Researchers at Alibaba introduced a new approach called ZeroSearch, which improves AI models’ search capabilities through simulations without the need to interact with actual search engines, according to a research paper published last week. Alibaba owns the South China Morning Post.
This method eliminates the high costs typically associated with routing queries through commercial search engines. It does so by enabling AI models – already trained on extensive knowledge bases – to generate quality content in response to queries.
By transforming a reference model into a search engine that can train other AI systems in answering queries, ZeroSearch cuts reliance on expensive external search infrastructure, making AI search training more accessible, especially for smaller teams with limited resources.
For example, sending queries to Google 64,000 times via an application programming interface cost about US$586.70, compared with US$70.80 when using a 14-billion-parameter AI model to generate responses for training, translating to an 88 per cent cost reduction.