
In an article in the French newspaper La Recherche, Nathanaël Fijalkow describes “DeepSeek: when an AI discovers how to reason”.
At the beginning of 2025, DeepSeek (a Chinese start-up) revolutionized the world of AI by publishing a series of language models that were as efficient as the industry leaders, yet consumed less energy.
Beyond their efficiency, it's their innovative training method that intrigues: their engineers claim to have a model capable of reasoning... without ever having shown it any reasoning. How is this possible?
Nathanaël Fijalkow explains in an article dedicated to the subject by La Recherche, which you can consult here or in the attachment below.