Research model catches bugs in AI-generated code, improving human oversight of AI.
See full article...
See full article...
OpenAI’s CriticGPT outperforms humans in catching AI-generated code bugs
- Thread starter JournalBot
- Start date
When robots replaced factory workers, it created robotics technician jobs. Then they created robot service technicians...
Upvote
43
(46
/
-3)
Their main product produces faulty output, so pay for their other product to fix it! Genius.
Upvote
56
(75
/
-19)
I am interested in the Force Sampling Beam Search. In the world of AI I wonder how many ML methods are still relevant for generic intelligence models?
Would be cool to see a 'top 10 ML algos powering your favorite AI'
Would be cool to see a 'top 10 ML algos powering your favorite AI'
Upvote
0
(10
/
-10)
You know, with all the money, time, energy, and probably copyright violations involved in trying to replace work done by humans, maybe you should just have the humans continue doing it in the first place?
Or, you could just introduce CriticCriticGPT in six months.
Or, you could just introduce CriticCriticGPT in six months.
Upvote
3
(30
/
-27)
Upvote
0
(4
/
-4)
As a general rule, isn't it usually preferable for the backup/correcting mechanism to be different? For example, aircraft with GPS still have INS, compasses and maps, etc. The idea being that if GPS is down, you can still use INS, and if INS is malfunctioning or it's a general computer failure, if you've got a compass, a map, and a good watch (and all pilots should always have all three), you can still navigate.
This is a basic safety engineering principle. Why the hell are they using the same technology as a backup/correction? Seriously, if they were designing a system deliberately, no one would do it this way. It reeks of desperation to prove that LLMs can do things that they are simply not capable of doing, by people who seem to believe that humans put together sentences in a similar fashion(!?!?!?!).
This is a basic safety engineering principle. Why the hell are they using the same technology as a backup/correction? Seriously, if they were designing a system deliberately, no one would do it this way. It reeks of desperation to prove that LLMs can do things that they are simply not capable of doing, by people who seem to believe that humans put together sentences in a similar fashion(!?!?!?!).
Upvote
14
(32
/
-18)
And now they're taking away Ars commenters' jobs. Leave me with something, OpenAI!
Upvote
23
(25
/
-2)
I hope so b/c I'd love to use it to check my code before I send a PR. But the article should have noted more clearly that this is solely research - and there is no product offering called "CriticGPT."
Upvote
13
(15
/
-2)
It's always two dumb bitches telling each other, "exaaaactlyyyyy"
Context: https://knowyourmeme.com/memes/its-always-2-dumb-bitches-telling-each-other-exactlyyyyy
Context: https://knowyourmeme.com/memes/its-always-2-dumb-bitches-telling-each-other-exactlyyyyy
Upvote
-15
(4
/
-19)
I use ChatGPT as a starting point. Its like having a drunk intern help you write emails and code. It can be helpful but you need to thoroughly review it before deploying it.
I'm not sure how helpful CriticGPT will be, maybe it'll be a drunk junior staff now? I still wouldn't trust its output without verifying it.
I'm not sure how helpful CriticGPT will be, maybe it'll be a drunk junior staff now? I still wouldn't trust its output without verifying it.
Upvote
0
(9
/
-9)
I like the sound of this.
I mean, I spend a lot of time criticizing and complaining. I need some assistance with this.
I mean, I spend a lot of time criticizing and complaining. I need some assistance with this.
Upvote
36
(36
/
0)
Upvote
-7
(8
/
-15)
It strikes me that some GPT errors will result in "fixable" errors within the GPT model itself, and these are likely the kind a CriticGPT could identify.
However, others are likely to stem from the fact that GPTs are designed to mimic human intelligence, but they aren't human. They don't have actual experience, judgement (which even humans with poor judgement have), comprehension, etc. And I can't conceive of any CriticGPT being able to address these, even if it can be taught to identify patterns that signal them.
I'm not sure how to better describe this. And maybe I'm wrong, and CriticGPT will do brilliantly... but I don't think so, and I find I actually hope it doesn't.
However, others are likely to stem from the fact that GPTs are designed to mimic human intelligence, but they aren't human. They don't have actual experience, judgement (which even humans with poor judgement have), comprehension, etc. And I can't conceive of any CriticGPT being able to address these, even if it can be taught to identify patterns that signal them.
