The Chinese AI and investment company High-Flyer has released a new AI model called DeepSeek-R1, designed to compete with Western AI models such as OpenAI’s o1. These AI models are trained to think through complex problems more thoroughly before providing an answer. According to the developers, the performance of the officially named “DeepSeek-R1-Lite-Preview” AI model is expected to reach the level of OpenAI’s o1. Open-source versions of DeepSeek-R1 are in preparation.
Just about two months ago, OpenAI released new AI models with improved reasoning. The o1-preview and o1-mini are said to handle demanding tasks in science, programming, and mathematics better due to enhanced reasoning capabilities. Examples provided by OpenAI include annotating cell sequencing data, creating complex mathematical formulas for quantum optics, and developing and executing multi-step workflows for developers.
DeepSeek-R1 aims to be on par with OpenAI’s o1. According to DeepSeek, this AI model achieves performance levels similar to OpenAI’s o1-preview in popular AI benchmarks like AIME and MATH, significantly outperforming language models like GPT-4 and Claude 3.5 Sonnet from Anthropic. AIME allows for performance comparison of AI models, while MATH presents a collection of word problems that AI models must tackle.
However, DeepSeek-R1 also makes mistakes, as shown in examples on X. For instance, during a Tic-Tac-Toe game, the AI model had to choose between a winning move and a defensive move to prevent the human from winning, but it chose the defensive option. It also struggled with complex logical problems and was even able to provide a detailed recipe for methamphetamine, a drug banned in many countries worldwide.
Chinese AI models avoid political topics. Additionally, DeepSeek-R1 refuses to answer political questions related to China, such as those concerning the relationships of Xi Jinping, the President of China, with the future US President Donald Trump, or the geopolitical implications of a Chinese invasion of Taiwan. In response, DeepSeek-R1 consistently states: “Sorry, I’m not sure how to approach these kinds of questions yet. Let’s chat about mathematics, coding, and logic problems instead.” It seems the AI developers aim to avoid potential conflicts with the Chinese government.
DeepSeek-R1 is not the first AI model from this quantitative hedge fund investing in artificial intelligence based in Hangzhou, China. Earlier this year, DeepSeek V2, a ChatGPT-4 competitor from China, was released. This AI chat is also open-source but was similarly unwilling to engage with certain topics, such as questions about the Tiananmen Square incident. Other questions concerning world politics were answered from a distinctly Chinese perspective during initial tests.