He claims to have a training method for his AI 88% cheaper

The Chinese company High-Flyer broke into the market in the market of the artificial intelligence (AI) At the end of last January. Deepseek, his proposal, has made its way among its competitors thanks to its open nature and its benefitsbut the authentic debate has revolved for several weeks Around the cost of training your models. And it is that according to its creators they were barely spent 5.6 million dollars In this process. Three and a half months later this figure is still unbelievable, so it is reasonable to contemplate it with distrust.

Anyway, Deepseek has put on the table the possibility of facing the training of the new models of investing much less money than the one spent by US companies OpenAi, Google or Anthropic in the tuning of theirs. Now it is the Chinese technological giant Alibaba who seems to be following The same path that has already traveled Depseek. And it is that it claims to have developed a AI models training system that reduces the cost almost 90%, which presumably will have a positive impact on AI search capabilities.

Alibaba’s jewel is called Zerostoch

The strategy that Alibaba engineers have devised to reduce the cost of training their AI models is ingenious. And instead of interacting with real search engines during this process, Zerostoch, which is what is called its technology, improves search capabilities carrying out simulations. To understand why this approach is much cheaper, we need to keep in mind that the costs associated with the redirection of commercial search engines are usually high.

According to Alibaba Send 64,000 consultations to the Google search engine through an API has an approximate cost of $ 586.70, while generating the appropriate responses for training by means of an AI model of 14,000 million parameters has an approximate cost of 70.80 dollars, which represents an 88% cheaper. In practice, what Alibaba engineers seem to have achieved is to tune a model that behaves as a search engine that is capable of training other AI models so that they are able to solve consultations. This scenario has a very evident advantage: training no longer requires interaction with external search infrastructures.

Alibaba, as we all know, is a gigantic company, but from now on this strategy can be used by much smaller companies to train their own AI models without the need to face a great investment in this process. In addition, presumably this technology will improve both the search capabilities of AI models and the ability with which they carry out the reasoning processes. At the moment Alibaba has used Zerostoch to improve the capabilities of his Quark model, which for just three days has been capable, always according to its creators, to combine the internet search and advanced reasoning capabilities to resolve inference and deliver precise responses to complex consultations.

Image | Markus Spiske

More information | SCMP

In Xataka | Samsung is preparing to give TSMC a bars where it hurts most: the manufacture of the chips for ia