BIG AI News : Open Source CRUSHES Everything, GPT-5 Paramters Leaked, AGI Could BeDecades Away?

From a new stunning open source model that can beat current state of the art models like GPT4O to revised timelines about the future of AGI, this video will show you what's happened in the past few days in terms of what you missed. So this is arguably one of, if not the biggest story that's happened this week in terms of AI announcements. This is Matt Schumer, the CEO of Hyperrite AI, and he tweets that I'm excited to announce Reflection 70B, the world's top open source model. Trained using Reflection Tuning, a technique developed to enable LLMs to fix their own mistakes, this model manages to surpass everything. And with 405 billion parameters coming next week, we expect it to be the best model in the world. Now it shows some further details, but the implications for this are staggering. I spoke a little about this in my school community earlier and I think this shows us what the future of AI is truly going to be. So on these charts, what we can see here is that we have many different AI systems. So on the right we have the llama 3.1, the 405 billion parameter model, which is of course the base one. Then we've got Gemini Pro GPT4.0, Claude 3 Opus and Claude 3.5 Sonnet. And remarkably, the only model that manages to come even remotely close to reflection 70B. That's 7.0 billion parameters. And that's pretty incredible because we do know that all of these other closed source, state of the art models are actually much larger than 70 billion parameters. But what we can see here is that Claude manages to surpass it on the GPQA and on the human eval, but it's not by much. Now it's pretty crazy because I think the most interesting thing that we have here is the way that this model was tuned provides us an insight into how an AI can actually get smarter in terms of its reasoning ability and of course its responses. So you can see right here that this is basically from the Reflection paper, which is basically where earlier I think last year we had a paper that basically spoke about when you have an AI system, give out a response, you don't just use that response, you get the model to respond to its own output and then ask it what it could have done better. And in doing so, the model kind of critiques itself and it manages to improve its own response. So right here you can see that the model is given a task. Write the first sentence of the preamble to Declaration of Independence, put, but do it in mirrored writing. So you can see right here it has thinking. This is something that I have seen in Claude 3.5 sonnet, which is probably one of the reasons that the model is so advanced in terms of its ability to perform correct tasks. And then we can see it outlines a plan and it says, let's break this down into steps. Then it says, let's go through this plan. And then this is where the magic actually occurs. We've got reflection. Before proceeding, I should clarify that there are two ways to interpret mirrored writing. Simply reversing the order of the letters right to left and reversing the order and flipping each character. For this task, I'll use method B as it is more challenging and true to the concept of mirrored writing. That's the end of the reflection. Then you can see it gives it. Then it reflects one more time and it says, I've double checked the mirrored text and. And it appears to be correct. Each character has been flipped and the order has been reversed. The punctuation marks remain as they would in a true mirror reflect. And then of course you can see after all of this internal thinking, it then gives a final output. So the user usually wouldn't see all of this stuff. If you've ever wondered what goes on in an AI recently, like I said before, we saw with Claude 3.5 that you were able to prompt it to see exactly how the model was thinking in terms of its step by step process. And this kind of gives us an insight into how these models get better responses. And here we have the output which says here is the first sentence of the preamble, yada, yada, yada. And it gives a really good response. You can see here it says it holds its own against even the top closed Source models, including Claude 3.5 Sonnet GPT4O, the top LLM in MMLU and other important benchmarks. And it beats GPT4O on every benchmark tested. And allegedly it clobbers Llama 3.1 405 billion parameters. And he speaks here about how it uses chain of thought. So we separate planning into separate step, improving the chain of thought, accuracy, potency and keeping the output simple and concise for end users. So you can use reflection 70 billion parameter here. And the thing is, is that they are currently experiencing high traffic. So if you're watching the video on the day of the release, you probably won't be able to use this AI system since they didn't expect to get that much traffic. But I suspect that as the week moves on, a lot more people are going to be using this model. Now, interestingly