DeepSeek: Moving beyond muscle car AI
The largely held belief that Nasa spent millions developing a space pen that could write in zero gravity, while cosmonauts just used a pencil, is a myth. But there are plenty of examples in recent history where big budgets and big tech are not always better.
Today’s AI datacentres are driving down the same route the US car makers took in the 1960s and 70s, with muscle cars and a growth path powered by bigger is better. No doubt president Trump’s “trump card” is the $500bn Stargate Project announced earlier in January, which will see huge investments ploughed into building US AI sovereignty.
But we only need to look back to the 1970s and how European car manufacturers reacted to an oil crisis by building highly efficient engines and arguably technically superior sports cars – to see what is likely to happen with AI datacentres in light of climate change.
The internet is awash with hypotheses regarding how China’s DeepSeek changes everything in the large language model (LLM) world. What is clear, and even Donald agrees, is that it is a wakeup call. The biggest isn’t necessarily the best. Sputnik 1 and Yuri Gargarin’s Earth orbit and Stuttgart’s 1970s Porsche 911 – when compared to the Corvette Stingray coming out of St Louis – shows us that alternative approaches can produce winners.
DeepSeek demonstrated that it is possible, with claimed development costs of just $6m, to build and train a large language model that can work as well as GPT-4o from OpenAI. It is also open source and costs significantly less – both in terms of hardware requirements and the cost of training and inference. In a research paper published on January 25, DeepSeek said: “Our goal is to explore the potential of LLMs to develop reasoning capabilities without any supervised data, focusing on their self-evolution through a pure reinforced learning process.”
The paper also looks at how larger models can be distilled into smaller models, resulting in better performance compared to the reasoning patterns discovered through reinforced learning on small models.
Wall Street reacted instantly to the publication of DeepSeek’s paper, wiping billions off the market value of major tech businesses including Apple, Google, Microsoft and Nvidia.
Its alternative approach to AI has got everyone excited. LLMs do not necessarily require huge amounts of expensive computational power. The fact that it is open source means anyone can download it and run it locally.
The approach the developers of DeepSeek have taken, should be something policymakers looking at sovereign AI capabilities, should definitely consider. They have lowered the barrier to entry. For an IT leader, they have also shown how advances in AI makes the tech more accessible to everyone.