Major AI NEWS#19 (GPT-6 News, Amazons New Q Model, Realtime AI Conversations and More)

So with Sam Altman being fired and with so much going on in the OpenAI saga, there was a lot of AI information that you actually might have missed. Because it wasn't just OpenAI that was doing developments. There were tons of other companies producing lots and lots of AI models and lots and lots of AI research. So that's what we're going to be showing you in this week's video of what happened in AI. So without further ado, let's get started. One of the first things that I did see, and this was a use case for GPT4, and this was something I. I can't believe I missed this, and I'm sure that some of you did as well, but this was something that was pretty, pretty crazy. So if you know, GPT, OpenAI, they basically released a GPT4 Vision, but they released the GPT Vision API, which means that people can do some really, really cool things. So essentially someone used GPT4 Vision and an 11 Labs Python script, so you can basically narrate what's going on. And this guy used David Attenborough to narrate his entire life. And I think if you're someone who's seen David Attenborough before, I think you should check this out and you'll understand truly how amazing this is.

Here's a quick demo of what happens when David Attenborough narrates your life. On the left, I'm going to start my camera, which every five seconds takes a photo from my webcam. On the right, I'm going to start David Attenborough. And now, as I move around here.

We have a remarkable specimen of Homo sapiens, distinguished by his silver circular spectacles and a mane of tousled curly locks. He is wearing what appears to be a blue fabric covering which can only be assumed to be part of his mating display. Look closely at the subtle arch of his eyebrow. It's as if he's in the midst of an intricate ritual of curiosity or skepticism. The backdrop suggests a sheltered habitat, possibly a communal feeding area or watering hole. Ah. In its natural environment, we observe the sophisticated Homo sapiens engaging in the critical ritual of hydration. This male individual has selected a small cylindrical container, likely filled with life sustaining H2O, and is tilting it expertly towards his intake orifice. Such grace, such poise.

Then, of course, we had someone who also used this same one to get someone to critique their designs directly in figma. And that person that I'm talking about was of course, Steve Jobs. I think something like this is really cool and Interesting because imagine you're someone who's maybe a software engineer and you're working on a product. Maybe you're someone who's working on product design and you can take the voice of the most noted person in your industry, ask ChatGPT to essentially talk in whoever it is style and have them narrate to you certain changes. I think something that is that technologically advanced that you can do is a real application use case. It might not be that good in terms of the pricing because GPT4 Vision is quite expensive, but I still feel like this just shows us how creative these models are and why this is really, really cool.

Image shows a computer screen with a web based application open. It appears to be an image gallery within a content management system or a digital asset management tool. The left side of the screen displays a navigation menu and there's a search bar at the top suggesting functionality for users to find specific images. The main panel on the right features a collection of vibrant, colorful artwork depicting various subjects such as portraits and pop art inspired designs. In assessing the design of this interface, there's a coherence in the color palette used within the artworks that gives the gallery a unified aesthetic. However, from a usability standpoint, it's unclear what the focal point is. The search functionality should be more prominently featured if it's an essential tool for interaction. Furthermore, the navigation on the left warrants a stronger visual hierarchy to help users intuitively move through categories or functions, simplify to amplify, let the content breathe, and guide users with a clear path through your design.

