After the phenomenal success of DeepSeek AI and Manus AI, China is now working on developing a “self-learning, self-improving” artificial intelligence (AI) model that will not require human interference for updates and maintenance. The scary concept straight out of a sci-fi flick could well be a reality in near future given China’s bludgeoning stride in AI race.
Earlier this year, China had disrupted the AI industry by launching its massively popular AI chatbot DeepSeek, capturing global attention due to its cheap operation rates. Now, Team DeepSeek has teamed up with Tsinghua University to upgrade its AI model, by converting them into an all-knowing mean machine that requires no human for assistance.
As per reports, researchers from DeepSeek and experts from Tsinghua University, a top school in Beijing, have written a paper about a new way to use reinforcement learning, a method where AI learns by getting rewards for favorable answers, to make AI models work better and faster while using less computer power.
What is self-improving AI developed by China?
Self-improving AI is an artificial intelligence upgraded model that can get better at what it does on its own over a period of time. Instead of needing humans to constantly update or fix it, this kind of AI learns from its mistakes, then it figures out how it can improve and adjusts itself to perform tasks more accurately or efficiently.
It’s like how a person might practice and get better at something without always needing a teacher.
The company stated that it is working on a new way to make its AI models as humans want. The firm is implementing a technique called reinforcement learning, which has worked well for training AI to do specific tasks quickly, such as playing games or solving small problems. But it’s been tricky to use this method for bigger, more general tasks.
To tackle this, DeepSeek came up with something called “self-principled critique tuning.” The technique helps the AI to improve itself by figuring out what is right and what is wrong.
As per the reports, their research paper has shown a new approach that proved to be better than other methods in tests by giving better results while using less computer power. In simple terms, it’s like teaching the AI to learn smarter and faster with less effort.
DeepSeek’s new AI model, DeepSeek-GRM, is designed to learn and improve in a broad, flexible way by getting rewards for doing well. The company plans to share these models with everyone for free by making them open source so anyone can use, study, or tweak them without needing to pay or ask permission.
However, DeepSeek hasn’t said exactly when they’ll launch their next big, main AI model. They’re keeping the release date unclear for now.
Furthermore, the US tech giant OpenAI is also planning to launch its o3 and 04-mini AI models in a couple of weeks. San Francisco-based Chinese tech giant Alibaba is also trying to improve its AI model after the launch of its new AI model, after the launch of its AI model called Qwen 2.5-Max, in January, which claimed to be better than OpenAI’s GPT-4o and DeepSeek-V3.
Recently, on 4th April, Meta has also launched its latest open-weight large language model (LLM), called Llama 4.
Also Read: What is Llama 4 Open Weight AI Model launched by Meta?
