The article introduces the concept of the "Jagged Frontier" of AI, a term coined by the author and co-authors in 2023. This refers to the uneven performance of AI across various tasks, where it excels in some areas like medical diagnosis and complex math but struggles with simpler tasks like visual puzzles or operating vending machines. The author argues that this jaggedness will persist, as even as AI capabilities expand, they may not align with human tasks due to limitations, such as a lack of memory in large language models (LLMs). The author cites Tomas Pueyo’s viewpoint that AI’s growth will eventually outstrip human capabilities, but raises concerns that jaggedness could prevent AI from fully overlapping with human skills. Bottlenecks in AI performance arise from both technical limitations and external processes. For instance, while AI can rapidly identify promising drug candidates, the pace of clinical trials and regulatory reviews remains a significant constraint. The article illustrates this with a study where GPT-4.1 outperformed human reviewers in synthesizing medical research but still required human oversight for edge cases, like accessing unpublished data. This demonstrates how jaggedness not only creates bottlenecks but also complicates the automation of tasks, highlighting the need for human involvement in certain scenarios. The discussion extends to "reverse salients," which are specific weaknesses that can stall progress. The recent launch of Google’s Nano Banana Pro image generation AI exemplifies this concept, as it addresses previous limitations in image quality that hindered other capabilities. The author notes that advancements in AI are often stalled by these bottlenecks, which, when resolved, can lead to significant leaps in performance. This interplay between jaggedness, bottlenecks, and reverse salients shapes the current landscape of AI development, affecting how quickly and effectively AI can be integrated into various fields.