OpenAI's Leaked INTERNAL AI AGENTS,,New Custom OpenAI Models, Google Research, A.I GENERATED GAMES,

So with another crazy week or day in artificial intelligence, this video is going to cover some of the most pressing stories that you most likely did miss. So let's not waste any time. One of the first things that was actually rather fascinating is the, I wouldn't say discovery, but the talk and the speak about the SORA model. Even though the OpenAI Sora model hasn't been released yet, there is still a lot of information surrounding that model that is shrouded in controversy. And one of those things is of course the training data for this model. So if you don't know when these AI models are made, they have vast amounts of training data that they usually scrape from the Internet. And in this article right here, you could see that it says OpenAI training Sora on YouTube videos would violate the platform's rules. In an interview with Bloomberg, YouTube CEO Neil Mohan doesn't confirm whether OpenAI used YouTube videos to train its text to video AI model, but he says it would be a clear violation of the platforms if it did. He states that, and I can't show the clip for copyright reasons, but he says that when a creator uploads their hard work to our platform, they have certain expectations. One of those expectations is that the terms of service is going to be abided by. It does not allow for things like transcripts or video bits to be downloaded. And then the crazy thing as well was that last month and the reason that why this is such, I wouldn't say a huge debate, but the reason I would say that a lot of people were discussing this is because the last time that this was a thing, okay. And there was a interview with Mira Marathi on the inner workings of sora. When asked about the training data, Mira Marathi said that she didn't know where the training data was from. She made a weird face. People have been making memes about it and it's been something that, you know, has just been widely discussed as like, I wouldn't say a shot at OpenAI, but I would say it's something that people are continuing to question because of course, as you know, before with Dall E3 and all these mid journey and you know, previous AI systems, people are wondering where did you get your artwork from? And you know, are you going to be impacting artists with this? Especially with video creators. It's like, you know, when you upload your content, you have a certain rights and you know, things that, you know, other people should abide by. And it's like, you know, if you're OpenAI and you're able to train an entire model of all of this video data. Where was that video data sourced from? So it's quite, quite, quite controversial at the moment. One of the things I don't think OpenAI would do anyways is release where they got the training data from. So I think even if OpenAI was going to, you know, get sued by someone, they would have to like, you know, get access to Sora, try and back test it against certain videos and try and recreate those videos with Sora to see if there was any, I guess you could say, similarities in those videos. And I guess that's probably going to be the only way. Now I don't think Google is petty enough to go after Open AI for this kind of thing, but you never know because as the, as the race heats up, it is really, really crazy. And we know that Elon Musk is currently going after OpenAI for various things. So I guess it will be interesting to see how this thing kind of plays out. Because training data, whilst it isn't the biggest debate, it is a big debate in Sora because the video model is really good and people are wondering how they got it to be. So, so with that being said, let me know what you think about that training data at the back. Next in AI news, we had Cohere introducing Command R, a state of the art retrieval, augmented generation optimized large language model to tackle enterprise grade workloads and speak languages of global businesses. Says our R series model is now available on Microsoft Azure and coming soon to additional cloud providers. You can see currently the benchmarks in RAG are state of the art, which essentially means that this is the best model out there. Now if you don't know what this model is and essentially how it's different from GPT4 for basically this model is something that is, I wouldn't say hallucination free, but it is, you know, it kind of steers into that area where businesses need to work with their private data offline or you know, for privacy reasons and essentially they just want to be able to have business solutions with that data that are really, really effective. So one of the key differentiators between this and GPT4 is that this optimization allows for in depth analysis of documents that enables the model to deliver millions of answers efficiently. And essentially, you know, a comm command R is built to specifically excel in RAG and tool use and RAG enhances the model's ability to access private data to power applications, while the tool use feature actually allows it to interact with external tools and databases, enabling A richer set of behaviors. Now, the performance and scalability of Command R are also pretty, pretty effective. We've seen that it is often more effective sometimes than the base model that GPT series usually has, like GPT 3.5 or GPT 4. It's usually quite a lot better than that. And in addition, there is also very, very competitive pricing. Essentially, Commander is usually cheaper than OpenAI's GPT4 Turbo, making it a really, really attractive option for enterprises looking to deploy AI solutions at scale without, you know, incurring those high, high costs that usually come with