The gap between what artificial intelligence appears to do and what it actually does has never been wider. In boardrooms and newsrooms, the prevailing assumption is that we are witnessing the birth of general intelligence, a technology that will shortly render human expertise obsolete across every domain. The reality is more interesting and more limited.

Large language models are, at their core, extraordinarily sophisticated pattern-completion engines. They predict the next token in a sequence based on statistical regularities learned from vast corpora of text. This makes them uncannily good at tasks that humans find tedious—summarization, translation, first-draft generation—and startlingly poor at tasks that humans find trivial.

The arithmetic problem

Ask a frontier model to multiply two four-digit numbers and it will often fail. Not because multiplication is hard, but because multiplication is not what these systems do. They do not compute; they interpolate. When a model produces the correct answer to a math problem, it is typically because similar problems appeared in training data, not because it has derived the answer through logical steps. This distinction matters enormously. A system that appears to reason but actually pattern-matches will fail unpredictably, in precisely the situations where reasoning would be most valuable—novel problems, edge cases, high-stakes decisions.

The knowledge boundary

Models do not know what they do not know. They cannot reliably distinguish between confident recall and confident fabrication. This is not a bug that better training will fix; it is an architectural feature. The same mechanism that allows a model to generate fluent prose about quantum mechanics allows it to generate equally fluent prose about quantum mechanics that happens to be wrong. Users who treat these systems as oracles rather than drafting assistants will be burned, repeatedly.

The planning deficit

Perhaps the most consequential limitation is the inability to plan over extended horizons. Models can generate impressive-sounding strategies, but they cannot execute multi-step plans that require maintaining state, adapting to feedback, and recovering from errors. They operate in a kind of eternal present, each response generated without genuine memory of what came before or anticipation of what comes next. This is why autonomous AI agents remain brittle despite years of research—the underlying models lack the temporal reasoning that planning requires.

Our take

None of this diminishes the genuine utility of current AI systems, which are already transforming how knowledge work gets done. But the transformation is augmentation, not replacement. The professionals who thrive will be those who understand both what these tools can do and, crucially, what they cannot. The hype serves neither investors nor practitioners. Clear-eyed assessment does.