Hinter den in aller Munde gehenden Papers von DeepSeek: Neben Liang Wenfeng gibt es noch einen 18-jährigen chinesischen Highschool-Schüler, der einst göttliche Prompt-Wörter geschrieben hat.

Throughout yesterday, the entire internet was flooded with news about DeepSeek making it onto the cover of Nature.

Sure enough, DeepSeek officially issued a response:

Compared with the early arXiv pre - print version, the version published in Nature this time contains richer implementation details and more detailed experimental analyses. We believe that these new additions will provide effective experience for the community to better understand and build advanced inference models.

In the past year, we have been fortunate to witness and participate in the rapid development of AI inference models. In the future, we will, as always, make unremitting efforts to continuously expand the boundaries of artificial intelligence research.

For detailed information about this paper, you are welcome to review the previous article on APPSO:

"DeepSeek Makes Its Debut on the Cover of 'Nature': A Chinese Large - Model Creates New History and Does What OpenAI Dares Not Do"

Speaking of which, Liang Wenfeng, as the corresponding author, is already quite eye - catching. But careful netizens also noticed two familiar figures in the contributor list of this Nature paper — Tu Jinhao and Luo Fuli.

Let's start with Tu Jinhao. Born in 2007, in 2024, he stood out from 563 university and enterprise teams worldwide with 34 points (the highest score in that session) in the newly - opened AI track of the Alibaba Global Mathematics Competition and directly won the first place.

His competition strategy was quite ingenious. Drawing on the idea of "self - debate", he let multiple large models conduct multiple rounds of "self - question - self - answer - self - verification", and finally let the AI select the optimal answer.

In addition, another well - known feat of this high - school student at that time was transforming Claude 3.5 into 'Thinking Claude' through a god - level prompt engineering.

It is reported that when Tu Jinhao was constructing the prompt for 'Thinking Claude', he fed the initial version of the prompt back to the AI and asked it to think and improve on its own. After multiple iterations and optimizations, the Claude 3.5 large model demonstrated deep reasoning ability similar to the o1 chain - of - thought released by OpenAI at that time.

As of the time of publication, this prompt had received 15,600 stars on Github, which is quite remarkable.

The full version of Tu Jinhao's god - level prompt is as follows👇

Here is the Github address:

https://github.com/richards199999/Thinking - Claude

When being interviewed by "WAIC UP!", the first publication under the World Artificial Intelligence Conference (WAIC), he once mentioned: "The 'thinking process' is the key factor determining the quality of reasoning. It is more important to let the model learn to 'slow down and check' when in doubt than to simply answer."

In Tu Jinhao's view, the hallucination problem of AI models needs to be viewed rationally. He believes that: "Although we know that models have hallucinations, humans also have hallucinations. To some extent, this can be regarded as a manifestation of creativity."

Regarding how to cultivate an interest in AI, this post - 2000s generation has his own unique insights.

Also in the interview of "WAIC UP!", he emphasized: "My experience is to use it more. Using it more is really very important. Because the model is not just a tool now. It is more like a friend or a partner. You can make it play whatever role you want."

"After using it, you will develop an interest in it and find it fun. At first, I just used it as a toy and found it very interesting. Then I would wonder why it can think. Driven by this interest, you will continuously explore in depth, read blogs, search for information, and gradually accumulate knowledge. This is a relatively good way of learning." Tu Jinhao summarized his AI learning experience in this way.

Public information shows that in many exclusive interviews with podcasts and technology media, he has successively discussed topics such as "hallucination creativity" and how to reshape models without an engineering background. He can be regarded as a representative figure among post - 2000s AI practitioners.

As for another contributor, Luo Fuli, we have reported on her before. She studied computer science at Beijing Normal University for her undergraduate degree and was recommended for a master's degree in computational linguistics at Peking University.

In 2019, while still a graduate student, she published 2 papers as the first author and 8 papers in total in the top NLP conference ACL. Subsequently, she received offers from top companies such as Alibaba Star, Tencent Technical Experts, and Baidu AIDU, and finally joined Alibaba DAMO Academy, leading the multilingual pre - trained model VECO in the AliceMind open - source project.

In 2022, she switched to Magic Square Quantitative, a leading quantitative private equity firm, to do deep - learning strategy modeling. In 2023, she joined DeepSeek, a large - model startup under Magic Square, and became one of the core developers of the MoE large model DeepSeek - V2.

After the release of V2 in May 2024, she wrote on Zhihu that its Chinese proficiency was "in the first echelon of closed - source models at home and abroad", and the API pricing was only 1/100 of that of GPT - 4, making it the "king of cost - effectiveness".

Google Scholar shows that Luo Fuli has a total citation count of 10,025, an h - index of 27, and an i10 - index of 34. Her academic achievements are mainly concentrated in the research and technical reports of the DeepSeek series of large models.

Representative papers include Deepseek - r1 (2025, cited 3596 times), Deepseek - v3 technical report (2024, cited 1809 times), DeepSeek - Coder (2024, cited 1129 times), as well as many papers such as DeepSeek - llm, DeepSeekMoE, and Deepseek - v2.

Since the end of 2024, many media have reported that "Lei Jun, the founder of Xiaomi, offered a multi - million - yuan annual salary to poach Luo Fuli", making her stand out with the label of "a post - 1995 AI genius girl".

But she quickly spoke up on her WeChat Moments, rejecting the genius persona, calling for "a quiet atmosphere to do things", and emphasizing that she "is not a genius and just wants to do difficult but right things".

It is worth mentioning that her Zhihu profile also reads "Please stop hyping and exploiting individuals by major self - media".

In February this year, her relatives confirmed to some media that she had joined a "new position", but did not disclose the name of the company. There is also no record of her in Xiaomi's internal system, so whether she has joined remains a mystery. The information on "Nature" only shows: "Independent researcher, Beijing, China."

Of course, one person can go fast, but a group of people can go farther. Every signatory of the paper has left their mark on it. The glory of this Nature cover should belong to the entire team, and even to the global academic community jointly exploring AI.

But at the same time, it also represents a trend: Chinese AI talents are appearing on the global top - level academic stage more frequently and in more core positions. As the world's first mainstream large - language model to undergo peer review, this time, Chinese AI has really made a strong presence on the world stage.

Summary of relevant links:

🔗 DeepSeek Nature paper: https://www.nature.com/articles/s41586 - 025 - 09422 - z

🔗 Supplementary materials: https://static - content.springer.com/esm/art%3A10.1038%2Fs41586 - 025 - 09422 - z/MediaObjects/41586_2025_9422_MOESM1_ESM.pdf

🔗 Peer review: https://static - content.springer.com/esm/art%3A10.1038%2Fs41586 - 025 - 09422 - z/MediaObjects/41586_2025_9422_MOESM2_ESM.pdf

🔗 Nature editorial article: https://www.nature.com/articles/d41586 - 025 - 02979 - 9

🔗 Nature news: https://www.nature.com/articles/d41586 - 025 - 03015 - 6

🔗 DeepSeek original arXiv paper: https://arxiv.org/pdf/2501.12948

This article is from the WeChat official account "APPSO". Author: APPSO, discover tomorrow's products. Republished by 36Kr with authorization.