5 ms·

> After going in circles a few more times, I decided that was it. It got close. It seemed to understand the problem, but it could not actually properly solve it

> After going in circles a few more times, I decided that was it. It got close. It seemed to understand the problem, but it could not actually properly solve it. I had this same loop issue with Chat-GPT. I had something I wanted to do with asyncio in Python. That's not something I work with much so I thought I'd see if Chat-GPT could help me out. It was actually good at getting me up to speed on ansycio and which parts of the library to look at to solve my problem. It got pretty close, but it can't seem to solve edge cases at all. I got into this loop where I asked it to make a change and the code it output contained an error. I asked it to fix the error so it gave me a slightly modified version of the code prior to the change. So I asked it to make the change again and the code it spit out gave the same error again. I went through this loop a few times before I gave up. Overall, it's cool to see the progress, but from what I can tell GPT-4 suffers from all the same issues Chat-GPT did. I think we're probably missing some fundamental advance and just continuing to scale the models isn't going to get us where we want to go. My biggest concern with the current batch of LLMs is that we're in for Stackoverflow driven development on steroids. There's going to be a ton of code out there copy and pasted from LLMs with subtle or not so subtle bugs that we're going to have to spend a ton of time fixing.

kneebonian 4y ago
This here is my fear to. I have a buddy right now that is getting his degree in CS and he is using ChatGPT for a lot of his assignements. I worry that the next generation of developers are going to grow up just figuring out how to "program GPT" and when they have an error rather than investigating it (because they can't because they aren't actually familiar with code in the first place) they'll simply tell GPT about the error they are having and tell it to spit out more code to fix that error, slapping more mud on the ball. Eventually these systems are growing larger and larger at a faster and faster pace, and no one understands what they are actually doing, and they are so complex that no one human could ever actually understand what it is doing. Imagine if every codebase in the world was like the Oracle DB codebase. In this future a programmer stops becoming a professional that works to create and understand things, instead they become a priest of the "Machine Spirit" and soon we are all running around in red robes chanting prayers to the Omnissiah in an effort to appease the machine spirit.
- insomagent 4y ago
  I mean, we already kinda do that with Stack Overflow and cargo cult programming.

kneebonian 4y ago

This here is my fear to. I have a buddy right now that is getting his degree in CS and he is using ChatGPT for a lot of his assignements. I worry that the next generation of developers are going to grow up just figuring out how to "program GPT" and when they have an error rather than investigating it (because they can't because they aren't actually familiar with code in the first place) they'll simply tell GPT about the error they are having and tell it to spit out more code to fix that error, slapping more mud on the ball. Eventually these systems are growing larger and larger at a faster and faster pace, and no one understands what they are actually doing, and they are so complex that no one human could ever actually understand what it is doing. Imagine if every codebase in the world was like the Oracle DB codebase. In this future a programmer stops becoming a professional that works to create and understand things, instead they become a priest of the "Machine Spirit" and soon we are all running around in red robes chanting prayers to the Omnissiah in an effort to appease the machine spirit.

insomagent 4y ago
I mean, we already kinda do that with Stack Overflow and cargo cult programming.