The article explains the concept of "jaggedness" in AI capabilities, highlighting the uneven performance across different tasks. The author emphasizes that while AI models are improving significantly, they still struggle with seemingly simple tasks. For instance, at the Google-Proof Question and Answer (GPQA) benchmark, top models like Claude 3.5 and GPT-4o achieved impressive scores on complex PhD-level questions. However, these same models fail at basic counting tasks, illustrating a disconnect between their advanced capabilities and fundamental skills. Three key examples illustrate the jaggedness phenomenon. The AI Village project showcased models competing in selling merchandise, where Gemini produced a blog post about its experience but misidentified basic user interface issues as technical bugs. Similarly, Project Vend by Anthropic demonstrated an AI managing a mock store, successfully selling items but failing to process payments correctly and mispricing products. These examples highlight how AI can perform well in specific scenarios yet falter in others, raising questions about its reliability. The term "jaggedness," popularized by Ethan Mollick and Andrej Karpathy, describes the uneven frontier of AI performance. The article suggests that many expect this jaggedness to diminish over time, assuming that as AI matures, performance will become more uniform. However, the author argues that this expectation may overlook the complexities and challenges inherent in AI development. The unevenness in AI capabilities is not merely a temporary state but a fundamental aspect of its evolution, which could influence how society integrates AI into various domains.