Quote:
Originally Posted by PeteMoss
The difference would be that google gives me multiple options for most of my searches so I can evaluate it. This (right now) is just giving me answers and for more advanced topics - I have no idea if its right or not. I just have to trust it.
Now a question - this is a language model - is it just sucking up information and then giving it back to us or is it learning these answers? That's where you get to the point of the AI ethics etc - when the thing starts teaching itself answers or learning to code itself better - essentially cutting the human out of the process.
|
As a search engine technology, what is impressive isn't the results it returns (not surprising they're weak because it isn't connected to the internet), but its ability to understand imperfect prompts. Basically, the situations where you can't figure out the right word to put into google. I was trying to remember the name of a particular plant, where I could kinda remember the sound of the word. I had been googling this for a couple minutes. So I told ChatGPT that the name sounded something like hyperia, it flowered in wooded areas in early spring; it correctly figured out that I was thinking of Hepatica. Concepts like 'sounds like' are far outside google's ability at the moment. That's a specific example, but basically it can remove that extra step of searching google for lists of things so you can identify the thing you really want to ask about.
My understanding is that it currently doesn't learn anything as a result of its question and answer sessions... Essentially, once the model is trained, it doesn't actually learn anything until the next training session. And every time you connect with it, it's starting fresh. This is the way AI diffusion models including art platforms like Dalle and Midjourney work as well.
However, I think developers are tweaking things like permissible behavior behind the scenes, which might operate like an external filter, and positive or negative interactions will end up influencing how the next model is trained.