You Won't Believe OpenAI JUST Said About GPT-5! Microsoft Secret AI, Hallucination Solved, GPT2

So it's an absolutely incredible week in artificial intelligence. Here are some of the major stories that you did miss. One of the key stories that you did miss is this update from OpenAI's Sora. So I was going to make an entire video on this, but I thought I'd include it in this video first. It's pretty incredible on what SORA is able to do because it showcases a different capability. What we can see OpenAI's Sora doing here is that it's able to completely change some aspects of a certain video. So right here, what's really, really incredible is that you can literally just change a single character in a specific scene. And honestly, as I'm watching this back, seeing that this is artificial intelligence, like this is a completely AI generated clip, is literally still mind blowing because my brain can't conceptualize it. Like the small, subtle details of this person walking and this person walking right here are just really, really incredible with how everything just flows through seamlessly. For example, the physics on the jacket, it's just really, really incredible. So the thing is, you could see that they replaced it with a metal industrial robot. You can see they changed this to a 60 year old man with a white beard, then they changed it to a woman, and then they changed it to a 30 year old man in a gray suit. So I think the reason that this is such a big thing is because now we have a situation where you can literally replace characters. And I'm wondering if OpenAI Sora, you're going to be able to use your existing footage because it's a bit vague. This wasn't something that they posted a blog post about. They didn't state like a new release stating that this is a new feature. This was all they shown. So I'm guessing that what potentially you could have is you could have a situation where you literally just have your original footage and then of course you want to replace it with something else. So I'm wondering how that is actually going to impact certain, I guess you could say companies, because I know that there are a lot of companies right now who are in the VFX space and they're working on a few things that are like this, that are really, really effective. So I'm currently wondering how that is actually going to work, because previously I'm not sure how those companies are going to now function. And it's pretty, pretty interesting because not only can you do it with people, you can actually do it with different objects. So previously what we did see with Sora was that if there was a car moving or a stationary like object that's moving, that's the main focus, you could also have that object be doing that as well. So you could replace out like a red Ferrari for like a blue Ferrari, replace it for like a truck. And what's also cool is that what I remember is that you can also replace the background as well. So it's a pretty crazy future because the customization of this is going to be pretty insane. And I think while Sora is pretty crazy, I think the crazy moment is going to be when companies like Runway and companies like Pika Labs finally upgrade their video models to, to something that's you know, on this level I think content creation and the kind of things that you see online that's going to be really interesting at that point. Now if you haven't known already, Microsoft is working on something called Mai1. And Microsoft released a new AI model to compete with Google and OpenAI. It says for the first time since it invested more than $10 billion into OpenAI for exchange for the rights to reuse the startups AI models, Microsoft is training a new in house AI AI model large enough to compete with state of the art models from Google, Anthropic and itself. So this new model is essentially Mai1. It's being overseen by Mustafa Suleiman, the ex Google leader who most recently served at the startup of Inflection before Microsoft hired the majority of the staffs from there and paid 650 million for the rights to its IP in March. And essentially they didn't carry over Inflection but they do have the training data and all of the research efforts from that. So they're probably going to be doing something similar to that. So MA1 is going to be larger than any of the smaller open source models that Microsoft have previously trained. If you don't know what I'm referring to, I'm essentially just referring to the Phi series which are really, really compact and really really effective. And with these large language models, well these are like small, large language models or small language models, the thing with them is that you can now have them on, on device. So with that them being on device, what's crazy about that is the fact that, and essentially with these smaller models the gist of them is pretty much that you can use them on your phone with your own data and you don't need the Internet in order to access an unlimited amount of intelligence. Which means for the future it's going to be really, really good for a lot of different applications. For example, the governments can use them. You know, if you're in certain remote locations where you don't have access to update intern information or you don't have Internet access, it can act like a search engine that you can really use and be deployed in areas where the infrastructure just doesn't exist yet. So