calling the model several times. Like I said before, it's basically, you know, basically hallucination three. I mean it's not, but it's pretty much there now. There was actually a very interesting paper that was called mixture of depth paper. And this thread kind of explains why it's so important. And the tldr, which is the too long didn't read is basically that this paper just allows you to use less compute for the same kind of questions. So essentially right here this guy breaks it down. He basically says why Google's DeepMind mixture of depth paper and more generally dynamic compute methods matter. Most of the compute is wasted because not all tokens are as equally hard to predict since for example, you need a lot more compute to figure out what is in the first token after full stop. If you think about it, planning the next sentence is a much harder problem than placing a punctuation. And this is pretty much true. It's kind of like similar to how humans think. And it says in the ideal case, the model should only allocate the exact compute it needs to accurately predict a token. There is long and rich research behind these ideas. And this is what Google DeepMind showed. But with LLM transformers, he said it's analyst to mixture of experts. That was essentially where you have, you know, smaller AI systems that have domain expertise in, for example, maybe math, maybe coding, and then you essentially have that all combined into one giant model and that is able to be very effective at pretty much any task because you have, you know, six smaller experts or eight smaller experts that are fine tuned on that specific area and it's able to do really well. And he basically says that it's pretty similar to that. And essentially it's able to skip layers and in the orange area right here you can see that this is where compute savings are had. And then this is where he talks about the savings are further compounded when paired with mixture of experts. And we're entering an era of scalable compute of LLMs tokens will not have fixed costs. The machine will take the time it needs to think. But massive improvements for both GPU rich and GPU poor. So essentially it's like, you know, certain questions are going to be weighted differently than other questions. And I've always thought that this does make sense because simple questions do like, have simple answers and those are a lot easier to predict. But you know, with a mixture of experts, of course you're saving compute because you're not using the entire LLM, you're just using a subsection of that. And then, you know, with this new thing where you're using a mixture of depths, you can see that, you know, as like the compute, you know, the compute cost is just continually, continually going down. Which is really, really good to see. Because like I said before, what people don't realize is that yes, whilst we do need fastest chips and fast AI systems, I can't remember where I read this ages ago, but it's basically like, I wouldn't say a hive mind, but just if you think about every single, you know, piece of software, every single GPU, all of those things around AI, all of those things are making maybe 5% improvements a month. And because all of these things are improving at small areas, you do get essentially a huge, huge benefit over the long time. And that's why these exponential growths can happen. Because if we now can have a system that, you know, uses less compute, maybe it loses 25%, 10% less, you know, you can now start to do even more things on top of that, find out more solutions. And it's just, it's just overall it's not self improving, but it is something that is compounding, which is really, really good because it now means that you're allowed to get more out in terms of output, in terms of the compute that you actually have. Then we had Microsoft's official blog and we had a breakthrough in quantum computing. I wouldn't say breakthrough, but they've just been getting really, really good at being able to do this. So essentially what we have here is Microsoft and Quantinum have made a big step towards quantum computing. They've basically created the most reliable logical qubit so far. And qubits are the basic units of information in quantum computing, similar to bits in regular computers, but much more powerful. Logical qubits are actually made from multiple physical bits. Logical qubits are actually made from multiple physical qubits and are actually less prone to errors, making them more reliable for complex calculations. And this is Actually pretty important because this development is a big deal, because it actually moves us closer to building supercomputers that can handle reliably complex problems that are much faster than today's computers. And these could include finding new ways to fight climate change, developing new materials, creating more effective drugs. You know, just a crazy, crazy level of discovery could really happen if quantum computing does get solved at the crazy level. Now, what was actually crazy about this as well was that by using a special system developed by Microsoft that checks and fixes errors in quantum calculations, they've managed to run over 14,000 experiments without any mistakes. And this achievement actually demonstrates a significant improvement in the reliability of quantum computers. So I'm actually intrigued to see, you know, what kind of breakthroughs we make in the future, because this kind of breakthrough could lead to a, you know, very, very exponential increase in what is capable because quantum computers are able to do so much more than a classical computer. And it really does just go to show that we really, really could be living in a sci fi