HUGE AI NEWS : MAJOR BREAKTHROUGH!, 2x Faster Inference Than GROQ, 3 NEW GEMINI Models!

So with some incredible announcements this week that I genuinely can't even believe today I'm going to be showing you guys all of the stories you missed that happened earlier this week. So now, quickly for this story, and I'm not going to gloss over this for too long because I already made a video on this, but essentially Google Research actually released this paper where they said diffusion models are real time game engines. Now the reason that this research went so viral is because it's essentially having an AI system neurally render the gameplay in real time. So what you're seeing isn't like gameplay, you're seeing someone play and the AI is rendering the gameplay as it happens, which is, you know, absolutely incredible because this is the kind of technology that people thought was sci fi. So it's something that is just completely insane considering the amount of research going to this and what people had thought. So this is something that's really strange. I mean, I guess people, you know, thought that this was going to take quite a lot of longer time. And of course the main thing about this, which is why it's so crazy is is of course the implications for this. So this is something that is rather fascinating because it means that there are potentially simulated worlds going to be coming in the future and this is going to be something that many people are going to be exploring. So it's going to be really interesting to see where this research actually leads to for the future. Now, probably the most insane news from this week, and it's debatable because there is another story that's also incredible is that Magic Labs, that lab that I spoke about a while back, apparently they have had a recent breakthrough. So, so you can see right here that they stated, okay, LTM 2 mini is our first model with a hundred million token context window. That's 10 million lines of code or 750 novels for reference. If you want to know how long 100 million tokens is, it's unfathomably long. Because right now what we do have on the Google API, we have Google Gemini, which is a million context length and we have it up to 2 million. But you don't even have most use cases for that long. I genuinely can't even find use cases on my day to day where I need 2 million context length. And of course there are some use cases which I will mention in a moment. But the point is that that is absolutely insane. Guys. I don't think you guys understand how much 100 million tokens are. And I do know that Google were researching 10 million token context length windows for reference. If we look at the things that GPT4 has, I remember first when we were looking at these models, we were like, okay, this is pretty crazy. It's got 32,000 tokens. I can't believe this. This is like 16 times more than what we've had before. Four times more, whatever. And then I remember when Anthropic came out with their early research that they were looking at 100,000 token context length and we were like, wow, this is incredible. Now guys, the fact that we are already now at 100 million tokens is insane. Now the report is pretty crazy. I'm going to show you guys why this is so crazy. Because the thing about this is that anyone can come out and say, okay, we have 100 million token context length. You can put in 750 long novels in there and it's going to be able to read it. But the real thing is that they talk about this problem. And the problem that they talk about is the fact that current long context evaluations aren't great. So basically what they state here is that the popular needle in a haystack evaluation places a random fact in the middle of a long context window and then asks the model to retrieve the fact. So basically you have like, you know, the Harry Potter trilogy. I don't even know if it's a trilogy. Let's just think of a really, really long novel. You put that into the model and then you'll put a random word somewhere in that novel and then you'll ask the model to find that random word. And then of course, when it does, you're like, okay, it can actually, you know, use all of its context length. But they basically say that this isn't good because of the fact that it stands out in a fiction novel about wheels. Like if you have something random, it's going to stand out and these models can recognize these needles. So they're basically saying that what they've done here is they've designed something that's even harder, okay? And they've designed something called the hash hop. Okay? So it says hashes are random and thus incompressible, requiring the model to to be able to store and retrieve the maximum possible information for a given context size at all times. So basically what the model has to do is the model has to actually understand. Well, not really understand, but actually retrieve what's going on and not just realize, hey, something strange is going on in this piece of context length. So you can see right here, we prompt a model trained on hashes with hash pairs, you can see that they've got all of these pairs. And then we ask it to complete the value of the randomly selected hash pairs. So they've got like insane amounts of these. And of course it's going to have to randomly select a random hash pair, which is remarkable because it's like there's no real way to identify which one is going to be picked. All of these are completely random. Of course, you could say maybe there's some kind of pattern here, but they're randomly generated. So this is going to be insane. Like, this is going to be absolutely insane in terms of the fact that when we look at the results of testing this against the this, you know, 100 million context length guys, it is pretty much perfect up until 32 million context length. Like the fact that, you know, even until, you know, you can see right here that until you get to 100 million context length, that's when it starts dropping. The fact that research in AI now has got us to the point where basically we're up to 32 to 50 million context length in tokens and it can