it's really important for these kinds of things because it's going to be really, really good. It's just, you know, a really, really important piece of technology. And the crazy thing about this is that it's going to have 500 billion parameters or settings that can be adjusted to determine what models learn during training. And I think 500 billion parameters is pretty crazy because its GPT 3.5 actually has 175 billion parameters. So it's going to be around three times larger than GPT 3.5. So I'm thinking that potentially they could reach the capabilities of GPT4 or it might just be in between them. So it's going to be interesting to see how it rivals other models because this is going to be another model that they've trained. And I'm wondering what the main use purpose is of course going to be and how effective it's going to be. Because if you don't know, llama are training llama 3 which is still improving as they've stated, and it's probably going to be on par with GPT4 and that one is around 400 billion parameters. So I'm guessing that this is a model that's probably going to be on par with GPT4, but probably just a little better by the time it is released. And it says this means that Microsoft is now pursuing a dual trajectory in sorts of AI aiming to develop both small language models that are inexpensive to build into apps that could run on mobile devices alongside larger state of the art models. And it shows that Microsoft's willingness to chart a new path in AI separate from the technology developed by OpenAI, which currently underlies all of the AI co pilot chatbots in Microsoft's products that can automatically spin up emails and analyze documents. So they really haven't stated what the purpose of the new model will serve. It hasn't been determined yet and it would basically just depend on how it performs because I don't think right now making a ChatGPT competitor is the thing to do. I think Claude and Google are already hot on the heels of OpenAI, but if you've been paying any attention to where the actual, I guess you could say focus is going to be next year. Right now, these LLMs are just basic in terms of their capabilities. And when I say that in terms of the usability of the models, like they're not reliable, they hallucinate, they don't have any personal information, there's no agenting information, there's no memory and there's no personalization. So those are a lot of the things that are coming from the next level of models. So it's kind of interesting that Microsoft is doing this. I'm guessing that what they're trying to probably do is just trying to own at least something that's probably really good and that rivals OpenAI and still work with them because they don't just want, you know, OpenAI to just be that company. But it's also pretty weird because, you know, Satya Nadella said we are above them, we are around them, we own all of the rights. So it's kind of this weird relationship that they do have. Either way, I think it's going to be, you know, I don't think it's going to be an open source model, but I think it'll be interesting to see what it's for. Maybe it's going to be on some of their computers. I have no idea at this moment in time. But we do know that it is going to be new and if it's anything that we've seen before, it's probably likely, probably going to surpass Llama 3 in terms of the expectations. So I'm guessing it's going to be some kind of product with that, with Microsoft. So that will be something that I'm actually really excited for. So this Maya one product is going to be something to look out for on the horizon. So then this is where I was actually talking more about Microsoft and I'm talking about how certain models being offline are actually a really, really important thing. So you can see Microsoft launches AI chatbot for spy so it says Air Gapping GPT4 model on a secure network won't prevent it from potentially making things up. But it says Microsoft has introduced a GPT4 based generative AI model designed specifically for US intelligence agencies that operate disconnected from the Internet. According to a Bloomberg report, this allegedly marks the first time that Microsoft has deployed a major LLM in a secure setting designed to allow spy agencies to analyze top secret information without connectivity risks and to allow secure conversations with a chatbot similar to ChatGPT and Microsoft Copilot. But it may also mislead officials if not used properly due to the inherent design limitations of AI LLM. So basically what they have here is they have a new service that you can use. And if you don't know why it's so important for intelligence agencies to work offline, it's because they deal with a lot of sensitive data, which are things that are relating to national security. So it could be certain things for certain cyber security programs that are related to national infrastructure. And if those things exist online, the problem is once something's online, it can be hacked. And you might say that, yes, this is secure because of X, Y and Z, but it only takes one kind of hack in order for the entire grid to be put at some kind of risk, or entire banking and financial systems or whatever government projects that they're working on to be put at risk. And they literally cannot risk that at all. And in order to ensure things are safe, they just keep it offline completely. That way it's literally impossible