Capability lists date quickly. Limitation lists date more slowly, because the things these systems are bad at tend to be bad for structural reasons.
Knowing what they do not know
A model produces an answer with the same confidence whether it is well grounded or invented. Calibration has improved and the underlying issue remains: the system is generating plausible continuations, and a plausible wrong answer looks exactly like a plausible right one.
Reliable arithmetic and counting
Better than it was, still unreliable at scale. These are pattern systems, not calculators, and the workaround in serious applications is to give them a calculator rather than to trust the pattern.
Consistency across a long task
Give a model a set of rules and a thousand items, and it applies the rules well early and drifts later. Nothing enforces consistency the way a program would.
Knowing anything after training
Models have a cutoff, and without a search tool they will answer questions about recent events by producing something that sounds right. The failure is silent, which is what makes it dangerous.
Genuine causal reasoning
Correlation in training data is not causation, and these systems learn correlation. They can describe causal relationships found in text and cannot reliably work out new ones.
Understanding physical space
Counting objects, judging distances, reasoning about what happens if something is moved. Multimodal models describe images fluently and reason about geometry poorly.
Why this list matters
Because every one of these failures is silent. The system does not stop and say it is out of its depth; it produces something.
Knowing where the edges are is the practical skill, and it is more durable than knowing what the current best model scored on a benchmark.