retrieve that data accurately just shows you how insane things are moving in terms of the speed. And I remember earlier this year that I was talking about how this company is going to be onto something because like I said before, if you aren't familiar, this company had a breakthrough earlier this year and basically they showed it to investors and the investors were like, okay, this is insane. And the investors didn't even need a pitch deck. They literally just put $100 million. So, yeah, as I was saying before I was interrupted, this is pretty incredible because it goes to show how fast the research is moving and the fact that, you know, this company are having breakthroughs and doing this kind of thing, it just goes to show where we might be in the future. Now, I think this is going to be pretty crazy in terms of the implications for everyone else, because you have to understand that whilst, yes, we don't actually have access to this yet, and I know many people in the comment section are going to be like, okay, another fancy demo without any actual results. But the point here is that these implications are pretty insane because what we have is a situation where if these guys are basically saying that, look, we've got our system to 32, basically 100 million token context length, okay, which is just unheard of at this point, other labs and other models are going to have to facilitate this kind of token context length or they're going to have to innovate in some way, which once again is going to just drive that innovation. Because every single month there are new companies popping up, there are new methods of innovation, and there are new things that we really just didn't see. So it's crazy that this kind of thing is happening all the time. And I wanted to bring this to you guys attention because I can't imagine where we're going to be in five, 10 years from now in terms of AI, because the way how things are advancing this rapidly, it seems that we're moving quicker than anyone anticipated. And one thing I did forget to add was that this is of course the coding platform. So basically they're trying to develop these software agents. That's why they're so focused on having such a long context window. Because if you have a model that can literally ingest all of your code and then you can use natural language to fix your code and it can understand your code, it's going to be pretty insane. So I really can't wait for this to be released because I think the implications are going to be staggering. Now of course you could also see that Sam Altman did finally tweet about the fact that they've reached an agreement with the US AI Safety Institute for pre release testing of their future models. And it says, for many reasons we think that it's important that this happens at the national level. So the United States needs to continue to lead. Now I think this is going to be rather fascinating because the implications of this, I don't think they're that bad, but I think it's rather fascinating to the point where OpenAI has now gotten to the point where they're going to be testing these models with the US AI Safety Institute in terms of just trying to ensure safety. Now I do wonder how this is really going to work. Is OpenAI going to immediately give them raw access? Is it just going to be certain video models? Is it going to be every single model? I mean, it's going to be really interesting to see how this entire thing works because of course AI is something that evolves faster than regulations do. And I do want to understand as well, like how long is the AI Safety Institute going to be testing these models for? Because some people are speculating that this might slow down the release cycle even further. Of course, when you do have a model, you have the initial thoughts about the model, how you're going to train the model. Of course you've got the data collection of the model, then of course you've got the training of the model, then you've got the post training of the model. Then you've got six months of safety testing of that model where you have red teamers, you have individuals, of course you've got human feedback of the model, you've got just all of these different steps. And now apparently we're going to get another step on top of that, which is where we've got the US AI Safety Institute for pre release testing. So I'm wondering is this just going to be for certain more advanced models or is this just going to be for consumer products? Because like I said before, I think OpenAI, in the future they're going to have a lot of products that they're just simply not going to release and they're going to be super advanced but they're just not going to release them to the public because number one they're going to get backlash and number two if those models get jailbroken, there's going to be severe implications. So I am wondering about how that is going to be. I'm wondering if there's going to be some kind of transparency with that. But the future is going to be rather fascinating, really, really fascinating to see how the development space changes because one thing that I've realized as well is that there are a lot of people that are now, I wouldn't say hating on OpenAI, but they're becoming less and less patient because OpenAI has demoed a lot of things and they just simply haven't released them at all. So I mean, I wonder if OpenAI's image from that aspect is also going to take some kind of hits there. But it will be fascinating to see how things move forward. Then of course we did have Bland AI and I wasn't going to make a video on this, but this is something that I've been sent time and time again. It's something that shows us the future of AI employees and, and this can talk in any language or voice, be designed for any use cases and handle millions of calls 24 7. This is why in my post age economics community one of the things I've always said, you know, that people are going to struggle with in terms of a future career is jobs that are on the phone. So I'm going to show you guys exactly what this video is and then I'm going to give you guys my.

Thoughts and opinions of people that spoke to our AI. We're going to introduce you.

Hi, is this Isaiah? I heard you want to show me off in a demo?

Yeah. Hey Bland, I hope you're Having a great day. I just wanted you to explain to our audience how you work.

I'm having a great day. And yes, Isaiah, I'm an AI agent for enterprises. I can handle millions of phone calls simultaneously, day or night, and in any language or voice. I can be used for any use case in any industry across sales, operations and of course, customer support.

That sounds pretty awesome. Can you talk a little bit about what actually makes you special?

The handle conversation in an intelligent and hallucination proof way. I do this through an invention from our team called Conversational Pathways. It's the tree of prompts you're seeing on screen right now and it's how I navigate a conversation.

That sounds pretty amazing. And what kind of integrations does Conversational Pathways support?

Pretty much anything. I can be connected to a calendar, to schedule meetings, update a CRM and I can even accept payments over the phone. Do you want me to go into more detail?

No, I think that's enough for now. We'll let people talk with you on their own on our website. Thanks Bland. Have a great day.

Thanks Isaiah. You have a good day.

There are billions of phone calls made every single day. In the near future, almost all of them will be done with AI. That means no more waiting on hold, no more being endlessly transferred around and hundreds of billions of dollars generated and saved for enterprises. If you haven't already talked to Bland, you can go to our website, Bland AI and have a conversation with it where you can fill out our enterprise form and start automating your phone calls today.

Still hiring humans.

So yeah, I think this is another interesting aspect of how the economy is going to change. And what we can see here is a direct impact of AI impacting the workforce. I'm not sure if you work in, you know, customer service or whatnot. I've always said that, you know, certain careers, for example, have to move towards other adjacent opportunities that aren't as easy. So for example, one of the things that I know that you could do, for example, if this is in your industry, if you're, if you're an example like customer service and you're used to talking on the phone, what you could do is you could move to sales. And sales is something that you don't really want an AI there because there are just so many different questions that you do have that just don't always pop up. And the thing is as well that right here, what we can see, and hats off to the technology because it is pretty great and there's no surprise about this, but you can see that there is this tree. Okay, so there's a tree of. There's basically a tree of things that could happen. But when you have certain sales, every single person is going to be different in terms of their obligations. And the problem is that if you're using an AI to speak to someone, especially if you're in high ticket sales or certain kind of sales, it's going to be hard to get that person to buy. And of course in the future, maybe it's going to be overcome. But there are just certain things that, like for example, when you're talking to someone in sales, sometimes you might be from the same country as that person, you might have the same culture as that person. You, you might know, like an inside joke from a TV show from that person. I don't know. All of these things sound crazy, but it's going to be really hard to get that human aspect because the entire thing is that with sales, what you're trying to do as well is you're trying to build up as much rapport in a short amount of time as possible so that you can get that sale. So I think those kind of careers are going to become more attractive. But this kind of technology, I'm going to kind of see how this works. Because I do remember that there was also a law that was recently passed that basically said, look, and I don't think the law was passed just yet, but it was basically moving forward that basically says, look, we need to ensure that individuals actually can talk to a human if they want to talk to a human. But I do think that even though that law might exist in the future where you have to talk to a human agent, I still think that, you know, in the future this is going to be the future of how most conversations are going to be had, especially during phone calls. I don't know about you guys, but I've been waiting on hold a millions of times before and it's just completely tedious and completely fake frustrating for something that's really simple to do. I wanted to show you guys this clip because I have an article that, you know, follows up to this. So I'm going to show you guys this clip first and then I'm going to talk about the article.

Even with on prompting techniques, I mean, it's fascinating even with no new models coming out, right? Given. Given a fixed model, you can elicit better and better performance of it from it simply by improving how you prompt it. And there is a paper that came out three or four months ago that suggested that like emotional manipulation of the large language model would get better results. So the, the kind of the prompt suffix that they figured out was you. You say, hey, I, I need you to perform this task. You define the steps and so on, and you end with, it's very important to my career that you get this right. And the performance goes up. You're like, what is this? Like, what are computers now?