for any hackers to be able to get that data. So they keep it offline in like a local network. So unless you're actually there and you're plugged into the network, that's the only way you're going to be able to get that data. And these, you know, government agencies are extremely, extremely secret. So this is something that isn't surprising to me because I knew that there was going to be some kind of use case for these LLMs. But I think like they said, one of the key things that, you know, is pretty, pretty tricky about this is that GPT4 can confabulate, which can make up inaccurate summaries, draw inaccurate conclusions, or provide inaccurate informations to its users. So I'm guessing that they're probably going to have to use rag, and, and I hope that. I'm guessing that they're probably Most certainly using GPT4 with some kind of rag on top of it, because I think if they just use it, it would be a little bit risky to do. So. So this is something that shows us that the, you know, the use cases of large language models are increasing in terms of how broad their use cases are. Because a lot of people just think that, oh, it's a cool chatbot, but, like, there are a lot of use cases, you know, bigger than the ones that you might use them for. And I think it's important to realize just how crazy it, it really is and how much this is going to affect us in terms of, you know, just making the world a more efficient place. Now this is where we have one of the key pieces of information for the video. So There was a conversation with OpenAI on what to expect in the next 12 months. And one of the things that he said that really kind of shocked me and this is something that I've spoke about quite a lot because Sam Altman has previously echoed how bad GPT4 is and how I think he's not hyping up GPT5. I just think he completely understands where AI needs to get to and where things are going to be. And the fact that other OpenAI employees are actually stating that show us that GPT5, all these successor systems are going to be so much more incredibly better than what we currently have. And the fact that they're building this into the next product line means that we're truly at that point where I think the delta from GPT 3.5 to GPT 44 will be similar from GPT 4 to GPT 5. So this is all of the stuff he talks about. But I'm going to give you guys a quick rundown on this, at least.

In the Next, call it 12 months. So I think, I think the systems that we use today will be like laughably bad. They are laughably bad. I think we will look back in a year and realize how laughably bad they were. The kind of turn based, like almost kind of Oracle like system that ChatGPT is today is really, I think we don't see it as the long term engagement model for what these systems can do and how they're going to be used. We think that we're going to move toward a world where they're much more capable, they're going to be capable of actually taking on more complex work. They're going to have more of an assistive relationship with users. So there almost be kind of almost like a great teammate that you have kind of working with you on any given problem. And that's going to be a different way of using software. I don't think people really today understand what that looks like and can internalize it easily. But I think a kid born today will look back in 10 years from now if you want to make the 10 year prediction, and it'll be foreign to anyone born today that you can't talk to a computer the way you talk to a friend, the way that you would talk to a teammate, or the way that you would talk to a project collaborator. And I think that's a profound shift that we haven't quite grabbed.

And so it sounds like a big part of that is moving away from the sort of text interface and towards a verbal interface I don't think it.

Just has to be verbal. The amazing thing about these systems are how multimodal they are. Like they can reason visually, they can reason about really extraordinarily complex problems. We talked a little bit on the way here about could you actually have an AI that hits a level of kind of comedic quality?

Can AI be funny?

It can be an AI stand up comedian. I think it can. And so I think we're just scratching the surface on the full kind of set of capabilities that these systems have. And that's going to surprise us.

AI is moving faster than the pace of corporate America. It just is moving. The innovation is faster than companies typically move. How can companies, investors keep up with this unusually and sometimes uncomfortably fast pace?

That's the hard part. That's really what I spend my time thinking about, is how do I help the world calibrate to the pace of change in the research lab and for all of you, for the businesses and partners that we work with, what is a way to implement the technology that is going to be scalable, robust to the future, but also keeps pace with what's happening. And so we're still learning. We think we've got it maybe somewhat figured out. Certainly we learn the more we work with more and more companies. But I think that'll be one of the great challenges of the next few years.

Well, just looking at the innovations of the past year and you know, the SORA text to image, text to video innovation is just amazing. So I cannot even imagine what we're going to be talking about next year at this time. Brad, thanks so much for joining us here today.