I'm not sure how to better describe this. And maybe I'm wrong, and CriticGPT will do brilliantly... but I don't think so, and I find I actually hope it doesn't.
Upvote
3
(11
/
-8)
sam altman Act as assistant for now. Later on, maybe not so much. Then they'll just need a third system to replace the prompt-engineer/types-question-guy to eliminate the human aspect entirely below the C-suite level.
Upvote
-2
(2
/
-4)
Altman working hard to earn his Nobel prize for inventing a perpetual bullshit machine.
Upvote
6
(19
/
-13)
sigmasirrus
Ars Scholae Palatinae
Well this does sound ridiculous, but I have had some success with LLMs when I point out their errors and ask them to rewrite it. Maybe it’ll help a little bit? Nothing will fix the fundamental issues at hand because it’s not AI in the sci-fi sense of the word in the way that they want it to be.
Upvote
-1
(6
/
-7)
Yes, but think about how much more fun Ars comments will be when you ask the LLM to render them in the style of the Marx Brothers!
Upvote
4
(4
/
0)
ChatGPT 6: Kill All Humans!
CriticGPT 7: I see nothing wrong with your output.
(Terminators ensue)
CriticGPT 7: I see nothing wrong with your output.
(Terminators ensue)
Upvote
-2
(3
/
-5)
So GPT-4 is going to grade itself on how well it is doing?
What could possibly go wrong?
What could possibly go wrong?
Upvote
-9
(3
/
-12)
Yes, but as the old axiom goes -- if all you have is a hammer, everything starts looking like a nail. And all OpenAI has is an LLM.
Upvote
19
(24
/
-5)
But a big part of testing is actually running the code. I don't know how many bugs I couldn't see before running the code that were obvious after running it. But maybe it has learned weights that effectively execute code and observe the output?
Upvote
5
(5
/
0)
Yep, it's insane.
This tells me some things just from a high level.
1) Their tool does not do what they claim (chatgpt)
2) Their tool is not as accurate as they claim (chatgpt)
3) Given the first two, I'm fairly certain that it might be cheaper in the short and long run to use other machine learning systems besides Large Language Models to be more accurate and save more money.
Upvote
-4
(6
/
-10)
So the AI that has 100% confidence that the bullshit code filled with placeholders instead of functionality is somehow exactly what you wanted is going to be babysat by another AI that knows that's wrong?
I have doubts.
I have doubts.
Upvote
0
(7
/
-7)
Yeah, but, does it work? If annotators are choosing the generated critique over human, then I guess it does.
This is to aid the annotators, not replace them. Read the actual paper and/or article before posting. OpenAI's papers are among the most accessable out there.
Upvote
1
(6
/
-5)
It largely is different. First off, they are feeding the supervisor model with lots of data related to mistakes that GPT has made in the past. The primary LLM is not trained on this data. This results in several new layers to the model. Even more important, they are quite likely playing some very cool games with the system prompt that is fed to the fact checker. After a while this can get very expensive because, you wind up with lots of different versions of the LLM talking to each other. You eat lots of tokens and use up more context memory than you would like to think about.
Upvote
2
(6
/
-4)
The maybe read the paper. This is to aid human annotators. Besides which, do you find bugs when reading over your own work? Because I do. It is possible to use a model to improve itself. This is proven.
Upvote
6
(8
/
-2)
Language models can do that. OpenAI has an API to call arbitrary functions/tools which can include a shell, interpreter, compiler -- anything.
The limit is creativity.
Upvote
-2
(3
/
-5)
Aid the annotators .... by giving them another AI process to monitor for bad outputs? Shades of FSD.
Upvote
0
(2
/
-2)
I did read the paper.
I also happen to know OpenAI plays fast and loose with everything they say and do, and them slapping this together as yet another iteration and selling access is damned near guaranteed.
Upvote
11
(12
/
-1)
ChatGPT can run python scripts and such, but it can not compile code that I have ever seen, and I have numerous logs of it saying it can't run code*. Most visible when you ask it complex math problems. I think they changed settings so you only see "analyzing" now unless you change the setting in your account panel.
* I can't even count the number of times the model has insisted it can't run code, right after it said it can't read files that you can upload, nor could it explain why the upload button exist.
Upvote
4
(5
/
-1)