CS50x 2025 - Artificial Intelligence

A short system instruction like "you are a rubber duck" induces playful, domain-specific behavior because the model follows explicit role-playing instructions when generating replies.

A system prompt shapes an AI model's behavior because the model conditions its generations on high-level instructions that bias responses toward a specified role, tone, or domain.

Using third-party APIs lets teams build AI features without training their own models because API providers supply powerful pretrained LLMs and infrastructure that avoid the cost and complexity of model training and maintenance.

Decision trees translate directly into imperative if/else code because each tree node is a conditional test and each branch corresponds to the code path that implements that test.

Exhaustive deterministic search breaks down for complex games because each additional move multiplies the branching factor, creating exponentially many sequences that are infeasible to enumerate.

Generative AI creates new text, images, or video because models learn statistical patterns from massive corpora and then predict likely continuations or samples consistent with those patterns.

Large language models produce answers probabilistically because they estimate token probabilities from massive text corpora and sample likely continuations rather than asserting guaranteed facts.

Large language models represent tokens as high-dimensional numeric vectors and use attention to compute dynamic weights between tokens, which lets the model emphasize relevant word relationships for contextual understanding.