AI prompt engineering in 2025: What works and what doesn’t | Sander Schulhoff (Learn Prompting, HackAPrompt)

In this episode, host Lenny speaks with Sander Schulhoff, a pioneer in prompt engineering and AI security. The discussion starts with the significance of prompt engineering and how effective prompts can greatly affect AI performance, with evidence suggesting that good prompts can boost output accuracy significantly. Sander shares his experiences and findings from the Prompt Report, a comprehensive study analyzing over 1,500 papers which outlines numerous effective prompting techniques. Topics covered include basic techniques like 'few-shot prompting', advancements in 'self-criticism' for improving responses, and 'decomposition' to break tasks down into manageable components. The conversation shifts to more advanced techniques and highlights the critical role of AI red teaming in identifying and mitigating vulnerabilities in AI systems. Sander also delves into prompt injection, illustrating how seemingly harmless queries can manipulate AI to produce dangerous outputs. The episode emphasizes the ongoing challenges of securing AI models, particularly as they become more integrated into everyday life, highlighting that while some security measures can improve performance, the inherent risks associated with AI will persist. The episode closes with thoughts on potential solutions for AI development, emphasizing that discernment between beneficial and harmful applications remains a key concern.