Article originally in the Financial Times.
The Chinese artificial intelligence laboratory has used innovative techniques and developed an artificial intelligence model that was trained with a limited human influence. As a result, there was a sudden moment of understanding that could change developers’ costs to create key applications based on this technology.
They modified competitors’ models
The published research report on the functioning of the “argument” model of R1 deepseek reveals how a group led by a billionaire from the Lianga Wenfeng Hoja Fund has achieved significant results by removing narrow sites in the development of artificial intelligence.
The study states how Deepseek has used a series of more efficient techniques for the development of the R1, which, like the Openai O1 Openai, generates accurate answers by “thinking” step by step longer than most large language models.
The breakthrough results of Deepseek come from the use of “learning in the form of remuneration” that reduces human participation in creating answers to challenges. The company also created smaller models with fewer parameters, ie variables used to train the artificial intelligence system and form its output, with the powerful abilities of thinking. This was achieved by adjusting large models trained by competitors such as Meta and Alibaba.
It’s just the tip of the iceberg
All of this has shook the whole Silicon Valley, because R1 achieves better results in some tasks than recently released models of Openai, Anthropic and Meta, but its development cost a fraction of funds.
On Tuesday, Openai said it found evidence that DeepSeek had used its technology and used outputs from its models to train its large language models with lower cost, which is common practice for academics and weaker financed beginners.
Despite this dispute, experts said that Deepseek showed a real innovation. Artificial intelligence researchers also appreciated its willingness to publish a detailed technical report describing how she created her model. The laboratory did so for the first time.
“I think it’s just the tip of the glacier when it comes to the type of innovation we can expect in these models,” said Neil Lawrence, a Deepmind machine learning professor at the University of Cambridge. “History shows that large companies have an innovation problem when they are getting bigger. For many of these large companies, we have seen the replacing of investments in computing technology with intellectual hard work. ”
Thumbs and a moment of awareness
Large language models are created in two phases. The first is called “pre -training”. Developers use huge data files that help to predict the next word in the sentence. The second phase is called “post-training”. Through it, developers teach the model to follow instructions, such as solving mathematical tasks or encoding.
One way to force chatbots to generate more useful answers is called “strengthening learning from human feedback” (Reinforcement Learning from Human Feedback – RLHF), a technique that is the first to introduce Openai to improve Chatgpt.
RLHF works by calling human annotators the answers of the artificial intelligence model to challenges and select the answers that are the best. This step is often laborious, expensive and time -consuming and often requires a small army of human data markers.
Deepseek is a major innovation of Deepseek’s final step by means of a technique called Learning (RL) learning, in which the artificial intelligence model is rewarded for doing the right things.
Deepseek first developed a powerful predictive text model called V3. Then she used RL to “reward”, for example, by giving him a thumbs up to generate the correct answer. The Chinese company found that if this process has been carried out enough, the model was able to spontaneously solve problems without human supervision.
This technique was also used by Google Deepmind to create an Alphago artificial intelligence system, which almost ten years ago defeated human players in the ancient table game and started the current boom of computing techniques of deep learning.
Deepseek said it found that the model had a moment of sudden awareness when it reassessed its answers and adjusted the processing time to address various questions.
“This moment serves as a strong reminder of the potential [RL] To unlock new levels of intelligence in artificial systems, which opens the way for autonomous and adaptive models in the future, ”wrote Deepseek creators in their research report.
Lewis Tunstall, a researcher of Hugging Face, who is engaged in research into artificial intelligence, said: “It seems that the secret ingredient to work is to have a very, very strong pre -trained model and then have a very, very good infrastructure to perform this process strengthening learning on a large scale. ”
Small and large models
While Openai and Google invest billions of dollars in building large language models, Deepseek has also built smaller models that can be launched on phones or web browsers. It has achieved this by “distillation” the intellectual capabilities of larger models.
Deepseek used its R1 model to create a relatively small file of 800,000 data points, and then, using these data created by artificial intelligence, modified models created by competitors such as Qwen company Alibaba and Llamma Meta.
Deepseek found that these demanded models were particularly strong in comparing arguments, in some cases overcoming flag models such as Claude from Anthropic. “Basically, it can solve most of the mathematical problems I have solved at the bachelor’s degree,” Tunstall said.
This development could be beneficial for applications developers who have a cheap and effective way of creating products. According to a Think-tank researcher Rand Lennart Heim, learning artificial intelligence models is much more efficient than the pre-training process that requires a high computing performance.
He added that this new paradigm could allow competitors to create competitive models with much less computing performance and spending much less money. However, without money on chips, “they simply can’t put it to the same extent,” Heim said.
Deepseek did not indicate how much it spent on creating the R1, but claims that the V3 model on which R1 is founded has been trained for only $ 5.6 million. This amount does not include additional costs, such as the likely procurement of thousands of graphics processors to train the model, or salaries, experiments, training and commitment, Heim said.
And although Deepseek was the first to use its specific techniques, it is expected to follow other laboratories of artificial intelligence, while Hugging Face is already working on the R1 replication.
American artificial intelligence companies have also worked to use the abilities of their large, state -of -the -art models in smaller, brisk models. Last year, Google launched the Gemma model, which is lightweight and is based on its Gemini.
“The recipe for intelligence is quite simple,” says co -founder and scientific director of Hugging Face Thomas Wolf, adding that Deepseek techniques have understood very well in this area. “Therefore, I expect many teams to repeat it.”
Additional news: Cristina Criddle in San Francisco and Madhumita Murgia in London.
© The Financial Times Limited 2025.All rights reserved. Ringier Slovakia Media is responsible for providing this translation. The Financial Times Limited is not responsible for the accuracy or quality of the translation.
