Chinese researchers, backed by a Hangzhou-based hedge fund, recently released a new version of an artificial intelligence model called DeepSeek-R1, which has capabilities as advanced as the most advanced American products, but reportedly does so with fewer computing resources and at a much lower cost.
High Flyer, a hedge fund backing Dipsik, said the model nearly matches the performance of programs made by US firms such as OpenAI, Google and Meta, but does so using only about 2.000 older-generation computer chips. The chips are made by Nvidia, an industry leader based in the United States (US), at a cost of just about $XNUMX million to train the system.
By comparison, Meta's AI system, Lama, uses around 16.000 chips and reportedly costs Meta significantly more money to train.
Open-source model
The apparent progress in China's AI capabilities comes after years of US government efforts to limit China's access to advanced chips and the equipment used to make them. Over the past two years, under President Joseph Biden, the US has imposed multiple export control measures with the specific aim of slowing China's progress in AI development.
Dipsik seems to have paved its way to some kind of success through innovation, developing new and more efficient algorithms that allow the chips in the system to communicate with each other more efficiently, thereby improving system performance.
Part of what Dipsik's developers have done to improve its performance is visible to observers outside the company, because the model is open source, meaning the algorithms it uses to answer questions are public.
Market reaction
News of Dipsik's capabilities sparked a sell-off in U.S. technology stocks on Monday, as investors wondered whether plans by U.S. companies to invest hundreds of billions of dollars in artificial intelligence centers and other infrastructure would preserve their dominance in the field. When markets closed on Monday, the Nasdaq, which tracks most technology companies, lost 3,1% of its value, and Nvidia's shares fell nearly 17%.
However, not all AI experts believe that the market reaction to the release of Dipsik is justified, or that claims about the development of that model should be taken at face value.
Mel Morris, CEO of UK-based Corpora.ai, told VOA he believes the market reaction has been overblown, even though Dipsik is an impressive technology. He added that more information is needed to accurately assess the impact Dipsik will have on the AI market.
“There’s always an overreaction to things, and there is today. Let’s step back and analyze what we’re seeing here,” Morris says. “First, we don’t really understand exactly what the cost or time period involved in building this product is. We just don’t know. They claim it’s significantly cheaper and more efficient, but we have no evidence of that.”
Morris said that while Dipsik's performance is comparable to that of OpenAI products, "I haven't seen anything yet that would convince me that they've actually been able to break a quantum leap in the cost of running these kinds of models."
Doubts about origin
Lennart Heim, a data scientist at the RAND Corporation, told VOA that while it's clear that Dipsik benefits from innovative algorithms that boost its performance, he agrees that the general public actually knows relatively little about how the underlying technology was developed.
Heim says it's unclear whether the $6 million training cost cited by High Flyer actually covers the company's entire costs — including personnel, data training costs, and other factors — or whether it's just an estimate of what the final training would cost in terms of raw computing power. If the latter, Heim explains, the figure is comparable to the costs of better American models.
He also questioned the claim that Dipsik was developed with just 2.000 chips. In a blog post written over the weekend, he noted that the company is believed to have existing operations with tens of thousands of Nvidia chips that could have been used to do the work necessary to develop a model that can function on just 2.000 chips.
"This extensive computational approach was likely key to developing their efficiency techniques through testing and to providing their models to customers," he wrote.
He also pointed out that the company's decision to release the R1 version of its model last week - after the inauguration of the new US president - appeared to be political in nature. He said it was "clearly intended to shake public confidence in US leadership in the field of artificial intelligence, during a pivotal moment in American politics".
Dean W. Ball, a researcher at the Mercatus Center at George Mason University, was also cautious about saying that Dipsik R1 had somehow changed things in the field of artificial intelligence.
"I think Silicon Valley and Wall Street are exaggerating a bit," he told VOA. "But at the end of the day, R1 means that competition between the US and China is likely to remain fierce and we need to take that seriously."
Debate on export controls
The success of Dipsik is being used by some experts as evidence for their claims that export controls established under the Biden administration may not have had the desired effects.
"This suggests that the U.S. approach to artificial intelligence and export controls may not be as effective as advocates claim," Paul Triolo, a partner at DGA-Albright Stonebridge Group, told VOA.
"The availability of very good, but not cutting-edge, GPUs—which a company like Dipsik, for example, can optimize for specific training and workloads—suggests that the focus of export controls on the most advanced hardware and models may be misguided," says Triolo.
"However, it remains unclear how Dipsik will be able to keep up with global leaders such as OpenAI, Google, Anthropic, Mistral, Meta and others, who will continue to have access to the best hardware systems," Triolo adds.
Other experts, however, argue that export controls have not been in place long enough to show results.
Sam Bresnik, a researcher at Georgetown University's Center for Security and Emerging Technology, told VOA it would be "very premature" to call such measures a failure.
"The CEO of Dipsik said that access to high-level computing resources is the biggest limitation they face," says Bresnik.
"If (DipSik) had as much computing power at its fingertips as Google, Microsoft, OpenAI and others, there would be a significant increase in their performance. So I don't think it's true that some people are saying that DipSik is an indicator that export controls are not working."
Bresnik added that the strictest export controls were not introduced until 2023, meaning their effects may only be starting to be felt. He said the real test of their effectiveness will be whether American companies are able to continue to outperform China in the coming years.
Bonus video:
