AI Drone Kills Operator, GPT 4.2 Leaks,Bard Gets SUPERCHARGED, And Much More [AI NEWS #5]

Another giant week in artificial intelligence. And let's get straight into the news because boy oh boy is there a lot to cover. So you know the company Meta, formerly Facebook, the one that Mark Zuckerberg owns, they have actually been very, very good in terms of AI development, although they've been a little bit under the radar because they've been overshadowed by companies like Google and of course OpenAI. In conjunction with Microsoft, they've still been releasing large language models that have been doing insane things. Now one thing that they released earlier this year was was they did release a model called Llama. And this was a lightweight model. It was a 65 billion parameter model. Now what was interesting was that recently they actually released something called Lima. Okay, so essentially Lima was a fine tuned version of Llama with only a thousand carefully prompted responses without any reinforcement learning or human preference modeling. Now what's interesting about Lima is that it's one of those models when it's trained using the LLAMA architecture, essentially it means that it's going to be much more scalable and, and can be ran on much smaller devices. What's great about Lima is that it actually achieved remarkably strong performance on a range of tasks and outperforming state of the art language models such as OpenAI's ChatGPT3 and Google's Bard. Now the reason Lima is such a groundbreaking piece of technology is because number one, it's actually much more lightweight and it's actually a much more scalable model. Now what's incredible about this model as well and Llama is that I've actually seen people running different versions of LLAMA on their laptops locally, meaning that you don't need any servers. It could literally just be run on a laptop, which of course, as we know, the future implications are huge. Now what's also interesting about this as well is that there are several implications for Lima's success for the future of natural language processing and artificial intelligence. Firstly, it suggests that pre training is a powerful approach for training large language models and that it may be possible to achieve strong performance without relying on reinforcement learning or human feedback. Which means that this could make it much easier and more efficient to train large language models for for a wide range of applications. Secondly, Lima's ability to generalize well to unseen tasks suggests that large language models may be able to learn to perform a wide range of language understanding and generation tasks with only limited instruction tuning data. Finally, Lima's strong performance in a controlled human study suggests that large language models may be able to produce high quality output that is equivalent or even preferred to human generated responses in some cases. And in these human controlled studies, responses from Lima were either equivalent or strictly preferred to GPT 4 in 43% of cases and as high as 58% when compared to Bard and 65% when compared to ChatGPT, which was trained with human feedback. Then what we had was yolo, an image recognition tool that was upgraded from previous versions compared to YOLO version 6, version 7 and version 8. Here we can see a demonstration of YOLO being used on a clip of a popular TV show called the Office. And it's easily able to identify every single thing in the image. Now the thing about this is that of course people aren't going to be using this for TV shows for memes, but this has many different applications. Maybe you're going to be using this for a company's warehouse, maybe you're going to be using this for identifying things in a public setting. Maybe you want to see if certain animals are still in their enclosure. It definitely is something that we haven't really seen too much when it comes to artificial intelligence. But image recognition is something that is going to be useful. And as we do know, as future AI systems are evolved, we, we do know that many different AI systems are going to rely on a multimodal capability, which essentially means that they will have the ability to process different kind of inputs and outputs. Now with YOLO here, an image recognition tool, what happens if we do have this kind of software embedded into a physical robot that is going to be powered by a multimodal AI? I do think this kind of software is going to be very, very useful for identifying many different robots things. And as we do know, many different companies are working on their artificial intelligence robots. Then we had something that was also quite interesting, something that I didn't see coming. We had this research paper called Prolific Dreamer. High fidelity and diverse text to 3D generation with variational score distillation. So what's really interesting about this paper is that they've nearly cracked the code for Perfect text to 3D generation. You see, previous versions of text to 3D generation suffered from many other problems such as over smoothing and over saturation. So this paper actually proposes a new technique called variation score distillation, which is a particle based variational framework that treats the 3D parameter as a random variable and infers its distribution. With this new technique, they're able to generate high quality 3D nerfs which look very, very realistic. You can see here Michelangelo style statue of a dog reading News on a cell phone, a delicious croissant. We've also got things like an elephant skull, a blue tip, a small cactus planted in a clay pot. We've also got things like a pineapple. We've also got things like a snail on a leaf. And if you don't know what you're seeing on the right hand side, and if you are confused by that, that is essentially the 3D model's normal map, which is essentially a map technique which allows the model to have certain characteristics which make it look more realistic once the actual texture is applied. Essentially it just gives it more depth and that texture tells the computer how to render it and how to infer certain characteristics of that 3D model. Now, there are other 3D models that they're also made and there are also other nerfs that they were able to generate that do look pretty realistic. We may actually be quite some time away from actually having these as usable assets. But I do think that this technology is still progressing very quickly and it does have a whole host of other different applications. What's also good about this model as well, something that I forgot to mention is that it does cover a wide range of diverse results, which means you can actually prompt