However, until now, I didn’t have concrete numbers on how long it takes to get information when prompted. Retrieving information is inherently challenging by design I saw an analysis by Zdenko Danny Vrancedic that highlighted the time cost of getting answers to the same question from ChatGPT, Google, and Wikidata: “Who is the mayor of New York City?” In the initial comparison, as Denny suggests, ChatGPT was running on top-of-the-line hardware available for purchase.
In the second case, Google used the most expensive singapore business fax list not available to the public. Meanwhile, in the third case, Wikidata was running on a single, ordinary server. This raises the question: How can someone running a single server deliver answers faster than Google and OpenAI? How is this possible? LLMs face this problem because they lack real-time information and lack a solid foundation in knowledge graphs, hindering their ability to quickly access such data.
More worrying is the potential for them to produce inaccurate responses. The key statement here for LLMs is that they will need to be retrained more frequently to provide up-to-date and correct information unless they use RAG and GraphRAG more specifically. On the other hand, Google meets this challenge because it relies on information retrieval processes related to page ranking.