Improving Software for Society

News | Blog Post : DEEPSEEK – WHAT’S ALL THE FUSS ABOUT?

RETURN TO THE BLOG POSTS PAGE

DEEPSEEK – WHAT’S ALL THE FUSS ABOUT?

DeepSeek – What’s All The Fuss About?

By Ian Barker BA, FIAP, MBCS

At the end of January millions of dollars were wiped off the value of US tech company stocks following the launch of a Chinese generative AI app called DeepSeek.

Silicon Valley venture capitalist Marc Andreessen described the latest DeepSeek model as ‘AI’s Sputnik moment’ in a post on X – referring to the 1950s crisis sparked by USSR’s launch of a satellite ahead of the US: X Post AI’s Sputnik Moment.

So what exactly is DeepSeek and why has it sent such tremors through the industry? DeepSeek is a generative AI tool in the mold of ChatGPT, Gemini and others. You can ask it natural language questions, get it to write code and do all the other things you’d expect of an AI tool.

The big difference is that DeepSeek can run on much less powerful hardware than rivals such as OpenAI’s o1. It’s designed to make use of lower-powered Nvidia chips designed specifically for the Chinese market. It’s cheaper too, DeepSeek says that its v3 model, released in December last year, has cost less than $6 million to train, less than a tenth of what Meta spent on its most recent system.

Add all of this together and you can see why DeepSeek is making waves. Particularly as it seems to have emerged as an unintended consequence of the US restricting the exports of high-end chips to China.

DeepSeek can be accessed on the web or downloaded as an app for iOS and Android. And the publicity has meant people are keen to try it out, DeepSeek has been riding high in the Apple AppStore and Google Play Store downloads charts in recent weeks, ahead of rivals like ChatGPT.

CHINA IN YOUR HANDS

There is, of course, a slightly more worrying issue and that is the level of Chinese state influence. Ask it about the status of Taiwan, for example, or the 1989 Tiananmen Square protests and you’ll get very different answers from those delivered by ChatGPT. In some instances it won’t answer at all, in others it adheres strictly to the government line.

DeepSeek’s privacy policy points out that user data, including conversations and generated responses, is stored on servers in China. As cybersecurity expert at NordVPN, Adrianus Warmenhoven points out, “Users need to be aware that any data shared with the platform could be subject to government access under China’s cybersecurity laws, which mandate that companies provide access to data upon request by authorities.”

This has sparked concern among western governments, with Italy being the first to ban it from app stores in the country. It’s likely others will be looking to follow suit. Taiwan hasn’t imposed a total ban but has stopped government departments from using DeepSeek.

DeepSeek is open source, but that doesn’t mean it works in a completely transparent way. Enterprise risk management expert Jake Williams, faculty member at IANS Research, and VP of R&D at Hunter Strategy, says, “It’s important to remember that open-source AI (e.g., DeepSeek’s R1) means something foundationally different than open-source code. With open-source code, we can audit the code and identify vulnerabilities. With open-source AI, we can do no such thing. There are also very real supply-chain concerns, R1 is fairly easy to jailbreak and it has far fewer guardrails than other commercial models. At the same time, if you’re looking at GenAI for a given use-case and cost is the prohibiting factor, this is worth exploring. The performance is really good for code completion, and it’s also great for code comprehension (based on my limited testing).”

TARGET FOR ATTACK

The popularity of DeepSeek has also seen it targeted by the cybercriminal community. A cyberattack on the 27th of January saw DeepSeek temporarily limit new registrations, saying it was the victim of a “large-scale malicious attack”.

Researchers from Cisco and the University of Pennsylvania have published a report: www.wired.com/deepseeks-ai-jailbreak, about the risks associated with the usage of DeepSeek. Using common ‘jailbreak’ prompts to trick the software into bypassing safeguards they were able to get it to respond with misinformation and content seen to be harmful or illegal.

“The results were alarming: DeepSeek R1 exhibited a 100 percent attack success rate, meaning it failed to block a single harmful prompt,” the report says. “This contrasts starkly with other leading models, which demonstrated at least partial resistance.”

There are other issues too, DeepSeek left a database open, exposing chat logs, API keys and other information to public view. This issue has now been fixed but it’s unclear how many people may have accessed the data.

WHERE NEXT?

So, where does all this leave us? DeepSeek works well – provided you stick to non-controversial topics – and the speed with which the developers were able to produce such a competitive tool is certainly impressive.

It also demonstrates that you can produce a sophisticated AI model with fewer resources. This is something that other companies are likely to be looking at closely, particularly given projected resource demands for data centres.

US President Donald Trump has described the rise of DeepSeek as being “a wake-up call” for the US tech industry. How the industry responds remains to be seen but underestimating DeepSeek would certainly be a mistake.