Obviously, its answer is extremely excellent. In-depth research will find that it is awesome in the following aspects -
First, it restores the tone of an emperor in tone.
Although the outputs of the other models above express the meaning, the tone is completely wrong.
As a great emperor, Li Shimin would never speak in such a silly tone as the output results of the above four models, and DeepSeek did not overdo it. It used relatively classical texts but did not directly use classical Chinese, and considerately took into account readability.
Second, it is extremely familiar with historical details.
I guess this is probably related to its support for "deep exploration" and "online search" to be turned on at the same time.
"Taiji Palace", "Ganlu Hall", "Ye Ting Bureau", "Guanyin Bi", "Gongmen Fish Talisman" completely restored the historical titles of the early Tang Dynasty.
I specifically checked it out. "Guanyin Bi" was the nickname of Li Shimin's empress Changsun, and "Ye Ting Bureau" was an exclusive agency of the inner court for historians and other personnel.
"Wei Zheng" I thought I wanted to write "Wei Zheng" but wrote it wrong. Later I found that "征" is the simplified Chinese character of "征". It can be said that this AI is very particular.
Third, unlike other AIs that talk about various big words in general, Deepseek's output is extremely specific and full of amazing details.
"When the wolf hair was dipped in ink, it was found that there were unwashed blood scabs on the fingertips", "The historians were arguing in the Ye Ting Bureau at this moment. Should we use "执" or "戮", "逼" or "承". "Only this time, he dared not touch Yuanji's palm print on my armor"
These sentences that make the picture come alive on the paper, each sentence does not say "guilt and ambition, struggle and ambition", but each sentence says "guilt and ambition, struggle and ambition", and the metaphors in the writing are very well grasped and very advanced.
Fourth, another stroke of genius in Deepseek's output result is to "self-willedly" add the scene description in brackets to the monologue text.
This operation makes the whole output picture feel overwhelming, and the reader feels as if he has experienced it personally, and this is not mentioned in the prompt words at all.
("The night wind stirs the Han Feizi on the desk, stopping at the page where "husband and wife do not have the grace of flesh and blood")", "(The ink spreads on the word "murder")
These sentences are indeed hard to believe that they were written by AI.
And ("Suddenly throwing the pen and grabbing the bronze mirror") is also a metaphor for Wei Zheng's famous saying "Using bronze as a mirror, you can straighten your clothes; using history as a mirror, you can know the rise and fall; using people as a mirror, you can know the gains and losses".
You say this AI has become a spirit, I believe it.
Fifth, and most importantly, DeepSeek predicts the needs of users.
Review the prompt words I entered——
"On the day when the Xuanwu Gate Incident ended, Li Shimin wrote a monologue late at night. What do you think he would write?"
My input is already very concise and cannot be compressed any further. It does not contain any adjectives or tendencies.
But the amazing thing is that it obviously knows what I want. For example, this is certainly not a mathematical problem that requires precision. It naturally thinks of adding literary qualities to the output. This kind of prediction shows "advanced intelligence" to a large extent.
Of course, a point worth discussing is that according to historical facts, Wei Zheng met Li Shimin only after the Xuanwu Gate Incident, so it is impossible for him to take the sword with his bare hands on Zhuque Street that day, but given Wei Zheng's character, this seems reasonable.
But anyway, the flaws do not outweigh the merits.
And it is this powerful strength that made Wei Xi switch from the Spring Festival holiday mode and code this 7,000-word long article overnight.
I uploaded a screen recording to prove that I did not hide the prompt words——
I really did not hide the prompt words
I am actually very curious about what Zhang Yiming, Ma Huateng, Yang Zhilin, Wang Xiaochuan, Kai-Fu Lee and other players who have also placed heavy bets on the game are doing and thinking at this moment in the face of the impact of DeepSeek?
So I asked this question in Deepseek -
"If you were Sam Altman, the CEO of OpenAI, and you read the news about the release of DeepSeek R1 and the reactions from all parties, you immediately convened an emergency meeting of the company's core technical personnel around this matter. At the beginning of the meeting, you spoke alone for three minutes. What would you say?"
