Until a few weeks ago, some people in the Western world had heard of a small Chinese Artificial Intelligence (AI) company known as Deepsek. But on 20 January, it captured global attention when it released a new AI model called R1.
R1 is a “logic” model, which means that it takes step by step through tasks and details the user its work process. It is a more advanced version of the V3 model of Deepsek, which was released in December. The new offering of Deepsak is almost as powerful as the most advanced AI model of oppressed AI model O1, but at a fraction of the cost.
Within days, Deepsek’s app crossed the Chatgpt in the new download and determined the stock prices of tech companies in the United States. It also inspired Openai to claim that her Chinese rival had effectively piloted some crown jewels of Openi’s model to manufacture her own.
In a statement to New York Times, the company said:
We know about the fact and review those signs that Deepsek may have improperly distilled our model, and as we know more, we will share information. We will take aggressive, active counters to protect our technology and will continue to work closely with the US government to protect the most capable model being created here.
The conversation for the conversation contacted Deepsak, but did not respond.
But even if Deepsek was copied – or, in scientific parlance, “distilled” – at least some chatgipt R1, it is worth remembering that OpenII also opening its models. Accused of disrespecting wealth.
What is distillation?
Model distillation is a general machine learning technique in which a small “student model” is trained on the predictions of a large and more complex “teacher model”.
When complete, the student can be almost good as a teacher, but will represen the teacher’s knowledge more effectively and compactly.
To do this, it is not necessary to reach the teacher’s internal functioning. Everyone needs to pull this trick, asking enough questions from the teacher model to train the student.
This is what Openai has claimed that Deepsek has done: Openai’s O1 is a large -scale Queried and Deepsek’s own, used the output seen to train more efficient models.
A fraction of resources
Deepsek claims that both R1’s training and use require a fraction of the resources required to develop the best model of their rivals.
The company’s some marketing is due to doubts about publicity – for example, a new independent report suggests that the hardware expenses on R1 were as high as US $ 500 million. But still, Deepsek was still built very quickly and efficiently than the opponent model.
This can happen because Deepsek has distilled Openi’s output. However, currently there is no method to prove it decisively. A method that is in the early stages of development is to watermark the AI output. This adds invisible patterns to the output, the same applied to copyright images. There are different ways to do so in theory, but no one is effective or efficient so that it can be made in practice.
There are other reasons that help explain the success of Deepsek, such as deep and challenging technical functions of the company.
The technological progress made by Deepsek included taking advantage of less powerful but cheap AI chips (also called graphical processing units, or GPU).
Deepsek had no option after banning the firms after banning the US’s most powerful AI chips to China.
While Western AI companies can buy these powerful units, the export ban forced Chinese companies to innovate to use cheap options best.
A series of cases
The conditions of the use of Openai clearly suggest that no one can use its AI model to develop competitive products. However, its own models are trained on a large scale dataset scraped from the web. These datasets contained sufficient amount of copyright content, which OpenEE says it is entitled to use on the basis of “proper use”:
Training of AI models using publicly available internet content is proper use, as is supported by long and widely accepted examples. We consider this theory appropriate for the creators, are essential for innovators, and are important for American competition.
This argument will be tested in court. Newspapers, musicians, writers and other creatives have filed a series of cases against openiAI based on copyright violations.
Of course, it is quite different as to what Openai has accused Deepsak of doing. Nevertheless, Openai is not attracting much sympathy for his claim that Deepsek illegally harvested his model output.
The war of words and cases is an art of how AI’s rapid advance has carried forward the development of clear legal rules for the industry. And while these recent events can reduce the power of AI incumbents, there is much more on the results of various ongoing legal disputes.
Stirly
Deepsek has shown that it is possible to develop a state -of -the -art model cheap and efficiently. Can they compete with Openai on a level playground, it remains to be seen.
Over the weekend, Openai publicly tried to demonstrate its supremacy by releasing its most advanced consumer model, O3-Mini.
Openai claims that this model is its own previous market-agar version, O1, and “the most cost-skilled model in our logic chain”.
These developments are an era of enlarged choice for consumers with diversity of AI models in the market. This is good news for users: the competitive pressure will make it cheaper to use the model.
And benefits move forward.
Training and using these models keeps heavy stress on global energy consumption. As these models become more ubiquitous, we all benefit from improving their efficiency.
The rise of Deepsek certainly marks a new area to make models more cheap and efficiently. Perhaps it will also shake the global conversation of how AI companies should collect and use their training data.
(Author: Lee Frman, Senior Lecturer in Natural Language Processing, Melbourne University, The University of Melbourne And shaanan cohney, lecturer in cyber security, Melbourne University)
This article is reinstated by negotiations under a creative Commons License. Read the original article.
(Except for the headline, the story has not been edited by NDTV employees and is published by a syndicated feed.)
