Wookyoung Kim · November 4, 2025
xAI's Grokipedia for better search
Grokipedia On October 27, 2025, xAI unveiled Grokipedia, an AI-powered online encyclopedia. Elon Musk says Grokipedia was created to remove Wikipedia's bias, but I think its real purpose lies elsewhere. For xAI, Grokipedia is a foothold for beating Google's Gemini.
Polymarket | Which company has best AI model end of 2025? Looking at the AI industry now, Google Gemini's standing is rising tremendously. On prediction-market platforms, the probability that Gemini will be the best AI model of 2025 is approaching 80%, and on LMArena, which evaluates LLM performance, Gemini ranks first in almost every area, including text, vision, image generation, and video generation. xAI's Grok, by contrast, is struggling in most areas.
But there is exactly one area where Grok ranks near the top. In search, where you would naturally expect Google to be number one, Grok holds first place.
Grok 4 Fast | xAI The Grok 4 Fast model, released in September 2025, focused (1) on reducing cost and (2) solely on search. I think this may be because xAI judged that proper AI search is more important than the raw performance of the LLM itself.
After GPT-3.5, what dramatically improved LLM performance was the emergence of Reasoning AI, often called Chain-of-Thought (CoT). By making AI think before producing a final answer, it became possible to sharply raise LLM performance. And when this Reasoning AI is combined with internet search, Deep Research models, where AI searches the web during its thinking process, are currently the highest-performing SOTA models.
In LLMs, search is therefore becoming enormously important alongside model performance. Even OpenAI uses Google Search to train ChatGPT and to help ChatGPT produce answers, so it is no exaggeration to say that search has become increasingly important in the AI race and is now a decisive battleground.
But doing search well is not easy. Even Google does not have the ability to see and index every piece of content. In fact, good search is so difficult that Google asks content creators to tell it whether their content is good through "metadata."
Then what if good content were prepared in advance so an LLM could find and use it when needed? Wouldn't that make better search possible than Google, which searches every time? I think that is what Grokipedia is.
Documenting web-search material in advance for LLMs, as Grokipedia does, has several advantages.
Because the same documents are used every time, KV-cache can be leveraged, allowing a larger context window at lower cost.
Because documents can be delivered through an API, they can be provided in a form better suited to LLMs.
Because the model does not have to find multiple links and compare them, costs can be reduced and resources can be focused on reasoning.
Grokipedia will be used so xAI can perform better search than Google and Grok can produce better answers.
Grokipedia will become the core infrastructure behind xAI's search advantage, and it may even bring down Google's Gemini.