LLM Foundations: How Large Language Models Work
- Generative AI
- Enterprise h2oGPTe
- H2O Cascade
This lesson recaps the foundational concepts behind large language models before the hands-on labs begin. It covers what makes a model an LLM, including emergent capabilities and the next-token prediction objective, along with the encoder and decoder architectures behind them. It walks through retrieval-augmented generation as a way to ground answers in a stored document corpus, plus prompt engineering techniques like zero-shot and chain-of-thought prompting, evaluation methods such as LLM-as-judge and perplexity scoring, and guardrails for keeping a deployed model on task.
Key takeaways
- What emergent capabilities are and why they only appear in large language models during training, unprompted
- How retrieval-augmented generation lets a model answer questions using a stored corpus of documents
- Why chain-of-thought prompting, asking a model to reason step by step, tends to improve its answers
- What guardrails do to keep a deployed model on task, reduce hallucinations, and limit bias
- How zero-shot, one-shot, and few-shot prompting differ in how many examples a model is given
Start owning your AI