What Large Language Models Are Good At, and Where They Fail

Strong at transformation

Rewriting, summarising, translating between formats and explaining unfamiliar code. In each case the source material is supplied and the model reshapes it, which is where accuracy is highest.

Weak at recall and arithmetic

Asked for facts from memory, a model produces text shaped like an answer whether or not it knows one. Anything checkable should be checked, and anything numerical should be computed elsewhere.

Designing around the failure mode

Give the model the source. Retrieval, tool use and citations turn a confident guesser into a competent assistant, because the answer is grounded in something you can verify.