world in the future, provided, you know, only a few breakthroughs are met. But it just goes to show that the future might be crazier than we do think. So I'm going to keep an eye out for quantum computing because I think whilst many people think that it's something that is far, far away, I think once we do get certain breakthroughs, it's going to provide us with a whole wave of untapped potential that is just sitting there. Now, this is actually AI news or tech news. And the reason I think this is something that's rather important but pretty hilarious at the same time. So this is a piece of news that I guess went completely viral because it's absolutely insane, because a lot of people did think that on the back end this was AI technology, but it just seemingly wasn't at all. So it says, Amazon ditches just walk out checkout as its grocery stores. It says Amazon Fresh is moving away from a feature of grocery stores where customers could skip checkout altogether. Now, if you don't know this, this is just absolutely crazy. So Amazon Fresh has this store where they piloted it and they said, look, we have this amazing store where all you have to do is you walk in there, you pick up your products, you leave and you will be automatically charged. They were like, it's the first of its kind. It's a technological breakthrough. It's just amazing. It's called just walk Out. So it was a checkout list, grocery stores, and I've used these stores, they were actually Pretty, pretty effective. You'd walk in, you'd beep your Amazon thing so it knows you know, that you're in the store and then as you leave, you don't need to do anything and you just get your groceries. It was something that was really, really effective. However, however, there was something that they revealed recently. It says, okay, that though it seemed completely automated, just Walkout relied on more than a thousand people in India watching and labeling videos to ensure accurate checkouts. And the cashiers were simply moved off site and they watched you as you shopped. And this is something that went completely viral. Initial tweet I saw, I think it had 30 million views or 11 million views, I'm not sure which one it did. But it was simply absolutely incredible. Because this kind of thing, you would think that it was powered by some advanced vision system, potentially, you know, like a vision model in AI, like some kind of, you know, accurate YOLO V9 or V10 or something like that with potentially maybe multiple cameras and multiple sensors. But it seems that they just had a thousand people. Now the reason I'm actually talking about this is because I think that in the future this is going to be really, really, you know, interesting because in the future we won't need a thousand people in another country watching you browse and do whatever. What's going to happen is that you're only going to need an AGI level system. You know, with artificial general intelligence, it makes things like this possible because if checkout, you know, if this checkout list store didn't work simply because they just didn't have the technology, an artificial general intelligence system could actually provide us with the capability to do this. So with that level of, you know, understanding, we could now start to, I wouldn't say predict, but look at how certain stores might actually operate this, because this is why I included this story. Because let's say like a thousand, you know, Indians or whatever, or a thousand people in another country, we know that currently, okay, paying a thousand people. Whilst yes, paying a thousand people in another country probably isn't as expensive as paying someone in America because the wages are significantly lower in those countries, I still think it's going to be a lot more expensive than just running a few AGI systems in a few vision model cameras and just making that more effective. And the reason I'm saying that is because it now gives us a look into the future with AGI, multiple different things that weren't possible before are going to be possible. Just like, you know, self driving, you do need an AGI system to be able to take a look at things and not just like a vision model and a text model. You're going to need like a AGI level system that is able to, you know, interpret multiple different things, have really good reasoning and come to a great response. And I think what this shows us as well, like if you have AGI, I think this might be the future of stores because, you know, if you have a bunch of people, you know, of course you're going to have to people that are going to need to pack the, pack the food, you know, undo the delivery trucks, all of that stuff. Yeah, you're still going to need people. But in terms of, you know, cashiers, people doing the checkouts and stuff like that, I think that might be largely automated. And these stores actually did run pretty, pretty effectively. They did have different people. But it just goes to show that in certain cases, you know, your role isn't going to be completely gone, but you definitely will be changing in that role. So people who were previously cashiers, they were probably just people who helped around, help people find things in the store and helped, you know, load and unload the store in terms of certain things and just keep it running and make sure that everything's operational. But of course that could be disappearing as well. But with AGI we never know because it's largely a singularity. Now Another thing that OpenAI have actually been doing is they've been, you know, training custom models for a variety of different use cases. And I think this is something rather fascinating because it gives us an insight into