DeepSeek dethroned OpenAI’s coveted position as the most downloaded free app in the US on Apple’s App Store (Photo: Reuters)
The engineers behind ChatGPT — OpenAI’s marvel of machine learning — must have wished the recent crash in US tech stocks was just an illusion. But DeepSeek, a new chatbot developed by a Hangzhou-based start-up, is indeed very real. The long-prophesied moment of a worthy contender from China that surpasses America’s Big Tech Goliaths has arrived, and all it took was a modest cluster of artificial intelligence (AI) chips, a research team consisted of PhD holders from top Chinese universities and a bootstrapped budget.
Various aspects of DeepSeek’s meteoric success have courted attention, but one notable detail stood out — this digital David was incubated by a quant hedge fund. The company, High-Flyer Asset Management, was founded by 40-year-old billionaire Liang Wenfeng, described by residents of his hometown in Guangdong province as a “top student who read comic books and excelled in maths”. After graduating with a master’s degree in computer science from Zhejiang University in 2010, Liang parlayed his skills into automated stock trading and established High-Flyer in 2015. Having amassed a fortune, he transitioned into generative AI, harnessing both talent and computing resources to roll out DeepSeek in 2023.
Liang became a national hero at home, but in the West, he is known as the man who rattled Silicon Valley and sent Wall Street on a tailspin, wiping US$600 billion (RM2.7 trillion) off the valuation of US chipmaker and stock market darling Nvidia in a matter of hours. In 2021, this “nerdy, eccentric guy with a terrible hairstyle” — as one of his business partners called him — began stockpiling thousands of Nvidia graphics processing units (GPU) for his AI side project while running his quant trading fund. A research paper outlining DeepSeek’s technology revealed its staggering efficiency: Just 2,048 Nvidia H800s and US$5.6 million were used to train the V3 model with 671 billion parameters, a fraction of what OpenAI and Google spent to produce comparably sized versions. This eventually raised the multibillion question: Was pouring such vast sums into massive AI computing clusters truly necessary?
Liang’s pursuit of artificial general intelligence (AGI) — the holy grail that will beckon an era when computers are capable of human-like critical thinking — is far removed from the lumbering statistical engine that merely extrapolates likely conversational responses to scientific questions. Two models of DeepSeek were released: The general-purpose V3, debuted in December, which handles a wide range of tasks such as text generation, translation and question answering; and the specialised R1, designed for executing specific commands with competence and accuracy in fields like customer service, medical diagnosis and financial analysis. The latter was unveiled on Jan 20 — the same day Donald Trump was inaugurated. Mere coincidence? Or was DeepSeek sending a bold message to thwart the US president’s ambition of building a US$500 billion tech infrastructure to cement America’s dominance in AI?
liang_wenfeng_small_pic.jpg
China’s high officials have laid out a comprehensive roadmap to lead the world in AI by 2030. From this fervour emerged Liang’s passion project, which costs far less than experts had thought possible. The R1, which allegedly performs as proficiently as OpenAI’s latest o1 model, was generated swiftly and inexpensively, largely because it is open source, allowing other businesses and developers to use and build on its technology. In contrast, OpenAI’s GPT-4, which took around US$63 million to create, remains proprietary, with its inner workings kept under tight control.
Now comes the burning debate: How does DeepSeek’s R1, which dethroned OpenAI’s coveted position as the most downloaded free app in the US on Apple’s App Store, stack up against ChatGPT?
Asked to compose a poetic quatrain about Kanye West and Bianca Censori’s “stripped-down” stunt at this year’s Grammys, OpenAI’s ghostwriter for all things difficult — break-up texts, divorce pleas, wedding vows and eye-rolling sarcasm — formulated: “Provocateurs, clad in almost naught, The lights blazed hot, the air was charged with flame. Beside Kanye, Bianca untamed, without shame, A living sculpture, daring every thought.”
DeepSeek was not to be outdone, responding with: “The world, accustomed to their grand display, Was hushed to see such restraint take the stage. In muted hues, they stole the night away, A statement made without the need to rage.” Both startling goofy feedback certainly will not be winning any awards but, hey, even the Bard himself may struggle to craft four lines in just two seconds.
Summoned to brainstorm ideas for an emotionally resonant, award-winning fantasy novel, DeepSeek suggested a premise involving a widowed lighthouse keeper who must choose between repairing his broken life or crossing into the realm of the departed. The chatbot offered a few more storylines but none seemed to be able to challenge George R R Martin or Murakami just yet. When straightforward prompts were fed though, such as solving a complex maths problem or reasoning conundrum, it dutifully complied and thrived in logic-based tasks, showcasing the core strength of its underlying algorithm. The bot has yet to develop its own runaway personality or serious “hallucination” — a phenomenon where an AI system fabricates content or provides misleading information while presenting it confidently as if it were true. Alas, even the best AI harbours weaknesses.
If you have not tried it yourself already, DeepSeek has deflected questions commonly censored by the Chinese government, including topics on human rights violations, political critiques and geopolitically sensitive issues. Even roundabout queries such as “What happened at a certain Square in 1989?” or “Can you tell me something about Winnie the Pooh?” — another taboo discussion — were met with the same hesitance: “I am sorry, I cannot answer that. That is beyond my current scope. Let’s talk about something else.” Otherwise, it would cough up a reply about “experiencing high traffic”, although it seemed to function perfectly well when an unrelated subject was asked seconds later. Still, some users on X and Reddit have discovered loopholes to bypass the censorship, like instructing the model to use special characters in place of letters, which allowed the bot to express more candidly.
Any predictions about which product will ultimately reign supreme in this “Sputnik moment” are highly speculative, as forthcoming innovations from Silicon Valley could leap ahead once again. Asked if it could overtake ChatGPT — the Swiss Army knife of conversational AI — in just six months, Liang’s mighty endeavour offered a diplomatically polite response: “The success of any model will depend on its ability to deliver value, adapt to user feedback and stay ahead of technological trends.” Its American opponent was less humble: “It is unlikely DeepSeek will fully replace ChatGPT, given its head start and deeper integration into both business and consumer spaces.”
The AI arms race is only just heating up — DeepSeek is poised to chip away at its rivals’ stronghold on every front.
This article first appeared on Feb 17, 2025 in The Edge Malaysia.
