Károly Zsolnai-Fehér breaks down OpenAI's new web-search feature for ChatGPT, which cites sources and synthesizes answers to complex questions. OpenAI released a dataset of high-correctness Q&A pairs to test models. Flagship models like O1 and GPT-4o show better consistency and confidence calibration than smaller ones, though hallucinations remain a real risk.
I hear more and more people say that they would love to use a chatgpt like model not only as an assistant, but also for searching the Internet. And OpenAI just gave a lot of scholars exactly what they wished for. We will talk about what it can do, why it is not always sunshine and happiness. And we don't stop there. No sir, they also did something in a new research paper that could make it work better than we think. I'll get to it in a moment. But now most of the world uses Google for searching the Internet and it's great for simple questions when it works. But check this out. A new icon should appear soon for ChatGPT users and there you can ask for something more complex. Planning a trip, no problem. But it gets better. And yes, like a good fellow scholar, it cites its sources to make sure you indeed get correct information and not some hallucination that the AI just made up. But this is not just regular search. You can also ask follow up questions like where to grab dinner according to your preferences or what the weather is going to be like, but on specific days, not just a huge table with a bunch of stuff you don't need. And all this with more sources. You can also search and see whether your Nvidia stocks have made you rich already or not and get a little additional context on what just happened. Also, I have to say I am a bit less impressed about the collection of news. You can get that anywhere. But you don't have to ask simple questions like that. Ask something more interesting instead, like how to fix up your backyard and get not just copy paste, but a nicely synthesized answer with images tailored to you and a guide made on the fly. Important part, it also leads you straight to the source in a very visible way. So I hope this will also reward the creators of this kind of information as well. But there is a problem here and that problem is called hallucination. This is where the AI makes things up. If you ask something, you are going to get an answer whether that answer is correct or not. Well, that depends. We are seeing less and less of those hallucinations now, but they still exist. However, if you use a search engine that hallucinates, that could almost be as bad as just getting a bunch of ads instead of an answer. Okay, so what does it have to do with AI research papers? A great deal. Have a look at this. Okay, I hear you asking what a data set? So a bunch of questions and answers. What does that have to do with hallucinations? How does this help us? Well, yes, this is A dataset published by OpenAI, but it is no ordinary dataset. These are challenging questions and answers with high correctness on a diverse set of topics. I mean, just look at that. Many of these are a touch more complex than what you can just get through search. You see, when evaluating an AI model with answers to these questions, not even that is without challenges. It may be flat out incorrect, but being incorrect gets a bit more insidious. Here, look at this. Oh yes, fellow scholars, this is an AI that is not sure and it is hedging. You know, if I say that the stock market may go up or down or it might go flat, I am completely right, but not very informative. Is that incorrect? Hmm, hard to say. So what are the results? Well, hold on to your papers, fellow scholars, because when they compared their flagship models to their mini variants, oh goodness, that is a huge difference. They are still wrong a lot. And our question now is okay, but are they like people? Do they know that they are likely wrong on many of these things? Are AIs aware of their limits? I would say the answer is surprisingly yes. When asking them how confident they are in their answers, whenever they feel more confident, they are more likely to be right. And when less confident, less likely to be right. Now if we are using this for search, we are hoping that the answers are also consistent. Are they? Let's have a look. By asking the AI models the same question a hundred times and are they consistent? I would say that the flagship models kind of are not even close to perfect. So careful there. But once again, their flagship Reasoning O1 and GPT4O models are remarkably consistent. Much more consistent than some politicians I hear speaking these days. And this is where this new dataset shines. It helps AI models become more confident when they should be confident and more consistent. And this will finally lead to less hallucinations and more accurate information. For us fellow scholars, it is still a long road, but this is one bloody good step in the right direction. What a time to be alive. So what do you think? What would you fellow scholars use this for? Let me know in the comments below. And if you feel inspired and you wish to run your own experiments on Nvidia GPUs, check out Lambda. Lambda is a cloud service for large scale AI projects and it is my absolute favorite. Seriously. Try it out now@lambdalabs.com papers or click the link in the description. If you use Lambda, you can run your own experiments on super powerful hardware and you also support two minute papers.
