While many have already commented on DeepSeek this week, I wanted to read their research papers to confirm my thinking and make sure I had a full understanding of what has happened this week to cause the disruption we have seen with NVIDIA market cap and more general US big tech.
The first thing I want to say, is that this has very little to do with US vs China.
Of course, the datasets used to train the deepseek models have had more local data from China (but have still used a lot of western datasets) – but that is not surprising if you know where deepseek is based.
Some commentary have raised concerns about data privacy (if you use their online versions) but others have correctly highlighted that a number of western foundational models / companies have been keen to capture user data to be used for future developments. Making this less of a US vs China and more the typical trade-off of privacy for functionality.
Some would highlight this is about open vs closed source models. While reducing the gap in performance between open and closed models is a factor for the market reaction it is not the key driver.
For those that have worked in the AI industry for decades, we have seen disruption many times before – and it’s the changing balance of power between hardware and software.
Over the last few years, with deep learning models moving towards billions of parameters, the need for better and better hardware was key to the success of generative AI and LLMs. However, a combination of better datasets, improved internal architectures and algorithmic efficiencies mean that improvements to the software both in the training and inference reduce computational load needed and thus the balance of power slightly moves from hardware to software.
This hardware software balance has been an ongoing theme for AI and in particular, neural networks, for a very long time. I remember back in the 90s some firms built dedicated hardware chips for neural network training and simulation, but they didn’t last long. Having said that, I believe there is still much need for the GPUs and hardware acceleration capabilities we have created over the last ten years or more. But advances in software will also deliver amazing results too.
The good news is that open-source models performing as well as the best commercial models, together with the lower demand for huge compute resources, opens the door to businesses that wish to benefit from and build their own AI models. It also means AI consultancies like mine have more options available to us to deliver the best solution to our clients that want to protect their IP and build something unique to them and their industry.
Deepseek have just opened the playing field, taking some of the power away from big tech and making the best performing foundational models more accessible to businesses of all shapes and sizes.
This isn’t US vs China, this is the next baby step towards AGI, as it’s the software and architectural design that will give us general intelligence, not just raw computing power (having said that, watch out for quantum to change the balance of power again).
And by the way … happy Chinese New Year to those that celebrate it.