Products > Programming

Is ChatGPT any good at writing code?

<< < (6/7) > >>

tszaboo:
I gave it a try yesterday because I wanted to automate something.
It made python code that converted some pdfs to png. The second code already did what I wanted to do.
Maybe it can still make useful code.

IanB:

--- Quote from: Psi on April 01, 2024, 10:33:36 pm ---LLMs will become a integrated tool in IDEs to make coding easier and quicker.
--- End quote ---

I think it's important to recognize that LLM/GPT tools are "dumb" in that they do not have any logical understanding of the output they are producing.

They basically work by making statistical predictions based on a very large database of source text.

If an LLM is going to work well at generating code, then it needs to be based on a very large database of relevant and high quality code samples. This leads to the question of where will such a database come from? The code corpus needs to be freely available without legal restrictions, and it needs to be of acceptable quality. There is a lot of bad code out there today, and if you have an LLM learn from bad code, it will generate bad code as output.

tggzzz:

--- Quote from: IanB on April 05, 2024, 03:01:17 pm ---
--- Quote from: Psi on April 01, 2024, 10:33:36 pm ---LLMs will become a integrated tool in IDEs to make coding easier and quicker.
--- End quote ---

I think it's important to recognize that LLM/GPT tools are "dumb" in that they do not have any logical understanding of the output they are producing.

They basically work by making statistical predictions based on a very large database of source text.

If an LLM is going to work well at generating code, then it needs to be based on a very large database of relevant and high quality code samples. This leads to the question of where will such a database come from? The code corpus needs to be freely available without legal restrictions, and it needs to be of acceptable quality. There is a lot of bad code out there today, and if you have an LLM learn from bad code, it will generate bad code as output.

--- End quote ---

Pretty much.

LLMs repeat "an average" of what they've ingested, with a few permutations.

That leads to one of the nightmares of the LLM fraternity: what happens when LLMs ingest other LLM output. If LLMs are as successful as their fraternity hope/claim, then most of the stuff they ingest will be LLM output.

radiolistener:

--- Quote from: cfbsoftware on April 01, 2024, 10:17:26 pm ---Is ChatGPT any good at writing code?

--- End quote ---

No. It can write some basic code (like examples from some book) compiled from his database, but if you're needs to do something specific, you won't get working code from it.

ChatGPT has some fundamental problem which prevent it to solve some issue in his code. So, if first sample code compiled from his database is not enough for you, you can't get something else. It just will propose the same code for any your request to change or fix something in his code. From this point of view, it seems to be programmed to use the broken record technique.

In short, it often provide you with a wrong or mistaken code (which at a glance looks good, but when you analyze it you will found issue) and when you're asking to fix it, it just repeat that code again and again with keeping issue in the code. And there is no way to learn it to solve it.

Smokey:

--- Quote from: radiolistener on April 24, 2024, 04:28:08 pm ---
--- Quote from: cfbsoftware on April 01, 2024, 10:17:26 pm ---Is ChatGPT any good at writing code?

--- End quote ---

No. It can write some basic code (like examples from some book) compiled from his database, but if you're needs to do something specific, you won't get working code from it.

ChatGPT has some fundamental problem which prevent it to solve some issue in his code. So, if first sample code compiled from his database is not enough for you, you can't get something else. It just will propose the same code for any your request to change or fix something in his code. From this point of view, it seems to be programmed to use the broken record technique.

In short, it often provide you with a wrong or mistaken code (which at a glance looks good, but when you analyze it you will found issue) and when you're asking to fix it, it just repeat that code again and again with keeping issue in the code. And there is no way to learn it to solve it.

--- End quote ---

While yes, it's better at basic tasks it's seen before, it can also synthesize those basic tasks together into something bigger.
I've had ChatGPT create basic python GUIs in tkinter based on a description of the layout and functionality I gave it as text in the prompt.  There was no way it had ever seen that exact layout and code before, but it got it working after a reasonable amount of bug fixes.

A buddy was telling me about a new LLM that writes code (sorry, forgot what he said it was exactly).  It actually runs the code in it's own sandbox and debugs it itself on the fly.  Pretty cool.

Navigation

[0] Message Index

[#] Next page

[*] Previous page

There was an error while thanking
Thanking...
Go to full version
Powered by SMFPacks Advanced Attachments Uploader Mod