There has to be more to it than that, right? I mean these tools can write working code in whatever language I need, using the libraries I specify, and it just spits out this code in seconds. The code is 90% of the way there.

LLMs can also read charts and correctly assess what's going on, can create stock trading strategies using recent data, can create recipes that work implying some level of understanding of how to cook, etc. It's kinda scary how much these things can do. Now that my job is training these models I see how far they've come in just coding, and they will 100% replace a LOT of developers.

[–] purpleworm@hexbear.net 8 points 22 hours ago (2 children)

can create recipes that work implying some level of understanding of how to cook

Being able to emulate patterns does not actually indicate some sort of higher level of understanding. You aren't going to get innovative new recipes, they are either just paraphrasing what they have read many people describe or they are cobbling together words.

[–] MayoPete@hexbear.net 2 points 22 hours ago* (last edited 22 hours ago) (1 children)

That may have been a bad example because for recipes it could just search the web and infer that vegetables go with olive oil for a stir fry. Where it's impressed me so far is in taking a piece of complex code and being able to refactor it, add features, write unit tests, and write up development plans. That text doesn't exist. It has to do some form of reasoning to interpret the code and come up with solutions for that particular problem.

[–] purpleworm@hexbear.net 3 points 19 hours ago

Syntax is syntax. I think from the standpoint of making a computer do something, it's really not that different from language processing. That, and just like when you ask it to make a new recipe or whatever else, it is liable to make up something nonsensical and fail to identify the problem unless you spell it out first.

load more comments (5 replies)