In seven days that changed the world, a small Chinese start-up went from quietly releasing a new AI chatbot to fending off cyberattacks, security leaks and accusations of theft and sanctions-busting.
In an event with significance on a par with the release of ChatGPT on November 30 2022, DeepSeek-R1 was formally released on 20 January, but its implications percolated only slowly into mass media. One reason for this could be the lack of technology literacy among the media: few understood the implications of a new model having the reasoning capability of OpenAI’s o1, the most advanced version of ChatGPT publicly available.
When news broke into the mass consciousness a week later that the model had been developed by 20-month-old company DeepSeek at a tiny fraction of the cost of market-leading chatbots, reactions ranged from delight to disbelief to desperation.
The reactions themselves tell us much about DeepSeek-R1. For one, it was not only a highly cost-effective model in terms of its development but will also be highly economical for its users, as well as developers who want to integrate it into their own tools.
It has been released as open-source software, meaning anyone can adapt it to their own purposes. Usage of its more advanced versions or with reduced restriction is also substantially cheaper than established AI platforms. This has delighted, among others, former Intel CEO Pat Gelsinger, who has declared he is adopting it for his church-oriented start-up as a result of its cost-effectiveness.
Congrats to DeepSeek on producing an o1-level reasoning model! Their research paper demonstrates that they’ve independently found some of the core ideas that we did on our way to o1.
— Mark Chen, chief research officer at OpenAI
Disbelief is the pervasive response of current AI leaders, who refuse to accept DeepSeek’s detailed reports of how it trained its models. The start-up said it had come up with several innovations, the ability to teach an AI model how to reason without any human supervision. It also revealed a four-step approach that progressively treated the model as a baby, child, adult and mature adult.
Researchers on the ground applauded them, with even Mark Chen, chief research officer at OpenAI, declaring: “Congrats to DeepSeek on producing an o1-level reasoning model! Their research paper demonstrates that they’ve independently found some of the core ideas that we did on our way to o1.”
However, investors in OpenAI were less sanguine. Neal Khosla, son of legendary venture capitalist Vinod Khosla, who has put $50m (about R933.6m) into OpenAI, acted almost as a proxy for the investor community when he railed on X that “DeepSeek is a CCP state psyop and economic warfare to make American AI unprofitable” [sic] and that “they are faking the cost was low to justify setting price low and hoping everyone switches to it damage AI competitiveness in the us [sic]”.
One could describe use of the term “psyop”, or psychological warfare, as the last resort of the American conspiracy theorist, especially since no evidence was provided. But then OpenAI itself fuelled that narrative, telling the media it had evidence that DeepSeek had stolen OpenAI’s work in order to train its models cheaply.
Aside from none of this evidence being revealed, it was a rather rich accusation, considering that ChatGPT was trained through scraping — the term for “stealing” content off websites — practically all text available on the internet.
The true anger, rather than the disingenuous responses that were commonplace this week, is probably happening behind the scenes. The venture capital firms, wealthy investors and even sovereign funds who have pumped many billions into AI start-ups cannot help but notice the stark disparity: DeepSeek-R1 cost $5.6m (R104.6m) to build, compared to OpenAI spending $6bn (R112.03bn) to develop ChatGPT. This would be a frightening thought to custodians of that money, who may suddenly imagine that they've been pouring cash down a black hole.
Little wonder that they are looking for scapegoats, and there is no better scapegoat for Americans right now than the Chinese Communist Party (CCP).
This was hardly the most absurd reaction to the developing news. That was reserved for ordinary share investors, who adopted their habitual herd mentality in the face of breaking news and sold off any stock with a hint of AI.
Nvidia, the darling of investors for the past year since it makes the computer chips best-suited for AI applications, fell a huge 17%. By the end of Wednesday, it had clawed back a quarter of this drop. And it probably still represented a buying opportunity for those who understood that cheaper AI meant more demand for AI and more demand for chips.
The biggest surprise this week was that DeepSeek came as such a surprise in the first place — or that Americans were so shocked that a Chinese company had caught up so “fast”. For some years now, research has shown that China was catching up to the US in high-quality academic research papers on AI.
One of the criteria for measuring the quality of such papers is their share of citations in other papers. On this score, China overtook the US in 2020, according to Stanford University’s AI Research Index — reported as long ago as its 2021 edition.
However, as it is easy to inflate numbers by pumping out papers, one needs to measure the quality of papers by those most cited.
As far back as 2019, analysis by Semantic Scholar of more than 2-million academic AI papers showed that China was poised to overtake the US in the most-cited 20% of papers in 2020, and in the 1% of most-cited papers by 2025.
It seems 2025 arrived too fast for the American AI industry.
• Goldstuck is author of 'The Hitchhiker's Guide to AI'
Would you like to comment on this article?
Sign up (it's quick and free) or sign in now.
Please read our Comment Policy before commenting.