New AI breakthroughs show that bigger isn’t always better
The success of DeepSeek shows an interesting example where funding is the means to research and innovation and not directly the end goal.
Main content
An abridged version of this piece was first printed in Klassekampen.
This month, a new way to train AI models by DeepSeek broke the illusion that Silicon Valley has a monopoly on breakthroughs in generative AI.
The market for LLMs has been dominated by big tech, OpenAI’s ChatGPT being the first and most popular among them. These companies have largely been allowed to set the narrative, contending that AI can only be advanced with a very large quantity of resources, both in terms of hardware and engineering hours.
This whole expensive ecosystem has been seen as an unavoidable side effect of the AI revolution, and thus AI has been seen as something that can only be worked on by the tech giants in Silicon Valley. The damage of that narrative is that AI research and development is something only the wealthiest and largest countries can afford. For the rest, we are here to be consumers.
DeepSeek – a small lab in China – has now showed us that advancement can be done in a different way; using less resources and making your work open source.
The open-source approach decentralises software development and allows open collaboration. Open collaboration especially seemed like completely dismissed in the first wave of generative AI, as the tech giants rushed to put their own apps up for funding.
DeepSeek did not surprise the scientistic community in AI, but it did delight us. One large challenge is that we who do not work for the Big Tech or on generative AI are typically seen as supplementary to the “actual” actors in AI.
It’s not because we lack brainpower. Norway has 50 research environments willing to go into AI. There is so much to discover, learn and build.
China is still a very large market with a lot of resources compared to Norway. Generative AI may still be more resource demanding than what we are willing to accept. Privacy concerns in the free to use apps remain. The regulation necessity and issues around AI remain. But it has become a lot harder to say it is not worth trying to do AI better.