Large Language Models and The End of Programming - CS50 Tech

Page 1 of 1 [ 10 posts ] 

The_Face_of_Boo
Veteran
Veteran

User avatar

Joined: 16 Jun 2010
Age: 41
Gender: Non-binary
Posts: 32,886
Location: Beirut, Lebanon.

13 Dec 2023, 4:42 am





Scary.



Fenn
Veteran
Veteran

User avatar

Joined: 1 Sep 2014
Gender: Male
Posts: 2,465
Location: Pennsylvania

18 Dec 2023, 8:55 pm

Have you ever tried to use ChatGPT to write code? Not quite the end of programming yet - but a bigger game changer than the IDE.


_________________
ADHD-I(diagnosed) ASD-HF(diagnosed)
RDOS scores - Aspie score 131/200 - neurotypical score 69/200 - very likely Aspie


naturalplastic
Veteran
Veteran

User avatar

Joined: 26 Aug 2010
Age: 69
Gender: Male
Posts: 34,124
Location: temperate zone

19 Dec 2023, 6:00 pm

An interesting and important pair of vids. We all assume that computer programmers are gonna put the rest of us out of work. But its counterintuitive to think that computer programmers are gonna put themselves out of work as well.

But not being a computer geek its hard to get through the first vid. I even put the young lady on three quarter speed to slow down her rapid speech. Lol!

Maybe Ill make another crack at it later.



funeralxempire
Veteran
Veteran

User avatar

Joined: 27 Oct 2014
Age: 39
Gender: Non-binary
Posts: 25,486
Location: Right over your left shoulder

19 Dec 2023, 6:47 pm

I'm just waiting for when I can tell a computer 'Make me a copy of NHL 95, but modular to make it easy to mod.' and it delivers.


_________________
Watching liberals try to solve societal problems without a systemic critique/class consciousness is like watching someone in the dark try to flip on the light switch, but they keep turning on the garbage disposal instead.
戦争ではなく戦争と戦う


Fenn
Veteran
Veteran

User avatar

Joined: 1 Sep 2014
Gender: Male
Posts: 2,465
Location: Pennsylvania

20 Dec 2023, 9:55 am

I have been working with chatGPT to try and translate some code to python and it is slow going - it is both brilliant and ignorant simultaneously. It will suggest things that just don’t work. It can also translate a page of code in seconds. You can give it code samples and it can explain what it does in a bulleted list. It spotted a syntax error in code that had me stumped for a few hours in seconds. It can be addictive but it isn’t like in the movies. It knows the math to find the intersection of two lines in 2space and can translate it it into code, but cannot understand instructions like “move the first square to the left of the second square”. It is often right but wrong often enough that you have to test everything. It is a very weird experience trying to program with chatGPT. It knows Python syntax and modules better than I do but I have to fix it’s code and finish it often. Sometimes it will insist I cannot do something then suggest the same thing the following day. There is a token limit (it explained to me) so longer questions require shorter answers- and as your code grows this happens more and more, so you have to discuss small chunks of code and keep track of the big picture yourself. Like pair programming with s brilliant moron. 5 years from now the state of the art should be interesting.


_________________
ADHD-I(diagnosed) ASD-HF(diagnosed)
RDOS scores - Aspie score 131/200 - neurotypical score 69/200 - very likely Aspie


belijojo
Veteran
Veteran

User avatar

Joined: 4 Dec 2023
Age: 20
Gender: Male
Posts: 903

20 Dec 2023, 10:35 am

Im learning data analysis, one of the most replaceable technologies, is happy to see the world change for the better, but not me.


_________________
Enjoy discussing politics, sometimes have to beat around the bush.
Often use wrong words, really sorry


Stormyweathers
Tufted Titmouse
Tufted Titmouse

Joined: 20 Dec 2023
Age: 53
Gender: Male
Posts: 45
Location: Forney, TX

20 Dec 2023, 1:11 pm

Don't run for the hills just yet.

