Hacker News Reader

Comment by teekert | Hacker News Reader

I have ChatGPT make up some code every now and then. It’s really nice and when not obscure usually directly useable. By you need to understand what it produces imo. I love that it also explains the code and I can follow code it generates and judge its quality and applicability. Isn’t that important?

View on HN

Last year the output was poor. The year before then, GPT essentially couldn't write code at all...

You're not wrong about its quality right now, but let's look at the slope as well.

on the other hand, gpt-3 was trained on a data set that contains all of the internet already. A big Limitation seems to be that it can only work with problems that it has already seen

I mean a huge amount of code I see is stuff like "get something from API, do this, and pass to API/SQL" so I'm assuming a lot of that could be automated.

OrmGPT!

> on the other hand, gpt-3 was trained on a data set that contains all of the internet already

This fact has made me cease to put new things on the internet entirely. I want no part of this, and while my "contribution" is infinitesimally small, it's more than I want to contribute.

The whole model is wrong. We don't need an AI that just spits out words that look like words its seen before. We need it to understand what its doing.

Midjourny needs to understand that its drawing a hand, that hands have 5 fingers.

Nobody will use Bing chat for anything important. Its like asking some some random guy on the train. He might know the answer, and if its not important then fine, but if it is important, say the success of your business, you're going to want talk to somebody who actually knows the answer.

The new models similar to Midjourney (Stable diffusion) and notably Deliberate v3 can perfectly draw anatomy now and you can even choose how many fingers you want.

Reply on news.ycombinator.com