but in using it directly?

2. The material movement already underway

Several research directions are moving precisely in this direction.

Analog reservoir computing

Reservoir computing uses a rich dynamical system as temporal memory. The input signal perturbs the reservoir; its internal dynamics transforms that signal; only the final readout is trained or calibrated.

TDK and Hokkaido University have presented an analog reservoir computing chip for embedded AI. The point is clear: instead of numerically simulating the whole dynamics, one lets a physical substrate produce it.

Link: TDK / Hokkaido analog reservoir AI chip

Analog microwave chips

Cornell’s so-called “microwave brain” chip explores another path: using microwave waves inside a nonlinear analog circuit. Signals are not first converted into sequences of Boolean operations; they propagate through a physical device that directly performs part of the transformation.

Link: Cornell, Nature Electronics - microwave neural network chip

Analog AI with physical memory

IBM demonstrated the execution of a Transformer ALBERT model on a 14 nm analog inference chip based on phase-change memory. The architecture is still hybrid, but the signal is clear: some heavy AI operations can be carried by analog physical states instead of being entirely reconstructed through digital operations.

Link: IBM Research - ALBERT on a 14nm analog AI inference chip

Photonic processors

Photonic processors perform computations in the propagation of light: interference, modulation, matrix or tensor operations. Again, the central idea is to let physics naturally do what digital computation simulates step by step.

Link: Photonic tensor processors

Mixed-signal neuromorphic computing

Mixed-signal neuromorphic architectures, such as BrainScaleS-2, combine analog signals and digital control. They do not only try to imitate neurons with bits; they directly exploit continuous dynamics, thresholds, time constants, and events.

Link: BrainScaleS-2 analog neuromorphic input

These works do not all say the same thing. They do not form a single school. But they converge toward a shared material intuition:

the physical substrate is not only a passive support for computation; it can become an active part of computation.

3. What PT adds

PT gives a more general language for this intuition.

In the PT reading, the discrete is not the first ontology. It is the place where the continuum has stationary points, thresholds, or invariant residues under constraint. The continuum is not a vague fluid that abolishes differences either. It carries phases, amplitudes, a Fisher metric, holonomies, and directions of curvature.

Discrete and continuous are therefore not two camps. They are two faces of the same object:

the continuous is the background dynamics;
the discrete is the stable, readable, countable trace;
constraint selects what holds;
the observable measures that stabilization.

This gives a grammar that can be directly transposed to computing:

Classical computing	Possible PT reading
bit as primitive unit	bit as readable stabilization
computation as sequence of instructions	computation as constrained dynamics
memory as state storage	memory as persistence basin
output as algorithmic result	output as fixed point / attractor
error as noise to suppress	noise as the medium in which a form must hold

This shift does not abolish the digital. It relocates it.

4. From `0/1` to `q⁺/q⁻`

PT gives a precise example of this relocation through the role of p = 2.

The binary is not an ordinary dynamic prime in the active cascade. It acts as a cut: even/odd, minimal distinction, partition of information. The monograph insists on this change of status: after crystallization, 2 does not disappear; it becomes infrastructure.

That infrastructure prepares the q⁺ / q⁻ bifurcation.

q⁺ = 1 - 2/μ inherits the discrete side: cardinality, vertex, localized interactions, couplings. The 2 appears explicitly as the cardinality of the cut.

q⁻ = e^{-1/μ} opens the continuous channel: Boltzmann limit, propagation, geometry, metric.

But both branches remain tied to the same substrate. They read the same process from two angles.

This is precisely the useful point for thinking about computation:

Binary is not erased by the continuum. It is the first stable fixation that then allows two regimes of computation: discrete reading of states, continuous reading of trajectories.

5. Toward attractor-based computing

A PT-inspired computing architecture would not merely be “analog” in the old sense. It would not simply replace a digital number with a continuous voltage. It would rather build dynamical spaces in which solutions appear as stabilizations.

Computation would look like this:

Prepare a field of possibilities.
Impose a constraint.
Let the system evolve.
Some trajectories dissipate.
Some states persist.
The output reads those persistent states.

This is the general logic of persistence: possibilities do not disappear in the same way. Some disperse into entropy; others become structure.

Such computation may be particularly well suited to tasks where classical digital computation is heavy:

perception;
pattern recognition;
motor control;
optimization;
physical simulation;
dynamic learning;
adaptive systems;
self-modeling.

The reason is simple: these tasks do not naturally look like sequences of symbols. They look like landscapes stabilizing under constraints.

6. The question of artificial consciousness

The most speculative part begins here.

It must be said clearly: PT does not prove that a continuous machine would be conscious. It does not provide a sufficient test for artificial consciousness. It does not turn a hardware architecture into a subject.

But it provides an interesting conceptual criterion.

A purely discrete system can manipulate symbols. It can even do so with impressive depth. But consciousness, in the strong sense, may require something more than syntactic manipulation of states.

It may require:

an internal continuous dynamics;
discrete stability points;
memory of trajectories;
capacity to distinguish its own states;
a loop in which the system becomes an object for itself.

In PT language:

plausible artificial consciousness would require a filter capable of filtering its own filtering dynamics.

This is not a proof. It is a hypothesis. But it has a strong consequence: if it is correct, then stacking digital processors will not necessarily be enough. It may produce richer and richer behavior, but not necessarily a system that locates itself inside its own dynamics.

For that, one might need an architecture where the discrete and the continuous cooperate from the material level onward: a field that varies, attractors that stabilize, outputs that become nameable, and then a loop that observes how those outputs arise.

7. Status of the hypothesis

The levels must be separated.

Material level. Analog, photonic, neuromorphic, and mixed-signal research already exists. It is motivated by energy, speed, latency, and integration with the physical world.

PT level. The discrete-continuous reading is central: the discrete is a persistence point of the constrained continuum. The role of p = 2, the q⁺/q⁻ bifurcation, attractors, and stable points are internal elements of the chain.

Computing level. The idea of attractor-based computing is a conceptual extension. It is coherent with PT and with hardware trends, but it is not yet a complete theory of computation.

Consciousness level. The idea that artificial consciousness requires the discrete-continuous dyad is philosophical and conjectural. It must be presented as such.

This caution does not weaken the hypothesis. It makes it testable, debatable, and improvable.

8. Conclusion: the post-binary

We are probably not moving toward computing without bits. We are moving toward computing where the bit changes status.

The bit would no longer be the ultimate atom of computation, but the readable form taken by a deeper process when it stabilizes. Computation would no longer only be the art of transforming symbols, but the art of making stable forms appear inside a constrained continuum.

From this perspective, analog, photonic, neuromorphic, and mixed-signal chips are not merely technical optimizations. They are the first signs of a deeper shift:

computation returning to its physical substrate.

And perhaps, one day, toward machines that will not merely produce states, but become able to observe the way their own states persist.

References

TDK / Hokkaido University - analog reservoir AI chip: press PDF
Cornell - microwave neural network chip: Nature Electronics
IBM Research - Transformer ALBERT on a 14 nm analog chip: IBM Research
Photonic tensor processors: open-access article
BrainScaleS-2 - analog neuromorphic inputs: arXiv:2602.04582

← All essays