Plain-language explainer
What LLMs can't do
What are the real limits of large language models?
The honest list: they cannot reliably know what they don't know, so confidence and correctness come apart. They have no memory beyond the context window unless a product bolts one on. Exact arithmetic and counting are unreliable without a calculator tool. They cannot check facts against the world, only against patterns in training text. And they cannot act, browse or run code by themselves; every 'agent' is a harness of tools and permissions wrapped around the same predictor.
Last reviewed
Reading is the slow way. Start with a free lesson you can operate right now.
Start free: Predict the next word →Free, no code, no signup.
Then go deeper: The edge of the map: what LLMs can't do Locked
What people get wrong
- The limits mean LLMs are unreliable at everything. With tools, retrieval and checks, they are dependable inside well-defined lanes.
- Scale will erase all of this. Some limits, like calibrated self-knowledge and grounding, are structural to prediction and have survived every scale-up so far.
- A wrong answer proves the technology is fake. Same system, different task fit. Knowing which tasks fit is the actual skill.
Where you see it in real products
- Chatbots ship with calculators and code interpreters because raw arithmetic is a known weak spot.
- Agent products ask permission before acting: the harness, not the model, is the safety layer.
- The 'AI can make mistakes' disclaimer under every chat box is this page, in one line.
Common questions
- What are LLMs genuinely bad at?
- Reliable arithmetic and counting without tools, staying consistent over very long tasks, knowing what they do not know, and anything that depends on information nobody wrote down. Tools patch several of these. None of them are solved by a bigger model alone.
- Will scaling fix these?
- Some have improved steadily with scale. The ones that come from the shape of the system, no persistent memory, no way to check a fact, no experience of the world, are not the kind of thing more parameters address. Products fix those with tools and retrieval.
- What should I never delegate unchecked?
- Anything where being confidently wrong is expensive: legal, medical and financial specifics, numbers you will act on, and anything irreversible. Use a model to draft and to explore, then verify the parts that carry consequences.
Related explainers
More in Building on it, and trusting it
- What is retrieval-augmented generation (RAG)?
- What are evals, and how do teams know an AI feature actually works?
- What is a vector database, and why does every RAG stack have one?
- What does LLM as a judge mean, and does it work?
- Should I use RAG or fine-tuning?
- What is a data flywheel, and why is it called the moat in AI?
One idea at a time, in your inbox
New lessons and explainers, written the way these pages are. Now and then, not daily, and never a sales sequence.
We email you a confirmation link first. Unsubscribe in one click, any time. Privacy.
Part of See How AI Works, a free interactive course, where you learn how modern AI works by operating it, not watching videos.