Then of course we had Microsoft releasing its alternative to custom chatgpts. Essentially what you can do here as you can build standalone copilots which are essentially GPTs with a no code tool. And essentially what this allows you to do is to collect your GPTs to your Microsoft data sources. Quite like how Google Bard has access to all your Gmail and all that stuff. This is Microsoft's version of that. They did release this short trailer which you're seeing on screen now and I think like we said, the style of getting more customized AI models to our computers is definitely going to be a direction that we're heading in. It's going to make them more useful, it's going to make them far more effective. And this is just the beginning on what we are starting to see because even some of the GPTs that we had before and the stuff that with custom GPTs sometimes we're seeing that they aren't always accurate and don't always listen to our Instructions. So we with Copilot, Microsoft actually did do an entire event. I think this is essentially something that you could use with Microsoft 365. So if you're someone that regularly uses documents, if you regularly use PowerPoint, if you regularly use Hotmail or any of Microsoft's products, then this is going to be something that you are going to want to use. Because as you know, AI with your data is just literally insane. Just trust me, I've used it, not this particular software, but I've used AI with my own data and it is far superior than just asking ChatGPT and having to input the data in every single time. Then of course we had something which was actually quite unexpected. But this was a development that I did know was brewing. And this is of course Google DeepMind combining their capabilities with YouTube. It says today we're sharing a sneak peek at our first set of AI related music experiments. Dream track for Shorts and music AI tools built in collaboration with Google DeepMind. Now previously we did cover some stuff that was by Google that was essentially music. And when I did look at the previous models that were done by Google, they were actually really, really effective and honestly they were quite surprising. So it's clear that Google is still working on music AI and that Google DeepMind have been working on this for quite some time now. I will try and show you a clip from the trailer to see if it actually does work, but I, I think this is going to change everything. The reason I say that is because YouTube shorts are an inherently viral platform. So what do you think is going to happen when people don't want to use copyrighted music, they start using AI generated music. Are we going to see that shift away from and then maybe we're going to get music tracks that are essentially AI generated, be the ones that are going to be consistently viral. Now DeepMind does have a page about this where they talk about transforming the future of music creation. And, and you can see that announcing our most advanced music generation model and two new AI experiments designed to open a new playground for creativity. From jazz to heavy metal, techno to opera music is luck. And today in partnership, we're announcing Google DeepMind's Lyria, our most advanced AI music generation model to date and two AI experiments for creativity. So essentially the reason I do believe that this is really cool is not just the fact that it's going to be AI generated tracks that you can use and it's actually going to be soundtracks. Then people can create themselves. People can literally start humming a soundtrack and then immediately get themselves some music. Now I think it's good that they're testing this out on YouTube shorts because what we're going to see is how it actually does work across that section of YouTube. Now what I also do find interesting that people might not, but I do find this really interesting is that they talk about water masking AI generated audio with Synth id. So it says our team is also pioneering responsible deployment of our technologies with the best in class tools for watermarking and identifying synthetically generated content. So any of the content published by our LARIA model will be watermarked with Synth id, which is essentially something that they also use on Imagen and Google's Cloud Vertex AI. So basically what it does is with every single soundtrack, it adds a visual spectrogram to add a digital watermark. And essentially that means that if there is AI generated audio, you're going to be able to know exactly if that is AI generated or not. Then we have something cool from Ethan Mollick. We essentially have this paper which is really interesting. It's about GPT4 for self driving. Now the reason that this is really cool is because essentially this paper showcases using GPT4's vision abilities to use a car to self drive. Now essentially what they did was they essentially had a prompt and then essentially it says describe what you see and what you plan to drive next. And they gave it a bunch of different, different scenarios because they wanted to see exactly how GPT4 with vision would analyze these scenes and how it would score. Currently at nighttime it does do really well. You can see that it manages to analyze the scene and then predict exactly what it would do in the car. And then you can see here that it brings the conclusion. So for the capabilities of autonomous driving, it essentially says that GPT4 exhibits capabilities that have the potential to surpass those of existing autonomous driving systems. And in aspects such as scenario understanding, intent recognition and driving decision making in corner cases, GPT4 leverages its advanced understanding capabilities to handle out of distribution scenarios and can accurately assess the intentions of surrounding traffic participants. And I do think that that is something where it's going to excel because where GPT4 is trained on millions of millions of images and millions of different scenarios, especially with its LLM, autonomous driving isn't, it's not trained in that similar way. So that's why GPT4 with vision excels in area. It says moreover, it can infer the underlying motives behind these behaviors. As highlighted in section 4, we witness the performance of GPT4 with vision in making continuous decisions on open roads and it can even interpret the user interface of navigation apps and guiding drivers in their decision making processes, which is clearly superior than an AI that can literally just drive. Then of course we do have some limitations because with every single large language model and AI system that's currently available there aren't without limitations. It does sometimes struggle to recognize left and right. There is sometimes struggle with traffic light recognition. Spatial reasoning also is a issue because you need to be able to move the car in the effective enough space. But overall I do think that something like this is really interesting because it showcases that with more multimodal capabilities we're getting things that are more and more increasingly capable in ways that we genuinely didn't think then of course what we do have is Anthropic releasing their model claw 2.1 which offers an increased 200k context window and a 2 times decrease in hallucination rates, their system prompts, tool use and updated pricing. Now one thing I did find interesting about this model was the fact that they did something where they essentially nerfed the model. Now people might not agree with me, but I believe that there is a large majority of people that do. Because essentially what you can see here is that Claude 2 is in the purple and then Claw 2.1 is in the dark purple. I really did wish that they used completely different colors because that just doesn't make sense. But you can see that although 2.1 gets many less questions wrong, it also declines to answer many different more questions, which leads to the usability of the model being not as good. And I did do some browsing on Reddit and I did find a really, really interesting post where they do talk about how bad Claude is. But I do want to say that we actually did just drop a complete guide on Claude 2.1 contains absolute everything that you're going to need to know if you do want to use Claw 2. Because I looked around online and there weren't many deep Dive tutorials that give you absolutely everything including every prompt you're ever going to need, how to use it to code, how to use it to make stories from, how to use the API. So that was something that we did do. We'll leave a link to that in the description or you can just find it on the channel as one of our recent uploads. However, with the 200K context windows there is actually less hallucination in this model, so I guess that it is pretty decent for something. So if you do want to use this it's definitely still something that is usable, provided that you do know how to use it. Then of course we have something called Playht 2.0 Turbo. Essentially this is something that I've really wanted from AI for quite some time and I think we might finally just have it. Essentially real time AI conversations are here. So essentially this is something called their conversational AI text to speech model with less than 300 milliseconds latency, which is blazingly fast. So essentially why this is so cool is because not only can you clone any voice in accent, you get something that is really, really fast, which means it has more applications. One of the problems with ElevenLabs, although it does do very well in terms of the coherence of the speech, and we know just how good and accurate the voice cloning is, the problem is, is that it does take quite a bit of time to get that audio from the source. But with this, imagine this in an API, you're going to be able to have real time conversations if you manage to combine this with an LLM like ChatGPT. Very, very quick.

