HUGE AI NEWS : Strawberry/QStar RELEASE DATE, Reflection 70B EXPOSED, New AI Agent SOTA

So with a crazy week in AI in terms of news, let's take a look at the stories that you might have missed. Now one of the things that I think most people did miss and even I missed this was this crazy, crazy technology. So there's this new AI software from Halu AI and essentially this is absolutely incredible. So you can see here that this is a new text to video software. But the prompt and the video that comes out of it is absolutely incredible. So it says an over the shoulder shot of a woman close up. At first she is laughing, then she becomes sad. Sad. Then she starts to cry and then she covers her faith with her hand. Now I think this example is absolutely insane because this shows us exactly the kind of quality in terms of AI. Like, I mean that's just wild. Like she's laughing and then we see that her emotion changes and she actually looks really sad and she puts her hands on her head. I mean, this is absolutely insane because I'm not sure if they upgraded the model or whatnot, but this model wasn't as good as it is now. What when I last tested the model. So this is a video model which is text to video, which is rather fascinating because not only is this video doing the rounds on social media in terms of how realistic it is, we also have this video on Reddit that's actually showcasing just how good this tool is. Now I do want to state that I think the best thing that this model does do is that this model actually does clips at the current rate. Now you might be thinking, what do you mean at the current rate? Well, currently, for some reason, a lot of these text to video models, when they output videos, they're usually in slow motion. And this is largely one of the first times that I've seen an AI model actually use, you know, normal speed. And it's kind of weird, but it adds to the level of realism that we do see from these clips. And I do have to say, when I've been looking at these clips on social media, I truly don't think it's going to be that long to before we get full fledged AI movies and AI TV shows. And whilst yes, that might not mean entire industries are going to collapse, I do think that it's going to mean that, you know, if you're someone who's an aspiring creator, filmmaker, maybe you've always wanted to make your own TV show, but you've had a limited budget. I'm sure something like this is going to be something that can help you flesh out an idea. Or even validate the proof of concept. So this is going to be something that's really cool. And I do know as well that I'm thinking of the back end of this. And I do think, like many TV shows where they ended on a strange season or other things like that, people might just be even picking up these TV shows in their own way. I mean, it's going to be a really fascinating future. And I think this is one of the pieces of technology where we can really start to see that, you know, the future is going to be super uncanny and many industries are going to be ripe for disruption, which is fascinating to say the least. Now, there was also this information regarding the reflection 70b model. Now, if you aren't familiar with the 70b model, I do want to cover this quickly because I did make a video about that model. And essentially in that video they basically state that we fine tuned a version of Llama 3.1, 70 billion parameters to basically have chain of thought combined with reasoning steps, and it basically smashes everything. Now, there's information on the Internet that kind of suggests that reflection the API is just a Sonnet 3.5 wrapper with a simple prompt, and they are currently disguising it by filtering out the string claude. So essentially you can see here that when this individual says write the word Claude, use plain text, no tags. You can see that when it says to write the word, the word CLAUDE disappears. Now, there's been a lot of criticism surrounding this model because many people haven't been able to replicate the benchmark results. And this has been. So now there was some recent information regarding this. And of course, like I said before, there were many different tweets that have been, you know, pretty much debunking the reflection 70 billion parameters. And recently you can see Matt Schumer has come out here and said, I got ahead of myself when I announced this project, and I am sorry, that was not my intention. I made a decision to ship this new approach based on the information that we had at the moment. I know many of you are excited about the potential for this and are now skept. Nobody is more excited about the potential for this approach than I am. For the moment, we have a team working tirelessly to understand what happened and we'll determine how to proceed once we get to the bottom of it. Once we have all the facts, we will continue to be transparent with the community about what happened in the next steps. I'm mostly not trying to jump to conclusions here, but, you know, there's a lot of speculation going on the Internet about whether Matt was just trying to, you know, hype up his investment in Glaive AI, because he actually built the entire thing and hosted it with Glaive. So some people were saying that, you know, he's got like a personal investment in that area, so it was basically just good PR for him. And, I mean, when you think about it, it did get a lot of attention. Currently it seems like the post is deleted, but it did actually have over 9 million views. So currently I'm just going to await more details until we know the facts. But I got to be honest, it doesn't look completely good if you're, you know, saying all of these benchmarks, you come out with them and then everyone tests them and then they can't manage to repeat results, and then it was undisclosed that you're invested in that platform. It is a little bit shady, but I'm not, you know, writing completely off Matt Schumer. I'm just waiting for more information regarding the story because, you know, he. They do say here that, you know, at no point was I running any models from other providers as the API, and these guys basically are saying, you know, that there's a lot of work to do. So, I mean, it will be interesting to see how the full situation pans out. But I will say that, you know, your reputation is something that, you know, you have to try and maintain, and if this is something that turns out