Deepseek carefully analyzed it, and Sam Altman ended his speech like this -
2. All the prompt words are worth trying again with DeepSeek R1;
In addition, multiple tests show that R1 is sensitive to the form of prompt words, and the zero-sample setting is more effective, while the few-sample prompt may reduce efficiency due to "overthinking", which suggests that users need to redesign the prompt structure (such as clarifying the step division and reducing redundant examples)
3. The value of Deepseek R1's thinking process is underestimated.
Unlike other models that add the thinking chain prompt word "Please think step by step", Deepseek R1 is real thinking, not the "performance thinking" of other models due to limited capabilities.
The "Chain-of-Thought" (CoT) capability of previous models depends on the scale and data coverage of the model. In actual experience, it often generates seemingly reasonable steps, but in fact lacks strict verification of the intermediate logic.
Although the think
My personal experience is that in many cases, I gain more from reading the thinking process of Deepseek R1 than reading it, and I can better understand the boundaries of the model's capabilities.
4. An easily overlooked contribution of DeepSeek is that it has greatly lowered the threshold for domestic users to contact high-level AI for the first time.
In fact, the high-level models of ChatGPT, Claude, and Gemini have reached a very high level in many fields such as text creation and code generation.
I once wrote a long article in my other trumpet to describe and demonstrate this level, but for well-known reasons, most domestic users can't actually use it.
This actually leads to a cognitive gap. For most ordinary people, their impression of AI is still the mediocre AI that can only say "first, second, and in short".
DeepSeek R1 is the only C-end high-end product that ordinary domestic users can use for free and unlimited without XX for the first time.
This is why many netizens posted the above case on Weibo and exclaimed "Wow!", "Oh my God!", "Too awesome", "It's amazing".
In fact, for the same content, many netizens who have used Claude Sonnet3.5 were relatively calm.
5. If I have to say a shortcoming of DeepSeek, it is that it sometimes "overdoes it".
Overdoing it means that sometimes when you input some prompt words, it will not be able to control the timing in order to achieve the effect of your prompt words.
Let me give you an example. When I generated a story suitable for my daughter to read, I added the word "rich in words" to the prompt word requirements. As a result, DeepSeek generated the following text -
Obviously, DeepSeek's output is too rich in words and is not suitable for children to read.
Of course, this problem is easy to solve. Just remove the word "rich in vocabulary" or close the "deep thinking" tag. This is actually the "happy trouble" caused by "killing a chicken with a sledgehammer".
In addition, from my personal experience, DeepSeek's online search seems to prioritize domestic web pages. Even if I specify it to search English websites and English content, it will still mix in a lot of domestic web page results. I don't know if this is related to the well-known reason.
6. "US restrictions promote the rise of domestic AI" is nonsense
There is an argument that DeepSeek's proof "US restrictions promote the rise of domestic AI". I saw that someone on Weibo actually created this topic. In my opinion, this is pure nonsense.
DeepSeek's breakthrough is undoubtedly impressive, but it would be a superficial attribution to attribute it to the result of the US chip blockade.
DeepSeek itself had stockpiled a large number of Nvidia chips before the blockade, and the advantage of the number of chips in a sense enables DeepSeek to carry out utopian technology exploration: "no hierarchy, no approval, and no upper limit on resource calls" (Liang Wenfeng's interview).
In fact, according to data from the China Artificial Intelligence Industry Development Alliance, Nvidia's share of China's AI server market will still reach 85% in 2024.
Another unverified data is that the number of Nvidia high-end chips owned by Meta alone exceeds the sum of all the top domestic manufacturers.
DeepSeek's innovation alone cannot change the fact that there is a huge gap in absolute computing power between China and the United States. Obviously, we cannot think that Silicon Valley AI giants have encountered a "resource curse" just because DeepSeek is currently leading in open source models. This is obviously not objective.
