Implicator
According to a new research paper from Tongyi Lab and Alibaba Group, researchers have developed ZEROSEARCH - a system that trains AI language models to search effectively without accessing real search engines.
The innovation cuts search costs dramatically. While Google searches cost about $587 for 64,000 queries during training, ZEROSEARCH costs just $71 using high-end GPUs.
The system works by having one AI model simulate a search engine while training another to use it. In tests, their 7B parameter model matched Google's performance, while their 14B parameter version worked even better than Google Search.
The researchers solved two key problems that plague current methods: the high cost of repeated API calls to search engines, and the unpredictable quality of search results that can make training unstable.
"The quality of documents returned by search engines is often unpredictable, introducing noise and instability into the training process," the researchers write. Their solution gives precise control over the quality of search results during training.
Why this matters:
- The breakthrough could make advanced AI search capabilities more accessible by eliminating expensive search API costs
- The system works better than Google Search while costing just 12% as much to train
Read on, my dear:
Robert Brown
Tech journalist. Lives in Marin County, north of San Francisco. Got his start writing for his high school newspaper. When not covering tech trends, he's swimming laps, gaming on PS4, or vibe coding through the night.
