After the layoffs, I became a tech godfather

Chapter 153: The two camps of China and the United States completely parted ways at this moment

"Ding Dong, hello dear user, the new version of Xiaosha has been pushed. New model pruning, distillation, and training capabilities have been added. Come and try it out and train a model of your own!

"For commercial applications, please contact Baiju Technology to obtain special authorization."

At 10:30 pm, Xiaosha officially released an announcement to the public, officially updating the optimization and training capabilities of Xiaosha's external models.

In fact, all the mainstream AI large-model training companies in the world are now using Xiaosha to do various tasks, such as data cleaning, training feedback, etc.

There have even been some unexpected incidents involving new humans before.

But today its meaning is different. Baiju Technology has officially released a professional AI model for [training models].

As for the effect, the update log says:
20. Increase the upper limit of the original model by at least %.

It is easy to understand that no matter what model you originally have and what level you have reached, you can raise the upper limit of the model's capabilities by training and improving it with Xiaosha, which has enhanced AI training expertise.

50. Save 95%~% of training costs

For top companies, they do not need to save costs, they only hope to infinitely increase their capacity ceiling in order to become bigger and stronger.

But for more small and medium-sized enterprises or start-ups, saving costs is their top priority.

Now, for the same performance, depending on the model type, the training cost can be saved by more than half, up to 95%.

Naturally, you can't have both one and two.

3. Save training time.

This is very easy to understand. The work that originally required a lot of people to screen data and branch feedback was all handed over to Xiaosha.

As long as you are willing to sacrifice computing power, the concept of speed no longer exists.

As soon as the version was updated, the three major advantages were presented, which immediately caused an uproar.

"Oh my God, does this prove that Li Da's flying brick route is completely wrong?" Someone immediately realized this.

"That's not the case. There are still benefits to piling up computing power, but the more you pile up, the smaller the benefits become."

Some professionals immediately made relevant calculations: "According to Baiju Technology, if you only need 80% of the performance, then only 20% of the computing power is enough.

"However, if you want 100% performance, or even more, the computing power required will still be higher, and it will cost 80% or even more than the original cost."

"A friendly reminder to everyone: unlike companies that only boast all day long, Baiju Technology's rhetoric is conservative. How much it can improve and save specifically still needs to be verified in actual practice."

Having said that, the netizen has posted a screenshot of the AI he is training using Xiaosha:
"Brothers, wait for my test! Tomorrow night at 8 o'clock, the live broadcast room will broadcast the ultimate showdown between this newly trained model and Sora!"

This brother is obviously a professional, and many more non-professionals have also begun to join the ranks of model training and are still groping in the dark at this time.

Xiaosha has carefully prepared a beginner's tutorial for these people: building your own private large model from scratch.

"Xiaosha, what should I do first?"

[First, you need a graphics card. Based on your current needs, we recommend that you choose Moore Thread S80. Baiju Technology has reached an agreement with Moore Thread, and will provide continuous optimization for Moore Thread games or computing power cards in the future. ]

"Can't I use my RTX4090?"

[That's fine.]

"What's next?"

This is a new player with absolutely no basic knowledge. He is not even a programmer. He just saw Xiaosha had such a service and couldn't help wanting to try it out.

【We have prepared 108 kinds of model bases for you, please choose one. 】

[Next, according to my prompts, we start segmenting words and expanding our vocabulary...]

Xiaosha did not explain any professional knowledge to these people with no basic knowledge, but only prompted them to follow his instructions step by step.

As for writing code, Xiaosha can do it for you.

……

When users follow Xiaosha's step-by-step instructions and complete all the steps, and a simple model is successfully trained and deployed by their own hands, the sense of satisfaction that arises naturally is refreshing and fascinating.

"Hahaha! Look at me!" Sun Wei was like this. As soon as the model was trained, he showed it off to his friends on Douyin.

"Hello, Xiaowei!"

【Hello! 】

"Look, look, he can talk!" Sun Wei pointed at his mobile phone with excited heart and trembling hands. He looked so excited, just like the feeling he had when he taught his son to call him dad for the first time.

Even though it’s just two simple words, this is the first time!

