Time flies, and it is already 2024.

At the end of last year, Chen Pingjiang gave people an impression that was very different from usual.

The whole person is confused and even a little neurotic.

It was not until December 12 that he regained his vitality.

It turns out that the past is past, and now this timeline is the only one.

He cannot explain such mysterious issues as time travel and rebirth, and perhaps no one will be able to explain them in the next five hundred or one thousand years.

Later, he asked this question as a joke to a Nobel Prize winner who worked at Hua Xia University.

The other party’s answer was: “Apart from the setting of the multiverse, the timeline is unique. There are no multiple timelines in the world, and timelines cannot interfere with each other. On the timeline, your future self and your present self are the same you. It’s just because the time period exists that your future self can see your past self, but you in the same time period are still unique. If you live to 2050, you will be the you who traveled back from that time.”

“Even if you travel to the future, you may not be able to see your future self. During this period of time, you may die accidentally and disappear in the future! The premise is that the meeting in the past did not change history!”

In short, Chen Pingjiang was confused after listening to it.

When the last stone in his heart fell to the ground, the high-spirited Chen Pingjiang came back again.

……

these four years.

The confrontation between the two superpowers is an unsolvable knot.

Even if Trump steps down, even if a new president is appointed.

And this year Chuanzi is making a comeback.

The sanctions are becoming more and more numerous and intensive, covering all aspects.

The difference is that this time we are stronger than in Chen Pingjiang’s memory.

The total market value of the group companies under Shanhe Holdings has long exceeded 20 trillion yuan. If we calculate GDP, it can be compared with the top few powerful countries.

The valuation of Orange Technology exceeds 5000 billion US dollars, Renren is over 6000 billion, ByteDance is over 4000 billion, Renrengou is over 4000 billion, and Jiuzhou Auto is over 2000 billion, not to mention groups such as Zhizhen Catering, Renren Entertainment, and Galaxy Aerospace.

Chen Pingjiang’s personal finances had already expanded to an exaggerated level.

He had long lost count of how much money he had.

“He controls half of China’s Internet and has a monopoly on all emerging technologies.”

Even if you don’t want to admit it, this is the fact.

What is strange is that even though Alibaba and Tencent have encountered antitrust penalties, the companies under Shanhe Holdings are safe and sound.

Many strange rumors have emerged inadvertently, such as that XX boss is Chen Pingjiang’s father-in-law, and that Chen Pingjiang is actually a third-generation red.

Chen Pingjiang always just smiled indifferently at such rumors.

My father-in-law, Yan Liming, retired last year. When he retired, he also secured the treatment of a senior official, but that was all.

Everything Chen Pingjiang achieved was only because of cheating and his own hard work.

Among the Internet tycoons, Lei Jun has also started making cars and appears on camera frequently, seemingly intending to start his own personal IP.

Lao Ma has left the center of the stage and has been vacationing around the world for many years. His back looks indescribably lonely yet relaxed.

Brother Xiao Ma remains low-key and is rarely seen for years.

Talented people emerge from generation to generation, and a new generation replaces the old.

The more popular companies now are technology companies such as DJI and Yushu Technology.

Chen Pingjiang’s next generation is also growing up happily.

Chen Yu, the little kid who was born in February 11, is now 2 years old.

Perhaps because he was naturally smart, or perhaps because of the strict education from his mother Ge Mengshi, Chen Yu skipped a grade and is now in the third year of junior high school. His academic performance still easily ranks first in the provincial key high school.

Although she is not as naughty as she was when she was a child, the genes of a “violent girl” are still trapped in her small body. She is fascinated by computer hacking technology, robots, and drones, but disdains piano and dance.

This also made Ge Mengshi worry a lot.

What made her even more uneasy was that the young Chen Yu had already declared that he would join the army after graduating from university.

The remaining few little radishes also grew up.

Ge Zichen is now 11 years old. Perhaps because he has a strong mother, he lacks some masculinity. Coupled with his simple personality, he makes Chen Pingjiang very distressed.

After all, Chen Yi inherited the personality of her mother Zhang Xinyue, and is as proud as a white swan.

Chen Jiayi was given a laissez-faire attitude by his mother Cong Nini. Although he was not a bad student, he was a troublemaker in school. He was once given a disciplinary punishment of probation and was later severely beaten by Chen Pingjiang.

Only Chen Yifei can be trusted by his father. He is neither hasty nor arrogant, polite to others, and thinks twice before doing anything.

In addition…

Wan Yuqian did end up killing someone.

She took a year off and went to another place to give birth, where she gave birth to a daughter.

……

Liang Wenfeng appeared in Chen Pingjiang’s office again.

This time, his performance was cold but with a hint of excitement.

“Well, after waiting for so many years, the time is finally ripe.”

Just yesterday, Chuanzi, together with OpenAI founder Ultraman and Facebook founder Zuckerberg, held a press conference and announced the launch of a project called Stargate.

