ChatGPT

[Polymathic]

GPT3-5 is so extraordinary it has made waves the world over, incl. causing Google to declare code red and Microsoft to scramble to integrate it into all their products. What exactly is so disappointing to you?

It has a failure rate of over 50% probably closer to 60% of the tasks I give it.

 

[animal531]

It has a failure rate of over 50% probably closer to 60% of the tasks I give it.

If that's true you're giving it the wrong tasks and/or commands.

 

[Bryn]

It has a failure rate of over 50% probably closer to 60% of the tasks I give it.

It’s far more likely that you just need to work on how you ask it questions.

 

[gotaway]

If that's true you're giving it the wrong tasks and/or commands.

ChatGPT did worse than class average in a college level astrophysics exam...
It is not a super intelligent being/thing/device/AI (yet)! Just good at regurgitating knowledge that has been captured electronically in a conversational manner.

 

[Polymathic]

If that's true you're giving it the wrong tasks and/or commands.

It’s far more likely that you just need to work on how you ask it questions.

There were occasions where I'd ask it to do a task one day and it will do it then a couple days later, I'd ask it to do the same task and it wouldn't be able to do the task correctly.

To me it seems like ChatGPT has gotten worse at doing tasks now compared to the earlier days.

 

[Bryn]

ChatGPT did worse than class average in a college level astrophysics exam...
It is not a super intelligent being/thing/device/AI (yet)! Just good at regurgitating knowledge that has been captured electronically in a conversational manner.

That guy could have let ChatGPT try again once or twice and I'd be surprised if it remained wrong on the questions that tripped it up.

Also, a chatbot is not typically capable of passing a college-level exam. Furthermore, for all we know, astrophysics isn't an area that GPT3-5 has a particularly strong model of. Compared to, say, programming where it seems to have been trained on all public GitHub repos.

As a software engineer, I have found ChatGPT to be a game changer. But not in having it generate my code for me. Most of the time it's like having the ultimate know-it-all professor available for in-depth conversations about information systems, computational efficiency, algorithms and data structures, general problem-solving etc.

Do I ever copy its output verbatim into my notes or code editor? Very, very rarely. But does it improve the quality of my work and the scope of my thinking? Every single day.

There were occasions where I'd ask it to do a task one day and it will do it then a couple days later, I'd ask it to do the same task and it wouldn't be able to do the task correctly.

To me it seems like ChatGPT has gotten worse at doing tasks now compared to the earlier days.

I have no idea what domain you're referring to, so it's hard to comment further. My experience has been that ChatGPT's quality has improved, not declined. But that is largely due to initial issues with longer answers, especially those involving code, and doubly so when referring to the context of the conversation's history.

 

[gotaway]

That guy could have let ChatGPT try again once or twice and I'd be surprised if it remained wrong on the questions that tripped it up.

Also, a chatbot is not typically capable of passing a college-level exam. Furthermore, for all we know, astrophysics isn't an area that GPT3-5 has a particularly strong model of. Compared to, say, programming where it seems to have been trained on all public GitHub repos.

As a software engineer, I have found ChatGPT to be a game changer. But not in having it generate my code for me. Most of the time it's like having the ultimate know-it-all professor available for in-depth conversations about information systems, computational efficiency, algorithms and data structures, general problem-solving etc.

Do I ever copy its output verbatim into my notes or code editor? Very, very rarely. But does it improve the quality of my work and the scope of my thinking? Every single day.


I have no idea what domain you're referring to, so it's hard to comment further. My experience has been that ChatGPT's quality has improved, not declined. But that is largely due to initial issues with longer answers, especially those involving code, and doubly so when referring to the context of the conversation's history.

It knew of all the models, etc but came to the wrong conclusions a few times.

Best to treat chatGPT as a single answer Google search with a bit of pot luck as to which result to pick.

Handy, sure. Hopefully better than a rubber duck to explain your code to as a debugging tool

 

[Bryn]

It knew of all the models, etc but came to the wrong conclusions a few times.

Best to treat chatGPT as a single answer Google search with a bit of pot luck as to which result to pick.

Handy, sure. Hopefully better than a rubber duck to explain your code to as a debugging tool

Well, regarding the code aspect specifically, the trick is to not ask ChatGPT for code that you couldn't write yourself. Otherwise you'll wind up in a horrible position sooner or later, stunt your own development and potentially cost yourself a job.

If you can easily QA and modify ChatGPT's output - and obviously identify whether it's even worth using at all - then it can make an all-day task take a few minutes.

It generated plugin code for me the other day that took barely 10 minutes to clean up and deploy. The code quality was acceptable and it saved me a ridiculous amount of time.

I think many devs make the mistake of providing unreasonably surface-level descriptions of their requirements and then blame ChatGPT for lousy code or inconsistent attempts. Which is to be expected if someone doesn't understand a problem well enough to itemise the solution.

 

[_kabal_]

Well, regarding the code aspect specifically, the trick is to not ask ChatGPT for code that you couldn't write yourself. Otherwise you'll wind up in a horrible position sooner or later, stunt your own development and potentially cost yourself a job.

If you can easily QA and modify ChatGPT's output - and obviously identify whether it's even worth using at all - then it can make an all-day task take a few minutes.

It generated plugin code for me the other day that took barely 10 minutes to clean up and deploy. The code quality was acceptable and it saved me a ridiculous amount of time.