to be completely false, unfortunately, the reputation is going to be quite damaged considering the fact that these were big claims and everyone was truly excited for this kind of announcement. Now, in probably the biggest news of the video, and trust me, there's a lot, we actually have news that OpenAI Strawberry is actually going to be coming within the next two weeks. So this is actually rather surprising considering the fact that OpenAI has been delaying release after release, product after product, and pretty much keeping everything behind a waitlist besides GPT4O, which even that model people still don't have access to the advanced voice mode. So you can see here that it says strawberries. You know, OpenAI's reasoning focused AI is coming sooner than we thought. OpenAI plans to release Strawberry as part of its ChatGPT service in the next two weeks, earlier than the original fall timeline we had recently reported said to people who have tested out the model. Release timelines are always subject to change, of course, but we have a few other new details about the product. We should explain that while Strawberry is a part of ChatGPT, it's a standalone offering. So that is going to be rather interesting because it seems that it's going to be kind of like this separate service that you can use and I'm wondering if it's going to be included in, in the subscription. So you can see right here it says exactly how it will be offered is unclear. One option is for Strawberry to be included in the dropdown of menu of AI models customers can pick from to Power ChatGPT. And it's quite different to the regular service with some advantages and shortcomings. Of course, what differentiates Strawberry from other conventional AI is in its ability to think before responding, but rather than immediately answering a query, the thinking model, the thinking stage usually lasts 10 to 20 seconds. So I think this is going to be not a revolution, but a rather interesting element of the model because we're going to get a model that's potentially, well, largely going to be much smarter, but people are going to have to wait a lot longer for their responses. Now, I do wonder if this is going to be kept in terms of the output token context length, like if this model is going to be able to code certain things or if this model is going to be able to predict certain things. Things, because when you think about the kinds of reasoning that we're going to be using a Strawberry kind of model for, it's going to be things that involve many different variables and many different scenarios. So I think the main thing that I'm going to be most excited about with regards to this model, the one that takes 10 to 20 seconds, which when you think about it, compared to how long current models take to respond, that is quite the eternity in AI time. So that's going to be rather fascinating to see if that's going to be something that can truly have incredible intelligence compared to other models. And it says. But there are other key differences. For one thing, the initial version will only be able to take in and produce text, not images, which means it isn't yet multimodal in the way that other OpenAI models are. And as most large language models released today are multimodal, this seems to be a noticeable shortcoming. And the decision release it text only could reflect the pressure OpenAI is feeling to release products as it face more competition. Now, I do know that this is quite true because allegedly there is going to be some other models being released from other top frontier labs within the next two to three months as we manage to close out the year, people are starting to realize that OpenAI is a behemoth, but it's a giant that can be beaten with many different methods and strategies. And we've all seen recently how Claw 3.5 Sonnet has taken the reins in terms of the model that people use for their most intelligent query. So you can see it also talks about pricing. Strawberry is likely to be priced differently than OpenAI's chatbot, which has free and subscription pricing tiers. We're not exactly how Strawberry will be priced, but it will likely have rate limits restricting users to some maximum number of messages per hour, with the potential for a higher price tier that's faster to respond according to another person with knowledge of the product. So essentially it looks like once again, like GPT4 when it was released, we might be limited to, you know, just 25 messages every three hours or maybe every few hours or so. We are extremely rate limited. Now this is something that does happen quite a lot and I think that these companies should actually start to have higher tiers for people that just want to use these models based on the costs proportional to their chat. Like, I know that there are so many people that that would actually, rather than having rate limits, just allow people to purchase tokens and then just continue their conversation. So it will be interesting to see if they do incorporate something like that. And I think one of the most important things that we're going to see is of course, the pricing. So we do know that, yes, it's going to be higher price, but one of the most important things I'm looking at is how much higher will it be priced at? Because what this is going to give us an insight to is to how expensive Frontier Intelligence is going to be in the future. Because if we see that this model is maybe only double or maybe even triple and it's like 10x the performance, then we're going to see exactly how price scales with intelligence. The reason this is going to be so fascinating is because we know that the price of intelligence has continually started to drop. And I'm wondering if these companies are going to simply go in the other way, considering the fact that we're starting to get to the point where these costs are starting to mount up to quite a lot. Now we can also see here that what's going to be fascinating is that it says that we would also expect paying ChatGPT customers to have access to Strawberry first before it's released to bigger free tier users. I would expect the free tier users to get this model probably two years later, like they did with the other models. But that's something that isn't really the main focus. But what we can also see is that Strawberry is also expected to be easier to use than GPT4O for complex or multi step queries. Currently, customers