Building and using AI models responsibly
summary
The session shows concrete tooling and numbers: Microsoft ships Azure AI Content Safety (text and images, four categories: hate, sexual, self‑harm, violent) with configurable severity labels, blocklists, monitoring, and integration into Azure OpenAI. A Prompt Flow in Azure Machine Learning runs retrieval‑augmented generation, evaluation with GPT‑4, and prompt variants, improving groundedness from ~3.6 to 4.7 and pass rate from 70% to 94%, then deploys and monitors the model in production.
AI-written summary · about
chapters (25)
- 0:00Introduction
- 1:51Goals
- 4:08Innovation
- 4:44generative AI
- 8:35content safety
- 9:06content categories
- 10:10new domains
- 10:40Coup
- 11:12Demo
- 16:53Abuse Detection
- 19:53Measurement
- 20:13Meta Prompt
- 22:43Business Critical Measurements
- 23:19My Demo
- 28:03Evaluations and Measurements
- 29:22Evaluation Pipeline
- 30:40Bulk Test
- 31:37Evaluation Run
- 32:36Prompt Engineering
- 38:03Deploy
- 38:25Test
- 38:55Monitoring
- 40:04Summary
- 42:40Dashboard
- 45:25Recap
related talks
48mDeploy large language models responsibly with Azure AI
31mAzure AI Innovations at Microsoft
43mOperationalize AI responsibly with Azure AI Studio
45mSecuring and Governing Generative AI: Learnings from Microsoft
01h 00m5 Levels of AI in Microsoft Azure
50mAzure AI, Power new possibilities for every organization
18mTop-of-mind on AI with Mark Russinovich
47mSafeguard your copilot with Azure AI