The plan intends to invest 5000 billion in AI research and development.

The whole world was in an uproar.

Liang Wenfeng said excitedly: “We started to build a large position half a year ago and have been laying the groundwork for a long time. We have even bought a lot of put options recently.”

Bu Housheng standing by could not hide his excitement either. This kind of big deal might not come around once in many years.

Although the Americans have tried their best to deny the bubble in their own AI computing power, the world is not blind.

Many data center projects involve duplicate construction, and some even begin financing and construction without clearly identifying customer needs, resulting in a waste of resources.

The current bottleneck in the global AI industry is not insufficient computing power, but how to efficiently utilize existing resources.

Data shows that the GPU utilization rate in the data centers of many technology giants is less than 50%, and a large amount of computing power is idle. However, the United States continues to expand and purchase high-end chips at high prices, pushing up industry costs and exacerbating the risk of overcapacity.

AI is also a battle of national destiny for the Americans. Now they must go all the way and maintain their “self-perceived” lead to avoid causing the collapse of the U.S. stock market and the dollar.

They want to give life to this aging system.

Therefore, it is not surprising that this kind of ALL-IN gamble will create a bubble.

It’s just that the Americans’ general artificial intelligence currently has no competitors, so people don’t care about this bubble, and some even explain it as early construction.

Its investment in AI computing power stems more from anxiety about the “computing power arms race” rather than actual market demand, and there are even few applications that can be put into practice.

At best, everyone can only raise questions on ChatGDP.

……

2025 1 Month 1 Day.

Deepseek officially released two versions of the large model, V3 and R1.

V3 is a general NLP model, while R1 is a specially optimized advanced reasoning model. Each model has its own focus and performs better in specific areas. Deepseek’s V2 model was released quietly before, and because it lagged behind GPT4 in performance, it did not attract much attention.

But this time it’s different!
Renren, WeChat, Weibo, and Douyin platforms are promoting both Deepseek V3 and R1 versions.

It is rare for Chen Pingjiang to publicly sell his products on Twitter.

It attracted a large number of people to try it out at once.

Who knew that I couldn’t stop using it once I started.

“Oh my god, is that Chinese guy bragging? It outperforms top models like GPT-4o and Claude Sonnet 3.5, but the training cost is extremely low. The entire training was completed on a cluster of 2048 Nvidia H800 GPUs, costing only about $557.6 million?”

“If this is true, the world will be shocked.”

“Oops, I need to sell my Nvidia stock.”

“I tried it and I have to say it feels great. I think it works better than GDP.”

“I used to laugh at the fact that Chinese LLMs are not good, but now I’ve been slapped in the face.”

“Most importantly, it’s almost free. Even if you call AIP, the price is shockingly low.”

The popularity on Twitter is growing, and the same is true in China.

Are people in the technology circle and netizens collectively confused?
I just slept for a while, not for a year, right?

How come when I woke up, the domestic LLM model became so awesome?

And who is this Deepseek?
Why haven’t I heard of it before?

How did it come out?
“AI is better at editing manuscripts than me. My job is in danger!”

“Open source and free, developers are ecstatic!”

“His Chinese comprehension is excellent! When analyzing city policies, he even got the dialect puns!”

“The training cost is only 3/1 of Llama10? No wonder the giants are getting restless…”

“It feels like a dream. We were so weak yesterday, but how come we are ahead today? Pinch me.”

“You are not dreaming. The Internet is going crazy. All the media in the United States are reporting on this.”

“I used R1 to write the short video script. From topic selection to storyboarding, I finished it in half an hour. The client praised it for its ‘internet sense’. I used to be bald from staying up late, but now I get off work on time. It even understands the hot memes on Bilibili better than I do!”

“When translating diplomatic jargon, R1 was able to interpret the subtext in conjunction with historical events! Although the knowledge is limited to 2023, the output quality after manual data feeding is much better than that of a certain G model.”

After years of refinement and preparation, today’s Deepseek R1 is more powerful than the version Chen Pingjiang remembers.

A day later, DeepSeek ranked first on the free list of Apple App Store in China, and it also soared to first place on the free list of Apple App Store in the United States, surpassing ChatGPT.

It not only shocked the global AI circle, but also quickly attracted widespread attention and discussion from the international community.

Eight out of ten trending searches are for Deepseek.

Jim Fan, senior research scientist at NVIDIA, said on social media: DeepSeek is the “biggest dark horse” in the field of open source large language models in 2024.

The DeepSeek-R1 model achieved 80.5% accuracy on a math ability benchmark, which is higher than OpenAI’s o1.

In the Codeforces assessment in the programming field, it achieved a score of 2441, which is higher than 96.3% of human participants.

Open source enables overtaking on curves!
The focus of discussion on DeepSeek R1 is on three aspects.

High performance, low training and usage costs, and open source.