this with many different things, unlike other previous generations. So one of the papers that did get released earlier this year was Google's Music lm and it was honestly quite shocking. So essentially what you were able to do was you were able to generate audio from rich captions. So think of it like this, it's simply text to music, but it isn't text to music as we traditionally think about music. It's more so text to background music, I would say, or text to music samples that can be used in full scale productions. So for example, you could use music like this, the main soundtrack of an arcade game. It's fast paced, upbeat, with an electric guitar riff. Then we have another example of reggaetron and electronic dance music. Then we have some other examples here and there are some other examples here too. Now we already did cover this in another video, but what was actually interesting was that this tool actually started to roll out public use. Now you do still have to sign up for early use. I'm not sure what the wait list is, but I do know that many more people are now getting accepted. If you head on over to AI Test Kitchen and you do sign up and you are accepted, this is what you're presented with. It says how to make a good prompt. Be very descriptive. Electronic or classical instruments sounds best. Mention the vibe, mood or emotion you want to create. The certain queries that mention specific artists or vocals will not be generated, which is something that many people do want to emulate, for example, their favorite musician. That's obviously something that has its issues largely because of copyright. And what's also interesting as well is that the model actually allows human feedback. So essentially it gives you a selection where it says which track is better, give it a trophy. So let's say, for example, I wanted to generate lo fi music that I might hear in an elevator and I click generate and it's going to generate me two different music tracks. And once I see those two different music tracks, I can then play both of them and see which one sounds better. So let's play track number one. Now these tracks are only around 12 seconds long, which is currently what's been set, but I do think that they're still very decent. Okay, so we're gonna go ahead and we're gonna click the trophy which gives us track number one, feedback. And then of course, if I want, I can literally just click here, click shareable link and then I can also click download. Now I did find that this was really, really good. And this was another example that I did in another video. And then we can hear track two as well. So yeah, definitely something that we can actually use as background music. And I've got to be honest, if I did hear this music, I wouldn't instantly be able to tell, especially since I would be focusing on whatever the main piece of content focus would be. So I think we are a little bit off when it does come to music generation, simply because for now we are in that stage where we are actually giving it human feedback. So we need to just constantly give it these trophies. And as more and more users give it this human feedback, eventually the AI model is likely to improve over time and likely to output longer pieces of songs. But it's also important to note that the music LM is definitely capable of doing five minute songs and a whole host of many different things, such as conditioned melodies where you give it a specific song prompt and then it uses that driving song in order to make a generated audio. Then we had this browser agent which was actually quite effective. It's called Multi on. It's something that works as an autonomous agent. We've seen this before with GPT4 and GPT3's AutoGPT feature, which many people just decided to code. But this is, I guess you could say a fine tuned version of that. So the user has asked to book a Delta flight from June 11 to June 14. And you can see right here, this is still an early version, but this is what, what the AI agent is doing. So it tells you exactly what it's doing. In the bottom right hand corner, it's actually searching Google and then it's going to click on the matching desired dates and it's going to search which ones it wants for you. As the AI agent continues to go around, it's going to continue to visit various different websites, sort through many different things to book, and things like these are quite useful. You see, these tasks are what you describe as something that are quite tedious because many people, when doing things that we have to do honestly spend and waste, I would say, a lot of time because we have to do these mundane tasks that aren't really productive. I mean, it would be much more effective if we could have an AI that could sift through millions of different prices for us and then find us the best one. And that's exactly what we do have here. Now, once this is fine tuned, it's going to be interesting to see if these are what we're going to be using in the future as normal, because I do believe that many different tasks are going to be completely automated and, and we're going to have AI scrolling the web and booking us various different things then. What was also interesting is that ChatGPT has been getting a bunch of subtle updates. Essentially, if you've been using ChatGPT on a day to day basis, you may not know of some of the small changes that ChatGPT actually has had. One of the more interesting things is that in ChatGPT you can now actually share the conversations that you have with ChatGPT. Now this is actually really great because it allows you to share prompts and it allows you to continue conversations that already have an extensive history. And I know that this is something that I'm going to be using when I want to share a conversation that I've had. And it's very, very good because you're now able to share dialogues that otherwise people could have faked. So for example, if someone says that ChatGPT has said something, they can now simply share this feature. And this was something that, interestingly enough, OpenAI's CEO Sam Altman did say that he was surprised that ChatGPT was such a success because he actually said at YC Combinator that that literally every startup he advised them to have some kind of viral sharing feature in which apps would be able to explode. And it's funny that he didn't add that feature. And now with ChatGPT, he's slowly just sharing out the sharing feature, which I generally haven't seen that much on social media, but it is a cool feature that will be useful for the workspace. Then we have right here, which is an interesting chatgpt leak. Now remember, this is currently speculation and this was from