LARRUZ shows how ChatGPT improves when it reflects on its own mistakes across factual, coding, and decision-making tasks. The method pushes accuracy on HumaneVal from 80% to 91%, and includes live demos of Code Interpreter running Python simulations and data analysis.
Today we are going to take ChatGPT and it will be given a stern talking tool. Yes, we will try to help it to learn from its own mistakes and do better. Can it? Well, let's see together. But first and foremost, how do we do this? Well, these large language model assistants can typically be improved from the inside by reprogramming them. However, not in this paper. Here is a crazy idea. Let's try to improve it by reprogramming it from the outside instead with linguistic feedback. Yes, get this. It will get a stern talking to. Now I hear you asking, doctor, that is crazy. How do we do that? Well, here is an example. First we ask it a history question and we get an answer that is incorrect. Now we ask the AI to reflect on its mistakes and now it learns that the answer should have been a series of wars, not just one war. And now with this in mind, we can ask it again and bam, there we go. And this concept works for decision making, programming and reasoning style questions as well. That is fantastic. But how much of an improvement can we expect using this technique? Well, when running OpenAI's HumaneVal dataset, which is a test with more than 150 programming problems, these problems are handwritten to make sure the AI cannot have seen these problems before. So how does it do. Whoa, look at that. Suddenly it is now the best performing AI model. We go from 80% to 91% accuracy. That is a huge improvement. If this were GPT 4.5, that would have been a reasonable result and we didn't need to pop the hood. All we needed to do here is ask it nicely to essentially write a journal and learn from past mistakes. How cool is that? In his amazing science fiction novels, Isaac Asimov wrote about robo psychologists, scientists who were called when a robot was malfunctioning to find out what is going on in the brain of this little AI. And he wrote this about 73 years ago. And today, here we are. The first robo psychologists are appearing in our real world. How incredible is that? What a time to be alive. Now, in the meantime, ChatGPT and GPT4 also have some new really cool applications. For instance, I was stunned to find out that not every one of you fellow scholars know code interpreter now also known as advanced data Analysis. This can use Python programming to solve your tasks and even takes file uploads. I love seeing how diligently it writes code when I ask it something. This is a super simple example. Monte Carlo integration. My job here is essentially to direct the work of this little AI and make sure to check it thoroughly. Or wait a second, what if I just ask it to test itself? Hmm, it gives some nice proposals for me to do. Wait a second. How about you do all of these little AI and I am absolutely loving how it is going to town and checks the validity of its solutions from several different angles. We even get a beautiful little plot of this classic experiment. This tool is an absolute powerhouse and I feel that as it improves programming, computers will be available for everyone in the world. Loving it. This is just incredible. And don't forget, this can take files from us and it can also give files back to us transformed. Thus, we can even ask it for a little movie editing. Or you can upload your favorite Spotify playlist for it to analyze. It undergoes statistical analysis and if you like explanations better than visualizations and bar charts, you can even ask it to explain all this in plain English. How cool is that? Wow. So as you see, ChatGPT hasn't even been out for a year and it has improved leaps and bounds. You can have this incredible intelligence trained almost on the whole Internet right at your fingertips if you're looking for inexpensive cloud GPUs for AI. Lambda now offers the best prices in the world for GPU cloud compute. No commitments or negotiation required. Just sign up and launch an instance and hold on to your papers. Because with the Lambda GPU Cloud you can now get on demand H100 instances for just 1.99per hour. Yes, 1.99. And they are one of the first cloud providers to offer publicly available on demand H100 access. Did I mention they also offer persistent storage? So join researchers at organizations like Apple, MIT and Caltech in using Lambda Cloud instances, workstations or servers. Make sure to go to lambdalabs.compapers to sign up for one of their amazing GPU instances today. Thanks for watching and for your generous support and I'll see you next time.
