China’s DeepSeek scores over OpenAI with its R1-Zero model
27 Jan 2025
Researchers at China’s open-source AI platform, DeepSeek, have succeeded in creating a free and truly open-source AI model that can challenge OpenAI's most advanced reasoning systems. And they did it by letting the AI teach itself through trial and error, according to a research paper.
According to the research paper, DeepSeek’s R1-Zero model has been trained with the least human interference. While most AI models involve constant human supervision and repeated fine-tuning, the DeepSeek-R1-Zero was trained through a self learning process, like in the case of humans, says the research paper.
DeepSeek-R1-Zero has been trained via large-scale reinforcement learning (RL), without the use of supervised fine-tuning (SFT), leaving the learning process to itself.
In the beginning, however, the machine knew nothing, and had to be prompted by voice commands as to what would be the desired result. Through successive attempts it learned what is the desired result.
The `Reinforcement Learning’ method uses a learn-by-doing model where success comes after a lengthy trial and error process. Under this method, good decisions get rewarded while bad ones get punished.
After a series of decisions and ranking of results, it learns to follow a path that was reinforced by those results.
According to the researchers, Reinforcement Learning added new capabilities to DeepSeek-R1-Zero, like enhanced reasoning and self-verification and reflective action without programming for it.
It also learned to allocate more processing time to complex problems and developed the ability to reevaluate problems and find own mistakes.
DeepSeek -R1-Zero scored 79.8 per cent success rate in AIME 2024 mathematics problem solving tests, surpassing the level achieved by OpenAI's o1 model.
On standardised coding tests, it achieved a rating of 2,029 Elo on Codeshares and outsmarted 96.3 per cent of human competitors.
In fact, as reports have it, China’s new tech companies such as DeepSeek and ByteDance are looking beyond the chatbot models and are aiming for newer and better applications for AI.
Baidu, which developed its own search engine and the Ernie chatbot, last week said its generative AI platform Wenku is capable of creating documents like powerpoints and that its user base has expanded to 40 million while revenue increased by 60 per cent year-on-year.
Chinese AI companies are also increasingly finding corporate consumers, Gartner estimates that more than 10 per cent of Chinese companies are using AI, against 8 per cent eight months ago.
