Opening screen of DeepSeek’s chatbox (Screenshot by Damian Bertrand/Carolina News & Reporter)
Did you like the headline? It was made partly by an AI chatbox that has soared in popularity.
DeepSeek, a Chinese tech company, topped the charts late January in the app store with its Artificial Intelligence model, R1.
It differs from other AI models, such as OpenAI’s Chat GPT, by being open-source. It allows anyone to use and modify its code without permission and for any reason.
Users also have the opportunity to create and train their own version of a chatbox at a lower cost.
That is why it became so popular, said Richard Heimann, the state of South Carolina’s first director of AI, who wrote a book about understanding Artificial Intelligence.
“People get to look at (the work they just created) and see how it was done” by studying the code, Heimann said. “Whereas OpenAI, … people don’t really get to view behind the scenes. That’s all kind of hidden away from the user. So I think they’re excited about the openness and the transparency of DeepSeek.”
The model has another exciting twist. It operates at a fraction of the cost of leading AI models, according to DeepSeek.
The cost of training the DeepSeek model was about $6 million, DeepSeek self-reported. It cost more than $100 million to train ChatGPT-4, Open AI CEO Sam Altman told Wired.com.
But DeepSeek’s cost might be much more than a mere $6 million. The provided number doesn’t account for the money spent on research, algorithms, data or experiments.
Its investment is probably more than $500 million, according to a new report by SemiAnalysis, an independent research and analysis company specializing in AI industries.
“But everything said and done, DeepSeek seems to be more efficiently trained than other comparable models and performance,” Biplav Srivastava, a professor at the University of South Carolina’s AI Institute, said in an email.
Stock from Nvidia, an American company leading in AI computing, crashed after the release of the AI model, dropping nearly 17%.
Investors were worried that the low-cost Chinese model would threaten Nvidia’s hold on the AI market, prompting a mass exodus from their stock, according to an article from Reuters.
Heimann said he is skeptical about Nvidia’s stock crashing because of the Chinese company.
“I’m not convinced that it was because of DeepSeek,” Heimann said. “I think it was a coincidence or overlapped with the DeepSeek market.”
DeepSeek running at a lower cost than rival companies gives hope to other AI experts, said Srivastava, whose AI Institute promotes AI use on campus and at other universities.
“The AI world is very excited about it, and the reason for that is it has shown that with not too much investment, you can train new models and do more things,” Srivastava said.
Open-source models benefit from each other, replicating and making new AI models rapidly, Heimann said.
Companies take advantage of open-source models to create their own AI brand, he said.
Alibaba, the Chinese online marketplace company, revealed a new AI model days after the release of DeepSeek’s model.
“Models change so fast. … There’s novelty there,” Heimann said. “This is what’s great about open-source. It diffuses innovation pretty fast.”
DeepSeek’s model will be replaced soon enough, he said.
“I don’t think we’re going to look back in a couple of years and say, ‘That was a defining moment for DeepSeek,’” Heimann said. “I think it will be, ‘That was a defining moment for open-source technology.’”
The model is doing well in performance but is falling behind other AI systems in many areas, Srivastava said.
“It is not giving good information, or it is not giving unbiased information,” Srivastava said. “So there is room to improve.”
Newer and better AI platforms will supersede Deepseek, but its cultural impact will not be forgotten, he said.
“It has given confidence to academia and other institutions who were always thinking we need big computational investments to do anything meaningful in AI,” Srivastava said. “Now everyone is saying, ‘We can do it at a lower cost.’”
