AI Safety and Robustness: Recent Advances and Future Directions
Explore recent advances in AI safety and robustness in this 44-minute INSAIT Tech Series talk by Prof. Zico Kolter. Delve into the challenges of preventing undesirable outputs from large language models (LLMs) and learn about the built-in "guardrails" designed to enforce developer-specified policies. Discover how adversarial attacks have historically circumvented these safeguards and manipulated LLMs for unintended purposes. Examine the latest breakthroughs that have significantly improved the practical robustness of LLMs, including a recent competition where attackers failed to breach a deployed LLM over a month-long period. Gain insights into the current state of AI safety, ongoing challenges in the field, and the future prospects for developing safe AI systems.