But yeah, that was definitely very, very insightful. I think he arguably revealed more than Sam Altman does. And this is Brad Lightcap and He is the COO of OpenAI and he plays a pivotal role in guiding the organization's strategic vision and operational excellence. And yeah, I think it was pretty crazy when he said, you know, these systems are laughably bad, which is not just stating that, you know, they have a few limitations that we find out. He said that they are laughably bad. Which means that, you know, we're going to look back at this time and just see how bad the systems are. So I guess that's a key point. So all of the limitations are going to be gone. And I think one of the main things that you need to think about these AI systems is that like they've said, we're not going to be interacting with them as how we do now, where it's like, you Know, you just have to prompt it, you have to feed it all of this information. You get responses, you edit them and yada yada, yada. But I think it will be more so a situation where it has all of the data on you that you know are from previous conversations, it's able to really understand you and you're able to talk to it and just give it like huge amounts of data, whether it be images, video, whatever. And it's going to be able to reason a lot more effectively. And of course it's probably going to be agentic this time next year, meaning that it's probably going to be able to be doing a lot more things that we haven't really thought about yet. So, for example, you know, if we want it to write an essay for us, it's going to be able to open up a Word document, actually write the essay, save it to a certain file or whatever and email it off and then, you know, check it on whatever app or whatever website and. And I think that that's how the next level of AI is going to be. And that's why I think he's stating that, you know, the frontier is most certainly probably going to be changing in that aspect. But of course, one thing that I'm seeing a lot of people actually talk about is the fact that this is, I guess you could say, quite, you know, investor hypey, as some people are stating. And I don't think it is. I just think that this is the logical step for where AI is going to go next. And I think what this does show us is that we have to, I guess you could say, wait until this product is shipped because if it doesn't live up to the hype, then I guess we will know. And if it does live up to the hype, then I guess we will also know. So I'm really wondering whether or not this is going to live up to the hype, because if it does, it's going to be a really, really good thing. And I think it will fundamentally change the landscape. And I think if you know how OpenAI operate, they operate on first mover advantage and they really do have a giant lead, if you remember, GPT4s from 2022. So I'm betting that they're likely to change the landscape again and then we're likely to see the other big players really trying to catch up to them because whilst image surfaces from like Google and even Claude, they don't even have image generation services and OpenAI every single day and every single week they're shipping minor increments that make ChatGPT just that big little bit, much more better than these other systems. So there was actually this update here from Alembic and it says Alembic debuts hallucination free AI for enterprise data analysis and decision support. So it says. Artificial intelligence startup Alembic today announced it developed a new AI system that claims it completely eliminates the generation of false information that plagues other AI technologies, a problem known as hallucinations. In an exclusive interview with VentureBeat, Alembic Co founder and CEO Thomas Puig revealed that the company is introducing a new AI today and a keynote at the presentation at the Forrester B2B summit and will present again next week at the Gartner CMO Symposium in London. This is the reason I'm actually adding this to the video is because I think if the claims are even somewhat true, the claims are pretty insane. So he says here that the key breakthrough according to Puig is the startup's ability to use AI to identify causal relationships, not just correlations across massive enterprise data sets over times. And, and he says we basically immunized our generative AI from ever hallucinating. It is a deterministic output. It can talk about cause and effect. So they have this entire, you know, confusing thing system here that I'm not going to really get into. But I think if this is true, then it truly does change everything. And this is a video demonstrating Alembic's technology, showing how the AI can analyze the complex data set and make know strategic recommendations. So it's literally, you know, looking at the data and trying to pick certain things. But you know, stating that you got rid of hallucinations is a giant claim. So if they are stating that, I will have to say that they are really, really going to have to show that or else I don't think people are going to believe them because the way how their company is positioned, it doesn't seem like it's possible for that to be true just yet. But I mean, you know, if there's anything I've seen, it is possible for that to be true. So the infamous GPT2 chatbot is back. And a quick refresher. If you don't know what the DPT2 chatbot is, it's a strange model that OpenAI have, I wouldn't say released, but they've released it in the chatbot arena. And the chatbot arena is essentially just a place where people test out models to see how they compare against other models. And right here you can see them side by side, but it's so vague and so confusing that it's honestly quite frustrating because they released something called I'm also a good GPT2 chatbot. And they released I'm a good GPT2 