Loading slide
Hallucination and bias are the deep ones. But there is a shorter, more practical list of things these tools are simply bad at, and most of them you can now explain yourself, because each traces back to something earlier in this course.
Counting and exact arithmetic. Remember from the words module that the model sees tokens, not individual letters or digits, the same reason it stumbled on the r's in "strawberry." A number like 4,096 may arrive as one or two chunks, not four separate digits it can line up and add. It often gets sums right because it saw similar sums in training, but it is pattern-matching arithmetic, not calculating it, and on an unfamiliar one it can confidently produce nonsense. For anything that must be exact, use a calculator, or ask it to use one.
Knowing what it does not know. We just covered this: there is no inner gauge of certainty, so it is as fluent when wrong as when right. It cannot reliably raise its hand and say "this part I am unsure about."
Anything after its training stopped. The weights froze the moment training ended, so the model's knowledge has a hard back edge, an unmoving line in the past. Events, prices, releases, anything past that line does not exist to it, unless a tool fetches the information and pastes it onto the board.
Holding instructions across a long conversation. Straight from this module: the board is finite. A rule you set at the start can slide off the top as the conversation grows, and once it is gone the model stops following it, with no idea it ever agreed to.
Careful step-by-step logic. Long proofs, intricate puzzles, tight chains of deduction where one slip ruins everything: this is where predicting the next plausible token shows its seams most clearly. It can look like reasoning while quietly skipping a step.
None of this makes the tools weak. It makes them shaped. Knowing the shape, where the tool is strong and where it is thin, is exactly what lets you reach for it at the right moments and reach for something else at the wrong ones.