Intelligence as Aimed Progression Under Semantic Feedback

Current AI systems, in particular large language models, excel at generating plausible continuations, but lack a principled mechanism for determining whether a continuation should be pursued, revised, or abandoned. Prediction alone is insufficient for reliable long-term interaction. We propose a process-oriented view in which intelligence is understood as the capacity to sustain aimed progression under semantic feedback. In this perspective, prediction generates candidate continuations, while semantic feedback evaluates whether they can be coherently sustained over time. When relevant invariants are violated, interaction enters a state of rupture that requires repair or revision. To formalize this perspective, we introduce a round-based framework of viable computation, in which progression is regulated by invariant supervision rather than judged solely by final outputs. This leads to the {\em Intelligence Thesis:} intelligence is not the ability to produce correct outputs, but the ability to sustain aimed progression over time. Explainability, robustness, resilience, and alignment emerge naturally as structural consequences of maintaining coherent continuation under constraints. The same perspective also reframes notions such as superintelligence and singularity. Rather than representing a qualitative break or a new computational principle, they correspond to an expansion of the range of conditions under which aimed progression can be sustained. The framework can be summarized by a {\em Principle of Viability:} a system must avoid what is forbidden, maintain what is required, and revise its invariants when continuation is no longer possible, at a non-negligible cost.