Each of these would be fatal to competitors, let alone all three combined!
Different from OpenAI’s “massive data feeding” method, DeepSeek uses algorithms to summarize and classify data, and then feeds it to the big model after selective processing. While improving training efficiency, it also reduces the cost of DeepSeek.

In this regard, Andrej Karpathy, a founding member of OpenAI, said that super-large-scale GPU clusters may not be needed in the future!
DeepSeek-R1 has become the most downloaded large model on the open source community Hugging Face, with 10.9 downloads!

This means that developers around the world are trying to understand this model to assist in their own AI development.

Liang Wenfeng, founder of DeepSeek, said that open source is more like a cultural behavior rather than a commercial behavior.

In the face of disruptive technology, the moat formed by closed source is short-lived. Even if OpenAI closes its source code, it cannot prevent itself from being overtaken by others.

Ritwik Gupta, an AI policy researcher at the University of California, Berkeley, said that the emergence of DeepSeek-R1 proves that “AI capabilities have no technological moat,” and said that China’s system engineers have a much larger talent pool than the United States, and they know how to make full use of computing resources to train and run models more efficiently.

Meta’s chief artificial intelligence scientist Yang Likun also said on social media that the biggest gain from DeepSeek’s success is the value of keeping artificial intelligence models open source so that anyone can benefit. This is the power of open research and open source.

Bring more possibilities to the AI ​​field!

Forbes, Financial Times, CNN, and BBC are all reporting on it!

……

In the early morning of January 2025, 1 Eastern Time, the White House was still brightly lit.

The coffee in the hands of the National Security Advisor had long since gone cold, and the desk was littered with briefing documents marked “Top Secret” – each page was printed with the “DeepSeek” logo.

“It took them less than three years to go from a follower to a rule breaker.” He pushed the tablet to his colleague, and the picture froze on the front page of the New York Times: “China’s AI model costs only 1/20 of that in the United States. Open source strategy shakes the foundation of Silicon Valley.”

At this time, a more intense game is taking place on Capitol Hill. Senator Tom Cotton waved DeepSeek’s white paper at the hearing: “This is a digital Pearl Harbor! China is using our technology to bite us back!”

……

Behind the glass curtain wall of Meta headquarters, LeCun slammed his coffee cup on the table. “Mr. Zuckerberg, we must abandon the closed-source route of Llama3.” He pulled up the data panel: in the past 72 hours, the number of Llama3 downloads on Hugging Face plummeted by 63%, and the developer community of DeepSeek-R1 has spontaneously derived 12 industry-customized versions.

At the same time, an eerie silence permeated the OpenAI headquarters.

奥特曼盯着代码屏幕，上面是工程师团队连夜分析的结论：“R1的稀疏激活算法使GPU利用率提升至89%，这意味着他们用2048块H800实现了我们4096块H100的性能。”

“Launch Project Osiris,” Altman dialed the intercom. “Move forward the release of GPT-5 to the end of the quarter, and tell Microsoft we need an additional $30 billion in computing power investment.” When he hung up the phone, he caught a glimpse of a sample issue of Time magazine on his desk. The cover headline stung his eyes: “When Eastern wisdom meets the open source revolution.”

This is really a headache!
Even if Google or Mete came up with a big model and even posed a challenge to OpenAI, Ultraman would not be so panicked.

Because he knows that everyone is essentially walking the same path.

But this time it was different. As long as the Chinese were proficient in something, they would immediately make it cheap, and then break the unspoken rules, making it impossible for anyone to make money.

What’s more, the R1 released by the other party at this press conference far exceeds GDP4.

It’s simply killing me!
Why did he appear without making any noise?

……

As alarm bells rang throughout the trading floor, announcing a 12% plunge in the Nasdaq, hedge fund manager Johnson slammed his phone against the floor-to-ceiling window.

The glass did not break, but the scene reflected in it made him devastated: the paper loss of their Nvidia stock positions had reached US$43 billion.

“Short! Short all!” His roar was drowned out by the alarms sounding from the Bloomberg terminal.

Obviously, Nvidia’s stock is bound to plummet, and no one can save it.

……

On New Year’s Eve 2025, Chen Pingjiang stood in front of the AR big screen of CCTV’s New Year’s Eve Gala.

When the clock struck midnight, he tapped his phone, and the greetings generated by R1 appeared among the fireworks on the big screen – and were presented to audiences around the world in 32 languages ​​at the same time.

On the other side of the Pacific Ocean, Altman suddenly announced on Twitter that the release of GPT-5 would be postponed indefinitely.

“We tried to replicate the R1 algorithm, but found that for every 1% increase in hardware utilization, we needed to rebuild 20% of the underlying architecture,” an unnamed OpenAI engineer told the Wall Street Journal.

He said frustratedly: “We lost the AI ​​competition because we took the wrong path from the beginning and were too obsessed with the advantages brought by hardware.”

[I will update the postscript tomorrow, it will be very interesting. ] (End of this chapter)