This talk focuses on the practical process of building AI application prototypes, distinct from deep machine learning research or full-scale product deployment.
Large language models are general-purpose, probabilistic systems that predict the next word, requiring careful guidance to perform specific tasks.
Prompt engineering involves defining a persona, describing the task, and using techniques like few-shot prompting or chain of thought to guide the model's output.
Retrieval augmented generation (RAG) is the process of pulling relevant, personalized context from various data sources to include in the prompt for better results.
Since AI models are probabilistic, performance is measured at scale using offline evaluation frameworks and online A/B testing with user telemetry.
Effective AI user experience design prioritizes user control, allows for error correction, and balances model accuracy with system latency.
AI interfaces range from inline suggestions and text transformations to chat, each with different implications for user workflow and control.
Structured, multi-step AI workflows like Copilot Workspace give users control at key stages, improving accuracy and trust over a simple chat interface.
Developers have a responsibility to question whether AI is appropriate for a given application, especially in sensitive domains like finance or healthcare.
25:17Scott Hanselman
35:16Ankit Patel
30:24Markus Walker
41:06Pascal Reddig, TJ Griffiths, Fabian Schmidt, Oliver Winzenried & Matthias Niehoff & Mirko Ross
42:38Krzystof Czieslak
31:12Daniel Töws
31:39Rainer Stropek
23:55Zan Markan
Jobs that call for the skills explored in this talk.
