Q&A: What is agentic AI today, and what do we want it to be?

The article presents a Q&A with Phillip Isola of MIT EECS, delving into the concept of agentic AI. Agentic AI is defined as artificial intelligence capable of performing actions, contrasting with generative AI which focuses on creating content like text or images. These actions can range from physical robotic manipulations to digital tasks like booking flights. Isola explains that agentic AI systems often build upon foundational generative AI models, incorporating wrappers and tools to enable them to interact with applications, websites, or the physical environment. While many current agents are digital, such as customer service bots, they share the characteristic of taking actions and retaining memory of past events. A significant challenge in developing more advanced agentic AI lies in the scarcity of diverse training data. The future of this technology might necessitate models that can process a wider array of modalities beyond text, including video, physical forces, and time series data, requiring fundamentally different architectures. The discussion also touches upon the balance between AI automation and human assistance, particularly in high-stakes or safety-critical fields like medicine, where full automation might not be desirable or ready. The article implicitly raises questions about the ethical implications and risks associated with increasingly autonomous AI agents.