#154 - Google Gemini, Waymo Collision, Smaug-72B, EU AI Act final text, image watermarks
Last Week in AI
00:00
Bypassing Safety Measures: Language Hacking GPT-4
This chapter explores a study revealing how OpenAI's GPT-4 can be manipulated using less common languages to evade safety guardrails. Researchers discovered this method achieves a 79% success rate, raising concerns about the effectiveness of current restrictions compared to traditional prompts in English.
Transcript
Play full episode
