OpenAI trained AI critics to find bugs in code written by other AIs. The critics caught more bugs than humans and were preferred over 60% of the time, though they still struggle with hallucinations and complex distributed errors.
We have a very exciting paper coming out of the OpenAI lab with a crazy idea. So what is that crazy idea? Well, since we can get ChatGPT and similar AI chatbots to write us a piece of code, it does it very well. In the case of the new Claude 3.5, it can even write us up a video game in a matter of minutes, which is absolute insanity. I love it because this opens up a new world of creativity to fellow scholars who know less about coding or maybe can't code at all. However, these AI systems still make mistakes. So how do we catch these mistakes? Well, here is a crazy idea. Why not get another AI to look at the AI's answer and and critique it? Wow, that is indeed crazy. So does it work? Dear fellow scholars, this is 2 Minute Papers with Dr. Zsolnai. Previously we talked a great deal about papers where 20 to 30 ChatGPT agents could work together and write a video game for just $1. So the concept is very appealing. I mean, an AI that would critique another AI sounds like a science fiction movie. Loving it. But does it really work? Well, let's train the new AI first. How? Well, we need a lot of bugs to teach it on, and there is a super fun way of doing that, which is, you know, taking something that works and making it not work. Have an existing app and break it in interesting ways by inserting bugs in its code. And then describe what we just did, hand it off to the AI AI and let it learn how things typically break. Now, I know what you're thinking. Don't worry. They are also looking at bugs that naturally occur out there in the wild. So does it work? Let's have a look together. And the first results are. Wow, this is stunning. These AI critic systems find a lot more bugs than people. That is remarkable. What is also remarkable is that more than 60% of the time, the AI written critiques are favored over human written critiques. This will help make existing code bases work better and potentially help tempering them against attacks. The paper showcases examples of that that I really liked. And now hold on to your papers, fellow scholars. And look, I did not expect this at all. It even seems that compared to AI, only humans plus AI critics give less comprehensive results. Hmm. So does that mean that we just use the AI? Do humans just hold back the system? Well, not quite. Comprehensiveness is not everything. Look. Oh my God. Hallucinations are still there. This is when the AI just starts making things up, like erroneously saying there is a bug somewhere where there isn't in reality. Unfortunately, that still happens even with modern systems. Less so than before, but still there. And when there is a human helping out, there is much less of that happening. And just imagine what we will be capable of just two more papers down the line. Well, talking about papers, I have to say I am very happy to see more papers with nitty gritty details being published about these systems. More knowledge for all of us for free. Very very happy. I would like to send a big thank you for that. Now this is still not a perfect solution. There are limitations and since there is a paper we can also explore them a little easier. Fantastic. So hallucinations still happen. Human experts are required to look through the results carefully. Also, there are errors that are not isolated to a single piece of code. Instead, they could be the result of multiple issues arising from various places in combination. Not just one big mistake, but many smaller ones all across the board. It doesn't do too well with those. And if you look at the comment section, I often see you fellow scholars asking how it is possible that some comments arrive so early, often a day before the video itself. Well, our supporters can watch these videos earlier through Patreon and you can even get your name in the video description as a key supporter. Have a look in the description if you are interested. Also, what do you think? What would you use this for? Let me know in the comments below. If you're looking for an Nvidia GPU cloud service for large scale AI projects, check out One Click Clusters by Lambda, featuring up to 512 Nvidia H100 GPUs with InfiniBand networking. And hold on to your papers because with one click clusters you can access all these GPUs on demand with reservations starting at just two weeks. Try it out@lambdalabs.com paper or click the link in the description.