enough, if this model actually does surpass state of the art systems for everyday tasks, and in terms of the coherency of the models and how smart it is, I think this is just going to completely change the game because people will have no reason to use Frontier models and pay a $20 a month subscription for something that's actually quite worse. I do think that, yes, there are certain scenarios where these AI systems do excel, such as tool use, and of course certain windows where you can have interactive demos. But for an effective AI system that actually goes through all of these steps, I think this is going to be something that is remarkably useful for, for the average person. And it's quite surprising that these benchmarks are here so soon. Now, something else that was rather fascinating was the fact that OpenAI is considering higher priced subscriptions to its chatbot AI. Now, I think that this one has one of the most staggering implications, because I don't think OpenAI would do this if monumentally powerful technology wasn't on the horizon. So it starts out by saying, how much would you be willing to pay for ChatGPT every month? $50, 60, $75? How about $200 or $2,000? And that's monthly. So it says that's the question facing OpenAI, whose executives we hear, have discussed high priced subscriptions for upcoming large language models such as OpenAI's Reasoning Focused, Strawberry and the new flagship LLM dubbed Orion. And they state that in early internal discussions, subscription prices ranging up to 2,000 per month were on the table. Said one person with direct knowledge of the numbers, though nothing is final and of course we have strong doubts that the final price would be that high. Now, I think there are two main reasons for this, because one of the main reasons that this might be, and I don't think that this is a main reason, but it's just one that could be there is because these models are actually quite expensive to train and run than priors. So you can see that it says, for instance, we've reported that when given additional time to think, think, the Strawberry model can answer more complicated questions or puzzles than OpenAI's current models can, that additional thinking or processing time could mean more computing power and therefore more costs. And if that's the case, OpenAI is going to want to pass along some of that to its customers. And it basically says here, a high price would also mean that OpenAI believes its existing white collar customers of ChatGPT will find these upcoming models that a lot more valuable to their coding, analytics or engineering work. Now, I think that this is not going to be an issue provided that OpenAI can actually showcase that these models are worth their weight in gold. If OpenAI can showcase that these models are truly advanced in either coding, engineering or whatever task that requires a lot of cognitive capabilities, then I think mainly companies and small businesses are going to be paying a lot for these models that can solve their tasks at 1/100th of the average worker. Now, I know that most people might be valuing these models like, okay, $200 a month, that's crazy. But I think that's not how you want to look at the value of these models. You have to think about it in the other way. Think about it. The fact that most salaries are costing like, you know, 3400, 4000, 5000, $6000 a month, and if you can replace that with an AI system that can cost $200 a month, companies are not going to bat an eye because that is a fraction of the price that you're currently paying. So when you think about it from that aspect, you can say, okay, now I can see why companies and individuals might want to actually pay more for these models if we know that they can actually do a lot more in terms of their ability. Now of course, you do have the cost issue which is facing many of these companies because these models are not free to train. It does require millions of dollars in compute and energy and chip infrastructure. But I think the fact that they're discussing this kind of stuff tells me that the next level of AI that's going to be here is going to be really incredible. I know that many people are doubting the abilities of Strawberry and of course Orion, but I'm not doubting that these models are truly going to put us in the next frontier of AI, considering the fact that literally earlier in the video we did see a model that just managed to surpass closed source AI as a 70 billion parameter model that was fine tuned on solving problems. So I think that the future models we're going to get are going to be absolutely incredible. Now, there's only one caveat to this. I am wondering if OpenAI does have some insane models that are quite expensive, how are they going to maintain those customers and those prices when we know that the price and cost of intelligence has been dropping quite a lot? And I'm going to show you guys a recent interview where the CEO of Perplexity actually talks about this issue more in depth.

