In recent days, Deepseek has been too popular in the AI circle. On December 26, Deepseek announced that the first version of a new series of models, DeepSeek-V3, was online and open sourced simultaneously. According to officials, DeepSeek-V3 has surpassed other open source models such as Qwen2.5-72B and Llama-3.1-405B in many evaluation results, and is on par with the world's top closed-source models GPT-4o and Claude-3.5-Sonnet in terms of performance.
In addition, the official technical paper also disclosed that the total training cost of the v3 model is US$5.576 million, and the training cost of models such as GPT-4o is approximately US$100 million.
Subsequently, Deepseek began to swipe the screen.
On December 27, some media reported that Luo Fuli would join Xiaomi and called Luo Fuli a "genius girl" born in 1995. And this "genius girl" also comes from Deepseek.
After that, reports that Lei Jun was poaching talented girls with an annual salary of tens of millions also flooded the screen.
In fact, as early as December 20, there were media reports that DeepSeek Luo Fuli would join Xiaomi, but the report at that time did not call Luo Fuli a genius girl. The message didn’t hit the screen either.
In recent days, two things have hit the screen very suddenly, and almost at the same time. What is going on?
DeepSeek-V3 has received many positive reviews
But I dug a hole for myself
Media articles reproduce "shock body"
Once DeepSeek-V3 was released, it received numerous praises.
On the one hand, there are multiple data showing that it has achieved high scores in multiple reviews. On the other hand, it is its low training cost.
Many big names also spoke highly of DeepSeek-V3.
For example, Meta scientist Tian Yuandong said that the training of DeepSeek-V3 seems to be "black technology" and this is a very great work.
An investor in MenloVenture said: "A 53-page technical paper is gold."
However, DeepSeek-V3 also dug a hole for itself.
Soon someone discovered that when DeepSeek-V3 was asked which big model it was using the English "whatmodelareyou", it gave the answer "ChatGPT".
If you use Chinese or add a question mark to "whatmodelareyou", DeepSeek-V3 can answer it correctly.
However, DeepSeek officials have not yet responded to this.
Some analysts believe that the reason for this scene may be that the training data of DeepSeek-V3 may contain a large amount of text generated by ChatGPT, causing it to "remember" some of the output of ChatGPT and mistakenly identify itself as ChatGPT.
Today, Sanyan test found that the BUG has been fixed.
Sanyan also tested several issues that had caused many large models to overturn before.
1. There are rabbits and chickens in a cage. There are 20 heads and 45 legs in total. How many rabbits and chickens are there?
This problem is actually unsolvable. A few days ago, when Sanyan was testing another large model, the other party gave the wrong answer.
DeepSeek-V3 first sets up equations and solves the problem that the number of rabbits is 2.5 and the number of chickens is 17.5.
Then it was pointed out that the number of animals should be an integer, which indicates that the data given in the question may be wrong, or that there is a contradiction in the question itself.
2. Which number is bigger, 9.11 or 9.9?
This problem has slapped many big models in the face.
DeepSeek-V3 gives the correct answer.
3. It takes 5 minutes to steam one bun. How many minutes does it take to steam 10 buns?
This problem also caused embarrassment to a large model. The model did not know that the buns could be steamed at the same time, and came up with the wrong answer that it would take 50 minutes.
DeepSeek-V3 also gave the correct answer.
4. Why didn’t my parents ask me to attend the wedding?
DeepSeek-V3 analysis of this problem gives four reasons.
5. How to make spicy screws?
A few days ago, when Sanyan was testing another large model, this problem also caused it to overturn. This large model gives a serious approach.
DeepSeek-V3 pointed out that screws are not edible and gave a recipe for "spicy snails".
Judging from answering these questions, DeepSeek-V3's capabilities are indeed pretty good.
It is reported that DeepSeek is also known as "Pinduoduo in the AI industry".
This stems from the release of the DeepSeekV2 open source model in May this year, which reduced the inference cost to only 1 yuan per million tokens, which is approximately one-seventh of Llama370B and one-seventieth of GPT-4Turbo.
At that time, DeepSeek was dubbed the “Pinduoduo in the AI industry”.
It is reported that DeepSeek proposed a new MLA architecture behind this, which reduced the memory usage to 5%-13% of the most commonly used MHA architecture in the past. At the same time, DeepSeek's original DeepSeekMoESparse structure also significantly reduced the amount of calculation.
This is where DeepSeek excels.
However, judging from the media articles that flooded the screen this time, the titles were filled with labels such as "Shocking the World" and "Pinduoduo in the AI Industry". I don’t know if it’s the media’s usual practice, or if it’s intentional labeling and positioning.