what OpenAI is doing on the back end. They are doing a lot of stuff and they have been hiring a significant amount of people. I mean, rumors have it that, remember when they had 774 employees, apparently they've hired as much as 500 people in offices around the globe. So it seems that OpenAI is trying to become a behemoth in the AI space. And one of the ways that they are doing that is by partnering with different people in different sectors and building custom trained models for different professions. Right here we've seen that Harvey partners with OpenAI to create a model for legal professionals. And essentially right here, if I just scroll down to this, this is going to show you exactly what's going on. So the prompt is what is a claim of disloyalty? And this is something that is different compared to GPT4. You can see the custom trained model was preferred 97% of the time. And this is a custom model that was trained on a bunch of legal data in terms of the way that it is formatted and, and in terms of the sources that it cites. And we can see that GPT4 has a very standard response to this. So what OpenAI are doing with a lot of companies, which is rather fascinating, is that they're building these AI systems that are fine tuned and custom like custom models that are going to be a lot more effective than we initially thought. Now I do wonder how this is going to work once GPT5 comes out. Are they going to have to retrain it again or are they just going to have to do it simply with the same process? But I think this goes to show that OpenAI isn't just trying to build the next GPT5 or GPT6. They are really trying to, I guess you could say, as I've said before, come after the entire stack in terms of what's possible with AI. They do want to be pretty much everywhere. And it's clear now that OpenAI is going to be a major player in the space. They aren't just going to be a company that has text, they aren't just going to be a company that has a vision. They are truly going after everything. And I think if OpenAI currently, since they have the market leader position, if they're able to continually train models like this and go after certain sectors, whether it be in healthcare, in legal, in accounting, they could really be like, I don't even know, a trillion dollar valuation. I mean, it could be absolutely insane the amount of economic output this company could get, because I mean, companies are already using GPT4 and if there is, you know, a custom version like this, all of that data is going to be, you know, going back to OpenAI. All of that revenue, not all of it, but like a lot of the revenue, the amount of, you know, calls are going to be going back to OpenAI's revenue. So I think it's going to be very, very fascinating. And what's crazy is that they said, what's next? Okay, what's Harvey tackling next? They said, of course, one of their main focuses is agents, or how to comb multiple model calls together into a single working output. This would simplify the user experience and reduce the amount of prompt engineering and typing users need to do. The vision is for Harvey to serve as a supportive member of the team because, of course, the volume of legal work is growing and associates spent countless hours on complex but routine tasks. The opportunity we have, not just with legal, but with all professional services, is to take care of the routine tasks so professionals can focus their time on client interaction. And this is rather true. A lot of legal work is just quite boring, just sifting through files, looking through, you know, I guess you could say sources and citations for certain cases. But yeah, it's very, very, very fascinating. But it seems that they've managed to figure out a solution and this is not the only customer solution that you know, OpenAI has on their blog. They have so many different things. I will show you that, that in a moment but, but it's absolutely incredible what OpenAI, like I said before, is doing on the backend and on another one right here. This Twitter user points out In a recent OpenAI customer stories post on their site, they hinted at GPT based agents soon taking basic actions online and in the real world. You can see that this is a company that is mainly focused on the future of email and they state that what's next? Superhuman imagines a world where GPT based agents will soon be able to filter, triage and respond to email automatically, scheduling meetings and appointments and taking basic actions online in the real world. The success of Instant reply proves this agent led email future is not too far off. And we know that OpenAI, their focus right now is of course AI agents. So the companies that are working with them stating that this is their next focus isn't a surprise at all. Now the next thing that we have here is of course some speculation, but I think it's pretty, pretty wild. I won't be making an entire video on this, but it's something that you should be paying attention to. So essentially there was some old domains from OpenAI. This is Jimmy Apple's an infamous OpenAI leaker and he says curious to see some people not understand all the testing, checkpoints, models and research that OpenAI and others do. We don't have X, we call it Y, model C hasn't finished sitting on the last checkpoint, etc. Their old domains give you an idea and you can basically look through their old domains to see where certain of course systems were in their phase for their training, their red teaming or whatever. Now essentially there were these domains that were to interest of some people because they said very interesting things. So for example, we saw some of these domains that said evaluation, leaderboard site.