Alibaba has unveiled “ZeroSearch,” a new AI training method that allows large language models (LLMs) to simulate searching the web without relying on commercial search engine APIs like Google. This approach uses reinforcement learning to teach AI’s to retrieve relevant information by generating simulated search results, leveraging the LLM’s pre-existing world knowledge. The method drastically reduces the cost and unpredictability of training AI assistants for information retrieval, providing developers with more control and making advanced AI development accessible to smaller companies. In trials, ZeroSearch matched or outperformed models trained with real search engines, and Alibaba has open-sourced the technology for broader adoption.

Venturebeat