Report from the Washington Post
In Brief – DeepSeek’s AI assistant app and paper describing its technical infrastructure, which claims to have used far fewer and less advanced computer chips than the largest AI developers, undermined the core assumption of top AI investors and policymakers, that the top quality AI applications demand hugely expensive and advanced semiconductor investments. That caused major financial swings against AI chip giant Nvidia and sectors linked to data center construction and electrical generation. However, the DeepSeek models were “trained” using much larger models, including from OpenAI, and are an example of how smaller AI models built using larger models are proving capable and efficient.
Context – The top takeaway from the DeepSeek phenomenon is to remember the role of innovation and human creativity in the technology ecosystem. The prevailing techno-religious belief system in places like the SF Bay Area holds that data analytics will keep getting smarter forever. It is based on the idea that there are three supply factors that feed into data analytics. (1) Data. (2) Processing power (now called “compute”). And (3) Innovation. The foundation of the ever-upward arc of digital progress is the belief that more of each of the three leads to better analytics. And all three will grow forever. That was the view when “Big Data” was the buzzword. It’s the same in the “Neural Network” era. The current AI giants, presumed by many to own the future, dominate the data and processing power factors. But DeepSeek is a reminder, shocking to many, that the third, innovation, is historically the most unpredictable, least corporatized, and most powerful factor. DeepSeek has probably not revolutionized AI or LLMs. It’s been known for some time that small models trained on very large models can produce results close to the large ones on a small incremental investment, and DeepSeek did that better than ever. Maybe putting financial numbers on small versus large was key. Or the China factor. Or just the surprise. But AI has had examples where mercurial “innovation” allowed the very small to leapfrog over the very large. Don’t forget tiny Clearview AI jumping over the same giants.