Chinese IT Giant, AI 11:57 Invented a revolutionary method to teach

Chinese technological giant Alibaba has developed an innovative method to significantly reduce large language models (LLM) teaching costs. According to the company, the new technology, called Zerosearch, can reduce data collection costs to 88% compared to traditional approaches based on the use of external exploration services. About it reports South China Morning Post (SCMP).

Strong artificial intelligence models (AI) training requires tremendous resources, including high -performance equipment, large amounts of data and significant financial investments. One of the expensive stages is the collection and preparation of the training data set that usually requires communication systems.

The essence of the Zerosearch approach is the creation of new data models to teach the existing, well -educated LLM company. These models effectively mimic the information search and collection process, but make it internal without the need to pay for access to a third party API or use the external search infrastructure.

The economic effect is very important. Alibaba gives an example: 64,000 queries with a standard search API can cost about $ 586,70 for developers. At the same time, the use of the 14 billion parameter internal and model to complete the same task is only $ 70,80. Thus, at this stage, more than eight times decrease in costs.

According to developers, such a breakthrough in reducing expenses will make it more appropriate to develop and improve AI systems, especially for small budgets or small companies and research groups without their own comprehensive infrastructure. Alibaba himself already uses hybrid approaches by combining the capabilities of the Qwen family’s advanced models with traditional search technologies to increase the accuracy of the answers given to complex user requests and the level of relevance.

Previous Nvidia I am hereHow Trump’s restriction on chips supply to China is to pass.