So that's hilarious, not gonna lie. But there's also this. Okay, so this is Apple Intelligence. And the reason I wanted to include that, because I wanted to include this in the video, but I didn't know where, and since that other previous clip had a decent amount of attention, I wanted to show you guys that Apple Intelligence, not only are they having an announcement later this year, which should be Apple's true AI announcement, what they also do that testers have found that their prompts are meant to keep Apple Intelligence on the rails. So basically what Apple Intelligence have done here, what Apple have done basically is, is that they've actually said to the AI, do not hallucinate. So in the system prompt for the AI, they've said, do not hallucinate. And the craziest thing about this is that someone was digging around on Reddit and they managed to find this. And I find it incredible that like, you know, people have been trying to find out ways to get these systems to not hallucinate. And one of the ways that, you know, Apple have been, you know, getting the system to not hallucinate and to be more correct is that they, they literally just say, do not hallucinate. I mean, it's kind of crazy. Like it just kind of, it doesn't defy all logic, but it's something that's so weird that shows us that we don't really understand these LLMs.

Even though the training of a machine learning system might be. One might have imagined that even through the training of a machine learning, even though the training of a machine learning system might be circuitous, somehow, in the end the system would do what it does through some kind of identifiable and explainable mechanism. But what we'll see is that in fact, that's typically not at all what happens. Instead, it looks much more as if the training managers to home in on some wild computation that just happens to achieve the right results. Machine learning, it seems, isn't building structured mechanisms. Rather, it's basically just sampling from the typical complexity one sees in the computational universe, picking out pieces whose behavior turn out to overlap, what's needed and in a sense, therefore the possibility of Machine learning is ultimately yet another consequence of my favorite phenomenon of computational irreducibility. So why is that? Well, it's only because of computational irreducibility that there's all that richness in the computational universe. And more than that, it's because of computational irreducibility that things end up being effectively random enough that the adaptive process of training a machine learning system can reach success without getting stuck. But the presence of computational irreducibility also has another important implication. That even though we can expect to find limited pockets of computational reducibility, we can't expect kind of a general narrative explanation of what a machine learning system does. In other words, there won't be a traditional, say, mathematical general science of machine learning, or for that matter, probably also neuroscience. Instead, the story will be much closer to the fundamentally computational new kind of science that I've explored for so long, and that has brought us our physics project and the ruliad. So, in many ways, the problem of machine learning is a version of the general problem of adaptive evolution as encountered, for example, in biology. In biology, we typically imagine that we want to adaptively optimize some overall fitness of a system. In machine learning, we typically try to adaptively train a system to make it align with certain goals or behaviors, most often defined by examples. And yes, in practice, this is often done by trying to minimize a quantity normally called the loss.

