What LLMs Can't Do

Language models can pass difficult exams, write code, and explain quantum physics. So it's tempting to think that they can do anything. A professional knows where the edges are. Here's an honest list of what goes wrong, and what to do about it.

• They pass difficult exams
• Tempting to trust them fully
• Know the edges

Limits and remedies

They make things up, so ground them in sources. Their knowledge stops at a cut-off date, so give them search. They're poor at exact maths, so give them a calculator. Their answers vary, so test them over many runs. They can be persuaded by the text that they read, so never let them hold power. And they reflect the biases in their data, so measure the fairness.

When not to use one

So when should you not use a language model? When rules are clear and exact. When a mistake could seriously hurt someone, and nobody's checking. When you need the exact same answer every time. When data can't legally leave your building. Or when a simple tool does the job. Saying no to an LLM is often the sign of a strong engineer.

Summary