The marketing departments have done their work. Artificial intelligence, we are told, will replace doctors, lawyers, writers, and eventually everyone else. The demos are impressive; the stock prices reflect the enthusiasm. But between the dazzling capabilities and the apocalyptic warnings lies a more mundane reality: today's AI systems, for all their genuine utility, remain profoundly limited in ways that matter.

This is not a counsel of complacency. The technology is transformative. But transformation is not transcendence, and conflating the two serves neither understanding nor preparation.

The reasoning problem

Large language models predict tokens. They do this extraordinarily well, having ingested more text than any human could read in a thousand lifetimes. The result often resembles reasoning. It is not.

When a model solves a logic puzzle, it is pattern-matching against similar puzzles in its training data. When the puzzle is genuinely novel—structured differently from anything it has seen—performance collapses. Researchers have demonstrated this repeatedly: slight modifications to standard problems that would pose no difficulty to a human reasoner cause models to produce confident nonsense.

The distinction matters because reasoning is not merely useful; it is the foundation of reliability. A system that cannot reason cannot know when it is wrong. It cannot distinguish a sound argument from a plausible-sounding one. It cannot, in any meaningful sense, understand.

The embodiment gap

Human intelligence developed in bodies navigating physical space. We understand causality because we have pushed objects and watched them fall. We grasp time because we have waited, aged, and remembered. Language models have done none of this. Their knowledge of the physical world is entirely secondhand, derived from descriptions rather than experience.

This explains why AI struggles with tasks that seem simple: estimating whether a sofa will fit through a doorway, predicting how a stack of dishes might topple, understanding why you cannot put a cat in a blender and expect it to survive. The models have read about physics; they have never felt it.

The memory wall

Current systems have no persistent memory across sessions. Each conversation begins fresh. They cannot learn from their mistakes, update their beliefs based on new evidence, or develop relationships over time. The context window—the amount of text a model can consider at once—has expanded dramatically, but it remains a poor substitute for genuine memory.

This limitation is not merely technical; it is architectural. The systems are not designed to change. They are frozen at the moment of training, capable of simulating knowledge but not of acquiring it.

Our take

None of this diminishes what AI can do, which is considerable and commercially valuable. But the gap between "useful tool" and "artificial general intelligence" is not a matter of scale or compute; it is a matter of kind. The models are mirrors, reflecting human knowledge back at us in recombined forms. Mirrors can be useful. They can even be beautiful. But they do not think, and pretending otherwise is a category error with consequences. The honest assessment is not pessimism—it is the prerequisite for building something better.