Andrew Ng Explores The Rise Of AI Agents And Agentic Reasoning | BUILD 2024 Keynote

Andrew Ng discusses AI's vast opportunities, focusing on the application layer. He highlights the accelerating speed of machine learning model development due to generative AI, enabling rapid prototyping and experimentation. Ng emphasizes the importance of agentic AI workflows, showcasing their superior performance in various tasks, particularly in visual AI, through demos of a vision agent capable of processing images and videos. He concludes by identifying key trends driving AI's evolution, including faster token generation, large language models optimized for tool use, the rising importance of unstructured data engineering, and the burgeoning visual AI revolution. The discussion focuses on how the rapid advancement in model building using generative AI has shifted the development bottleneck from data collection to evaluation. The traditional sequential workflow of data collection followed by model training is replaced by a parallel approach, where data collection and model testing occur concurrently. This segment highlights how generative AI drastically reduces the time required to build AI applications, from months to mere days, enabling faster experimentation and prototyping, leading to quicker product iterations and deployment. This segment showcases the application of agentic workflows to visual AI tasks, demonstrating how these workflows enable efficient processing of image and video data. The speaker presents a demo of a vision agent that counts players in a soccer game and extracts relevant information from videos, highlighting the potential of this technology to unlock the value of large visual datasets. This segment presents empirical evidence supporting the superior performance of agentic AI workflows. It uses a coding benchmark to demonstrate that agentic workflows significantly outperform traditional methods, achieving a much higher success rate in solving complex coding puzzles. The discussion then transitions to explaining the four major design patterns of agentic workflows: reflection, tool use, planning, and multi-agent collaboration. This segment introduces the concept of agentic AI workflows, a crucial technical trend in AI. It contrasts the traditional zero-shot prompting approach with the iterative, agentic workflow, which involves multiple steps like outlining, research, drafting, and revision, leading to significantly improved results. This segment highlights four crucial AI trends: the rise of agentic AI, advancements in token generation speed, the adaptation of large language models for tool use in agentic workflows, and the increasing importance of data engineering for unstructured data. The speaker emphasizes the transformative potential of these trends, particularly in visual AI, enabling the creation of applications previously deemed impossible. The discussion underscores the exciting opportunities for developers and businesses to leverage these advancements. Agentic workflows significantly improve large language model performance, especially in complex tasks like coding puzzles. Instead of a single, continuous prompt-response, agentic workflows involve iterative steps like outlining, research, drafting, and revision, mimicking human problem-solving. This approach leads to much better results, as demonstrated by a substantial increase in accuracy on coding benchmarks (e.g., from 48% to 95% in one example). ,