John Berryman is the author of the O'Reilly book "Prompt Engineering For LLMs" https://learning.oreilly.com/library/view/prompt-engineering-for/9781098156145/ Slides:
Slides, notes, links and more resources: https://parlance-labs.com/education/prompt_eng/berryman.html
*00:00 Introduction and Background*
John's career, from aerospace and search technology to GitHub Copilot.
*00:47 Understanding Large Language Models*
Definition and functionality of large language models. Importance of the "large" aspect. Historical progression: RNNs, attention mechanism, transformers. Emergence of models like BERT and GPT.
*05:33 Overview of Prompt Crafting Techniques*
Introduction to prompt crafting techniques. Focus on evolving techniques and recent trends.
*06:09 Few-Shot Prompting*
Technique: Controlling output with few-shot examples. Importance of setting predictable patterns.
*07:39 Chain of Thought Reasoning*
Addressing reasoning challenges in LLMs. Use of few-shot prompting to improve logical reasoning. CoT examples.
*10:36 Think Step by Step*
Simplification of chain of thought reasoning. Direct instruction to model for step-by-step thinking. Advantages: reduced need for extensive examples, prompt capacity management.
*12:25 Document Mimicry*
Technique of document mimicry in prompt crafting. Examples: transcripts, common document structures. Conditioning model with familiar patterns and formats like Markdown.
*16:01 Intuitions for Effective Prompt Crafting*
LLMs as "dumb mechanical humans." Use familiar language and constructs. Avoid overwhelming the model with too much information. Ensuring clarity in prompts.
*18:11 Building Applications with LLMs*
LLM applications as transformation layers. Converting user requests into LLM-compatible text. Process: user input, LLM processing, actionable outputs.
*19:33 Context Collection for Prompt Crafting*
Importance of context collection for prompt crafting. Steps: collecting, ranking, trimming, assembling context. Copilot example structure: file paths, snippets from open tabs, current document; document mimicry with comments. Importance of context relevance.
*25:27 Introduction of Chat Interfaces*
Shift to chat-based interfaces in LLM applications. Use of special syntax for role differentiation. Benefits of structured chat interactions.
*28:22 Function Calling and Tool Usage With LLMs*
Introduction and advantages of function calling. Structure: names, descriptions, arguments. Expansion of LLM capabilities with tool usage. Cycling through tool usage, tool responses, assistant responses.
*33:56 Example: Tool Calling in a Thermostat Application*
Detailed example: thermostat application. Process: user request, tool calling, context awareness. Iterative approach for better user interactions.
*38:14 Q&A*
Discussion on few-shot prompting best practices. Hyperparameter adjustments. Function calling complexities and solutions. Considerations for better code outputs and prompt tuning.