Stay informed of free updates
Just register at Artificial intelligence Myft Digest – Delivered directly in your reception box.
OPENAI says he found evidence that the start-up of Chinese artificial intelligence Deepseek used the proprietary models of American society to form its own open-source competitor, while concerns are developing on a potential violation of intellectual property.
The Chatgpt manufacturer based in San Francisco told Financial Times that he had seen evidence of “distillation”, which he suspects of Deepseek.
The technique is used by developers to obtain better performance on smaller models using larger and more capable outings, which allows them to obtain similar results on specific tasks at a much lower cost.
Distillation is a common practice in the industry, but the concern was that Deepseek could do so to build its own rival model, which is a violation of the conditions of use of Openai.
“The problem is when you [take it out of the platform and] Do it to create your own model for your own ends, “said a person close to Openai.
Openai refused to comment more or provide details on evidence. Its conditions of use of users of the state of service cannot “copy” any of its services or “use the exit to develop models that compete with OpenAi”.
The exit by Deepseek of its R1 reasoning model surprised the markets, as well as investors and technological companies in Silicon Valley. Its models built on UNSTRING have reached high rankings and results comparable to the main American models.
Nvidia’s shares fell 17% on Monday, wiping $ 589 billion on its market value, on fears that large investments in its expensive AI equipment are not necessary. They recovered 9% on Tuesday, as well as other technological actions.
OPENAI and its partner Microsoft investigated the accounts that would be Deepseek last year who used the OPENAI, or API application programming interface, and blocked their suspected distillation access which violated the conditions for Use, another person with added direct knowledge. These surveys were first reported by Bloomberg.
Microsoft refused to comment and Openai did not immediately respond to a request for comments on this detail. Deepseek did not respond to a request for comments. China is closed for the Lunar New Year Holidays.
Earlier, the AI of President Donald Trump and Tsar Crypto David Sacks said “it is possible” that the IP flight has occurred.
“There is a technique in AI called distillation. . . When a model learns from another model [and] In a way, knowledge of the parent model, “Sacks on Tuesday at Fox News told.
“And there is substantial evidence that what Deepseek did here is that they have distilled the knowledge of the Openai models, and I don’t think Openai is very happy,” added Sacks, although he did not provide evidence.
Deepseek said that he used only 2,048 NVIDIA H800 graphics cards and spent $ 5.6 million to form its V3 model with 671 billion parameters, a fraction of what Optaai and Google spent to form models of comparable size . Some experts said that the model had generated answers that indicated that it had been trained in OpenAi GPT-4, which would violate its service conditions.
The initiates of the industry say that it is common for AI laboratories in China and the United States to use the results of companies such as Openai, who have invested in the hiring of people to teach their Models How to produce answers that seem more human. It is expensive and highly intensity of labor, and small players have often pushed themselves back from this work, say the initiates.
“It is a very common practice for start-ups and academics to use commercial LLM results aligned by humans, such as Chatgpt, to form another model,” said Ritwik Gupta, doctoral student of AA at the University of California in Berkeley.
“This means that you get this human feedback step for free. It is not surprising for me that Deepseek does the same. If they were, stop this practice precisely may be difficult, “he added.
The practice highlights the difficulty for companies wishing to protect their technical advantage. “We know [China]-The companies based – and others – are constantly trying to distill the models of the leader of US AI companies, “said Openai in his last declaration.
He added: “We are engaging in countermeasures to protect our IP, including a careful process for border capacities to include in published models, and believe. . . It is extremely important that we are working closely with the American government to best protect the most capable models of the efforts of adversaries and competitors to take American technology. »»
OPENAI fights against the allegations of his own violation of the copyright of newspapers and content creators, including prosecution of the New York Times and eminent authors, who accuse the Company of forming its models on their articles and books without authorization .
