The bottleneck
Consider a cat in a room. The cat emits photons. The photons reach the retina, are transduced into action potentials, and propagate through the optic nerve to V1 and downstream associative areas. At every step of this chain the cat is no longer there. There is only electrochemical activity. Yet when you close your eyes, the cat is there. Something happened along the way that converted a distributed physical activity into an internal representation reconvocable in the absence of the source.
In artificial systems, the question is different. The cat does not leave the chain in the same sense. Pixels can be encoded, but a set of pixels never enters the system unaccompanied. It is tagged in a semantic context register, either explicitly (a class label in supervised learning, a prompt in multimodal systems, a caption in image text alignment), or implicitly through training (the dataset already contains the semantic association between regions of pixels and labels, the pretraining corpus already pairs images with text descriptions). The system does not discover the cat. It learns a correspondence between a distribution of pixels and a semantic tag that pre exists in the register in which the system was trained. A cortical column receives no such tag. It receives a stripped afferent signal. There is no cat label attached to the action potentials reaching its dendrites. The representation must be constructed without the help of an external semantic register. And once constructed, it can be reactivated when the eyes close.
This is the bottleneck. Biological cognitive systems possess a transducer that takes a tagless physical signal and produces an internal representation reconvocable without the source, including from a non contextual internal trigger. Artificial systems do not. They simulate the output of such a transducer through tag matching, classification, embedding, generation, without ever building one. In a biological system, the cat can be reactivated by an internal trigger (a homeostatic variation, an emotional state, a diffuse associative cue) that does not itself carry the cat semantics. In an artificial system, the cat never reappears without an external stimulus that already carries the cat context.
A bottleneck structurally hard to address
The problem is intrinsically transdisciplinary. It requires neuroscience to characterize the afferent signal entering cortical columns, artificial intelligence to construct candidate transducers, and philosophy of mind to formulate the underlying question with sufficient precision. There is no academic structure whose mandate covers this combination, and no funding mechanism designed to host research that produces a structural specification rather than a publishable phenomenological finding or a deployable artifact. Programs that sit at the intersection of these fields fall in an institutional gap.
A second factor compounds the first. Mainstream artificial intelligence has built its architectures on the paradigm of external stimulus and learned response. The system reacts to an input, learns from a reward or a label, and produces an output. There is no endogenous internal state from which behavior originates. As long as research operates within this paradigm, the question of transduction from a non contextual internal stimulus to a contextualized representation does not arise. It is not that researchers chose not to work on the problem. The problem does not appear as a problem within the dominant frame. This invisibility is the deepest layer of the bottleneck: a research question cannot be addressed if the paradigm in which the field operates does not let it become visible.
Hypothesis and pilot
Hawkins (A Thousand Brains, 2021) proposes a framework in which cortical columns build internal representations through movement, prediction, voting, and reference frames. The framework is a useful starting point on how a column constructs a representation from sensory input, although significant open questions remain on which dimensions structure these representations and on the basis on which these dimensions emerge within the column. The framework does not address how a representation, once constructed, can be reactivated in the absence of external sensory input. Hawkins offers an entry into how the column builds. He does not answer how the cat reappears when the eyes close.
The hypothesis is that two structural conditions are required for a system to display the full property. First, an architecture compatible with the Hawkins framework, capable of building situated representations from afferent signals. Second, an endogenous internal state that operates as a motor of reactivation, capable of triggering a contextualized representation from a non contextual internal signal (homeostatic, attentional, associative).
Work on the second condition (an endogenous internal state as a motor of behavior and reactivation) is already underway in independent research efforts. Work on the first condition exists in fragments through Hawkins inspired implementations. What does not exist is the structural specification of the afferent signal that would allow the two to be coupled and tested together. This is the bottleneck the pilot addresses.
The pilot is to characterize the structural properties that an afferent signal must possess to be translatable by a Hawkins style cortical column architecture (dimensionality, temporal organization, invariants, coupling to motor and homeostatic state), and to translate this characterization into a target specification usable by independent teams. Once the specification exists, candidate transducers can be built and tested by feeding their output into Hawkins inspired column models, and the resulting system, coupled with an endogenous internal state, can be tested for the full property of reactivation from a non contextual internal stimulus. The work is now tractable because the current generation of artificial intelligence systems provides components (pretrained models, open source frameworks, accessible compute, modular tooling) that allow such a program to be assembled rather than built from scratch.
What this opens
Lifting this bottleneck would open two paths. The first is a path toward artificial cognitive systems whose operation is grounded in biological mechanisms rather than in scaled approximations of their outputs. The second, more immediate, is a new class of human machine interfaces operating along a unidirectional channel, from machine to living system. Once equipped with the transducer, the machine could produce a non contextual signal capable of triggering, in a biological system already endowed with its own transduction mechanisms, a contextualized representation. The reverse direction remains structurally blocked, even if the machine itself acquired the full property. Two systems that built their internal representations through their own histories of inputs and interactions cannot directly exchange those representations, in the same way two cortical columns from different brains cannot read each other’s content. What is not transferable is not the capacity, it is the history of construction.