Hello world. Hi, how are you? Hi there. I'm calling in regards to the purchase you made last week. Hello, Play Support speaking. Please hold on a sec, let me just pull up your details real quick. Can you tell me your account, email or your phone number?

I can't wait to see what people are going to do with this because I do believe that this is really the next step because this was a big hurdle and hopefully we do get some kind of model like this from 11 labs where we do get that latency reduced quite a bit. So it has a lot more applications. Then of course we do have something very, very interesting. You can see right here. We can see here that very recently, chat GPT, OpenAI, what they did was they actually trademarked GPT7 and GPT6. And I do find that really, really interesting because prior to this they only really trademarked the large language models just before they came out. But you can see here with GPT4, just before it was released, they managed to trademark it and they did actually trademark GPT5 quite some time ago. But whatever happened recently, I don't know what's been going on at OpenAI. Maybe it was the new cube breakthrough. They thought, you know what, we know exactly what's going to be in GPT6 and in GPT7, we might as well trademark it from now. So I think that that is something that many people didn't pick up on. But you can see here that this is official. This is OpenAI CO. This is the US trademark thing, the trademark and patent filings. So this was something that they recently did file. I'm sure that they're going to get this trademark and yeah, I just find it pretty crazy. Not many people have mentioned that they trademarked this, but it's clear that we're going to be getting GPT6 and GPT7 fairly soon. I'm not sure when because we haven't even got GPT5, but they trademarked that around four months ago. So it was clear that they knew what was going on with GPT5 from then. So that's definitely going to be something that I am looking super forward to then. Of course we have Meta saying today we're sharing two new advances in our generative AI research. These new models deliver exciting results in high quality diffusion based text to video generation and controlled imaging with text instructions. This quality is really surprising and the reason I like this first video so much is because the way that it allows you to actually edit the initial image is going to be something that is game changing many of the times. We often get an image from midjourney or Dall E3 and we don't possess the ability to change certain objects in the image. I mean, you could do it with Photoshop's generative fill, but it doesn't just work as well. So a model like this could be really, really interesting if implemented into things like photoshop or Dall E3. If they manage to do that, then of course they do have their emu video. So essentially it's text to video. It's another text to model video and it does allow for diverse characters and a range of other different things. I think this is really, really cool because it shows us that text to video is moving pretty quickly and it was always one of those things that I thought, yeah, text to image was good, but text to video is not going to catch up because it's so hard. But one thing that I realized is Never underestimate an AI's capability to do something that you never thought it was going to be able to do. Because this technology, although you think that it might slow down, it doesn't seem like it for right now. So this is also some of the examples of EMU Edit where essentially you can see them editing a cat with a pink jacket. And I do want to say that not only is this really good, the accuracy of this is really, really good. Like the accuracy is not just something that you wouldn't use. The accuracy of this is something that I absolutely would use. So it's definitely something that's really, really cool. And I'm pretty sure they do use something with their Segment Anything model, which essentially allows an AI system to accurately get every single object in an image and then accurately identify exactly what that is. So it will be interesting to see how that's done. And you can see. Yeah, it does segment the spacesuit and then of course changes everything about that. So that is something that is really, really cool. And hopefully they do release access to this. So then we had something that was really co from a company called Be My Eyes and essentially they worked with Microsoft to build something that connects with people who are blind or have low vision with sighted volunteers and companies. So essentially you know that many people have vision loss and some people don't have full vision. Essentially Chat GPT now has vision. And what does that mean? That that vision can actually help you with GPT4. So they create this thing called be my AI and essentially you can use that to essentially get detailed description of images