advertisement


ChatGPT, the latest conversational chatbot from OpenAI

Sue Pertwee-Tyr said:
I wonder whether university essay and test questions might be redesigned to exploit the 'tendency to bullshit' as one way to tell a student's work from an AI facsimile?
Click to expand...

we are working on it.

The easiest and quickest mitigation is timed assessments using laptops on a closed network - with invigilation.

That doesn't really work very well if you are operating a module with ~1800 students on it.
 
GPT-4 is a big jump forward, not just in accuracy but in reasoning ability. No-one designed better reasoning into it, it's still a transformer and hasn't been given fundamental new logic - it's bigger and is trained better but the same basic concepts as 3.5 - that's why GPT-4 is driving an uptick in excitement and worry right now, these things have emerged, and even Sam Altman isn't sure how.
 
darrenyeats said:
GPT-4 is a big jump forward, not just in accuracy but in reasoning ability. No-one designed better reasoning into it, it's still a transformer and hasn't been given fundamental new logic - it's bigger and is trained better but the same basic concepts as 3.5 - that's why GPT-4 is generating more excitement and worry right now, these things have emerged, and even Sam Altman isn't sure how.
Click to expand...

It was pretty useless when I used it. Although I am using it to help write code for an old game language.
 
I have spent half this (vacation) week having ChatGPT write songs for me about every silly thing that comes to mind. The 'romantic' one it wrote for my wife was actually pretty good, but when you give it silly parameters to write around ("Write a black metal song to cheer up my friend who is having surgery. Make it involve lemonade, Lurianic Kabbalah and Goetic demons. It should be something you can whistle and clap your hands to.")

It wrote this one for our grandchildren. The parameters it was given should be obvious:

--
Verse 1:
I woke up this morning, feeling kinda strange
My foot was on fire, it felt so deranged
I looked down and saw flames, shooting up in the sky
My foot is belching flames, twenty five miles high

Chorus:
My foot is belching flames, oh what a sight
It's burning so bright, it's lighting up the night
I don't know what's happening, but I'm feeling alive
My foot is belching flames, twenty five miles high

Verse 2:
I don't know what to do, my foot's out of control
I'm running down the street, trying to save my soul
People are staring, they can't believe their eyes
My foot is belching flames, twenty five miles high

Chorus:
My foot is belching flames, oh what a sight
It's burning so bright, it's lighting up the night
I don't know what's happening, but I'm feeling alive
My foot is belching flames, twenty five miles high

Bridge:
I know it sounds crazy, but it's just the truth
Something's got a hold of me, and I'm on the loose
I'm gonna ride this wave, see where it takes me
My foot is belching flames, and I'm feeling free

Chorus:
My foot is belching flames, oh what a sight
It's burning so bright, it's lighting up the night
I don't know what's happening, but I'm feeling alive
My foot is belching flames, twenty five miles high

Outro:
My foot is belching flames, and I'm riding high
I don't know where I'm going, but I'm reaching for the sky
I'm gonna keep on going, and never look back
My foot is belching flames, and that's a fact.
--
 
darrenyeats said:
You seemed to change mind half way through that post, x21!

It's a bit of a moving target, to be fair e.g. https://www.nextbigfuture.com/2023/04/gpt4-with-reflexion-has-a-superior-coding-score.html

How are you using it, is it a poorly documented language or just unfamiliar to you?
Click to expand...
I actually used the bing version to help me convert a project I implemented in a batch file I made a few years back. With a bit of work it was redone in python. I expect it to work with 3 different games now. I also dropped the requirement of installing the game server as a windows service.
 
....and what happens when AI starts building AI? It wouldn't take many iterations of that to take us somewhere very far away from where we are now.

I can't see how this can be anything but uncontrollable and unpredictable.

On a more positive note, however, when I asked Chat GPT if AI would take over the world it said not. Then again, it would say that, wouldn't it? :(
 
My experience of it when it works well it seems spooky but when it doesn’t you can see how it’s just using what is the most likely thing to come next, it can’t say “I don’t know but here is a guess” when it does that I start to worry and when it says “I didn’t know but I tried this and it worked” I will be a bit more worried but as to when that will happen might be a while away, might not.
 
russel said:
My experience of it when it works well it seems spooky but when it doesn’t you can see how it’s just using what is the most likely thing to come next, it can’t say “I don’t know but here is a guess” when it does that I start to worry and when it says “I didn’t know but I tried this and it worked” I will be a bit more worried but as to when that will happen might be a while away, might not.
Click to expand...

This is worth watching:
He talks about its improved problem solving abilities and capabilities.
 
tedmanzie said:
This is worth watching:
Click to expand...

I like his quote "Beware of the trillion dimensional space" which cuts to the chase I think. We are just no good at coping with the sheer scale of these models and that and our human tendency to rationalise it into our own hopelessly inadequate view of how our own brains work tells us very little about what it is and how it works. All the reductive stuff about it being a model about what comes next or a simple outcome of reinforcement learning seem to be missing the point to me.

I also like to flip these thoughts on their head. What if our own language model is also just a really good way of predicting what comes next? That we refine and improve over a lifetime of education and experience and everything we say about how this "works" or how worked something out is just a simplistic post-hoc justification unrelated to how it actually works? And how much of our own intelligence is just a really sophisticated language model built out of a trillion dimensional statistical space with no other mechanism in our brain (excluding support systems like memory, vision, etc.).
 
matthewr said:
I like his quote "Beware of the trillion dimensional space" which cuts to the chase I think. We are just no good at coping with the sheer scale of these models and that and our human tendency to rationalise it into our own hopelessly inadequate view of how our own brains work tells us very little about what it is and how it works. All the reductive stuff about it being a model about what comes next or a simple outcome of reinforcement learning seem to be missing the point to me.

I also like to flip these thoughts on their head. What if our own language model is also just a really good way of predicting what comes next? That we refine and improve over a lifetime of education and experience and everything we say about how this "works" or how worked something out is just a simplistic post-hoc justification unrelated to how it actually works? And how much of our own intelligence is just a really sophisticated language model built out of a trillion dimensional statistical space with no other mechanism in our brain (excluding support systems like memory, vision, etc.).
Click to expand...

what I have observed is the “hallucination problem” where it doesn’t know the answer so it basically guesses and they look reasonable guesses and there is nothing wrong with guessing the answer, however what it does not do is, to quote Feynman, to compute the consequences of that guess or carry out an experiment to see if the guess is correct.
 
russel said:
what I have observed is the “hallucination problem” where it doesn’t know the answer so it basically guesses and they look reasonable guesses and there is nothing wrong with guessing the answer, however what it does not do is, to quote Feynman, to compute the consequences of that guess or carry out an experiment to see if the guess is correct.
Click to expand...

Aye. Although there was some stuff I was reading recently about getting it to "do science" or at least propose experiments.
 
You must log in or register to reply here.


advertisement


Back
Top