Algorithm system

Take actionAfter testing DS-LLM’s performance, Lin Yuan became interested in DS, this small company.

He scoured the entire internet and discovered that this company had only recently been established, and its initial purpose was surprisingly not to develop general AI.

Its original intention was to research how to apply AI technology to the financial trading market.

What could Lin Yuan even say about this?

A chef, in order to cook good food, personally forged an iron pot, and it turned out that the performance of this iron pot was no less than that of a professional blacksmith’s masterwork.

When the chef forged the iron pot, he never thought about engaging in the blacksmithing industry in the future; he simply wanted to make a good pot for cooking.

If the matter had stopped there, it would have only been slightly bizarre.

What truly surprised Lin Yuan was how little computing power DS-LLM required.

Although Lin Yuan had previously gone to great lengths to acquire two H100s deployed in the cloud for training, even submitting a report to his superiors for approval, compared to the large models like ChatGPT, this was simply not an issue.

Want to deploy ChatGPT locally? Let’s not even mention that OpenAI is now CloseAI, so you can’t get the latest full-blooded version of ChatGPT.

Even if OpenAI were to open-source ChatGPT, the terrifying computing power cost of the latest ChatGPT4 version means that even if you could afford the GPUs, you wouldn’t be able to afford the electricity bill.

So, the outline of this entire matter became: a chef, in order to cook good food, specially forged an iron pot, and it turned out that this iron pot not only had similar performance to one forged by a professional blacksmith but also required less iron and labor.

This was simply too much!

Lin Yuan then proceeded to test the DS-V2 version.

Overall, the full-blooded DS V2 version was already close in performance to the early versions of ChatGPT4, but there was still a considerable gap compared to the ChatGPT4-o series released early last year.

But this was already enough to surprise Lin Yuan, because such a small team of just over a hundred people had created such an astonishing AI Large Language Model.

In contrast, look at those Big Companies, especially Dudu’s, who are constantly thinking about how to improve search experience with AI.

They are clearly also a chef attempting to make a good pot to cook good food, yet the difference between the two is so immense.

This immense difference is not reflected in the two companies’ Large Language Models, but rather in their input-output ratio.

Dudu’s has claimed to be involved in AI for about ten years, even once shouting the slogan ”all in AI.”

For a period, it did indeed gather many Big AIs in the field.

As a result, it inadvertently became an incubator for other companies, becoming a semi-“Whampoa Military Academy” in the AI field.

The current example refreshed Lin Yuan’s understanding—it turns out that unintentional actions can indeed lead to unexpected success.

The goal can change, but the determination to achieve it must be firm.

Most commendably, DS also open-sourced its AI Large Language Model, which shows much more vision than CloseAI.

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

Because this small company perfectly met Lin Yuan’s selection criteria—it had sufficient technical strength, sufficient perseverance, was not strongly associated with foreign capital, and most importantly, was not yet widely known.

DS was precisely at the perfect balance point between “Big” and “Small.”

Cooperating with too Big Companies is not only troublesome, but Big Companies also have a host of rules and regulations, and Lin Yuan would inevitably get entangled in them.

Just the algorithm intellectual property alone would be annoying enough.

And then there are the endless meetings and various charters; Lin Yuan gets a headache just thinking about it.

After all, Big Companies have a complete set of risk control mechanisms for risks, and as a result, after Big Companies scale up, their innovation capabilities tend to decline.

To cope with this situation, Big Companies create project teams, similar to independent studios, to tackle difficult problems.

However, these types of project teams are hidden too deep within Big Companies, and Lin Yuan had no opportunity to directly cooperate with these internal organizations, like special forces.

And what about small companies? Too small companies often face the problem of lacking R&D strength.

Another even more headache-inducing issue is ambition.

A small company’s first priority is to survive, and it would not normally be willing to invest resources in general-purpose technology fields.

So, DS’s size was just right.

Perhaps this was fate.

Lin Yuan quickly found DS’s open-source project on GitHub and sent an in-site message to their account.

Although as an open-source project, Lin Yuan could have directly submitted a project contribution request.

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

Once a request is publicly submitted, everyone will be able to see that a certain account submitted a key code update to DS’s Large Language Model.

What if one day someone uncovers that the account that submitted this update was Lin Yuan’s?

Then, if by chance, Lin Yuan’s short-selling of NVIDIA was also discovered, and the two were linked, wouldn’t Lin Yuan become someone who had meticulously planned to short NVIDIA?

Although Lin Yuan wanted to short NVIDIA to retaliate against Musk and vent his anger, doing the deed is one thing, but leaving traces is another.

Once obvious traces are left, there will be a lot of trouble.

Who knows how the line between successful short-selling and stock manipulation is distinguished?

Therefore, Lin Yuan only sent a private message, and he did so using a temporarily registered account.

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

According to the computing power system’s in-depth scan of the current major AI Large Language Models, the system’s comment was: the training process overly emphasizes the importance of computing power, thereby ignoring the value of inference.

In layman’s terms: they just blindly work without thinking.

Lin Yuan’s understanding of this is like solving a math problem: once you know the rules of addition, subtraction, multiplication, and division, you can perform addition, subtraction, multiplication, and division operations on any number of digits based on this, which is the process of inference.

In contrast, this is like forcing an answer.

Of course, the flaws in the current AI Large Language Models pointed out by the computing power system would naturally not be so superficial or crude.

But the principle is roughly the same.

So, Lin Yuan processed the insights from the computing power system and sent them to the other party.

He certainly wouldn’t just send a vague statement about emphasizing the inference process; instead, he performed targeted optimization on a specific set of neural networks within the DS-V2 version code.

Lin Yuan used this set of neural networks as a reference, modifying the code based on it, and using examples to illustrate the computing power system’s argument about emphasizing inference.

After the modification, this set of neural networks gained a dynamic, automatically adjustable attribute.

This allows it to start from the logical chain when answering practical questions, demonstrating the model’s own inference process to the user.

This point is crucial, because if the AI’s feedback inference process is wrong, you can directly interrupt the AI’s answer without having to wait for the AI to finish answering and then judge by reading the answer.

This inference process is like the outline of an answer, allowing humans to perceive the AI’s inference details, thereby achieving a bidirectional connection.

Because the way AI wants you to ask questions is often not the way humans actually ask questions.

Furthermore, the human language system inherently contains a concept of continuously deepening the topic as a conversation progresses, meaning that conversations between people require context.

This is easy to understand in daily conversations: the initial stage of conversation between two strangers always involves establishing a shared context.

Since this is the case, conversations between humans and AI are also like this.

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