Speaker 1: China's AI company called DeepSeek has released a new model called R1, and the results are shocking. Until now, the best reasoning capabilities amongst all AI models was held by OpenAI's O1 model, whose algorithm is speculated to be a chain of thought variation, but the exact details are unknown. DeepSeek has now matched its benchmarks using a mixture of reinforcement learning and chain of thought. The model uses an algorithm called Group Relative Policy Optimization, which rewards the model for accurate answers being generated in the correct format. Two particularly interesting things are mentioned in this paper. One is that extending test time computation leads to improved results, which concurs with recent papers from Google Research. The other is the discovery of aha moments, where while solving a problem step by step, the model pauses, reflects, and retries the problem with a different approach. This is similar to how humans solve a problem, discovering different paths before coming to an optimal solution. Since this model is open source, the numbers are likely to be verified by the research community quickly. Thanks for watching, cheers.