What part do you not understand? Big AIs like grock, chatGPT, and Claude use massive amounts of power to run massive server farms filled with AI processors to spin up the terrabytes of data that is then processed to answer simple questions or do basic internet searches. This could be done on your local video card using LLMs at a cost similar to running a video game on the same video card.
Here is groks answer:
A reasonable estimate is 70-85% of typical questions asked to big AIs (ChatGPT, Claude, Gemini, Grok, etc.) could be answered at a comparable or sufficient quality level by a capable local LLM (e.g., recent open-source models like Llama 3.1/4 variants, Mistral Large, Qwen, DeepSeek, or Gemma in the 70B+ parameter range) when equipped with internet access via tools like search APIs, browsing, or retrieval-augmented generation (RAG).
