Taking the cue from the previous episode on the history of AI, all the way to ChatGPT, this episode looks into the concept of multi-modal AI. We explore how this technology integrates text, images, and audio to mimic human brain processing. We discuss fusion mechanisms that combine these modalities, allowing AI models to comprehend and respond to complex inputs. These mechanisms are crucial for practical applications, such as extracting information from PDFs or answering questions about images. Subsequently, we transition to reasoning models that can be prompted to provide sequential reasoning. Reasoning models, like DeepSeek’s r1, are designed to automatically reason through problems and manage the reasoning effort based on complexity. This approach distinguishes itself from prompting techniques such as “let’s think step by step” or “chain of thought,” which aim to enhance accuracy through structured reasoning. Group Relative Policy Optimization (GRPO) emerges as a reinforcement learning method employed to train models like DeepSeek R1. GRPO incentivizes model improvement through rewards, such as correct answers in mathematical problems. This approach facilitates self-supervised training without human intervention, enabling the emergence of extended thinking chains and enhanced responses. In the concluding segment of the discussion, we address the reduction in training and inference costs, even as companies invest substantial resources in GPUs for training large models efficiently. Algorithmic advancements and hardware improvements facilitate the training of smaller models, thereby increasing AI’s accessibility to enterprises and startups. Agentic AI, model context protocols, and smaller language models represent emerging trends that will shape the future of AI. These advancements will render AI more practical and efficient for real-world applications. Produced by Sharmada Venkatasubramanian
続きを読む
一部表示
2025/05/0518 分
48 分