Did the emerging Chinese technology firm DeepSeek replicate ChatGPT to create an artificial intelligence system that caused a stir on Wall Street recently? That’s the question being raised by OpenAI, the creators of ChatGPT, alongside top AI advisors to U.S. President Donald Trump. While neither has provided concrete evidence substantiating claims of intellectual property infringement, such remarks could lead to a closer examination of the situation surrounding DeepSeek’s rapid advancements in AI technology.
“There’s a significant amount of evidence suggesting that DeepSeek distilled knowledge from OpenAI’s models,” stated David Sacks, a prominent AI advisor, during an interview with Fox News on Tuesday. “I don’t believe OpenAI is pleased with this development.”
DeepSeek, along with the hedge fund from which it originated, High-Flyer, has yet to respond to inquiries sent via email, coinciding with the onset of China’s extended Lunar New Year holiday. OpenAI claimed in a recent statement that it has observed China-based firms consistently attempting to extract information from the models of leading U.S. AI companies, although they did not specifically name DeepSeek as a target.
OpenAI has prohibited the practice of ‘distillation,’ a learning technique used for building new AI models by repeatedly questioning an established, larger model. The organization, in collaboration with Microsoft, aims to track down accounts that may be attempting to engage in distillation, subsequently banning them from accessing their services. Microsoft has not provided any comment on this matter.
OpenAI further indicated that it would work closely with the U.S. government to safeguard its advanced models from competitors and adversaries attempting to utilize U.S. technology. Meanwhile, OpenAI faces lawsuits from media companies, authors, and other entities accusing it of copyright violations, which are currently navigating through various courts in the U.S. and elsewhere.
Tech investor Lutz Finger, also a lecturer at Cornell University, pointed out the irony of Big Tech calling out distillation techniques when their own practices might align with similar violations. “Although distillation breaches most terms of service, it’s hypocritical for Big Tech to condemn it,” he remarked. Finger, a former employee of Google and LinkedIn, noted that while it seems probable DeepSeek adopted distillation methods, proving it could be challenging because such techniques can be easily disguised.
DeepSeek has referenced its use of distillation in its publicly available research papers, acknowledging its dependence on openly accessible AI models developed by Meta (the parent company of Facebook) and Chinese tech giant Alibaba. However, there is no mention of OpenAI, which restricts access to its models, apart from illustrating how DeepSeek compares in performance.
Even prior to the commotion caused by DeepSeek, many users testing the company’s AI model observed that it frequently identified itself as ChatGPT or referenced OpenAI’s regulations. “When queried about its identity, it often responded, ‘I’m ChatGPT,’ which likely stems from the direct incorporation of data gathered from extensive interactions with ChatGPT into DeepSeek’s training dataset,” explained Gregory Allen, a former official at the U.S. Defense Department who now leads the Wadhwani AI Center at the Center for Strategic and International Studies.
Such mentions don’t necessarily constitute evidence of intellectual property theft, as chatbots frequently generate false information. However, DeepSeek presents its technology as “open-source” yet does not reveal the specifics of the dataset utilized for training its model. “The reasoning behind this decision is quite evident: they likely harvested data from ChatGPT for training purposes,” Allen added.
Analysts are grappling with various aspects of DeepSeek, scrutinizing the startup’s public research papers regarding its latest model, R1, and its earlier versions. One startling revelation for Wall Street was DeepSeek’s claim that the expense for training its primary v3 model amounted to a mere $5.6 million, an astonishingly low figure compared to the billions spent on developing ChatGPT and other leading chatbot technologies.
It’s notable that the $5.6 million figure pertains solely to the actual training of the chatbot, excluding any preliminary research and experimentation costs, as mentioned in their paper. Nevertheless, this figure and DeepSeek’s relatively low pricing for developers raise questions about the immense financial and energy investments fueling AI development in the United States.
Operating under certain limitations, including U.S. export bans on powerful AI chips, DeepSeek indicated its reliance on a comparatively lower-capacity AI chip from Nvidia, which is still permitted for sale in China. Interestingly, a 2022 social media post from High-Flyer disclosed its acquisition of a cluster of 10,000 more sophisticated Nvidia chips just months prior to the imposition of export restrictions to China.
