Natural Language as a General-Purpose Programming Medium
What our work with AI Agents and Large Language Models, especially in the context of specification-driven development for highly regulated enterprise use cases, has gradually made visible is a deeper hypothesis about the direction of modern AI. At its core, modern AI appears to be moving toward the transformation of natural language into something closer to a general-purpose programming medium, one that can be interpreted, operationalized, and executed by computational systems. This prospect is powerful, but it immediately raises a foundational difficulty. A medium that begins to function in this way cannot remain at the level of loose conversational usefulness alone. It must eventually support a sufficiently disciplined interplay of syntax, semantics, and pragmatics, so that interpretation becomes controllable, execution becomes reliable, and at least part of its behavior can be subjected to verification.
It is precisely at this point that our work began to reveal a deeper conceptual tension. The traditional expectation would be that such executability can emerge only once semantics has been sufficiently stabilized, with pragmatics treated as a secondary contextual layer added on top of a prior core of meaning. Yet our experience increasingly suggested that this picture may be incomplete, or at least no longer adequate as a general architectural assumption. In practice, current AI systems often achieve substantial pragmatic effectiveness before they possess anything like a fully explicit, inspectable, and globally coherent semantic-rational core. This suggests not merely an engineering gap, but a more fundamental deficit in the way pragmatics has been formalized in the computational tradition. Some inherited assumptions about the relation between semantics and pragmatics may therefore need not just refinement, but partial reversal.
Pragmatics Before Full Semantic Stabilization
The recent progress of Large Language Models has made it increasingly difficult to preserve one of the older background assumptions of formal computer science, namely that semantic determination must, in a strong architectural sense, come first, while pragmatic use enters only later as a secondary layer. In the classical picture, a language becomes executable because its meaning has already been sufficiently stabilized. Once that settlement exists, the machine can compile, interpret, verify, or optimize expressions under a fixed regime of meaning. This picture remains one of the great achievements of computer science, and nothing in the present proposal seeks to dismiss it. However, the emergence of powerful language-based AI systems shows that this model is no longer sufficient as a general account of how useful machine intelligence operates in practice [MEANING-SEP] [PRAGMATICS-SEP].
What current AI systems demonstrate, often quite spectacularly, is a form of pragmatic competence that precedes any fully explicit, inspectable, and globally coherent semantic-rational core. They can answer, summarize, translate, reformulate, plan, classify, and sometimes even support complex technical work. Yet the basis of this usefulness is still only partially disciplined with respect to contradiction management, provenance, scope control, assumption tracking, and evidentiary justification. This is precisely why the same systems that can be remarkably useful can also be unreliable, overconfident, or difficult to audit. The literature on hallucination and broader model risk has made this problem impossible to ignore: fluent output is not the same thing as epistemically controlled reasoning [BENDER-2021].
The concern that motivates Meta-Rational Pragmatics (MRP) begins exactly at this point. We do not introduce it in order to deny semantics, nor in order to celebrate vagueness or replace formal rigor with fashionable language. On the contrary, the proposal begins from the claim that rigor must be extended, not abandoned. The problem is that, in realistic settings, the central difficulty is often not only to reason correctly within a fixed semantic regime, but to determine under which interpretive policy a fragment should be read, formalized, checked, deferred, revised, or left intentionally ambiguous. Many real tasks do not begin from well-formed premises. They begin from partial knowledge, underspecified goals, conflicting evidence, unstable vocabularies, and several plausible ways of carving the problem. In such conditions, it is no longer enough to ask what an expression means under a fixed semantics. We must also ask under what policy of interpretation it should be treated at all [PRAGMATICS-SEP].
Wittgenstein, Use, and Situated Sense
This is the point at which the later Wittgenstein becomes especially relevant. If the early Wittgenstein is often associated with the hope that logical form may sharply delimit sense, the later Wittgenstein, especially in Philosophical Investigations, moves in a different direction. Meaning is no longer approached primarily as a static correspondence relation captured once and for all by ideal logical form. Instead, meaning is inseparable from use, from practice, from the language-game in which an expression functions, and ultimately from a form of life within which such use becomes intelligible [WITTGENSTEIN-1953] [BILETZKI-2023]. The philosophical importance of this move for our purposes is considerable. It suggests that sense is not always something fully settled prior to use; rather, in many cases, sense emerges through situated participation in rule-governed yet open-ended practices.
For AI, this Wittgensteinian shift has a direct computational implication. If meaning is deeply conditioned by use, context, and practical embedding, then one should expect that systems operating close to natural language will often display competence before they can offer a clean and unified semantic reconstruction of their own behavior. But this should not lead us into complacency. The lesson is not that anything contextual is therefore acceptable, nor that reasoning can be reduced to conversational fluency. The more demanding lesson is that a serious AI architecture must represent and govern the transition between contextual interpretation and more disciplined forms of formalization. It must know, at least partially and operationally, what kind of interpretive regime is active, how strong its assumptions are, what kind of evidence it presupposes, and when it should refuse premature collapse into a single reading. In that sense, Meta-Rational Pragmatics can be understood as an attempt to formalize not just reasoning inside a frame, but reasoning about the suitability, limits, and revision of the frame itself.