have to type all kinds of additional words into ChatGPT to get the answer that they want, such as telling the chatbot to walk through its intermediate reasoning steps to arrive at its final answer, otherwise known as chain of thought prompting. And Strawberry's capabilities are supposed to help customers avoid doing that or other hacks to achieve smarter results. So basically what, what we have here is the model which is going to have an internal chain of thought or an internal reasoning steps. And I think this is rather useful because many times what you have to do with these models is you have to iterate again and again and again in order to get your final output. And this is something that not only is time consuming, it's sometimes quite difficult because there are steps that you miss and you might not get the best in terms of the intelligence, what is truly possible out of that specific model. So it's quite intriguing to see how OpenAI's internal workings are going to be. And that's something that I'm actually going to be talking about in the next part of this video because there was a research paper that actually shows some of these details. So it says that this means that Strawberry will not only be better at math problems and coding, but also at more subjective business tasks like brainstorming product marketing strategies. And as we previously reported, this model will provide suggestions that are more specific to a user's company and more detailed, like generating a week by week execution plan. Now, what could be a major red flag in terms of A warning for OpenAI is that some people who've used Strawberry the prototype have complained that it is slightly better responses compared to OpenAI's currently released GPT4. Oh, aren't worth the extra 10 to 20 seconds of waiting, the person said. So I do think that whilst yes, when we start, you know, currently there's going to be 10 to 20 seconds over time that is likely to go down. But I am wondering the fact that there are 10 to 20 seconds, if that is going to remove the actual usefulness of this model because you have to understand the average user doesn't want to wait 10 to 20 seconds for anything to load. But I do think that what OpenAI might have a real success at is marketing this Strawberry product as some kind of not God tier intelligence, but higher level intelligence. So people are truly waiting for the responses because they need reliability and they need creativity as well. As the ability to perform complex math problems and of course the ability to actually perform better on coding. Now, I know most people are talking about reasoning and this and that, but the main thing that I'm going to be looking at is to see if this new story model can actually code better than Claude 3.5 sonnet, as there is a single person I've spoken to that doesn't use Claude 3.5 Sonnet for their coding project. So what that does mean is that likely within the next two to three weeks, which is a very short time, we could be getting a new state of the art model that people could use that is going to be better than Claude 3.5 Sonnet in terms of coding. And I'm truly starting to wonder what are people going to be able to build if this model is marginally better or even largely better than Claude 3.5 Sonnet because there are so many people that are now building software and applications with these platforms that they didn't have access to before. So then we also had research published by Google DeepMind and apparently this reveals OpenAI's strawberries approach. So basically it searches at inference time which allows it to reason better. So it says test time compute can be used to outperform a 14 times larger model. Now basically what they're talking about here is that if you give a model time to think, the model's ability to respond in a way that is better, just improved on an insane level. Like you can see, test time compute can be used to outperform a 14 times larger models. So it does seem like this is the approach that OpenAI is currently using in order to get better reasoning at their models. And like I said before, it seems like we are now favoring reasoning over speed because of course you would rather a correct answer more than an answer that is just extremely fast. Now we do know that for future models it's likely that this trend will continue, that the frontier of intelligence will always likely be slower than the current level of intelligence, which is available to individuals for free, because we've largely figured out how those models works and there are open source versions and you know, much smaller distilled versions. So I think that if we're always looking into the future, these models are always going to take a little bit longer to respond because it seems like as well what they're going to be doing is searching over a large space of possibilities. So it seems like we're truly moving towards that stage where we're going to get models that are truly a lot more reliable and a lot more intelligent. Now the legendary Jimmy Apples has said some recent predictions that give us a insight to the timeline of future model releases. Apparently we should have a four times model, maybe still called 4.5, which is of course the upgrade from GPT4 apparently in October, which would mean that we get this model next month. And of course the big boy, which is GPT5, could be as early as December. Now I do find this fascinating that if we get a model like GPT 4.5 just around the corner and then after that we then get GPT5 in Q1, that would mark an insane three to four months, which is why you probably should be prepared for this time, because this prediction of GPT5 in around Q1 to Q3 2 of 2025 is actually quite accurate. You might be thinking. What makes you say that? Well, while back like a really long time ago, and I say a really long time ago, it was only probably a year ago when GPT4 was released. I actually looked at the timelines and then I gave a prediction based on that, which I'm going to show you guys now. And I probably should have included, you know, Q1 dates or Q2 dates, but essentially this was the timeline. So of course GPT5 did actually finish training earlier, 2,024. So we know that GPT5 is actually quite likely to be early 2,025 or even possibly in December. Now I do think that GPT5 might actually be released as early as December, but I think it actually hinges on whether or not.