DeepSeek R1 0528 8B vs 671B (Live Test)
AI Summary
The video explores the performance differences between the new Deepseek R10528, distilled down to the Q138 billion version, and the full Deepseek R10528 with 671 billion parameters. It starts with a logic test in a skyscraper scenario where the AI has to navigate to the 30th floor using five buttons, each with specific functions. The presenter meticulously illustrates how the models approach the problem, emphasizing the reasoning process, trial-and-error methods, and optimization strategies employed by the AI. The video highlights how the newer version shows improved causal reasoning compared to the original model, detailing the AI’s thought process as it arrives at solutions and evaluates them. Ultimately, the video concludes with a detailed analysis of the approaches the AI takes to determine the optimal button press sequence, demonstrating the strengths and limitations of the Q138 billion model compared to the full version.