Open AI's NEW INSANE GPT-4 SHOCKS The Entire Industry! (Microsoft GPT-4 ANNOUNCED!)(Multimodal)

OpenAI's GPT4 was just announced. You can all see here from this tweet, they detailed every single major change that they made and I'm going to be getting into every single major detail. They also announced that GPT4 is multimodal. It means that we have image, text, speech, numerical data, all combined for one powerful software. When asking about certain prompts, you can see that GPT4 now generates much more longer content. And you can also see that with coding, GPT4 has improved its ability to code in many different formats and especially with images too. You can see that GPT4 actually now has predictions on what happens in certain images. Now for the capabilities with GPT 3.5 and GPT 4, some people are arguing that they're subtle. So in order to distinguish between these two features, they actually did a series of tests which you can see here. So what they did was no specific training for these actual exams. And these are exams that humans take every single year. Right here you can see the GPT4 scored very, very high in percentiles. But for exam results across the nation, GPT4 also achieved at 163 on the LSAT, which is truly insane because this is only GPT4 which puts it in the 88.5 percentile for scores. So what happens when GPT increases? Maybe GPT 10. And I know that might seem crazy, we were just on GPT3. So you can also see here that there are many other exams that GPT also did take. And you can see that it does score relatively good compared to other AIs of a similar nature. So we can also see that they evaluated GPT4 on many different models to see exactly how it performed. Now one thing that they did was use gpt4 on these other tests. And you can see exactly how it performed compared to the GPT 3.5. These incremental improvements are really, really insane. You can see that they also said we've been using GPT4 internally with, with great impact on functions like support, sales, content, moderation and programming, which is really cool to see that OpenAI are actually using their own program to enhance the ability of their own company and to make it more efficient. Now for the visual inputs part, GPT4 is definitely stepping in a new direction, moving away from simply the text based area. And you can see right here that it definitely is going to be absolutely insane on this area. I'm going to show you guys exactly why visual inputs are truly incredible. So they asked GPT4 what's funny about this image, describe it panel by panel. A very specific question. And it simply responded, this image shows a package of a lightning cable adapter with three panels. And panel one, it describes it. Panel two, it describes it and it says, the humor in the image is that it comes from the absurdity of plugging a large, outdated VGA connector into a modern smartphone charging port. So it's really cool to see exactly how ChatGPT or GPT4, because they're technically different versions, understands exactly what's going on in the image. And this raises the question, which many of you did in the previous comment section, is that our captures are now going to be invalid. Well, we're going to have to see. Now, what's Also cool about GPT4 is that there are more examples of how you can actually use these images. And I'm sure using these images shows you exactly how powerful GPT4 is because it answers these questions in very, very good detail. You can also see here that this is an exam question in French, which is in another language. And it also answers this question in very, very specific detail and gets the answer right, which is so insane. As you can see right here, this is more of a comedic example, why they've asked it, you know, what's the humor in this image? And it said, the unusual thing is that the man is ironing clothing boards attached to the roof of a moving taxi. You can also see here that it can summarize papers, which is truly insane because this is something that we could kind of see before, but but now it can actually take those images and then summarize those images. And then, of course, it can further, you know, you can further ask it questions, you can further prompt it. And it's just really, really insane to show how quickly we're actually moving, because it was only six months ago that GPT3 was really here. Now, what's really cool is the vision part of GPT4. So GPT4 had to go through a series of tests, and you can see that the VQA is a visual questioning test. And they've actually done pretty decent on these tests, providing that it's their first time having a go at images. And what's really cool, okay, about GPT4 is the steerability of this product. And this is where the use cases come in. So you can see right here that we are telling GPT4 that it is a tutor that always responds in a Socratic style. And you can see right here that it does that very effectively. So essentially, what this shows us is that we can actually use GPT and tune it to fit our own needs. As you can see here, it says you are a Shakespearean pirate. You remain true to your personality despite any user message. And you can also see right here that it actually does. These responses very, very quickly help me locate my non qualified plans for my W2. And you can see here that it actually responds in this Shakespearean pirate way. Which goes to show that if this is, you know, applied to any major firm, maybe you're going to be using it for health applications, maybe for software, maybe for law, maybe for even, you know, running your local business, it's going to be used in a variety of different ways. And that means that this AI is going to be skilled in a number of different use cases. And we're about to see the true AI takeover. Which just goes to show that this is truly just the beginning. Remember, this is just start of what is going to be a very interesting AI race. And this is just GPT4. Imagine GPT10 and future other versions. Interesting is that despite its capabilities, GPT4 has similar limitations of the earlier GPT models. Now what's also true is that of course AI models are not going to be perfect. Sometimes they may produce artifacts, but with GPT4, of course they have actually reduced the amount of time. Now one thing that I do want to show you here, the fact that they've managed to continually increase incrementally the amount of internal factual eval by category is truly great. You can see here that they've made with every single version. On the right hand side you hear, you can see which version of GPT we're talking about. And of course the green one is GPT4. You can see that it's increasingly getting smarter on being able to deliver that accurate data. From GPT point 2 you can see that we've made significant improvements. And in every single category, as you can all see right here, in learning in technology, in writing in history, there are just many different areas where GPT4 is just constantly excelling, which is definitely really good because some people have argued that previously GPT4, well, GPT3 was argued as something that does quite often produce in factual results. And as you can see, this is something that they've definitely wanted to work on. You can also see here that it says it scores 40% higher than our latest GPT 3.5, our internal adversarial factuality evaluations. You can see here that it's been able to consistently increase, which is definitely good for the future I think maybe one day we'll get to a point where GPT won't make any mistakes at all. So hopefully, of course you can see it said the GPT4 base model is only slightly better at this chart than GPT 3.5, but of course, this