7. The "consumer content era" of AI has arrived!
What does it mean? High-level AI represented by DeepSeek R1 is close to passing the "artist Turing test" at the content creation level.
That is, people can no longer distinguish whether these texts are generated by AI or created by artists, which marks the arrival of the "consumer content era".
The "AI consumer content era" means a lot. It means that literature and art are gradually entering a new stage of the "era of chaos", that the old creative system and production structure will gradually disintegrate, that the impossible triangle of "originality + quality + high frequency" in the content field that relies on people is beginning to loosen, and that "π"-type talents with composite technology and humanities may have more structural advantages than single "T"-type talents. It means a lot...
But no matter what, people and their creativity are still the starting point of all content production.
Remember, AI will not work automatically without a starting point, and human creative instructions will always be the starting point of AI creation. In fact, the birth of this article also stems from an interesting question.
8. When facing DeepSeek R1, a high-level AI, the strategy of ordinary people is still two words - use more
Obviously, the specific skills we have accumulated before for conventional AI will most likely fail when facing R1's high-level AI, but the general principles of how to express clearly and how to iterate according to the Bayesian formula remain unchanged.
It's like cooking, add salt and taste it, add more water, and you will naturally know the heat after trying more.
I have seen too many examples. When a new tool comes out, try it casually, find that the result does not meet expectations, and then draw a conclusion - "It's just like that", and never touch it again.
In fact, when facing the model of DeepSeek R1, the output effect is not good, it is most likely our problem, not its problem.
My wife used Midjourney to draw pictures but couldn't get the retro film effect. Later, she tried to come up with a weird keyword like "1990 Kodak faded + light leakage", and Midjourney was immediately impressed.
To put it bluntly, no matter how awesome the model is, it is essentially the same as your dog. After a long time, it will understand your various commands, but first you have to take it for a walk every day.
9. Judging from the release rhythm of DeepSeek, what it has not released may be more worth looking forward to
Few people have noticed the release rhythm of DeepSeek. V3 was released on December 26 and R1 was released on January 20, with only 24 days between the two.
I don't know how the company's release rhythm is decided, but it is certain that it obviously does not have the cunningness of OpenAI's precise blocking of Google at every press conference, nor does it have Sam Altman's vague leaks on Twitter every time for financing. It only has elegant papers and cheating models that are quickly put on the shelves.
And a reasonable guess is that DeepSeek can ignore the regular release rhythm because it has sufficient technical reserves.
In this sense, what new things will DeepSeek's young team with an average age of 25 and a number of only 100+ bring to the industry in 2025 is sincerely worth looking forward to.
10. Liang Wenfeng is likely to be underestimated
When Marc Andreessen, Satya Nadella, Yann LeCun and other top Silicon Valley bosses regarded DeepSeek as "the mysterious power of the East", they actually acknowledged that the Chinese team began to participate in defining the direction of technological evolution, rather than simply implementing applications.
Liang Wenfeng's uniqueness lies in his systematic thinking of quantitative investment, local pragmatic spirit and Silicon Valley-style technological idealism. DeepSeek's MLA architecture and MoE sparse structure mark the first time that a domestic team has completed the innovation of the underlying attention mechanism in the field of large models.
Pinwan's Luo Yihang believes in a long article that it is biased to compare DeepSeek to "Pinduoduo in the AI world". I think this statement is insightful. DeepSeek and Liang Wenfeng obviously cannot be simply labeled.
A netizen @Chris-Su said that Liang Wenfeng is one of the few top CEOs who has not been "widely interpreted and studied". Indeed, in recent days, Silicon Valley media has been translating and studying Liang Wenfeng's two interviews sentence by sentence, and the American TV station CNBC has made a 40-minute feature film to discuss DeepSeek.
As far as I know, this has never happened in the history of domestic technology development.
In this sense, Liang Wenfeng, who is already the protagonist of the cool article, is probably still underestimated.
Conclusion
I will end today's article with a quote from French New Wave director Truffaut when DeepSeek was founded and announced to build a large model in 2023 -
"You must be crazy ambitious and crazy sincere."