Reddit. However, it's been all over Twitter and it seems like it's public news. And right now, apparently a user that predicted previously leaked some of the ChatGPT updates before has come out with this screenshot where essentially you do have this my profile area and this my files area where which it could mean that you're going to be able to upload files to ChatGPT and you're going to be able to have ChatGPT remember certain preferences about you and certain information. Now I've got to be honest here. I've done some digging and I've seen this upload file feature on ChatGPT before. You see, there was a user on YouTube who actually made this tutorial in which you could actually submit a file to chatgpt. And it was a really cool workaround, but it just goes to show that this is something that is actually possible. So I wouldn't be surprised if these features were added because a lot of the times when you do start a new chat, you would like ChatGPT to remember certain things. And there are many third party applications like Chat PDF in which you can actually talk with PDF files and you can actually upload many different files to ChatGPT. So this is something that I would expect. And just for reference, this was the actual original video here. It actually has over 500,000 use and it's something that you can still use today. Then of course we had a major update, which was one that we really couldn't wait for, which was Adobe Firefly being released to the public. So if you don't know what Adobe Firefly is, it's essentially a tool like Midjourney, but it's created by the company Adobe. If you don't know who Adobe are, they created Photoshop, which is used by many different people. And of course many people who are familiar with the Creative Suite know about Illustrator and of course Premiere Pro and of course After Effects. Now, Adobe Firefly is really good because it has many different features that Midjourney doesn't have, and it has many different creative features that many of the creative individuals who work in the creative industry already do need. Now, some of these features aren't actually in the Adobe Firefly app just yet, but they still are currently being worked on. So many different things, such as text to 3D are still currently in production. But the video that was released by Adobe does actually showcase a bright future, which for those content creators who are looking to create content very, very quickly along with AI assistants. And of course that brings us to something that Adobe actually did release, which surprisingly took everyone by surprise. And I've definitely seen it a lot on social media because it has very, very effective effects. And of course this is Adobe's Generative Fill feature. If you don't know what that was, it was basically a revamp of an old feature that they had. But when using the AI technology, you actually able to generate a completely unique thing within that image that was completely, seamlessly added. This was something that many people wanted originally and for it to be simply embedded into Photoshop was a complete lifesaver because you can simply outpaint in paint, add objects. And although There are some AIs out there that can actually do this right now, none of them were able to do this as effectively as Photoshop and as natively as Photoshop did. So Adobe did a great deal by releasing this. And in this video you're going to see me actually messing around with it. And, and generally it was actually effective. It wasn't just some kind of interesting fad, it was something that I can guarantee you people are using because it saves you hours and hours of editing work. Generative fill is something that many people have used to expand pictures, old pictures, to just simply reframe things. And of course it's something you can do essentially called out painting with things like stable diffusion and things like neonado AI. But this is something that many people do like now because it's just in Adobe and it seems really, really effective. And the thing with Generative Fill as well is that it also gave you three different output versions that you could use. Another thing that was released, well, I wouldn't say released, but it was a research paper that was really, really interesting because this research paper talked about emergent correspondence from image diffuse. So essentially they talked about an emergent ability from text to image, which like we talked about in a previous video, where we Talked about how GPT5 would be extremely risky and how AI does present a whole host of new emergent capabilities. This research paper showed us that this is actually true even in a text to image. So essentially you know how when you look at two pictures and you can point out things that are similar between them like maybe they both have dogs or trees or a big blue sky. You're able to understand that those things are the same, even if the pictures are a bit different. Now imagine we want a computer to do this, so we want it to look at two pictures and understand what's the same and what's different. That's called finding correspondences between images in the world of computers. So in this paper, the smart scientist found out that computers can actually do this really well when they use something called image diffusion models. This is like a superpower the computers use to diffuse or spread out the image and understand its details. What's amazing is, is that the computer can do this without needing any specific instructions or supervision. It just kind of figures it out. And this superpower can be used to create special features which are like little hints or clues about what's in the picture. So they named these clues diffusional features, or DIFFT for short. These diffs are super helpful because they can be used to compare the two images and find out what they have in common. And the best part, the computer doesn't need any extra help or fine tuning to do this. In fact, when scientists tested this, they found out that DIFFT was even better than than other methods at finding these correspondences when looking for the same meaning in different pictures, what they call semantic correspondence, DIFFT performed better than some other popular methods. So, in essence, these scientists discovered a way for computers to be super smart at looking at pictures and understanding what's the same and what's different about it. And they don't need any help or instruction. Now, what's also cool as well is that they managed to use this on video data, even though video data it wasn't trained on. So I do think that this is an emerging capability which is actually really interesting. And the webpage