AI/ML Learning

Structure prompts with five components: role, context, task, format, and constraints. Treat prompting as an iterative loop — run the prompt, identify what fell short, revise one component, and re-run until output is reliable.

Start with intuition before math: understand supervised, unsupervised, and reinforcement learning through real examples, then build up to neural networks, gradient descent, and transformers. A layered curriculum prevents cognitive overload from jumping straight into formalism.

LLMs generate statistically plausible text rather than retrieving verified facts, and have no reliable sense of uncertainty. Be skeptical of specific numbers, dates, citations, and anything after the training cutoff, and verify high-stakes claims independently.

A transformer lets every word attend directly to every other word using an attention mechanism with Query, Key, and Value vectors. This solves the memory degradation problem of RNNs, which processed text strictly left to right and forgot earlier context.

Always read generated code before running it, since it can be plausible but wrong, insecure, or mismatched to your context. Trust it less for recent events, precise numbers, narrow domains you know well, and any high-stakes output.