His “little” update already lifts a stir

The Startup China has just given another great blow to Chatgpt, Gemini and Claude with a “small” update of Deepseek R1. Now several researchers speculate that at least part of the data used come from Google Gemini.

In this same situation the Startup China shortly after launching Deepseek R1 at the end of January and After staggering the foundations in which the chatgpt business model had been sustained.

The “little” Deepseek R1 update, which has been able to approach O3 of OpenAi And even overcome Gemini in some tasks, has again an impact on the Genai industry, being also Open Sourcefree and without limits of use.

Sam Paech, a Melbourne -based developer who is responsible for creating emotional intelligence evaluations for AI, published what he states that it is the evidence that Deepseek’s latest model He was trained with Gemini results.

The researcher maintains that R1-0528 prefers words and expressions similar to those used Gemini 2.5 Pro from Google. What is not an irrefutable proof, as it maintains Techcrunch.

There is another under pseudonym creator, known for its tool to evaluate freedom of expression called Speechmap, which maintains that The “thoughts” that Deepseek generates They have similar aspects to those of Gemini.

The Startup China has not revealed the source Of the data with which he has trained his updated model, so he leaves space for possible speculation and returns to the load with these accusations that have already occurred in January of this year.

At the moment, Google has not declared anything, but Openai in January told Financial Times that he found evidence that the Startup China would have used a technique called “distillation”which is based on extracting data from models from the elderly or more capable.

And here a fundamental aspect comes into play with which all major intelligence models have been trained: Web or Internet data extraction. The open web is precisely where IA companies get most of their training data.

That a model of AI or another shares phrases, senses and similar words is more than common. And for nothing it is helping that this open website is being “contaminated” by the one that creates clickbait and bots that seek to misinform with false data or news.

Also, AI companies themselves They are protecting their models Faced with distillation, and OpenAi already requires that organizations complete a process of identity verification to access the most advanced models.

Deepseek R1, with its update of a few days ago, has stood out because it already reason deeply, it reaches OpenAi O3 in performance, generates sessions of up to 60 minutes of deep thinking and in 24 hours reached 500 million tokens of use by users.