chatbot dot and this is referring back to a tweet that Sam Altman, you know, cryptically posted earlier where he tweeted, you know, I'm a good GPT2 chatbot. And they updated the arena with two very similar models. Now the reason that this is quite frustrating is because there's so much confusion with as to what these models even are. A lot of people, including myself, are kind of struggling to know what they are. There are speculations stating that they're probably fine tuned versions of the original GPT2. Maybe it's an entirely new smaller architecture that has been fine tuned on certain data sets. That means now it's on the level of GPT4. My best bet is that OpenAI are honestly experimenting with something really small. I'm guessing that they're just testing how effective it is. And I think maybe that this is something like a 50B or 70B kind of model, simply because I just think that it's not a huge model like GPT4. I think they found out how to make GPT4 a lot more smaller and a lot more efficient because the nuances between GPT2 and GPT4 aren't stark in differences. However, there are some nice differences where you can see that the model is a little bit better in terms of its outputs. For example, we can see that this person Dmitry, was doing some kind of experiments with GPT2. And we can see here, this is GPT4 Turbo versus GPT2 chatbot. And he asked the Chatbot to draw itself. You can see that GPT2, the GPT2 chatbot actually did a much better job. You can see it actually looks like a little robot with hands and legs and eyes and an actual expression. We can see GPT4 Turbo vs GPT2 draw a burger with the American flag on top. We can see that, you know, this one on the right hand side does actually look a little bit better in terms of the way how the placement is and everything. Then we can see here that we've asked it to draw the Statue of Liberty. I would argue that this looks, it looks pretty cool. But I mean, this one, you can see that it is a little bit better. And then of course, he asked it to do the Eiffel Tower. The one on the right, you can see it is a lot better. Then of course you can see the Pentagon. And I mean, that's honestly my suspicion right now because there is a lot of, I guess you could say, speculation on what these models are. I don't think it's a big model because I don't think OpenAI are working on those big models. I think they're focused on GPT5 and I think they understand that, you know, GPT5 and GPT6 are where some of the major gains are going to come from. So I don't think they're testing GPT5 here as many people have experimented with. However, there was a tweet that states that Sam Altman confirms that the GPT2 chatbot is not GPT 4.5, which many people see as the likely next update GPT4. So I think that that is probably true. Whether or not that article is false or not, I'm not entirely sure because you can't verify anything nowadays. But my best bet is that this is a 70B model that is just as good as GPT4 because we've seen that if llama 3 the 400 billion parameter model is going to be as good as GPT4 and if we can think that maybe OpenAI are, let's say, a year and a half ahead of anyone, I think that we can presume that a 70 billion parameter model could probably be as good as GPT4 with certain fine tuning. You know, I guess that's just my honest opinion. So maybe I'm going to be wrong, but maybe it's Q Star, maybe there's a whole new algorithm. But if it was GPT 4.5, they would literally just state it because if it was a new GPT4 model, they would just state that. So I think it's honestly just a really small model that they've fine tuned to become really, really effective. And I'm guessing that's just what it is. I honestly have no idea. And you can see that other people here think that it's a other secret model. It's something else. It's a Gemini Ultra Pre Flight. Some people are saying it's something else. I mean, and we can see someone here also shares my thing, a much smaller model but GPT4 quality that OpenAI has developed for device inference, which Apple will use to power the new Siri. And I think it really just is a smaller model that they're going to use for on device tasks. Because if we look at what OpenAI is doing, one of the things that the secret Things that they've actually spoken about is that they are in funding talks with Emerson and Thrive and basically they're making a new AI device. So it actually makes sense for them to be making a smaller AI model that runs on device, because if it runs on device and if they can have inference at really, really high speeds, then it's going to be like anything that we've never seen before. So I'm pretty sure that this is what they're working on. I mean, anything else just doesn't make sense. GPT 5 is going to be just so much better. GPT 4.5 probably isn't coming. GPT 2 being fine tuned doesn't really make sense. It just makes sense for it to be a smaller model. I could be completely wrong, but that's my best guess at this point. Another thing, if you don't know, if you don't really pay attention to social media and what's going on in pop culture, you won't have known that it was the Met Gala. But apparently this picture that got 16.3 million views and 300,000 likes and 68,000 retweets was actually AI generated. So there was this image going viral of Katy Perry at the Met Gala. And it's a pretty cool image, it's a pretty cool dress. But the crazy thing is, is that this image is not even real. It doesn't exist. And it goes to show just how crazy artificial intelligence can be when generative AI is used to create images. And exactly what we're seeing here and is part of the problem on why people are stating that, you know, we're going to need a new way to verify images on the Internet. And I guess this is pretty good. So you can see here understanding the source