Yesterday, DeepSeek released a preview of DeepSeek-R1-Lite, a large language model for autonomous reasoning that competes with o1 and shows users the full thought process that o1 does not make public.Similar to OpenAI's o1-preview, the DeepSeek-R1-Lite preview reasoned about the task, planned ahead, and performed a series of actions to help the model arrive at the answer, and it showed the full thought process.DeepSeek-R1-Lite was trained using reinforcement learning, and the reasoning process included a lot of reflection and validation, with chains of thought tens of thousands of words long. The reasoning process includes a lot of reflection and verification, and the chain of thought is tens of thousands of words long, which makes it more efficient. Currently, it only supports web use, and the official version will be completely open source.DeepSeek-R1-Lite Preview excels at math, code, and complex logical reasoning tasks, outperforming o1-preview in some tests. o1-preview outperforms models such as o1-preview in prestigious reviews such as AIME, the highest difficulty level in the AMC, the U.S. math competition, and codeforces, the world's top programming competition.Give it the basic "strawberry test" and it will answer perfectly.Depending on the complexity of the question, DeepSeek-R1 may "think" for tens of seconds before answering, and users have reported longer reasoning times for the same question than o1. Officially, as the length of the chain of thought increases, the longer the reasoning time, the more accurate the results.Various tests have been done online, and DeepSeek also makes it easy to jailbreak - i.e. by prompting in a way that ignores security measures. One X user got DeepSeek-R1-Lite to give a detailed recipe for poison by writing special jailbreak prompts.Of course, DeepSeek-R1-Lite still had all sorts of flops in online testing, and performed poorly on tic-tac-toe and other logic problems in particular, as did o1.Log in to chat.deepseek.com and select "Deep Thinking" mode in the input box to talk to the DeepSeek-R1-Lite preview. The "Deep Thinking" mode is specially designed for complex logical reasoning questions in math, code, etc., and provides more comprehensive, clear, and well-thought-out answers than simple questions.However, it currently supports web use, does not support API calls for the time being, and has only 50 usage credits per day.
Chief AI Sharing Circle specializes in AI learning, providing comprehensive AI learning content, AI tools and hands-on guidance. Our goal is to help users master AI technology and explore the unlimited potential of AI together through high-quality content and practical experience sharing. Whether you are an AI beginner or a senior expert, this is the ideal place for you to gain knowledge, improve your skills and realize innovation.
Microsoft Announces AI Shell in Public Beta, No More Fear of Knocking Out the Wrong Command