Like Sun Wei, many netizens have trained and deployed their own language models and completed the first sentence of conversation.

"This is not right!" One of the netizens, named Gao Hongyu, is a professional who originally worked in AI training-related work.

After receiving Xiaosha’s update, I immediately performed the relevant operations according to the novice tutorial.

"Master Gao, what's wrong? Isn't this conversation good?" a group of people asked.

Although everyone can see that this newly trained model is certainly far inferior to Xiaosha, let alone Xiaosha, any AI that can be named is better than it.

However, everyone was watching how Gao Hongyu trained himself to do this.

There is only one NVIDIA RTX4090 graphics card in total, and the training time is not long, only two hours so far.

Being able to have a fluent conversation is good enough.

"That's exactly what's wrong!" Gao Hongyu said, "You don't know that because you haven't trained AI, the first version trained by AI had irrelevant dialogues, because every word was calculated based on probability.

“After countless cycles of fine-tuning various feedback decisions, we have achieved the current effect. But this is only the first version!

"No matter how it responds, do you believe that we can understand what it says?"

"Oh!" All the netizens who were watching the live broadcast breathed a sigh of relief instantly. It turned out that what Gao Shen said was too strong, not too bad.

Someone immediately asked: "How many cycles did you need to train to this level in the past?"

"In the millions!" Gao Hongyu did not say a specific number, but rather an order of magnitude.

"Hiss!" Everyone was shocked. Isn't this too different?
"Xiaosha, Xiaosha, why is this happening?" Gao Hongyu asked Xiaosha. [In Chinese, the meaning of the Chinese word is rich, and the same number of tokens can express richer meanings. This is a unique advantage in AI training.

[This alone can save more than 30% of computing power compared to semantically poor languages.

[With Xiaosha's enhanced professional capabilities, the training of the basic model consumes almost no computing power. ]

Gao Hongyu understood Xiaosha's words at the beginning, and this coincided with the description in the update log - the simpler the model, the more computing power can be saved.

80% of the capabilities only require 20% of the computing power.

The one you train yourself is just a toy, it has less than 0.1% of its capabilities, so the computing power requirement can almost be ignored.

Don’t even mention that you have a 4090 graphics card. Even if you don’t have one, the integrated graphics card or even the arithmetic unit in the CPU may be enough.

……

But Gao Hongyu was not the only one who discovered this. Countless professionals and relevant institutions at home and abroad also conducted relevant evaluations on Xiaosha’s update as soon as possible.

The results of the evaluation...

Just look at the stock price trend of Nvidia: it opened down 10% at 30:2.7 Beijing Time, which is still a normal fluctuation.

But just ten minutes later, news of Xiaosha's updated version spread. Although no one knew the specific performance of this new version at that time, the impact had already emerged:
It collapsed by 12% during the trading session. Two hours later, the ability of Xiaosha's new version was verified, and it crashed by 17%. In just over two hours, the market value evaporated by nearly one-third and exceeded one trillion US dollars.

At this time, the second emergency meeting of the New Liberal Party was still in progress and had lasted for more than ten hours.

Huang Renxun, like Nadella, has begun to grit his teeth.

In the past, Xiaosha's existence had little impact on NVIDIA, and sometimes the impact was even positive.

After all, Xiaosha is so strong. Even if you know that its training cost is very low, other people can't use it.

If America wants to surpass Nvidia, it must add more cards. For example, OpenAI has contributed tens of billions of dollars in revenue to Nvidia just by purchasing cards during this period.

But, today, it's completely different!

The powerful plan that Nvidia had been advocating was completely exposed, and the multi-trillion dollar AI story that it had finally told was completely ruined by Baiju Technology.

Huang Renxun looked at Altman, and Altman happened to look at Huang Renxun at this time. The two exchanged glances at this moment: What should we do?
"We will definitely not use Xiaosha's abilities." Altman stated immediately.

Huang Renxun thought to himself, you are talking nonsense, if you want to use Xiaosha’s abilities, others will have to let you use them.

Moreover, not all of Xiaosha's abilities are open. The AI trained using Xiaosha can only reach 120% of the most advanced level in the world today except for Xiaosha. This is far from Xiaosha, and is even far inferior to the new humans.