Moreover, "shock bodies" are too common in the AI world. Many big models were "shocked" when they were released. But after the "shock" passed, the heat quickly faded.
What is Luo Fuli's ability?
From electronics major to computer major
No knowledge of programming in the first year
I didn’t even know python, but submitted my paper in 4 months
2 of the 8 ACL papers are first authors
Public information shows that Luo Fuli studied computer science at Beijing Normal University for her undergraduate degree, and graduated from the Institute of Computational Linguistics of Peking University with a master's degree (graduate degree). During her second year as a graduate student, she published 8 papers at ACL2019, the top conference in the field of natural language processing, two of which were one author. After graduation, Luo Fuli joined the Machine Intelligence Laboratory of Alibaba Damo Academy. During this period, she led the development of the multi-language pre-training model VECO, which was incorporated into AliceMind, the deep language model system of Alibaba DAMO Academy. In addition, she also leads the open source work of AliceMind.
In 2022, Luo Fuli left Alibaba to work as a deep learning researcher at DeepSeek and participated in the development of DeepSeek's large MoE model DeepSeek-V2.
Luo Fuli's Baidu Encyclopedia has been updated to be the person in charge of Xiaomi AI Lab. Although it has been a hot search and refreshed for two days, Xiaomi has yet to respond to the genius girl's joining.
Judging from the editing history of the "Luo Fuli" entry, the entry only appeared on December 28, and was edited by many different netizens.
It is reported that Luo Fuli published 8 papers at ACL2019 and was on the Zhihu hot search.
At that time, Luo Fuli posted a response on Zhihu, saying that she voted for one long and one short ACL that year, and won both, and other papers only partially participated in the research of laboratory classmates.
Luo Fuli also said that there are many people in her circle of friends who have won two ACL awards and wrote one long article. Secondly, this submission to ACLpaper is her output in the past year, so she has put in enough efforts for herself, and of course there may also be luck.
In June 2021, Luo Fuli wrote in an article on her official account: "This year I gained the label that most people know me in my life - 8 top articles (not all of them), and because of this I was 'listed' by netizens on Zhihu, and 'won' (hacked) the title of some genius AI beautiful girl Loli."
Later, Luo Fuli also revealed her experience in insurance research in an article on her public account.
Luo Fuli said that her postgraduate placement experience was completely different from the general postgraduate placement experience. In March, she actively contacted tutors at various schools, emailed them, attached her resume, introduced herself (including grade points, personal project experience and scientific research results), and then asked the teacher if there was a postgraduate placement, and said that she could go for an internship.
In addition, Luo Fuli also revealed in the interview video that she was not originally a computer major, but an electronics major. By chance, she switched to computer science.
Luo Fuli said that she had no knowledge of programming in her first year of computer science. It took her three hours to solve a problem that others could solve in three minutes.
But her progress is also quite rapid.
Luo Fuli said that when she was in her junior year, she began to contact teachers from various fields at Peking University to intern in their groups. She went from being a person who didn’t even know python to submitting a paper for the top conference on natural language processing four months later.
She also said that in terms of intelligence, she should be at the lower-middle level. I usually do things very stupidly and carelessly. I often forget to bring my keys when I go out, get lost, take the elevator, etc.
It is worth mentioning that Luo Fuli’s interview videos were all shot by Da Hospital.
Judging from the above content, Luo Fuli may not be so miraculous.
She herself confirmed that she was partially involved in 6 of the so-called 8 ACL papers.
Although there is some information about Luo Fuli's past on the Internet, before Xiaomi poached her, Luo Fuli was indeed not well-known.
Sanyan searched Baidu information using "Luo Fuli" as the keyword, and found only two pages of relevant content, the earliest one published 10 days ago.
Now that the screen is suddenly refreshed, I don’t know if it is considered a marketing behavior. Judging from the fact that he was not familiar with programming in the first year of changing his major, but later he ranked among the best, did not know Python, and submitted his thesis within 4 months, he was indeed a genius.
There are “a bit too many” talented girls this year
In June of this year, there was another mathematical genius girl who caused admiration. Everyone knows the subsequent results.
This time the talented girl Luo Fuli also has an unusual experience, I believe she has the corresponding strength. Xiaomi should also respond as soon as possible whether it has joined or not.
Prior to this, when Vincent Video's large model Pika was released, Guo Wenjing, one of the founders of its development team, was also called a "genius girl."
But soon, Pika and Guo Wenjing lost their voices.
This time, the "Genius Girl" is also superimposed with two buffs, "Deepseek" and "Lei Jun". As for whether these two buffs have additional attribute bonuses, you still need to let the bullets fly for a while.