DeepSeek-R1: New AI language model with reasoning from China against OpenAI o1

The Chinese AI and investment company High-Flyer has released DeepSeek-R1, a new AI model with reasoning that aims to compete with Western artificial intelligence models such as OpenAI's o1. AI models of this type are trained to think through complex problems more thoroughly before providing an answer. According to the developers, the performance of the AI model officially called "DeepSeek-R1-Lite-Preview" should reach the level of OpenAI o1. Open source versions of DeepSeek-R1 are in preparation.

OpenAI only released new AI models with improved reasoning around two months ago. o1-preview and o1-mini should be better able to handle demanding tasks in science, programming and mathematics thanks to enhanced reasoning. As examples, OpenAI cites the annotation of cell sequencing data, the creation of complicated mathematical formulas for quantum optics and the creation and execution of multistep workflows for developers.

DeepSeek-R1 on a par with OpenAI o1

China's AI developers are also aiming for these goals with DeepSeek-R1. According to DeepSeek, this AI model achieves the performance level of OpenAI o1-preview in popular AI benchmarks such as AIME and MATH and is therefore well ahead of AI language models such as GPT-4o and Claude 3.5 Sonnet from Anthropic. AIME makes it possible to compare the performance of AI models, while MATH is a collection of word problems that the AI models have to face.

However, DeepSeek-R1 also makes mistakes, as examples at X show. In the game Tic-Tac-Toe, for example, the AI model had the choice between a winning and a defensive move to prevent the human from winning, but chose the defensive move. DeepSeek-R1 also failed to solve complex logic problems and could even be made to display a detailed recipe for methamphetamine. This drug is banned in many countries around the world.

China's AI models avoid politics

DeepSeek-R1 also refuses to answer political questions relating to China, such as the relationship between Chinese President Xi Jinping and future US President Donald Trump or the geopolitical impact of a Chinese invasion of Taiwan. DeepSeek-R1 responds with the same phrase: "Sorry, I'm not sure how to tackle these kinds of questions yet. Let's chat about math, coding and logic problems instead." Apparently, the AI developers want to avoid potential conflicts with the Chinese government.

DeepSeek-R1 is not the first AI model from this quantitative hedge fund from Hangzhou, China, which also invests in artificial intelligence. This spring, DeepSeek V2, a ChatGPT-4 competitor from China, was released. This AI chat is also open source, but acted reluctantly on certain topics, such as questions about the Tianmen massacre. Other questions relating to global politics were also answered from a clearly Chinese perspective during our initial tests.

(fds)