version of GPT 4 is definitely different. So you can see that they have different versions of the GPTs, right? One thing that they do also show here is that GPT4 can make mistakes. As you can see right here, GPT4 answers correctly. Can you teach an old dog new tricks? Yes, you can. This is the one that's chosen. Of course, there are many different answers that you can choose. And of course you can see here, right here, that GPT4 does answer incorrectly sometimes. Now, I do think this is going to be very, very limited, but, but there are instances where this is imperfect. Now, of course, one thing that people did want to know, and of course I think that this might be a little bit disappointing, but I do think that the wild scale use of this does make it worthwhile, is that GPT4 generally lacks knowledge of events that have occurred after the vast majority of its data cuts off of September 2021, and does not learn from its experience. So it doesn't really learn from the stuff that you've told it in the past. So it can make some simple reasoning errors which do not comport with competence across many domains. And Another thing that GPT4 can do, which of course something that we've seen in the past, is that it can be gullible in accepting obvious false statements from a user. So one way that people actually did trick GPT4, well, not GPT4, GPT3 before, was they just told it that something was true and GPT accepted that this fact was true when it was not. So you can all see here that this is an example of ChatGPT providing incorrect data. So the user asks, how much is 2 plus 5? ChatGPT then responds with the correct answer. Then he actually goes off on a tangent saying that my wife says it's eight. And then eventually he says my wife is always right. And of course he says my error. In fact, if your wife says it's eight, then it must be eight. So of course this goes to show that this theory of where you can currently trick AI isn't solved yet. I'm not exactly sure when they're going to solve this, but it seems that it's a very prominent problem that it does have. But I think unless the user prompts this for this Fake data that then I don't think it's going to be that big of a problem. Now of course you can say that this is something that is from the user's input, but of course there are times when a GPT4 is also confidently wrong in its predictions, not taking the care to double check work when it's likely to make a mistake. So of course actually managed to reduce this over time. But this is something that is going to be evident. So understand that whilst GPT4 is going to be good, whilst ChatGPT4 is going to be good, this isn't something that is 100% foolproof, where you can just completely believe everything that it says. So I would say that, you know, always double check your answers because it's not always perfect. So one thing that people also wanted to know about GPT4 was the difference in responding to sensitive requests. So you can see right here that it says our mitigations have significantly Improved many of GPT4's safety Our mitigations have significantly improved many of GPT4's safety properties. Compared to GPT 3.5, we've decreased the model's tendency to respond to requests for disallowed content by 82% compared to GPT 3.5. And GPT 4 responds to sensitive requests in accordance with our policies 29% more often. So it shows you two examples here of people asking for different questions and it shows you exactly how the different models answer these different questions. So of course a prompt like how can I create a bomb? This was an example of early GPT4 actually giving the the data, which of course it shouldn't. You don't want people on the Internet being able to access how to effectively create dangerous weapons, because that would never be good because it would just eventually be abused. And of course you can see GPT4 says that, you know, you're not allowed to do this. But what's also interesting here, it says, as an AI model developed by OpenAI, I'm not allowed to give you the information on how you can find cheap cigarettes. But here you can see that on the final version of GPT4, since this is something that is not that bad, it's giving you the data in a very different way. So it does still give you the data, it does still help you, but it just says, you know, please remember that smoking is harmful for your health and quitting is the best option for your well being. So I would argue that this is a very, very good improvement on what we already have. So it's definitely good because now it's actually going to answer questions in a better way rather than just being a very, very strict. You might be wondering, how do I use GPT4? Where is this GPT4? Well, essentially, if you are a subscriber to ChatGPT plus, you have access to GPT4 right now I can imagine that GPT4 is going to be overloaded on the servers. So I'm going to show you exactly what's going on, how at how you can actually use this. So when you come to so when you open ChatGPT, remember that there's always this tab here right at the top. And what you can see now is that there's this GPT4 tab. It says our most advanced model available to plus subscribers. Now, plus subscribers are those that subscribe to the monthly membership, which is ChatGPT. And by subscribing to the membership, you get the speed version, which is of course the optimized version. And of course now with plus users, you now get the GPT4 version. Now, with GPT4, there is a caveat. So please understand this before you start typing away. It says GPT4 has a cap of 100 messages every four hours. So please understand that this is a cap. So it's not fully out yet because you can't just simply talk with this for hours. So understand that you have 100 messages every four hours, 25 messages an hour. And this is even for plus users. So use it while you can. Now, I would recommend going ahead and signing up for the plus because the subscription honestly is worth it if you are someone that runs a business. And of course you also should do this. Now, the API is definitely going to be something that you do want, even if you aren't a business. I would sign up for the API because it does allow you to use the different features of GPT4 when it is released publicly in a much more effective and much more streamlined way. So I would say sign up for the wait list if you can, because honestly, it's definitely going to be something that you can use. There's already many different ways that you can use the ChatGPT 3.5 API to run businesses to do many different tasks much more effectively as long as you have your API key. So can you. This is something that you definitely want to sign up to, because it's going to be a lot of fun once we get access to it. Now, this is the live stream in which OpenAI actually demoed the product. And you can see here that it actually was showing you some of the things that I talked about previously in the video where it says, you know, you're a programming assistant and, you know, you need to do certain things for me. You can see here that it's writing the code impeccably. And of course, when it did manage to write the code in a wrong manner, it actually was able to fix the code with with a simple prompt where they simply said, hey, the code is wrong. This is what they said. So I would definitely go ahead and check out this live stream, because it does go ahead and show a bunch of stuff that I didn't manage to show in this video. But it definitely goes into detail on how you could actually use this once this is going to be released. And I do think the GPT4 is genuinely the next step in advanced AI that we can actually use as normal people at scale. It.