Structural Limitations of LLMs

The most important thing to understand about LLM's is that they don't know things, they just predict words.

When a model gives you a confident, detailed answer, it is not retrieving a fact.

It is generating the most statistically likely sequence of words based on patterns from its training data. This is why hallucination is not a bug you can patch.

Beyond that, LLMs operate within a fixed context window, have no memory between sessions and have no grounded sense of time. Anything outside the context window simply does not exist for the model. Yesterday's conversation is completely gone unless you engineer a way to bring it back. Anything that happened after the training cutoff is unknown for live data, prices, or current events, no prompt will fix that. You need external systems to supply what the model structurally cannot.

Two more limitations are worth keeping close.

1. precise reasoning math's, logic, multi-step calculations is brittle in LLMs because token prediction is not the same as computation. Route deterministic tasks to deterministic tools.
2. LLMs cannot reliably signal when they are guessing. A confident-sounding wrong answer looks identical to a confident-sounding right one. You cannot trust apparent certainty as a proxy for actual accuracy.