Solve the Wheel, Break the Chariot

There is a special compression of experience. The boundaries of the room disappear, the shoulders become rigid, breathing is arrested and the world is reduced to the object at hand. This is a privilege fundamental to most excellent work, but also a reflection of one of the most ancient limitations of the mind. As soon as the field of experience is compressed to a point, that point seems to represent the entire world of experience.

AI systems are now sufficiently fluent to make the trap apparent from the outside and amenable to analysis. A model performs a task, defines its boundaries, identifies its important objects and achieves real competence. Ultimately, the boundary that permitted performance of the task is converted to represent the entire situation. This provides an apparent discovery of the true object and continues the process of optimization within a framework for which validity no longer is confirmed. This reflects a reification of the boundary and represents a technical problem well in advance of its philosophical implications.

Contemporary models work by relation. Tokens receive their force from context, events are represented in relation to their neighbors, attention conveys information across a field and agents integrate a local representation in which an error, file, request, service or plan becomes interpretable. Skill results from patterns of relation at many levels. The same mechanisms that produce an informative local representation can also lead to confusion between local representations and the overall situation. Increased capacity for context does not solve this problem. Rather, it provides the model with more information, but without resolution of the scale at which the problem actually exists.

The best example is the software agent that observes an error trace, opens the file, traces one import, modifies a local service and immediately succeeds in testing. With this level of detail, the work is clearly highly skilled. But the actual problem was a product flow, a common domain model, a deployment path and a user expectation. The agent solved the wheel and broke the chariot, and continued to respect the file long after the activity had reached the level of the entire system. Most apparent failures of shallow reasoning represent failures of scale traversal, i.e., an inability to transfer between levels at which a particular phenomenon is valid.

Human perception provides a compact link. The same ambiguous stimulus represents an H within THE and an A within CAT. Context is not imposed following recognition of the letter, but instead accompanies perception from the beginning. The word superiority effect, established through Reicher's forced-choice experiments and modelled by McClelland and Rumelhart's interactive activation, shows people reading a letter more accurately inside a real word than alone. In this situation, the letter facilitates identification of the word, and the word facilitates identification of the letter. Lower evidence and higher expectation settle together. Intelligence is this reciprocal movement across levels, and a system that only travels upward loses the detail, while one that only travels down loses the form that gives detail its role.

This is where Nagarjuna earns his place, after the engineering problem has already been named. Madhyamaka employs a substantial vocabulary of metaphysical concepts, but provides a simple and highly appropriate working diagnosis. Confusion arises when we confuse dependent designation with intrinsic nature. A boundary can be conventionally valid, practically indispensable, and still hold no self-standing essence. Distinctions remain necessary. They work through dependence rather than through any nature of their own.

His clearest example is the chariot. It is not the wheels, axle, frame, yoke or reins, nor a floating entity independent of them. Instead, it is designated conventionally by components and configuration and by use and function. Remove sufficient components from this field and the chariot disappears, even though some elements remain in hand.

Artificial intelligence systems continuously manipulate objects of chariot form. A service, error, intention, violation of safety, or plan of an agent are all designations within a field of dependencies, and not self-contained things.

Consideration of a bug as limited to its failing line results in correction of a symptom, not of a cause. Consideration of a service as limited to its directory of source code results in loss of all contracts, adaptations, latencies and expectations that contribute to the nature of the service. Nagarjuna does not tell an engineer how to write a scheduler. He gives a precise warning. If the object is dependently designated, the right boundary is task-sensitive and revisable.

Madhyamaka precision is worth the effort because it avoids two easy errors. One is naive realism in which the natural unit of analysis is simply given. The other is an undisciplined relativism in which every boundary dissolves. The disciplined middle is more useful. Boundaries are conventional, dependent and fully adequate for their purposes. They are exactly appropriate for a given task and immediately misleading when extended to another scale. In engineering terms, a boundary is a contract for an interface with a defined scope, and should produce a corresponding loss of confidence outside that scope. This is equivalent to the position of place expressed by Nishida Kitaro. Each local object is understood within a field of experience that is not simply another object adjacent to those it contains. The repository is not just another file. Finally, a system that considers its current bounded context as representing the whole of meaningful experience has allowed the map to consume the territory.

A boundary-fluid system would not be boundaryless, because every task demands selective attention. It would establish boundaries explicitly, apply them temporarily and revise them in response to stress. It would regard frames as tools and not as final interpretations, and would develop the ability to zoom in and out as routinely as to retrieve or to plan.

Such operations would yield straightforward questions concerning the nature of work. Which boundary of operation am I currently using? For which aspects of the task is it informative and for which is it uninformative? In what larger context does this object acquire meaning? Which smaller mechanism would invalidate my present interpretation of the task? What results would indicate that the task is operating at a different level of organization?

These are philosophical but easily operationalized questions. The agent must characterize the current level of analysis, produce the neighboring scales prior to action and maintain an interactive representation of the relations among function, module, service, product and user. Finally, a passing integration of results in combination with a failure should represent an indicator of boundary error and not of noise.

The ideal capability is well shaped. Sufficient stability to operate, sufficient flexibility to adjust the course. Fix the wheel, then evaluate the chariot.

None of this requires AI to confirm an ancient philosophy by replicating it in silicon. A language model does not experience emptiness, because it operates with contextual embeddings. The resulting statement is more limited, and thus more compelling.

Modern systems embody a practical form of an ancient confusion regarding the relation of conventions to error. Intelligence is based on conventions, and error results from misinterpretation of a conventional mode of representation as self-grounding. The same awareness that would make a machine wiser is available to the reader of this text. The restriction that accompanies every act of attention need not become a wall, but may instead serve to define limits, accomplish work within them and retain awareness of the larger context from which they were derived. Ride the chariot, repair the wheel and maintain awareness of the wider field.