that they have also does have some interactive demos. So I think this is a real win for computer vision. And it just shows that even on the back end, even on certain categories in which you might not think there's a lot of AI progress, there are small, small incremental changes and hundreds of different research papers being published weekly in which these small changes are going to amount to a huge gradual change over the next coming years. The paper also did show us image editing with difft, where essentially, if you can plot certain points on certain images, then of course you can translate this across into many different other things. And like I said, I think this has a vast majority of different applications, and it will be useful to see one day. Once all of these AI tools come together and produce a highly complex AI system such as AGI or even asi, there was also a small update to Bard, which actually increased the effectiveness by 30%. So Sundar Pichai, the CEO of Google, said, we're updating Bard with a new technique called implicit code execution. Now, it runs code in the background when it detects computational prompts, improving the accuracy of word and math problems by 30%. And this was something that Bard did struggle with for quite some time, and it was literally one of the main reasons as to why nobody really used Bard over ChatGPT. So with this issue being solved, it's likely to give Bard some more ground in terms of usability. There is also a longer blog post which they talk about how Bard has improved its logic and reasoning skills. And they said that essentially they have a new method. So the new method allows it to execute code to boost its reasoning and math abilities. And this approach takes inspiration from a world study, dichotomy and human intelligence, notably covered in the book Thinking Fast and Slow. So system one is thinking fast, intuitive and effortless, and system two is slow, deliberate and effortful. And essentially what they've done is they've actually combined these two systems in order to improve the accuracy in Bard's responses. Now, what I always found cool about Bard was that it always was super, super quick. So if it's much quicker than ChatGPT and is now going to be much more effective with three different prompts, I think that over time, Google could eventually gain some ground here. Then we had a stunning research paper by Google's DeepMind in which AlphaDev discovers faster sourcing algorithms. Essentially, they figured out a way to use new algorithms to transform the foundations of computing. And essentially this is going to mean that systems over time get more efficient and more streamlined. There was also Microsoft's orca, which was released, which I'll describe now. Now, it might seem complicated at first, but this is actually really simple. All Microsoft did was they published a research paper in which they managed to build a system that was basically nearly just as good as GPT4, but a lot less intensive, meaning that we could have a lightweight, easily scalable version of GPT4 that as just effective. Now, the paper does go into how good this new language model is called orca, and honestly, it is quite effective. Now, I think what this research paper shows is that the way how they trained this new AI, they essentially just trained it based on the explanations that were from GPT4. And basically what they did was they got these complex explanations, got GPT4 to explain them in super simple manners, then trained the language model on that, and there were only a thousand of these. And then essentially that language model is very, very effective. So what we're seeing is rapid advancement in the scalability and the level of these language models, which means then in future we're going to get large language models with less parameters, more lightweight, more scalability, which means eventually all these large language models which are hosted across tons of servers, won't be the norm. And ORCA is pretty much on par with one of OpenAI's earlier models, which is very similar to ChatGPT. So it will be interesting to see where this model is in the future and just how quickly these models are able to surpass GPT4. What's also cool was that Boston Dynamics robot actually did manage to get an upgrade, which actually allows it to see various different things and allows it to also be be an additional worker in terms of many different industries for many different applications. You can see the video right here. And it just goes to show that many of these actual robots that are being integrated with AI are a lot more effective than just these large language models that we can prompt, because they actually have real daily use in reaching things that we can't reach, in seeing things that we can't see, and in analyzing certain situations with the many different sensors in their hardware. What is very interesting was a recent test which was also done with the US Air Force, in which they actually deny running a simulation in which an AI drone killed an operator. So according to this article by the Guardian, it goes into details on how an AI drone that was deployed used highly unexpected strategies to achieve its goal. So the article starts out by saying that the US Air Force has denied it's conducted an AI simulation in which a drone decided to kill its operator to prevent it from interfering with efforts to achieve its mission. Carl Tucker, Cinco Hamilton described a simulated test in which a drone powered by artificial intelligence was used, was advised to destroy and enemy air defense systems, and ultimately attacked anyone who interfered with that order. The system started realizing that while they did identify the threat, at times the human operator would tell it not to kill a threat, but it got points for killing that threat. So what did it do? It killed the operator. It killed the operator because that person was keeping it from accomplishing its objective. So it goes on to say that we train the system. Hey, don't kill the operator. That's bad. You're gonna lose points if you do that. So what does it start doing? It starts destroying the communication tower that the operator uses to communicate with the drone to stop it from killing its target. Now, of course, no real person was harmed, and Hamilton, who was an experimental fighter pilot, has warned against relying too much on AI and said the test showed you can't have a conversation about artificial intelligence, intelligence, machine learning and autonomy if you're not going to talk about ethics and AI. Now, the article continues to state that the US Air Force spokesperson denied any such simulation had taken place.