All we know is that despite us not really Trying to grow, it's growing. Look at the trend. The growth rates are clearly there. Yes, the cost per query is very high. That's why we want to raise money. If it was not high, we don't need that much money. So we are betting on the fact that this technology is on an increasing curve. The models are going to keep improving, the cost per query is going to go down tremendously. And since the day we said it, it's gone down by 100x. And so with the same amount of money, you'll be able to serve a lot more users over time. So you're betting on something that's really going to work. And when the cost per technical query goes down and the models are going to get more and more capable and smaller sizes, then we are heading towards a world where whatever is like, you know, one in a 10 queries hallucinate today would be one in 100 in a year, one in a thousand, in two years, one or 10,000, three years, it's going to like increase its quality exponentially. So you're like, you cannot understand that world today. It doesn't exist yet. So you're betting on the potential of getting there.

And in terms of the world that we might not understand in the future. This weekend the X AI team brought on Colossus, which is 100,000 H1 hundreds in their data center online from start to finish. This was done in just over four months. Colossus is the most powerful AI training system in the world and moreover, it will double its size to 200,000 H2 hundreds in just a few months. This is absolutely incredible and I think this is why many people continue to tell you, including myself, that you shouldn't underestimate Elon Musk, because he's shown time and time again that he's able to consistently disrupt certain industries with his ability to rapidly put together different teams of specialists in different fields and build great products that people actually want and love. If you look at what he's been able to do with SpaceX compared to Boeing, it's absolutely incredible that SpaceX has managed to beat Boeing to the punch on several different occasions. Even as to, you know, bring back these astronauts that are currently stuck in space, that's a whole nother video. But for those of you who know what I'm talking about, you'll know that Elon Musk isn't someone to underestimate. So I can't imagine how crazy the future is going to be now. Whilst, yes, I think the future is going to be a Bit crazy, some individuals think not. Now, Andrew Ng, machine learning pioneer, stated that AGI is many, many decades away and maybe even longer. And companies that say it is only a year or two away are using non standard definitions to lower the bar. Take a look at this because this is thought provoking and may rethink your timelines, but I'll give my opinions in a moment.

Regarding AGI, the standard definition of AGI is AI that could do any intellectual tasks that are human. So when we have AGI, AI should be learn to drive a car or learn to fly a plane, or learn to write a PhD thesis in the universe. But that definition of AI think where many. I hope we'll get there in the life. One of the reason that there's hype about AGI in just a few years is there's some companies that are using very non standard definition. And if you redefine AGI to be a law, then of course we could get there in one or two years by using the standard definition of AGI of AI to do any intellectual task. I think we're still many decks, but I think it'd be great. Regarding AGI, the standard definition of AGI is AI that could do any intellectual tasks to the human. So when we have AGI, AI should.

