DeepSeek’s breakthrough is a pivotal moment for the democratization of AI
Subscribe to enjoy similar stories. The Center for AI Safety (CAIS) and Scale AI collaborated to create what they called ‘Humanity’s Last Exam,’ a test stuffed with 3,000 PhD-level questions from mathematics, humanities and life sciences. A new Turing Test for the age of AI, only three models got scores close to 10%: OpenAI o1, Google’s Gemini 2, and an unknown Chinese model called DeepSeek R1.
This Chinese model is far from unknown today. It has risen to the top spot on Apple’s app store and single-handedly caused the Nasdaq to slip by 1.5% on a single day, with Nvidia, the bellwether of AI stocks, crashing 17%. That DeepSeek’s latest AI model performs almost as well as OpenAI’s or Google’s creations is not why it is being hailed as the second coming of AI.
The reason is how little time and money it took. Built as a ‘side project’ by High Flyer, a Chinese hedge fund founded by Liang Wenfeng, it took just two months and $5.6 million, as opposed to the hundreds of millions spent by OpenAI and others. As claimed, it used last-generation H800 GPU chips, having been denied Nvidia’s latest by US sanctions; it needed just 2,000 of them, compared to the 100,000 or so that big US models required.
The shock and awe does not end there. As it is a small-sized model, it can even run on a high-end gaming computer on the edge, rather than on the machines of a gigantic electricity-guzzling data centre. Most importantly, the model is entirely open-sourced, with all its technical details published in a white paper, which makes it easy to work on for any company anywhere in the world.
Read more on livemint.com
