GPT 5 was trained to Read

Description

As AI models become more capable, the instructions we use to guide them must evolve. What worked for one generation of AI can have strange, literal, and unintended consequences on the next. OpenAI's Brian Fioca shares a perfect example from The Roo Cast, detailing how an old system prompt—likely written for a less "helpful" model—sent a new model (GPT-5) on an exhaustive, file-searching tangent. It's a fascinating look at the new challenges of prompt engineering and the literal-mindedness of modern AI. Watch the full conversation with Brian Fioca on The Roo Cast, hosted by Roo Code. #roocode #developertools #softwareengineering #aicoding #llm #largelanguagemodels #generativeai #api #aimodels #machinelearning #artificialintelligence #claude #opensource #opensource #computerprogramming #coding #windows #linux #openai #chatgpt #gemini #vscode #agi #asi #tech #googledeepmind #google