Websites are great specialized tools for particular problems and LLMs are great general tools for natural language conversations. But websites require learning a specialized interface for every new website you visit and every new problem you have, and language models only return text, which is very limiting as a problem solving tool.

A few attempts have been made to bridge this gap but they've been lackluster. ChatGPT Plugins, such as for Wolfram Aalpha, requires developers to write a new application designed specifically for OpenAI. The application then has to be approved by OpenAI,and the user has to then install the plugin and then enable it for their chat session. The number of barriers to entry are so high that plugins are used by less than 1% of ChatGPT users. Even with plugins installed the best ChatGPT can do is generate an image or text as the output of the plugin, which doesn't create desired side effects like opening a bank acount, booking a flight, or filing taxes.

A second iteration was Browsing for ChatGPT, but it has the same problem of only returning a static image or text and is not an interface you can continue to interact with using natural language in addition to other problems like not storing preferences or sessions the way websites would do with cookies.