Xueba's computing system

Chapter 269 Take action
After trying the performance of DS-LLM, Lin Yuan became interested in this small company, DS.

He searched the Internet and found that the company was founded not long ago, and its original intention was not to develop general AI. Instead, it was to study how to apply AI technology to the financial trading market.

What the hell is this... What should Lin Yuan say?

In order to cook delicious dishes, a chef made an iron pan by himself. It turned out that the performance of this iron pan was as good as that of a professional blacksmith's.

When the chef was making the iron pot, he didn't think about working in the iron industry in the future. He just wanted to make a good pot for cooking.

If the matter ended here, it would only be a slightly bizarre story. What really surprised Lin Yuan was that the computing power required by DS-LLM was so little.

Although Lin Yuan had previously specially got two H100s deployed in the cloud for training, and even submitted a report to his superiors for approval, compared with those big guys like chatGPT, this is not a big deal at all.

Want to deploy chatGPT locally? Not to mention that openai is now closedai, you can't get the latest full-blooded version of chatGPT at all. Even if openai really open-sources chatGPT, the latest version of chatGPT4 has a terrible computing power cost. Even if you can afford a GPU, you can't afford the electricity bill.

So, the outline of the whole thing became: a chef specially made an iron pot in order to cook better dishes. It turned out that this iron pot not only had similar performance to that made by a professional blacksmith, but also required less iron and labor hours.

This... is just too much.

Lin Yuan then tried the DS-V2 version. After overall testing, the DS V2 full-blooded version was close to the performance of the early version of chatGPT4, but there was still a big gap compared to the chatGPT4-o series released at the beginning of last year.

But this was enough to surprise Lin Yuan, because such a small team of more than 100 people actually created such an amazing AI language model.

In contrast, look at those big companies, especially Dudujia, who are thinking about improving the search experience with AI all day long. It is obviously a chef who wants to make a good pot to cook good dishes, but the gap between the two is so huge.

The huge gap is not reflected in the two companies' large language models, but in the input-output ratio.

Dudujia claimed to be engaged in AI about ten years ago, and even shouted the slogan "all in AI" for a time. For a period of time, it did gather many big names in the field of AI. As a result, it accidentally became an incubator for other companies and became half of the "Whampoa Military Academy" in the field of AI.

The example before him refreshed Lin Yuan's cognition - it turns out that an unintentional planting of willows can really lead to a forest of willows. The goal can be changed, but the determination to achieve it must be firm.

What is most commendable is that DS company has also open-sourced its own AI language model, which is much more advanced than closeAI.

In view of this, Lin Yuan decided to cooperate with this small company.

Because this small company completely meets Lin Yuan's selection criteria - it has sufficient technical strength, sufficient perseverance, and has no connection with foreign capital, and most importantly, it is not well known.

DS is a company that strikes the perfect balance between big and small.

Not only is it troublesome to cooperate with a company that is too big, but it also has a lot of rules and regulations, and Lin Yuan will inevitably be deeply involved in them. The algorithm intellectual property rights alone are annoying enough. And there are endless meetings and regulations, which make Lin Yuan feel overwhelmed just thinking about it.

After all, large companies have a complete risk control mechanism for risks, so as large companies grow in size, their innovation capabilities will decline. In order to deal with this situation, large companies will set up some project teams similar to independent studios.

However, these project teams were hidden too deeply within large companies, and Lin Yuan had no chance to work directly with internal organizations like these special forces.

As for small companies, they often face the problem of lack of R&D capabilities. Another more troublesome problem is the ambition issue. The first priority of small companies is to survive. According to common sense, they are not willing to invest resources in general technology fields.

So, the scale of DS is just right.

Maybe this is fate. Lin Yuan quickly sent a private message to the other party's account through DS's open source project on github. Although as an open source project, Lin Yuan can directly submit a project contribution request.

However, Lin Yuan still wants to keep a low profile.

Once the request is made public, everyone can see that a certain account has submitted key code updates to DS's large language model. What if one day someone finds out that the account that submitted the update is Lin Yuan's?

If by chance, Lin Yuan's shorting of NVIDIA is discovered again, and the two are connected, wouldn't Lin Yuan be considered to have deliberately shorted NVIDIA?

Although Lin Yuan wanted to vent his anger by shorting NVIDIA, it is one thing to do, but whether or not to leave traces is another matter. Once obvious traces are left, there will be a lot of trouble.

Who knows where to draw the line between successful short selling and stock price manipulation.

Therefore, Lin Yuan just sent a private message, and he sent it using a temporary account.

In the private message, Lin Yuan provided the other party with a key idea, a key idea based on reasoning and evolution.

Based on the computing power system's in-depth scan of the major AI language models currently available, the system's comment is: the importance of computing power is overemphasized during training, thereby ignoring the value of reasoning.

In simple terms, it means: you just act recklessly without using your brain.

Lin Yuan's understanding of this is that it is just like doing math problems. Once you know the rules of addition, subtraction, multiplication and division, you can do addition, subtraction, multiplication and division operations on any number of digits based on this. This is the process of reasoning.

The opposite of this is to force an answer.

Of course, the defects of the current AI language models on the market pointed out by the computing power system are not so shallow and poor. But the principle is similar.

Therefore, Lin Yuan processed the insights of the computing system and sent them to the other party.

Of course, he did not just post a random statement emphasizing the reasoning process, but instead made targeted optimizations to a certain set of neural networks in the DS-V2 version code.

Lin Yuan used this set of neural networks as a reference, modified the code based on it, and used examples to illustrate the argument of the computing system that emphasizes reasoning.

After the modification, this set of neural networks comes with dynamic and automatic adjustable properties, so that when answering practical questions, it will start from the logical chain and show the model's own reasoning process to the user.

This is crucial because if the reasoning process fed back by the AI is wrong, you can directly interrupt the AI's answer without having to wait for the AI to finish answering and then make a judgment by reading the answer.

This reasoning process is like a guide to the answer, allowing humans to perceive the details of AI's reasoning and thus achieve a two-way collaboration.

Because the way AI wants you to ask questions is often not the way humans actually ask questions. Moreover, the human language system has a concept of deepening the topic as the conversation progresses, that is, conversations between people need to be contextualized.

This is easy to understand in daily conversations. The early stages of a conversation between two strangers always involve establishing a common context. If this is the case, then the same is true for conversations between humans and AI.

Emphasizing reasoning means emphasizing that AI should grasp this logical chain of reasoning.

(End of this chapter)