Just when we thought we had our minds around ChatGPT, another large language model is causing havoc. This time it’s a Chinese one called DeepSeek, which was released on 20 January 2025 and on 27 January caused stockmarkets worldwide to crash. The reason was that an obscure Chinese AI startup showed that it could train AI models to perform almost as well as the most sophisticated Western ones like Google and OpenAI, with only a fraction of the computing power, and so a fraction of the cost and energy consumption.
The market reaction was brutal. Semiconductor giant, Nvidia used to be the most valuable listed company in the world, but on 27 January its share price fell by 17%, wiping out $600 billion of market value − the biggest one-day loss in the history of America’s stockmarket. If training models require less computing power, then fewer chips will be sold. Other players in the AI supply chain, from data centre hardware to energy-generating turbines to networking components, suffered a similar fate. In the bloodbath all the big names were hurt, ranging from Siemens Energy to OpenAI and Dell, to Microsoft, Alphabet and Amazon. Analysts called it China’s “Sputnik moment”. Meanwhile consumers flocked to DeepSeek’s chatbox, and within a few days it was the most downloaded free app ever in the USA.
So how did DeepSeek do it? According to reports, DeepSeek’s founder built up a store of 50 000 advanced Nvidia A100 chips, which had been banned from export to China since 2022 as a result of USA export controls. Experts believe that this collection allowed him to build a powerful AI model by pairing these chips with cheaper, less sophisticated ones and reprogramming them.
And if you thought that was good, USA researchers now claim to have gone orders of magnitude better, training their LLM, s1 for just $6. In comparison, DeepSeek took 2,7 million hours of computer time to train while s1 took seven hours. To do this they challenged the conventional wisdom that bigger datasets and more computing power lead to better performance. Instead, they focused on smaller amounts of high-quality data. By forcing the model to ‘think longer’ before answering, they significantly improved performance on complex maths and science tasks.
It looks like this could be a disruption to the current AI business model, which relies on expensive chips and huge computing power − and hence energy. The tech giants have poured huge amounts of money into developing AI, and they are not happy. In fact DeepSeek has been called a “technology thief” and the big players are in full ‘not invented here’ mode. For example, OpenAI suggested that DeepSeek tapped into its own data through distillation, and criticised it for violating its intellectual property. I find this a bit ironic when the top AI companies in the USA have often been accused of abusing intellectual property rights.
However, there are some unanswered questions about data and privacy management. DeepSeek collects large amounts of personal information from users, which is then stored in ‘secure servers’ in China. There are fears that peoples’ data could be harvested for intelligence purposes. The US Navy has banned its members from using DeepSeek’s apps, as has NASA. Australia has banned it from all government devices and systems, while Italy has blocked all its services citing insufficient transparency about personal information. Meanwhile users in the UK and USA have shown no such caution. DeepSeek has rocketed to the top of the app stores in both countries, with over three million downloads since its launch.
When you think about it, OpenAI, Google and Meta are already bending rules and collecting massive amounts of data with little transparency. I had a go at downloading the Deepseek app and the personal questions weren’t very different to what you have to go through to get Whatsapp, whether you like it or not.
So is all this more hype? Probably. Deepseek has been hailed as a game-changer, but at this stage it’s not quite clear what game it’s changing; but something that’s not only good but is also produced cheaply and released as a free open-source model has got to have a future. DeepSeek could result in greater use of AI, along with everything that comes with it. Underneath the jockeying for position among the major players the real winners could be us consumers.
However this comes with the tradeoff we have all experienced − the security of your personal information. Chatbots are powerful tools, but the price is your privacy. Who knows where DeepSeek will be − not next year, but next month?
| Tel: | +27 11 543 5800 |
| Email: | [email protected] |
| www: | www.technews.co.za |
| Articles: | More information and articles about Technews Publishing (SA Instrumentation & Control) |
© Technews Publishing (Pty) Ltd | All Rights Reserved
printer friendly version