What is today's AI genuinely good at?

A large language model is, at its core, a system trained to predict likely text. That sounds narrow, but it turns out to cover a lot of useful work. Models are strong at tasks where the answer is mostly about shaping language: drafting an email, summarizing a long document, translating between languages, rewriting in a different tone, explaining an idea at the right level, and writing or fixing code.

They are also good at pattern recognition over text you give them. Hand a model a messy block of notes and ask for the action items, or a contract and ask which clauses look unusual, and it often does well, because the relevant information is right there in front of it rather than dredged up from memory.

The honest picture is that the strengths and the weaknesses come from the same place. A model that predicts plausible text will produce fluent, helpful answers most of the time, and it will also produce fluent, confident answers when it should not. The failures are not random glitches. They cluster in a few predictable areas, and the rest of this topic walks through them one at a time: making things up, not knowing recent events, not truly understanding what it writes, and stumbling on exact math and counting.