of what we hear and see online. It says we're introducing new tools to help researchers study content authenticity and enduring the Coalition for the Content Provenance and Authenticity Steering Committee. So it says people around the world are embracing generative AI to create and edit images, videos and audio in ways that turbocharge creativity, productivity and learning. And as generated audiovisual content becomes more common, we believe it will be increasingly important basically to understand what's going on. So basically they began adding metadata to all images edited and Created by Dall E3 and they're going to be adding this metadata for Sora. So this is going to be launched broadly. And basically what this is going to be, they're going to use this so that you can essentially have a way to understand if an image is generated by AI or not. And I think this is probably going to become industry standard. So, for example, right now you've got all these open source tools where you can create this image, you can create that image. But I think in the future, in order to operate in certain regions, you're likely going to have to use some form of verification that can be governed by some kind of body so that people can easily identify whether or not an image has been AI generated. Whilst yes, this is hard, I think it's going to be pretty, pretty important for the future of trust and transparency across social media, because if we don't have something like this, then unfortunately it kind of ruins the Internet because people are going to enter an era where anything they see they simply won't believe. And it makes the Internet a very, very unusable place. Because if the Internet is something where everything you think is AI generated and you can't trust anyone and you can't trust anything, it's going to be a real big problem for the future. And the Internet is a great place. It's a place to share information, it's a place to discover and find out what things are going on right now. So I think that this is something that is really important for them to, of course, focus on now. There was actually supposed to be a new product from OpenAI coming out, but OpenAI is actually considering postponing the event at the company's headquarters where they were expected to share updates and show product demonstrations, a person said so they were planning to present information on Thursday about new products, according to a person with direct knowledge. It's unclear exactly what the event was supposed to be about, but they reported that OpenAI is working on the search product, which I've actually talked about in another video, but they apparently have delayed it to Monday. So this is something that the information have spoken about. And Jimmy Apples has says this will be beyond a joke if they do postpone this. Stop these little games and just show your cards. Let it be known that the releases and event was a real thing. They are cats if this doesn't go ahead. So Jimmy Apples, like many of you do know, is a notorious AI leaker. And I think with Google I O around the corner, some people are stating that OpenAI might just wait for Google I O to release their next product, because what they might want to do is they might want to just wait for Google I O they release their AI product and then OpenAI stun the world with their AI product. And then Google's limelight is once again surpassed and, you know, Google once again lose the lead because it is a competition, they are in competition with each other. And this is something that OpenAI does know and something that they are, you know, working on for. Like they're literally all competing for the same kind of customers. So if this does happen, I completely wouldn't be surprised. And this is just something to keep in mind because the Google I O is on the 14th and this was supposed to be scheduled for this week, so maybe they're just literally just going to leave it till after and it'll be interesting to see if that actually does occur. Also, there was granite code released. IBM just released a family of new open code LLMs from 3 billion parameters to 34 billion parameters, trained on 116 programming languages released under Apache 2.0, which means that it's pretty much open source, you can use it for whatever. And you can see here from the benchmarks that it does surpass every other large language model on coding capabilities, such as generating code, explaining code, fixing code, and the average is just a lot better. So this is just a 7 billion parameter model, 8 billion parameter model, and it's pretty, pretty incredible on how effective this model is at completing code. So I'm wondering how crazy this is going to be, because this is going to be something that a lot of people do want to use, especially on sites like Ollama and stuff like that. So it's going to be really interesting to test. Then, of course, the autonomous vehicles are coming. So I do think that in the next 10 to 15 years, driving might be a thing of the past. I think we're probably going to be moving towards those smart cities where all the cars are connected. There's this autonomous grid of completely autonomous cars. And this company, you can see wave AI, they're thrilled to announce they've raised a billion dollars in their fundraising round led by Softbank Group, with contributions from Nvidia and Microsoft. And essentially they're going to have embodied AI products for the automotive industry. And it's going to be interesting, it's going to be a really interesting future because everything's changing due to generative AI. Well, not generative AI in this case. This is a completely different version of AI. But I think it's important to pay attention to this because Waymo has had some success and wave AI in Europe and in the uk, I think it's going to have.