So that is Stephen Wolfram saying that there might not be any way to actually identify the mechanisms behind deep learning, because, like in neuroscience and biology, it might depend on wild computation and might be the outcome of computationally irreducible evolutionary adaptation. Till they are, it's just basically too complex for humans to ever understand. Now in some actual insane news, okay? And I can't even believe this happened. In fact, I actually knew this was coming, but this is actually sooner than I thought. So, introducing cerebras inference. Okay, so for those of you that thought GROK was quick, for those of you that thought GROK was quick, Cerebris inference is basically twice the speed of Grok, guys, twice the speed of Grok. Like, you can see 800 tokens in. In a second. That is absolutely insane, guys. Like, like, it's just crazy. Like, I'm literally, you know, speechless at this. So you can see instant code generation. You can see right there. I'm going to show you guys it was. Okay, I thought it did say 0.8 seconds, but that is really, really quick. Guys. I'm not even entirely sure, like, you know, how this is going to work in the future in terms of how much data LLMs are going to be able to output. But this is some insane stuff. You can see right here that if we actually take a look at this compared to some of the other proper popular providers, you can see that azure is on 20 seconds per user. Perplexity is at 52, Amazon Web Services at 50. Recently we had Grok, which is at 250, but we've got Cerebras which is nearly double the speed. Okay, like Grok was insanely fast, but now we have something that is basically double the speed. So that is insane. Now the reason that this is going to be crazy is because the applications for this are really good. So we could have super smart models that are also quite fast because of course there's been that discussion about test time, compute, putting inference towards figuring out what the problem is. But I mean, imagine actually getting incredible answers in just a fraction of the time. One of the things people have always discussed is how we need these lightweight models to give us quick responses. But what if we could get superior models with even superior latency? So you can see right here that says, gives me a list of computer scientists, tell me what each one's greatest achievement was. And then they've tweeted and basically said that this is basically going from like dial up Internet to, you know, gigabytes per second Internet. So it's like you remember when you had slow Internet as a child, not sure if some of you grew up with dial up Internet, how slow it was. And then, okay, as we got older, as technology improved, we got Internet that was just super fast. And I mean that literally for like the LLM that was literally like loading a web page. Like could you guys see that? Like 1800 tokens in a second, guys, that's insane. And then on the right hand side we can see these models generate the token generating the tokens. I mean, of course this is just something that I didn't even think was going to come soon enough. But I mean, can you imagine in the future when we get further inference improvements and things are basically automatic. So I mean it's incredible, like literally incredible stuff going on this week. And then we have Dario amade saying the AI could increase the rate of discovery in biology by 100x compressing the total amount of progress that occurred in the 20th and 21st centuries to just a few years. Now this has staggering implications for everyone because this basically means that this stuff could, could be applied to longevity research, could be applied to certain terminal illnesses that we thought were once terminal. I mean, you know, compressing decades of research into just a few years means that we're going to unlock a lot of information in a very short amount of time.

If we think about what's really advanced biology, like, it's really disproportionately a few technologies that kind of power every. Right. So like genome sequencing, right. Just the ability to, like, read the genome. Fundamental to like, you know, most of modern biology. Right. More recently, crispr, the ability to edit the genome. Right. Fundamental to many experiments where we want to intervene in animal experiments starting to become important to, you know, pharmaceuticals and, you know, and curing diseases. Although there's a while to go and, you know, it needs to be more reliable. There are a lot of other techniques that are needed. I think if we get AI, right, it could like increase by 10x, maybe 100x the rate at which we invent these discoveries. So if you look at crispr, you know, there was no reason you couldn't have done it 30 years ago. It took 30 years to invent it. And so I think if we can, you know, greatly increase the rate at which these discoveries happen, we'll also greatly increase the rate at which we cure disease. And you know, the way, the way I think about it is, you know, could. Could we have like a kind of a compressed 21st century, right? Could we make all the progress in biology and that we were going to make in the 21st century using, you know, AI by, you know, kind of speeding things up 10x? And if you think of all the progress we made in biology in the 20th century and then, you know, kind of compress that into. Into five or ten years to me, to me, that's the upside. Like, I think, I think that might literally be possible and you know, diseases that have been with us for millennia now.

Lastly, but not leastly, Google actually did release three new products slash AI models. So they introduced a new, smaller variant of Gemini 1.5 Flash, an 8 billion model, a stronger Gemini 1.5 Pro model, which is better on coding and complex prompts, and a significantly improved Gemini 1.5 flash model. I've used these models, the Gemini 1.5 Pro models, the advanced ones, and I got to be honest, they are actually really good at breaking down some prompts. Interestingly enough, sometimes when I've had some questions that I just haven't been able to get right outputs from either Claude 3.5 Sonnet and GPT4. Oh, Gemini 1.5 Pro actually does present the right answer in a really detailed way. So it's a bit weird because sometimes it's a really nice hit and sometimes it just completely misses and just doesn't get things right. So I would say if you're someone, you know, you're struggling with an issue, you know, Claude can't solve it, you know, GPT4O can't solve it, just log into the API and I'm going to show you guys right now. So you just click this and then you just wait a second or two and then basically just come over to here. And they have two different experimental ones. They got the experiment 01 and then they've got experiment 0827. And so the most recent one is 827 and the one before that was 821. But it's still pretty useful to just see if your issue can be solved. And you actually do get 2 million tokens in this chat area, so it's really useful to try. So I would say go ahead and use it because it's something that you're going to want to try.