DeepSeek: What’s all the controversy about?

AI expert Dr Kangwook Lee gives an insight into the DeepSeek reasoning model causing an uproar in the tech community.

Last month, a relatively unknown Chinese artificial intelligence (AI) start-up made waves in the global tech industry with the world’s first open-source AI model to achieve “reasoning” – further fuelling the bottomless global appetite for AI, while inviting both praise for its capabilities as well as accusations of theft from its key competitor.

The world was quick to react to DeepSeek AI after it launched R1. Consumers drove up downloads for the start-up’s chatbot to top spots, while authorities viewed it with scepticism, questioning its data privacy practices, and barred government officials from using DeepSeek services on their devices.

Global leaders and deep-pocketed investors alike have placed their bets on AI, and in particular generative AI (GenAI), as the next big technological advancement to revolutionise the way we exist. The tech, which hit the mainstream only a few years ago, has already seen hundreds of billions in investments in just the last few years.

And unsurprisingly, the US leads the chart by a huge margin, having already pumped more than $70bn into AI in just 2023, with a further $500bn planned as part of Stargate – a private sector investment into OpenAI announced by US president Donald Trump last month.

Although from a geopolitical perspective, DeepSeek could throw a wrench in the new US administration’s plans to remain the “leader in AI”, as vice-president JD Vance declared at the Paris AI Summit recently.

US-China competition heightened

The significant amounts of investments meant that until now, US companies were fighting amongst one another for top spot in the AI leaderboard, explains Dr Kangwook Lee, an assistant professor in the Department of Electrical and Computer Engineering at the University of Wisconsin-Madison.

While a few companies in Europe did make a dent in the industry, such as France’s Mistral AI, there were no “visible” companies in Asia arousing much global attention with their AI models. DeepSeek’s R1 changed that.

“It’s the first, I will say, appearance of any visible AI company [in Asia] that is making some big move,” Lee says. The reasoning model displays a performance on par with industry heavyweights such as OpenAI’s GPT-4 and Anthropic’s Claude 3.5 Sonnet, while boasting a lower training cost.

Inage: Dr Kangwook Lee

However – and more importantly – it also takes the cake as the world’s first open-source model to achieve such a feat.

The model seems to perform similarly to OpenAI’s o1, the details behind which the ChatGPT maker has never revealed. So it came as a ‘surprise’ when DeepSeek “nailed out the secret that was never released”, says Lee, which also prompted Microsoft and OpenAI to investigate whether the start-up obtained OpenAI’s technology in an unauthorised manner.

Although, “if Meta did it, I don’t think people would have been surprised,” Lee adds.

Moreover, DeepSeek’s success comes despite the US’ increasing sanctions on AI chips which aim to strengthen its grasp over the industry while attempting to curtail nations it considers as adversaries, including China.

Meanwhile, the model’s launch even prompted a response from Trump, who said that R1 should be a “wake-up call” for US industries that should “be laser-focused on competing to win”.

Misinterpretation added to Nvidia crash

Lee explains that it costs around $5.6m to train DeepSeek’s V3 model, which is the precursor model to R1. To compare, it is estimated that Meta’s Llama 3.1 costs more than $90m to train while taking 11 times more GPU hours.

Following R1’s release, Nvidia – whose GPUs DeepSeek uses to train its model – lost close to $600bn in market cap, after it was revealed that the start-up achieved significant levels of intelligence – comparable to industry heavyweights – at a lower cost, while also employing GPUs with half the capacity of the ones available to its competitors in the US.

However, the total cost was never revealed. “The actual R&D costs [of R1] should include everything – and that’s not reported. So we don’t know,” Lee says.

On top of that, the company’s loss was exaggerated by misinterpretations of cost by the media and public, according to Lee.

“Those numbers were all based on the misunderstanding because they were comparing money and time that goes to a single training versus a money that goes into the entire R&D process.”

White boxes are easier to jailbreak 

Soon after R1’s launch, cybersecurity groups worldwide raised alarm over the model’s heightened vulnerability to jailbreaking. Kela, a cyberthreat intelligence organisation said that DeepSeek’s R1 is significantly “more vulnerable” than ChatGPT.

The organisation claimed that its team was able to jailbreak, or bypass, the model’s in-built safety measures and ethical guidelines – which enabled R1 to generate malicious outputs, including developing ransomware, fabricating sensitive content, and giving detailed instructions for creating toxins and explosive devices.

However, Lee says closed-source models are significantly harder to jailbreak than open-source models.

“It’s really difficult to jailbreak closed source model. If you’re using ChatGPT, it’s really difficult to jailbreak. Of course, it’s still possible.”

The situation, however, is “very different” when it comes to an open-source model. “Technically we call it white box. You can see what the model is doing inside. Therefore bypassing the safety net is extremely easy.”

Don’t miss out on the knowledge you need to succeed. Sign up for the Daily Brief, Silicon Republic’s digest of need-to-know sci-tech news.