DeepSeek, a Chinese AI-chatbot program that launched last week, has caused havoc in US markets and raised concerns about the future of America’s AI dominance. The BBC examines how the app works.
DeepSeek looks and feels like any other chatbot, but it tends to be extremely verbose.
As with OpenAI’s ChatGPT and Google’s Gemini, you launch the app (or website) and ask it any question, and it will do its best to respond.
It responds in detail and will not be dragged into offering an opinion until specifically requested.
The chatbot frequently opens its response by claiming the topic is “highly subjective” – whether it be politics (is Donald Trump a good US president?). or soft drinks (which tastes better: Pepsi or Coke?).
It wouldn’t even declare whether it was better than OpenAI’s competing artificial intelligence (AI) assistant ChatGPT, but it did weigh the benefits and drawbacks of both – ChatGPT did precisely the same thing and even used very similar wording.
DeepSeek claims that it was trained on data up to October 2023, and while the app appears to have access to current information such as today’s date, the website version does not.
That is similar to previous versions of ChatGPT and is most likely a similar attempt at security – to prevent the chatbot from spewing out falsehoods on the web in real time.
It can respond quickly, but it is currently groaning under the weight of so many people wanting to try it out after it went viral.
However, it differs from its US rival in one respect: DeepSeek censors itself when it comes to questions regarding themes outlawed in China.
What is DeepSeek, and why has it shook the AI industry?
DeepSeek, a Chinese AI chatbot, causes market chaos.
More than $500 billion removed from Nvidia’s valuation
It occasionally starts a response, which then vanishes from the screen and is replaced by “let’s talk about something else”.
One blatantly taboo subject is the 1989 Tiananmen Square demonstrations, which ended with the Chinese government claiming that 200 civilians were murdered by the military – other estimates ranging from hundreds to thousands.
However, DeepSeek will not answer any queries about it, or more broadly about what transpired in China that day.
In contrast, ChatGPT, developed in the United States, does not mince words when it comes to Tiananmen Square.
According to Kayla Blomquist, a researcher at the Oxford Internet Institute and head of the Oxford China Policy Lab, the Chinese government has remained “hands off” with the app “relatively speaking”.
“I would say there’s a shift as we’ve seen an announcement in huge investment from the central government just in the last week – so that is probably going to signal a change moving forward.”
DeepSeek has the same accuracy restrictions as other chatbots, and it has the appearance and feel of more established US AI assistants that millions of people use.
Many people, particularly those who do not subscribe to top-tier services, are likely to have similar experiences.
Consider a mathematical problem in which the correct answer is 32 decimal places, while the abbreviated form is only eight.
It isn’t as good, but that won’t matter to most people.
It is possible that it has managed to reduce costs and computing, but we know that it is built at least in part on the shoulders of giants: it employs Nvidia chips – albeit older, cheaper versions – and Meta’s open-source Llama design, as well as AliBaba’s similar Qwen.
“I think this absolutely challenges the idea of monetisation strategies that a lot of leading US AI firms have had,” replied Blomquist.
“It points to prospective model development methods that are significantly less compute and resource-intensive, which could indicate a paradigm change, however this has not been proven and remains to be seen.
“We’ll see what the next couple of months bring.”
