Train a transformer on enough text and it starts to do more than predict the next word. Large language models generate essays, answer questions, write code, and translate between languages. Scale, both in parameters and data, drives the capability.

These models learn statistical patterns across billions of tokens. They do not store facts in a database. Knowledge is distributed across weights, which is why they can be fluent and wrong at the same time. Hallucination is a natural consequence of the architecture, not a bug to be patched.

Common capabilities

• Text generation and summarization
• Question answering and reasoning
• Translation and rewriting
• Code generation and debugging
• Conversational interaction

Deployment raises issues. Training costs millions of dollars. Inference is expensive at scale. Copyright and privacy concerns accompany the training data. Alignment work tries to steer behaviour, but no method guarantees the model will always respond as intended.