Large Language Models and Generative AI

How does a large language model transform text into internal numerical representations? In what sense is a language model predicting the next token, and why is that powerful? How do mechanisms like attention and context windows shape what the model can remember and relate? What are the limits of token-based prediction when trying to model human meaning and understanding?

Which kinds of tasks do large language models excel at, and which remain fundamentally hard for them? Why can language models appear to understand but still produce confident nonsense or errors? What are hallucinations, and how can users recognize when an answer is likely to be fabricated? How should we interpret benchmarks and demos when judging what these systems can really do?

How does the phrasing, structure, and context of a prompt influence a model's behavior? What are effective strategies for steering models toward safer, more reliable outputs? Why is aligning AI behavior with human values and intentions such a difficult technical and social problem? How much control should ordinary users have over model behavior, and what guardrails are essential?