> I've Been Vibe Coding For The Last Few Days, Maybe A Total Of Eight Hours, And I Got ChatGPT To Build Something Fairly Useful

My project is getting more impressive.

I need a good name. I think I will brainstorm this with ChatGPT on names.

wrote:

Generally speaking, what would you say GPT5 could improve your productivity?  A little bit, somewhat, considerably?

I estimate my $20 subscription provides $1,000 a month in value. Maybe more in some months.  I am tempted to pay for the $200 a month plan. I just wish they had a plan between $20 and $200. If they had a $100 plan I’d sign up immediately.

Sockpuppet wrote:

I estimate my $20 subscription provides $1,000 a month in value. Maybe more in some months.  I am tempted to pay for the $200 a month plan. I just wish they had a plan between $20 and $200. If they had a $100 plan I’d sign up immediately.

I tried using GPT but found it unreliable. You always have to ask it for references when it suggests something and then check those references. It makes shyt up, A LOT.

Samples_Boi wrote:

I tried using GPT but found it unreliable. You always have to ask it for references when it suggests something and then check those references. It makes shyt up, A LOT.

False.

That was like 2 years ago. You obviously haven't used shyt lately.

I have it do my EE work as well. You can give it a picture of a circuit and it will help you figure out what it is doing. It can't draw a circuit worth a shyt, though.

But it will look up all the chip specs, tell you where to put zener diodes and why, etc. It has made mistakes so you need to double check it but not if you give it the spec sheets to work with. If is going off-the-cuff without the actual documents, it can get things wrong.

You are probably using it wrong.

SamplesBoi wrote:

What do you expect the $200 plan would give you that the $20 plan does not?

You are paying for more compute so you will get better, faster answers.

Made more progress today.

The proxy now has a UI for configuration.

Security layer is now in place.

About 98% of the code has come from ChatGPT with me tweaking a few things here and there. One big problem I had today was I couldn’t get it to produce a readme.md file. It would generate about half of it in markdown and the rest in normal text like all of its other responses. I eventually told it to produce the file for download instead of directly in the UI. This worked.

I also gave it my .editorconfig and it mostly follows it. I also told it to stop putting in inline comments. I really fuqing hate code comments.

Sockpuppet wrote:

I’ve been doing it since early 2024. But never a full application from scratch. It just wasn’t good enough before.

It's been way, way more than good enough since just 4o (and lots of people were using a Claude model but I hate it). 4o is still my favorite model even though 5 is technically better and I use 5 daily.

5 is better at math, specifically.

You sound way, way behind.

AI generally wants to use a standard architecture which is too complex so it is best to come up with an overall architecture yourself. It is basically a team of really excellent programmers. You still need to manage them. You also need to understand how to feed it data and prompts. If you know how to do it, it won't make mistakes. If you are clueless like SampleBoi and ask a model that doesn't know shyt about some technical component without giving it the reference to the component you want to use, its like asking someone to figure something out where they can't look at the specs. You either let it look at the specs/libraries/etc. or you give it the specific ones and references it you want it to deal with. Then it is golden.

And the internet has bad info on it. Yesterday, I would have fried a board I had if I listened to GPT5 since it thought something was 3.3V when it was 5V and 5V would have fried the pin. Someone was feeding the component 3.3V specifically to get the output down to 3.3V but you run it at 5V for it to be efficient.

It is all about feeding it the right tokens. That takes a while to figure out perfect, just like using a new CAD program.

I can literally make anything I want to do in a short period of time and it does literally everything for me 1000X better than any programmer can do coding on their own.

It's like when Magnus plays chess against a computer. He says it makes him feel stupid.

All programmers are stupid now compared to synthetic AI programmers. And every day the gap widens. I don't need a team of programmers to do something any more. One person can do the work 10 or more and can do it 10X better and faster.

wrote:

Skill issue.  Lack of comments is the sign of a low performer.

I’ve seen no correlation between developer skill and comment volume.

In fact I am reminded of a low performer I interacted with recently. He wrote some shytty code but some great comments.

Sockpuppet wrote:

I’ve seen no correlation between developer skill and comment volume.

In fact I am reminded of a low performer I interacted with recently. He wrote some shytty code but some great comments.

Trusted advisors should understand that documented code, via comments, is important.

Samples_Boi wrote:

I used ChatGPT to identify a spider last night and it did very good.

However it warned me that if I upload personal information it will be used to train future LLMs until the end of time.
I'm pretty sure Spanky's bank accounts and health problems will be part of sample code data sets when kids get homework in college.

But his numbers are better than anyone here

wrote:

Who owns the IP for that code?
\

* crickets *