DeepSeek’s low-cost AI chatbot is a win for the whole industry, Northeastern expert says
“The lower costs — and the demonstration that an upstart team can enter cheaply — will also attract other entrants,” says David Bau, a Northeastern professor in the Khoury College of Computer Sciences.
The Chinese startup DeepSeek has sent shockwaves throughout the AI world with the release of its less-resource-intensive AI chatbot, calling into question the amount of power and financial investment needed to develop the technology.
The company started making waves this week when its app surpassed ChatGPT as the most popular free app on the Apple App Store. The news comes just a week after it released its R1 reasoning model, which it said cost just $5.6 million of computing power to develop.
For comparison, OpenAI’s GPT-4 cost around $100 million to develop.
A Northeastern expert on artificial intelligence says this news is a positive development for the industry and will likely have a cascading effect. Notably, the company’s models are open source, meaning they are free to be downloaded and modified by other parties.
“We’ve been tracking DeepSeek since they started to release their work, and we agree that their recent R1 drop is a big deal,” says David Bau, a Northeastern University professor in the Khoury College of Computer Sciences and the principal investigator of the National Deep Inference Fabric, which is working to investigate the “black box” components of AI chatbots.
Wall Street reacted sharply to DeepSeek’s rise. Microsoft and Google saw their stock prices sink. Nvidia was hit particularly hard, losing about 17% of its market value, which is close to $593 billion.
But Bau says this news shouldn’t be seen as a failure on Nvidia’s part, noting that DeepSeek’s architecture is based on Nvidia’s chips, just fewer of them.
“Many are asking whether DeepSeek will make American Nvidia chips less relevant, but we observe that historically, as you make technology cheaper, then you increase the aggregate demand for it,” Bau says.
For Nvidia’s part, Jensen Huang, the company’s co-founder and chief executive officer, welcomed DeepSeek’s new model, calling it “an excellent AI advancement.”
OpenAI CEO Sam Altman originally had high praise for DeepSeek’s technology, but the company in partnership with Microsoft has since announced that it is investigating whether DeepSeek illegally used OpenAI’s data in the development of its models. David Sacks, President Donald’s Trump’s AI czar, has made claims that there is “substantial evidence” that points that it did.
Bau says DeepSeek’s work may also encourage others to join the fray.
“The lower costs — and the demonstration that an upstart team can enter cheaply — will also attract other entrants,” he says.
But he emphasized that there is still much to explore.
“Every time we grow AI capabilities, we grow the scientific mystery,” he says. “More investment in the science of AI internals and AI interpretability is needed.”
Nada Sanders, a Northeastern distinguished professor of supply chain management, says she does not find DeepSeek’s innovation surprising and undercuts the idea that huge amounts of money are required to build this technology.
Large language models function similarly to predictive models used by the supply chain industry, she explains. What she has found in her time working on those models is that the most elegant models are often the ones that were highly tailored and not necessarily built using high amounts of computing power.
Yet the hype around AI in Silicon Valley has created this idea that the more money and horsepower you throw at the technology, the better, she says.
“It really has become about stock prices, the amount of money that is being generated because we need bigger and we need better,” she says.
DeepSeek has brought back to the conversation the value of being nimble, creative and scrappy.
“I am hopeful that companies and other businesses are going to emerge out of this and understand that they can achieve success through smaller large language models and smaller neural nets,” she says.