He has reinvented it from scratch and that is the revolutionary

With R1Deepseek has achieved something that seemed impossible: to train an AI model with reasoning capabilities comparable to those of OpenAi … but without depending on huge data sets labeled.

Its approach based on reinforcement learning opens the door for many more groups developing advanced.

Why is it important. This advance changes the rules we assumed in the development of AI. Until now, creating models with reasoning capacity required huge amounts of data labeled and computational resources only within reach of giants such as OpenAi, goal or Google.

Deepseek has shown that there are A much more efficient alternative path.

Between bambalins. The Deepseek R1 training process is divided into two main phases:

1. First, R1-Zero Learn to reason exclusively through reinforcement learning, exploring solutions by test and error.
2. Then, R1 refine these capacities with a small amount of “cold start” data to improve aspects such as readability.

The model uses An expert mixing architecture (MOE) with 671,000 million total parameters, but only activates 37,000 million per consultation. This is what allows you to obtain a performance comparable to that of OPENAI O1 with a fraction of computational resources.

The contrast. While Openai invests hundreds of millions in labeled data and computing, Deepseek has achieved similar results with less than 6 million dollars (declared investment).

Its smaller distilled models, from 1.5 billion to 70,000 million parameters, have also achieved surprising performance.

The example. It is a silly example, but that is precisely why we wanted to verify its way of reasoning to the type of question you dislodge. We asked him the question “If Xataka was a Spanish football team, what would it be?”

His very long answer was autaeafirmo and then discarding the conclusion again and again.

1. First he just described Xataka and to make a superficial review of the main Spanish clubs.
2. Then it was raised if we would be Athletic, but understood that although its policy of “only Basque players” is unique, that is not comparable to Xataka’s innovation. He did something similar with Valencia, Barça and Rayo arguing different causes to rule them out.
3. He linked Real Madrid for us, Ahem, social mass leadership … but agreed that this is not linked to innovation.
4. He went through Eibar and Getafe, discarding both … but then returned to Eibar since “fell” in which they use analytics and technology. In the end he ruled it for being a small club.
5. He commented that Xataka’s leadership fits with an offensive style such as Barça de Guardiola or the Madrid’s Madrid counterattack …
6. … And finally he reached Villarreal and Girona.
7. After some reasoning, he stayed with Girona, arguing his strong culture of data, his innovative approach, his recent growth and his global vision (he is part of the City Football Group), in addition to adding something striking: “Both combine limited resources with intelligence Strategic: Girona maximizes its template with Scouting Advanced, while Xataka optimizes relevant and accessible content for a mass audience. “

His final conclusion was “the Girona FC embodies the essence of Xataka: modernity, technological adaptation and a fresh narrative that challenges the status quo“. 🚀⚽

Reading all your reasoning was spectacular.

Turning point. This development anticipates entry into a new era where Innovation in AI will not depend exclusively on access to large resourcesas it has been happening so far.

The learning techniques for reinforcement and distillation of models can level the pitch between large companies (or startups with investments of nine zeros) and much smaller equipment.

Deepen. This advance goes beyond simple incremental improvements. Deepseek has shown that it is possible to build models that reason autonomously without having to show them thousands and thousands of examples.

Reinforcement learning allows the model to discover effective reasoning strategies, similar to how humans learn to solve problems.

In Xataka | I have tried Deepseek on the web and in my Mac. Chatgpt, Claude and Gemini have a problem

Outstanding image | Xataka with Mockuuuups Studio