The Meta-Rational Layer Between Flexibility and Execution
The adjective meta-rational is therefore not used here in any mystical or anti-rational sense. It does not mean irrationality, nor some vague appeal to going beyond logic. It means that the system must be able to reason not only within a local inferential regime, but also about the status of that regime relative to the current task. The analogy with metacognition is deliberate. In psychology and education research, metacognition refers to the monitoring and regulation of one’s own cognitive processes, including planning, checking, evaluating, and revising strategies of thought [FLAVELL-1979]. We are not claiming that MRP is a psychological theory, but the analogy is productive: once interpretation itself becomes an explicit operational concern, an AI system must do something structurally similar. It must monitor how it is interpreting, how strongly it should commit, which assumptions are active, and when the current mode of reasoning has become too brittle, too expensive, too permissive, or too shallow [FLAVELL-1979].
This is why MRP matters for the present deliverable. It gives us a conceptual vocabulary for a missing layer between natural-language flexibility and disciplined execution. If one remains only at the level of ordinary prompting, the system tends to compress multiple competing interpretations into one smooth textual continuation. That is often useful, but it is a poor foundation for verification, structured decomposition, epistemic typing, and scientific traceability. Conversely, if one insists that useful execution can begin only after complete formal semantic stabilization, one excludes from the start many of the tasks for which contemporary AI is most promising. The architectural challenge is therefore to create a runtime in which candidate interpretations can be generated, typed, compared, revised, and selectively formalized under explicit policies, rather than silently collapsed into a single answer. This is the deeper reason why we are concerned with Meta-Rational Pragmatics: not because it is terminologically fashionable, but because it names a real engineering requirement that becomes visible once AI systems are used seriously in research, analysis, and knowledge-intensive workflows.
The VM and the Active Theory Frame
It is in this broader perspective that the Meta-Rational Pragmatics Virtual Machine should be understood. The VM is conceived as a runtime for policy-governed interpretation and execution. Its central computational object is not simply a prompt, a chat history, or a generic workspace, but what we call the active theory frame: a bounded, goal-conditioned assembly of theories, facts, constraints, hypotheses, tools, and admissible inference policies. This notion is important because practical failure in advanced AI systems often comes not from lack of inferential power in the abstract, but from activating the wrong subset of knowledge, combining incompatible assumptions too early, or failing to distinguish between tentative formalization and verified structure. The active theory frame is meant to be the unit through which such choices become explicit and governable.
From this point of view, the forthcoming chapters can be read as an effort to make this conceptual direction computationally operational. They will move from the philosophical and architectural motivation toward more explicit design choices concerning controlled representations, policy-governed interpretation, epistemic memory, bounded reasoning contexts, and the interplay between heuristic generation and stronger verification regimes. The purpose of the present chapter is therefore intentionally introductory but also programmatic. It does not claim that Meta-Rational Pragmatics is already a finished theory. It claims something more modest and, in our view, more defensible: that the current generation of AI systems has made visible a real structural problem, namely the gap between pragmatic performance and disciplined reasoning, and that this gap cannot be adequately addressed either by pure prompt engineering or by nostalgia for fixed semantic frameworks alone.
Programmatic Conclusion
A final point deserves emphasis. The proposal should not be misunderstood as a claim to universal superiority. On the contrary, one of the lessons from the theory of optimization and search is that there is no universally best procedure across all possible problem classes [WOLPERT-1997]. A viable architecture must therefore make its biases explicit, recognize structure early, and select regimes of interpretation and reasoning in ways that are aligned with the local problem. In this sense, Meta-Rational Pragmatics is not an escape from the limits identified by the No Free Lunch perspective. It is a disciplined response to them. Instead of pretending that one fixed semantic machine, one solver, or one style of reasoning will fit everything, it begins from the premise that intelligence becomes more robust when the governance of interpretation itself is made explicit.
For all these reasons, Meta-Rational Pragmatics is introduced here not as a decorative philosophical label, but as a serious attempt to define the missing middle layer between natural-language fluidity and formally disciplined execution. The VM discussed later in the series is one exploratory answer to that need. The chapters that follow will therefore treat this perspective not as an abstract manifesto, but as the conceptual opening for a more concrete architectural and technical discussion.
References
- [BENDER-2021] Bender, E. M., et al. (2021). On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?
- [BILETZKI-2023] Biletzki, A., & Matar, A. (2023). Ludwig Wittgenstein.
- [FLAVELL-1979] Flavell, J. H. (1979). Metacognition and Cognitive Monitoring.
- [MEANING-SEP] Speaks, J. (2024). Theories of Meaning.
- [PRAGMATICS-SEP] Bach, K. (2024). Pragmatics.
- [WITTGENSTEIN-1953] Wittgenstein, L. (1953/2009). Philosophical Investigations.
- [WOLPERT-1997] Wolpert, D. H., & Macready, W. G. (1997). No Free Lunch Theorems for Optimization.