Until a few weeks ago, few people in the western world had heard of a small Chinese artificial intelligence company (AI) known as Deepseek. But on January 20, Global Attention captured When he launched a new AI model called R1.
R1 is a “reasoning” model, which means that it works through step by step and details its work process to a user. It is a more advanced version of Deepseek’s Model V3which was released in December. Deepseek’s new offer is almost as powerful as the most advanced AI model of the OpenAi rival company, but to a cost fraction.
In a matter of days, the Depseek application surpassed ChatGPT in new downloads and established prices of actions of technology companies in the United States drop. Also led Openai to say that his Chinese rival had effectively stolen some of the crown jewels of Openai models to build his.
In Declaration to New York TimesThe company said:
We are aware and reviewing the indications that Depseek may have inappropriately distilled our models, and share information as we know more. We take aggressive and proactive countermeasures to protect our technology and continue working closely with the United States government to protect the most capable models that are being built here.
The conversation approached Depseek to comment, but did not respond.
But even if Deepseek copied, or, in the scientific language, “distilled”, at least part of Chatgpt to build R1, it is worth remembering that Openai is also accused of lacking respect for intellectual property while developing its models.
What is distillation?
The distillation of the model is a common automatic learning technique in which a smaller “student model” is trained in predictions of a larger and more complex “model” model.
When completed, the student can be almost as good as the teacher, but will represent the teacher’s knowledge more effectively and compact.
To do so, it is not necessary to access the internal functioning of the teacher. Everything that one needs to achieve this trick is to ask enough questions to the teacher’s model to train the student.
This is what OpenAi states that Depseek has done: consult O1 O1 on a mass scale and used the results observed to train the own and most efficient models of Deepseek.
A fraction of resources
Veteran claims That both training and the use of R1 required only a fraction of the necessary resources to develop the best models of their competitors.
There are reasons to be skeptical of some of the company’s marketing bombings, for example, a New independent report He suggests that hardware spending in R1 was as high as US $ 500 million. But even so, Depseek was still built very quickly and efficiently compared to rival models.
This could be due to the fact that Depseek distilled Openai’s exit. However, there is currently no method to prove this conclusively. A method found in the early stages of development is Water brand AI outputs. This adds invisible patterns to the outputs, similar to those applied to the images with copyright. There are several ways to do this in theory, but none is effective or efficient enough to have it into practice.
There are other reasons that help explain Depseek’s success, such as the company’s deep and challenging technical work.
The technical advances made by Deepseek included take advantage of the less powerful but cheaper chips (also called graphic processing units or GPU).
Depseek had no choice but to adapt after The United States has banned companies From exporting the most powerful chips to China.
Although Western companies can buy these powerful units, the export prohibition forced Chinese companies to innovate to take the best use of cheaper alternatives.
A series of demands
Operai’s Conditions of use Explicitly indicate that no one can use their AI models to develop competitors. However, their own models are trained in masive data sets on the web. These data sets contained A substantial amount of copyrightwhat Operai says he has the right to use Based on “fair use”:
The training of AI models that use publicly available Internet materials is for fair use, as supported by the precedents of long data and widely accepted. We see this principle as just for creators, necessary for innovators and critical competitiveness for us.
This argument will be tested in court. Newspapers, musicians, authors and other creatives have filed a series of demands against OpenAi for the violation of copyright.
Of course, this is quite different from what Openai accuses a gransejo of doing. However, OpenAi is not attracting much sympathy For his statement that Deepseek illegitimately reaped its model production.
The war and demand war is an artifact of how the rapid advance of AI has overcome the development of clear legal rules for the industry. And although these recent events could reduce the power of IA holders, many depend on the result of the various legal disputes.
Shaking the global conversation
Deepseek has shown that it is possible to develop state -of -the -art models in an economical and efficient way. It remains to be seen if they can compete with Openai on a level playing field.
During the weekend, Operai tried to demonstrate his supremacy for Publicly releasing Your most advanced consumption model, O3-mini.
Operai states that this model substantially exceeds its own leading previous version in the market, O1, and is the “most profitable model in our reasoning series.”
These developments announce an era of greater choice for consumers, with a diversity of AI models in the market. This is good news for users: competitive pressures will make the models cheaper to use.
And the benefits extend more.
The training and use of these models place a massive tension on global energy consumption. As these models become more ubiquitous, we all benefit from the improvements in their efficiency.
Deepseek’s rise certainly marks a new territory to build models in a cheaper and efficient way. Perhaps the global conversation about how AI companies should collect and use their training data will also shake.
(Author: Read FrermannProfessor of Natural Language Processing, University of Melbourne, The University of Melbourne and Shaanan CohneyCybersecurity professor, The University of Melbourne)
This article was republished from The conversation Under a Creative Commons license. Read the Original article.
(Except for the headline, this story has not been edited by NDTV staff and is published from a union feed).
By
