total-news-1024x279-1__1_-removebg-preview.png

SELECT LANGUAGE BELOW

OpenAI Says DeepSeek Copied ChatGPT, But It’s Facing Copyright Claims Too

Until a few weeks ago, few people in the western world have heard of small Chinese artificial intelligence (AI) companies known as Deepseek. But January 20th, it I got global attention When a new AI model called R1 is released.

R1 is a “reasoning” model. In other words, the tasks work in stages and explain the work process in detail to the user. Deepseek is a more advanced version V3 modelReleased in December. Deepseek's new product is almost as powerful as the rival company Openai's most advanced AI model O1, but it is a little cost.

In a few days, DeepSeek apps have expanded Chatgpt by new downloads and set stock prices for high -tech companies in the United States. Tumbling。 In addition, we derived Openai Claim The Chinese rival built a unique model by effectively stealing some of the crown gems from the open model.

in Statement on the New York TimesThe company said:

We are reviewing that DeepSeek may be inappropriately distilled in the model and share information as you know the details. We take aggressive and aggressive measures to protect the technology, and continue to work closely with the US government to protect the most capable models built here.

The conversation approached DeepSeek for comments, but did not respond.

However, even if DeepSeek is copied, or “distilled” in scientific terms, or at least some of ChatGpt to build R1, Openai is neglected while developing models. It's worth remembering that you are accused.

What is distillation?

Model distillation is a general machine learning method in which smaller “student models” are trained in the prediction of a larger and complex “teacher model”.

When completed, students may be almost as good as teachers, but they express teachers more effectively and compactly.

To do so, there is no need to access the interior mechanism of the teacher. The need to do this trick is to ask a teacher model for enough questions to train students.

This is what Openai claims that DeepSeek did. We had a large -scale inquiry of O1 of OPENAI and used observed output to train Deepseek's own more efficient model.

Part of the resource

Deepseek claim Both R1 training and use require only a small part of the resources needed to develop the best models of competitors.

Some of the company's marketing hype is skeptical. for example, New independent report It suggests that the hardware spending to R1 has reached $ 500 million. But still, DeepSeek was built very quickly and efficiently compared to rival models.

This may be due to DeepSeek's Openai output. However, there is no way to definitely prove this. One method at the early stage of development AI Output Watermark。 This adds an invisible pattern in the output, as applied to images protected by copyright. Theoretically, there are various ways to do this, but there is nothing more effective or efficient as practiced.

There are other reasons to explain DeepSeek's success, such as the deep and challenging technical work of the company.

Deepseek's technical progress includes a more powerful but inexpensive AI chip (also called a graphical processing unit or GPU).

Deepseek had no choice but to apply afterwards The United States is banned from companies From exporting the most powerful AI chips to China.

Western AI companies can buy these powerful units, but the ban on exports has forced Chinese companies to innovate to make the most of inexpensive alternatives.

The United States prohibits exports of the most powerful computer chips to China. Gal/shutter stock

Series of lawsuits

Openai's terms of service It explicitly states that no one can develop competing products using AI models. However, your own model is trained in large datasets that have been cut off from the web. Included in these datasets A considerable amount of copyright materialOpenai says that it has the right to use it Based on “fair use”:

AI model training using the published Internet materials is fairly used, as it is supported by a long -term acceptance precedent. We believe this principle is fair for creators, necessary for innovators, and important for our competitiveness.

This debate is tested in court. newspaper,, Musician,, author And other creatives filed a series of lawsuits against the Open due to copyright infringement.

Of course, this is completely different from the accusation that Openai is doing DeepSeek. Nevertheless, Openaii I don't have much sympathy For the claim that DeepSeek has illegally harvested the model output.

The war between words and litigation is how AI's rapid progress has surpassed the development of clear legal rules in the industry. And these recent events may reduce the power of AI incumbent, but they depend on the results of various legal disputes during the progress.

Shake a global conversation

DeepSeek indicates that it is possible to develop state -of -the -art models inexpensive and efficiently. It has not yet been seen whether they can compete with Openai in an equal stadium.

On the weekend, Openai tried to demonstrate its advantage Public release The most advanced consumer model, O3-Mini.

Openai claims that this model is very excellent in O1, which is a leading version of the market, and is “the most cost -effective model in the reasoning series.”

These developments have a era of increasing options for consumers, and have the diversity of AI models in the market. This is good news for users. Competition pressure makes the model cheaper.

And the advantages are further expanded.

The training and use of these models will be placed A Large strain About world energy consumption. As these models become more ugly, we all benefit from improving its efficiency.

The rise of DeepSeek certainly marks new areas for building models in a cheaper and efficient way. Perhaps it will shake a global conversation about how AI companies should collect and use training data.conversation

(author: Lee FrelmannA senior lecturer in natural language processing, Melbourne University, University of Melbourne and Shanan ConnieCyber ​​Security lecturer, University of Melbourne)

This article has been reissued conversation Under Creative Commons license. Please read Original article

(Except for the headlines, this story is not edited by the NDTV staff and is published from Shinjikate Feed.)


Facebook
Twitter
LinkedIn
Reddit
Telegram
WhatsApp