Google makes almost all its money from search. These platforms are all there to reinforce its search monopoly. ChatGPT has obsoleted search. ChatGPT will do to Google search what the Internet did to public libraries - make them mostly irrelevant.
How has ChatGPT obsoleted search, when hallucination and the token limits are major problems?
It's (sort of) obviated search for certain kinds of queries engineers make, but not normies.
I say sort of, because IMO it's pretty bad at spitting out accurate (or even syntactically correct) code for any nontrivial problem. I have to give it lots of corrections, and often it will just invent new code that also is broken in some way.
Let's consider what Google did to the previous paradigm: libraries and books.
Books had editors and were expensive to publish, which imparted some automatic credibility. You might even have involved a librarian or other expert in your search. So a lot of the credibility problem was solved for you, up-front, once you got the information source.
Google changed the game. It gave you results instantly, from sources that it guessed looked reliable. But you still had to ascertain credibility yourself. And you might even look at two or three pages on the same topic, quickly.
Google has been mostly defeated now and often none of the links it suggests are any good. That trade-off seems to be done.
Here comes LLMs. Now it's transferring even more of the work of assessing credibility to the end user. But the benefit is that you can get very tailored answers to your exact query; it's basically writing a web page just for you in real time.
I think the applications that win in this new era will have to make that part of their business model. In science fiction, AIs were infallible oracles. In the real world it looks like they'll be tireless research assistants with an incredible breadth of book-learning to start from but little understanding of the real world. So you'll have a conversation as you both converge on the answer.
I've replaced almost all my usage of Google Search with ChatGPT. The only reason's I've had to use Google search is to look up current news, and do some fact checking. In my experience, GPT-4 rarely provides incorrect results. This includes things like asking for recipes, food recommendations, clarifying what food is safe to eat when pregnant, how to drain my dog tricks, translating documents, explaining terminology from finance, understanding different kinds of whiskey, etc.
This was true for me too, but I'm starting to find the data's cutoff date a problem, and it gets worse every day. I was reminded about it yesterday when it knew nothing about the new programming language Mojo or recent voice conversion algorithms.
The eventual winner will have a model that stays up-to-date.
It's mentioned in the article, but LoRA or RAG will enable this.
Phind is getting awfully close to this point already really. Integrating new knowledge isn't a bottleneck like we know from expert systems, I think it just hasn't been a priority for research and commercial reasons, till recently.
I asked ChatGPT to find Indian food in a tourist town. Googling verified that only one of its suggestions was a real place; the other four were hallucinations.
It's possible GPT-4 will be better; I haven't been able to test it because I remain on the waitlist.