and ask it further questions to a dedicated AI assistant. Imagine you're walking through a grocery aisle. Your vision is blurry because that's just how your vision is. One thing that you have to know that some blind people aren't completely blind. Being blind is a spectrum. So there's complete blindness and then there are levels. And I think depending on where you are on that scale, you're going to need some kind of vision assistant. And this is something that's really cool. You can literally take a picture of something, whether it be close, whether it be far, and then of course chat with it. You know, you could even use the voice thing. And then of course say, look what is going on here? And then it could easily tell you exactly what's going on. So one thing that many people don't realize about AI research is that it definitely does help those who are disadvantaged. And this is something that's really wholesome that I'm glad that the development of AI is actually contributing to. Then of course we have Inflection two, the next step up. So essentially this is is the second most capable large language model in the world today. And why this is crazy is because they did tell us that they were going to be training larger and larger models. They were going to be working with more and more data and eventually they want to train something that's far more powerful than GPT4. And you can see that it actually does beat Google's Palm 2. So we have Inflection 1 in the light green. Then we have Inflection 2 in the dark green. You can see that it does surpass Palm 2 on many of the significant benchmarks. So it's really interesting that finally companies are starting to ramp up their capabilities. You can see from the left to the right on this table, you can see that Inflection 2 is at second place, Claude 2 is at third, and then Palm 2 is at fourth. So it's really cool to see how these large language models are managing to evolve on different benchmarks. And it will be interesting to see if any company can manage to catch up to GPT4, because as these large language models have been developed, it seems like with every single one we do get a step closer to GPT4. And essentially we had this tool from CREA AI and essentially it talks about real time AI is here. It says with it a new generation of AI creative tools are coming. And I think this is really cool because it gives us a new way to edit images and a new creative way to create art pieces that we really didn't have before. So you can essentially use a prompt and then just use basic things to fine tune that prompt. And I really like this because a lot of the time we don't get to control where certain items are in that image. We don't get to control the background, the foreground, the colors. I mean, yeah, we do to an extent, but sometimes we have such a specific creative vision in our heads. And I believe a tool like this is going to allow us to really, really get our hands into that final, final creative area where we can truly master and bring anything to life in high quality and bring down the barrier to entry that is creating art, especially one that is as photorealistic as this. And there was something that I did find kind of interesting because Amazon released a AI model called Q. And I don't know if this was just a jab at OpenAI or whatever, because Amazon do exclusively work with, I think they work with Claude or Anthropic, but they don't work with OpenAI. And essentially they called this model Q. I'm not sure if that was done because everyone's talking about Q right now and they wanted to steal the limelight. But essentially this is their large language model. Now we don't actually know what's powering it. If I did have to guess, I would guess that it's some unique version of Claude because people are speculating online as to which model it is. But it does say here that Amazon and Anthropic announced strategic collaboration to advance generative AI. So Anthropic has selected Amazon Web Services as its primary cloud provider and will train and deploy its future foundation models on Amazon Web Services Trainium so essentially Amazon are benefiting from Claude and Claude's benefiting from Amazon. So I would say that this is likely what's behind the chatbot because many people have just forgotten about Claude. But some early experiments by Ethan Mollick essentially show that Q feels like a GPT 3.5 class model with lots of guardrails that is very narrowly focused on a knowledge base. So essentially he did a bunch of small tests on how to use Q and other stuff like that and he finds that it's not that good compared to of course GPT4. But then again this is just something that's just for business. So I'm guessing that this is a specific use case and isn't just a general knowledge chatbot. Likely what we're going to see is these chatbots roll out into Amazon's shopping experience. I don't know when, but definitely very soon.