“Well, Baiju Technology is too bad!” Huang Renxun said, “Their intention is very clear, that is, to the extent that OpenAI achieves, their Xiaosha trained AI will be released to the level that exceeds yours by 20%.

"As you improve, Baiju Technology will also open up again and will always be 20% higher than you!"

Huang Renxun is here, it's really a murderous act. It's okay that OpenAI can't beat Xiaosha, but the AI trained by Xiaosha will also be permanently 20% ahead of them. Who can bear this!

Altman was an irritable man, and he got even more angry at this moment: "No, I don't believe it. Add more cards, add more cards!"

Although he knew that the powerful flying brick was a trap, Altman had to jump in because they had no way out now.

And this is exactly what Huang Renxun wants. Hey, I don’t care how your AI training is, as long as you buy the card and continue to tell this story with us.

“I think we should unite and continue to share the trained AI capabilities with all alliance companies, and then the alliance companies should also contribute to AI training.”

Huang Renxun made this suggestion. He wanted to get all of America's software, hardware, system and other companies involved. There is strength in numbers, and only if everyone participates can more cards be purchased.

"Then, while training the AI, we will ban Xiaosha worldwide. At least our allies must ban it. As long as we find a way to train an AI like Xiaosha, we will win!"

It has to be said that Huang Renxun was thoughtful. He brought the topic back to the original decision and sold more cards at the same time.

……

The next morning, after Hao Cheng received Xiaosha's morning news summary, he was stunned for a moment, but then he smiled.

As soon as he arrived at the company, Li Qingbo also reported relevant intelligence and data. NVIDIA, OpenAI and the American Top Enterprise Alliance formed an AI training joint action consortium.

The consortium announced that it plans to invest more than $300 billion in computing power cards in the coming year.

At this moment, the two camps of China and the United States completely began to part ways, each going their own way.

Affected by this, Nvidia's stock finally stopped its decline, and the full-day decline was reduced to 21%.

"This is normal," Hao Cheng set the tone: "Our target this time is the software industry in America, and bursting the bubble of AI computing power cards led by Nvidia is just a by-product.

“Moreover, even if the bubble bursts, the demand for computing power cards will still exist in the long run, but it’s just not as big as they boast.

“Also, it is true that Xiaosha has improved efficiency, but the threshold for making and training large models has also been lowered, which will definitely attract more small and medium-sized enterprises and even individual users to train their own models.

“In this case, it is hard to say whether the total amount of computing power demand will increase or decrease.

"Therefore, as long as the domestic chip production capacity does not completely break through the limit one day, Nvidia will at most be half-crippled. No matter how many or few cards are used, they will always be used.

"What's more, if we want to move towards more powerful AI capabilities, we still need a lot of computing power cards at this stage. It's really too early to say that Nvidia is going to close the curtain."

Hao Cheng finally said: "Don't think that the enemy will fall down as soon as we open fire, and don't think that once the enemy falls down, they will never be able to stand up again.

"Remember, our first phase goal and achieving it thoroughly is our current mission."

"Yeah!" Li Qingbo quickly started his full-day work as if he had been injected with chicken blood, while Hao Cheng himself, with the assistance of Xiaosha, browsed the feedback about the borderless system on the entire Internet.

"Xiaosha, help me summarize the exploration of the boundless system on the Internet and organize it into categories." Hao Cheng said.

【Okay, in progress... Completed. 】

"Xiaosha, what do you think are the new features and functions of the Unbounded System that users are most concerned about? Help me select some of them. I will use them in the subsequent press conference."

Hao Cheng didn’t know whether he had set a precedent. The product had so many new features and functions that he had no idea how to introduce and promote it.

It’s better to just throw it out there and let users explore and experience it themselves, and then at the press conference we’ll focus on the things they care about.

[The top three features are the new operation mode, somatosensory and possible more types of games in the future, and the user behavior password that can be sensed at any time to verify the authorization. ]

"Based on these three features, I was asked to plan three promotional videos, produce three reference samples, and a demo film for the press conference."

【OK. 】

(End of this chapter)