I think many devs make the mistake of providing unreasonably surface-level descriptions of their requirements and then blame ChatGPT for lousy code or inconsistent attempts. Which is to be expected if someone doesn't understand a problem well enough to itemise the solution.

100% this.

This is why even something like CoPilot is great.

Of course I could write the code to e.g. calculate the average age of users created today, or, I could just quickly write the comment “//calculate the average age of users created today”, and then validate that the code produced actually does what it should and follows our conventions.

ChatGPT just takes this to another level.

IMO, AI will replace average/poor developers, but not in the way people think - skilled developers with business/customer competence will just use these tools to do more, in less time.

 

[AlphaJohn]

 

[FiestaST]

 

[telkomiskaI<]

saw a video of someone tricking CHATGTP into providing step by step how to target most government buildings and create a explosive device lol

 

[OnlyOneKenobi]

saw a video of someone tricking CHATGTP into providing step by step how to target most government buildings and create a explosive device lol

Yeah so unless some expert trained CHATGPT on how to do that, the fool who tries will probably blow themselves to kingdom come.

 

[Wut]

Yeah so unless some expert trained CHATGPT on how to do that, the fool who tries will probably blow themselves to kingdom come.

It was low-level stuff that's not actually "explosive" things such as thermite and napalm. Anarchists cookbook type stuff.

 

[Anthro]

saw a video of someone tricking CHATGTP into providing step by step how to target most government buildings and create a explosive device lol

Luckily they have their cellphone number and eMail address and can follow up on the complaint against authority..


Remember, Remember the 5th of November.. and all that

 

[OnlyOneKenobi]

Probably one of the more useful ChatGPT related plugins I've seen recently is this "Gimme Summary AI" Chrome extension which will summarise an article right from your browser toolbar. Not bad, especially considering that it's free.

Gimme Summary - Get summary using ChatGPT AI

Summarize articles on the web using ChatGPT AI

chrome.google.com

 

[OnlyOneKenobi]

So once again, I have to say I don't find ChatGPT to be particularly useful for coding or scripting as never seems to get things right on the first try. It's great for creative writing, generating ideas and descriptions and so on, but if you have to keep poking and prodding and re-writing prompts for it to get a simple script right then it's probably not worth the effort.

 

[Polymathic]

According to this ChatGPT has an IQ of 83, which explains a lot

 

[InvisibleJim]

So once again, I have to say I don't find ChatGPT to be particularly useful for coding or scripting as never seems to get things right on the first try. It's great for creative writing, generating ideas and descriptions and so on, but if you have to keep poking and prodding and re-writing prompts for it to get a simple script right then it's probably not worth the effort.

I've found it pretty good for quickly generating relatively simple scripts - most of the stuff I've asked it for has worked.

Have you tried the code completion beta that you get access to with the open ai account? You might find it a better workflow for developing more sophisticated code

 

[MagNorthDigital]

TL;DR - Stop reading this, drink a glass of water, and think of a way for "ai" to make your life easier, so you have more time to do whatever... sleep, eat, procreate, surf instagram... dunno, doesn't matter.


‐----------‐-----------------
Random thoughts / observations, while under the influence of fermented grapes...

Your brain on ChatGPT.

What if I told you that there is a force stronger than gravity, a force so powerful it has ruined cities, ended countless lives (yep, more than gravity) and literally stopped human advancement. That force is simply... laziness.

But not the type of laziness that makes you sigh because you don't want to... something... oh no, this form of laziness is actually an optimisation.

It is the brains' way of dealing with all the signal processing and the resultant output, its all those tiny shortcuts that keeps life flowing smoothly while you sit and watch Netflix.

Your brain has two modes of thinking, fast and slow processing. That cashier behind the till, you better hope she's using fast processing to scan your mince, mash and gravy microwave dinner (Cause otherwise she's not only slow at her job, she might actually have a learning disability)

And then we have slow processing, the type of thinking that gives us the ability to predict a possible future or learn from a past action.

Jesus, MagNorth, this is getting a bit long, can you say something i can disagree with in 20 words or less please?

Oh, okay, sure...

The country deserves nothing better than the ANC rubbish that has been running it into the ground.

Right where were we, brain processing speed and optimisation.

So, by using Google to "search" for an answer, we externalise a portion of fast processing, no more wondering or recall, copy paste save submit.

Now with chatgpt we are not only subbing out fast processing, but also learning and slow processing. And our brain could only imagine one thing better.

Offload everything, and give me chemical rewards!

Jho, MagNorth, that's some matrix kak, but I've seen the movie, it ends with Neo face down ass up... how did we go from Google searches to failing kung-fu?

Simple, and for the same reason no one can fix eskom, South Africa, most of the problems on this colourful rock in space.

These problems are the ones a quick Google search, or a chat with gpt cannot solve. It's not that there is no solution. Eskom worked at one point, South Africa did too. We seem to have gotten distracted by the new and shiny s__t, like a 5 year old with a Nintendo.

The only way to really advance, to really move the needle forward even one mm is to do the hard thing, to accept that problem solving, just like looking after your physical body, is a area to train, to learn, to grow.

Don't seek to offload your problems, your duties, your life. The journey is what matters, the moment of clarity it's own reward.

Right... how about that glass of clean water