The Chinese AI company Deepseek exploded into the news cycle During the weekend after it replaced Openai’s Chatgpt as the most downloaded app on the Apple App Store. Its commercial success followed the publication of several articles, in which DEEPSEEK announced that its latest R1 models – which cost significantly less for the company to make and for customers use – equals and in some cases exceeds, the best publicly available models from Openai.
So what did Deepseek do so deeply pocket openai? It’s hard to say with certainty because Openai was quite hidden about how it trained its GPT-O1 model, the previous leader of various reference tests. But there are some clear differences in the accesses of the companies and other areas, where Deepseek seems to be impressive progress.
Probably the biggest difference – and certainly the one who sent the shares of rag manufacturers like Nvidia disturbing Monday – is that Deepseek creates competitive models much more efficiently than its larger counterparts.
The latest models R1 and R1 zero “reasoning” are built on top of Deepseek’s V3 base model, which the company said trained for less than $ 6 million In computer costs using older NVIDIA hardware (which is legal for Chinese companies to buy, unlike the company’s state chips). By comparison, Sam Altman, General Director of Openai said GPT-4Not even the company’s best base model, cost more than $ 100 million to train.
Karl Freund, founder of the industrial analysis firm Cambrian Ai Research, said Manhattan-sized data centers.
“You can build a model quickly or you can do the hard work to build it effectively,” Freund said. “The impact on Western companies will be that they will be forced to do the hard work they were not ready to undertake.”
Deepseek did not invent most of the optimization techniques it used. Some, how to use data formats Use less memorywere proposed by its larger competitors. The image that comes out of Deepseek’s articles – even for technically ignorant readers – is a team that has entered every tool they could find to do training, requires less computer memory and drew their model architecture to be as effective on the Older hardware that was using.
Openai was the first developer to introduce so-called reasoning models that use a technique called chain thought, which mimics the test and error method of people to perform complex tasks, especially in mathematics and coding. The company did not say how it did it.
Deepseek, on the other hand, presented his process.
In the past, generative AI models have been improved by incorporating what is known as strengthening learning with human feedback (RLHF). People label the good and bad features of a multitude of AI -answers and the model is stimulated to emulate the good features, such as accuracy and consistency.
Deepseek’s great innovation in building its R1 models was to remove human feedback and draw its algorithm to recognize and correct its own mistakes. “DeepSEKR1 Zero demonstrates skills such as self-verification, reflection and generation
long [chains-of-thought]marking a significant milestone for the research community, “the researchers wrote.” Notably, it’s the
First Open Research to Validate those Reasoning Skills of [large language models] can be stimulated simply by [reinforcement learning]. ”
The results of the pure strengthening learning were not perfect. The results of the R1 model were sometimes difficult to read and exchanged between languages. So Deepseek has created a new training manifold, which incorporates a relatively small amount of labeled data to strip the model in the preferred direction combined with several rounds of pure reinforcement learning. The resulting model, R1, exceeded Openai’s GPT-O1 model about several mathematical and code problem sets designed for humans.
Bill Hannas and Huey-Meei Chang, Chinese technology and policy experts at the Georgetown Center for Security and Appearing Technology, said that China is closely monitoring the technological advances and practices of Western companies who have helped their companies find solutions to US policies as ragged Embargs that are designed to give an advantage to US companies.
Deepseek’s success, they said, is not a bad thing for the domestic industry, but it is a “awake call to US AI companies obsessed with Gargantuana (and expensive) solutions.” Do more with less “supports the approach to several Chinese state -Fifted laboratories.
