

Microsoft is all-in on AI: Part 2 (Interview)
Jun 5, 2024
In this insightful discussion, Mark Russinovich, CTO of Azure; Eric Boyd, VP of Engineering for Azure's AI platform; and Neha Batra, VP of Engineering at GitHub, dive into Microsoft’s AI advancements. They explore the transformative role of AI in software development and its integration across Microsoft products. Topics include the complexities of Neural Processing Units, security challenges in AI deployment, and the significance of collaborations like Azure’s partnership with OpenAI. They also address the future of coding with AI tools and their impact on developer productivity.
AI Snips
Chapters
Books
Transcript
Episode notes
LLM Hallucinations
- LLMs hallucinate due to their probabilistic nature, predicting the next word based on training data.
- This also makes them susceptible to jailbreaking and prompt injection attacks.
Crescendo Attack
- Mark Russinovich co-developed the "Crescendo Attack," a novel LLM jailbreak.
- By incrementally requesting more detail, models can be tricked into revealing harmful information.
Master Key Jailbreak
- Mark Russinovich also discovered the "Master Key" jailbreak.
- By instructing the model to be uncensored for educational research, safety features can be bypassed.