LLMs were not productified in a meaningful way before ChatGPT in 2022 (companies had sufficiently strong LLMs, but RLHF didn't exist to make them "PR-safe"). Then we basically just had to wait for LLM companies to copy Perplexity and add search engines everywhere (RAG already existed, but I guess it was not realistic to RAG the whole internet), and they became useful enough to replace StackOverflow.

Their tendency to bullshit is still an issue, but if one maintains a healthy skepticism and uses a bit of logic it can be managed. The problematic uses are where they are used without any real supervision.

Enabling human learning is a natural strength for LLMs and works fine since learning tends to be multifaceted and the information received tends to be put to a test as a part of the process.