Support NeoGAF

[CrankyJay™]

4.0 is miles better than 3.5. Really useful for debugging code.

It was pretty invaluable for a project I'm heading up converting out existing codebase from .NET461 to .NET6, as well as setting up some Azure pipeline scripts.

One trick I've learned, if you're working with anything that is versioned, to give GPT the version number so it's just not guessing. The more context you feed it, the better.

 

[murmulis]

I gave up arguing with that dumb copy-paste bot. For an AI, it sure can't learn, no matter how many times you prove it wrong (even using the data it provides).

That's because it has no memory. It doesn't learn from user inputs. Even if it agrees with you that the answer was wrong - it will give that wrong answer to other users.

The reason why they made it this way is that - if it learned from user inputs then sooner or later 4chan would turn it into GPT4Chan and ClosedAI definitely doesn't want that.

 

[GymWolf]

Not anymore, no.

This message was written by chatGPT

 

[Ar¢tos]

That's because it has no memory. It doesn't learn from user inputs. Even if it agrees with you that the answer was wrong - it will give that wrong answer to other users.

The reason why they made it this way is that - if it learned from user inputs then sooner or later 4chan would turn it into GPT4Chan and ClosedAI definitely doesn't want that.

But my problem is that, using the data itself provided not me, it kept saying that a composition from 1897 is last one a specific composer wrote, before and after listing a composition from 1898 from the same composer.

 

[FalsettoVibe]

Yep.

 

[LimanimaPT]

I've just used it today to generate some c# code to solve a problem I was having. It was absolutely amazed. I'm a programmer for 30+ years, so I don't lack experience but still I was amazed. The conversation I had was nothing short of incredible...

 

[Danknugz]

i was considering testing to see how well it would do building a simple web based front-back end application using .net or whatever

i started fooling around with unity a month or so ago and it was way more helpful than googling or trying to ask for help in discords

 

[Pidull]

I have found it invaluable. I use it for quick email drafts/proofreading, as well as to help me come up with questions I should ask based on certain situations.

It helps me think outside of the box my brain has built up over the years in some cases, or eliminate writers block.

 

[sendit]

I use GPT4 on a daily basis. Whatever fear mongering about AI I have to put up with for the rest of my life is 100% worth it knowing that I no longer have to waste my time on Stackoverflow.

Exactly this. I can now avoid having to filter out a bunch of search results, and avoid the casual snarky responses on stackoverflow.

 

[sendit]

Developers will be replaced with "Prompt Engineering"

 

[Pelao]

I have been using it on a daily basis to assist me with my work for about six or seven months now.
It saves me so much time it's ridiculous.

 

[casablanca]

I use it several times a week.

Mainly for correcting my English sentences at work or prepare big communication messages on slack etc
Also it's very helpful to get the right information on some products when the documentation is shitty.
Also learned a couple of python and .net functions that made my code faster and more efficient.

 

[murmulis]

But my problem is that, using the data itself provided not me, it kept saying that a composition from 1897 is last one a specific composer wrote, before and after listing a composition from 1898 from the same composer.

It's possible that the composition from 1897 is specifically mentioned as "the last one" in some texts that are in the training dataset. LLMs are no better than the data that they've been trained on. Garbage in - garbage out.

This is why the EU requirement to disclose AI training datasets is very good. So that it's possible to check where exactly its inaccuracies are coming from.

I encountered a similar thing with StableDiffusion. When I tried to generate a minigun - it generated something that resembled a cannon instead. Then I looked up "minigun" in the dataset that SD was trained on and found out that many images that were tagged as "minigun" didn't actually contain miniguns, but cannons and towed artillery.
It's surprising how good the results are despite the fact that the quality of captions in the training dataset is all over the place.

BTW - I also tried Dall-E and Midjourney - and they could not generate miniguns either. Seems like that they all use similar datasets under the hood.

 

[SJRB]

GPT gets profoundly annoying when it starts giving bad code and then goes in circles recommending the same erroneous code.

The highs are high, but man the lows are fucking low.

 

[reezoo]

I am a programmer and chatgpt is saving around 20/30% of my time for trival things. Along with ChatGPT I also use github copilot. They are very good tools for doing shallow / boilerplate things, save lots of time.