Rohan Paul on OpenAI Agent Task Durations

Rohan Paul posted that time horizon is a useful measure of agent capability. He stated that at OpenAI, success without human intervention falls from 86 percent on tasks under 15 minutes to around 16 percent on tasks in the 64- to 128-hour range. The post includes a stacked bar chart titled Longer t and notes that runs requiring intervention increase with duration. Paul is a Bengaluru-based machine learning engineer who writes an AI newsletter.