Have you actually tried getting one of these models to generate even a simple piece of code? I have with a few, and my experience is that it is no better than asking a room full brand-new students to write you a piece of code, and then you have to correct it.

Well, except that the students will usually give you something that sort of works whereas the AI doesn't even attempt to run it. So far, the models are all giving me garbage that takes longer to correct than it would to write it myself.

Just as one fast example, I asked MS Copilot for a Graph API snip of code that would delete an email from an Exchange Online mailbox. It gave me a bizarre mix of two REST API based management PowerShell modules and completely made up some parameters for "deleteMode" offering deletion options that don't exist.

Now me, I happen to know this because I knew the answer to the question before I asked it, but if I did not already have this skill ... all I would have is trash.

My honest feeling on the matter is that given enough time, maybe five or ten years more, AI will be able to help a skilled developer produce content more rapidly, much more rapidly, but just as Tesla has learned the hard way, there had better be a skilled driver with their hands on the wheel who knows when to take over.

In the short term, I bet where AI is going to shine is in producing documentation, something most developers ... do not commit to with the priority it deserves. :D



The_Face_of_Boo
Veteran
Veteran

User avatar

Joined: 16 Jun 2010
Age: 41
Gender: Non-binary
Posts: 32,886
Location: Beirut, Lebanon.

22 Dec 2023, 2:21 pm

I tried copilote and chatgpt for web development (Angular, Vue, Express node.js).

And it’s doing pretty well.

The thing is, if you read about this topic you will find out that AI is progressing *exponentially*.
I don’t remember the numbers exactly, but for example if next year it becomes x2 better, then the year after would be x4 better and so on…



The_Face_of_Boo
Veteran
Veteran

User avatar

Joined: 16 Jun 2010
Age: 41
Gender: Non-binary
Posts: 32,886
Location: Beirut, Lebanon.

22 Dec 2023, 2:24 pm

If the videos didn’t scare you enough
https://www.ml-science.com/exponential- ... wo%20years



Fenn
Veteran
Veteran

User avatar

Joined: 1 Sep 2014
Gender: Male
Posts: 2,465
Location: Pennsylvania

15 Jan 2024, 2:52 pm

Stormyweathers wrote:
Don't run for the hills just yet.

Have you actually tried getting one of these models to generate even a simple piece of code? I have with a few, and my experience is that it is no better than asking a room full brand-new students to write you a piece of code, and then you have to correct it.

Well, except that the students will usually give you something that sort of works whereas the AI doesn't even attempt to run it. So far, the models are all giving me garbage that takes longer to correct than it would to write it myself.

Just as one fast example, I asked MS Copilot for a Graph API snip of code that would delete an email from an Exchange Online mailbox. It gave me a bizarre mix of two REST API based management PowerShell modules and completely made up some parameters for "deleteMode" offering deletion options that don't exist.

Now me, I happen to know this because I knew the answer to the question before I asked it, but if I did not already have this skill ... all I would have is trash.

My honest feeling on the matter is that given enough time, maybe five or ten years more, AI will be able to help a skilled developer produce content more rapidly, much more rapidly, but just as Tesla has learned the hard way, there had better be a skilled driver with their hands on the wheel who knows when to take over.

In the short term, I bet where AI is going to shine is in producing documentation, something most developers ... do not commit to with the priority it deserves. :D



With ChatGPT I have found that TDD helps. Also if I run the code ChatGPT gives me and get a compiler/interpreter error, then just cut and past the error back to ChatGPT it actually does a good job if fixing the error most of the time. The trick is to be persistent.
And don’t expect it to do all your thinking for you. It is just one more tool like an IDE or a Search Engine.


_________________
ADHD-I(diagnosed) ASD-HF(diagnosed)
RDOS scores - Aspie score 131/200 - neurotypical score 69/200 - very likely Aspie