Be so I think this one was rather fascinating because many people have, you know, pushed back on this video because I think what it does is it says that AGI is, you know, something that's able to do a million different things. Now there's completely no disagreement with Andrew Ng in his prediction, but I think the problem with AGI is that there are so many different definitions of, of AGI that when the talk of AGI is here, so many people disagree on what that thing is, that we can't ever get the statement from anyone that seems to all converge in one single point. Like there isn't one single thing that we can all agree an AGI can do, which leads to various different definitions for what this system is going to be capable of. Now of course, many people always to refer back to deep minds and levels of AGI paper and this actually gives us a useful guideline for where AGI is. You can see that we already have superhuman narrow AI and you can see that we are trying to get to level 2, which is competent or at least 50th percentile of skilled adults. Now I think if we do have an AI that can manage to, you know, learn how to fly a plane, learn how to drive a car, do all of that thing I'm not sure with as to why that wouldn't be like level four, which is, you know, virtuoso AGI, or near level five, which is superhuman AI. There was actually one tweet I did agree with. It says Andrew Ng's definition of AGI seems to aim for a godlike, all encompassing intelligence rather than general human intelligence. Can you name a person who can drive a car, fly a plane, write a PhD thesis and do other very specific things? Why set the bar for AGI so high? Instead, let's define AGI as common sense intelligence across domains with specific expertise in certain areas, creating multiple AGIs rather than one. Now, I do think that this is rather important because if there was someone that could do all of these things, they'd definitely be in the 1% of the 1%. And if we say that AGI is only a thing that can do that, it then kind of confuses people in terms of how these definitions are. So I do think that with respect to Andrew Ng, what he's describing is probably something that is a virtuoso AGI, which is outperforming 99% of humans and is probably quite close to something that many could define as a superintelligent. But in speaking about superintelligence, if Andrew Ng thinks that superintelligence is far away, remember Ilya Sutskeva, the man who was integral to OpenAI's success? You can see here that it says exclusive OpenAI's co founder Sutskova. New safety focused AI startup SSI raises $1 billion. So this company is valued at $5 billion and these funds will be used to acquire computing power. And basically Safe Superintelligence Inc. Founded by Ilya Sutskova, is basically just going to focus on superintelligence. If you missed this basically on their web page, they, they basically said that this is something that they think they could do and it was something that they're going to be pursuing. You can see here that the most important sentence from this entire website is that superintelligence is within reach. Now, I don't think that if superintelligence wasn't in reach, they wouldn't be doing this. And whilst, yes, some people could say that, oh, it's just another company doing hype marketing, I think that the likes of Ilya Sutskova are not to be underestimated in terms of his genius to be able to do things that we might not think about. So I think it's rather fascinating because they say our singular focus means no distraction by Management overhead or product cycles, which is quite clearly referring to, you know, ChatGPT, where you need to do certain things in a certain time so you can release your model ahead of competitors and they're basically just going to completely focus on super intelligence and nothing else. So the fact that this company founded by Ilya Sutskova and is just focusing on super intelligence and not even focusing on AGI gives us the kind of insight to where things might be in terms of certain industry insiders and their differences of opinion. We also had this image brought to my attention by the legendary Jimmy Apples, which shows that GPT5 compared to GPT4. I know this is quite hard to see, but I will zoom in. But you can see the image of GPT5 in terms of the model size. It seems that it might be 3 to 5 trillion. Now, if you remember, GPT3 was 175 billion, GPT4 was 1.8, mixture of experts. So GPT5, it looks like it's going to be double the size in terms of how big the model is. Now, this is quite fascinating because this is the first time that we've got any sort of details with regards to the size of GPT5. And I do know that, yes, there is, there is quite a lot of misinformation out there with regards to how these models are in terms of their size. If you all remember that very disingenuous image that was posted across social media. And I'm referring to this image which shows that one model is just incredibly bigger than the other one. But this image was proven to be fake and I think visual diagrams like this are completely useful, but not useful when they present misinformation. GPT5 seems like it's only going to be twice the size of GPT4. But I think the size isn't the main thing. Like I said before, with models like Strawberry and Orion, the main thing that we want to really understand is, of course, the reasoning capabilities and the reliability. So then we have The CEO of 1x Robotics, the robot that was recently revealed in an interview, stating that they are targeting to scale by 10x every single year. Tens of thousands in 2026, hundreds of thousands in 2027 and millions in 2028. If this is true and accurate, this scale of humanoid robotics is going to be absolutely incredible, considering the implications for this.

We've had like an internal mantra for quite a while, which IS scale by 10. So we did like 10 EVEs, then we did three goes to 100 and then we're going to do thousands of NEOs, tens of thousands of NEOs. So this means thousands of NEOs in 2025, tens of thousand in 2026, hundreds of thousands in 2027, millions in 2028. You can do the math. That's hard. So far we've been on target, but it's really hard and it's been really painful. I mean it's not like we're there yet, but it's also not like we haven't done some of this before. So for Eve, for example, we peaked between 10 and 20 units per month on the previous line. And for Nio we're to going to basically 10x that.

Right.

And it's a lot of organizational pain. You need to build out how it is to execute as manufacturing company, which one thing is all the systems that needs to just be in place for you to be organized and efficient at a scale with respect to supply chain and materials and processes and traceability and all these things. It is step by step. Yeah, I think a mistake people often do is think that they can go manufacturing from 1 to 1 million and like clearly you can't. Right. You have to walk the steps and you can walk them pretty fast, but you still have to walk the steps. So we're taking, I'd say a bit of a humble approach to this. I think supply here, sorry, demand far outstrips supply, but you still have to walk the steps. So feel very confident right now since we're building pretty large batches of Nio now already on the new factory line, that we will hit our targets for next year. Very confident. About 2026, something kind of magical happen happens when you go from like tens of thousands to hundreds of thousands of millions. And it's traditionally like you see a lot of companies fail. It's very painful and just you need to make sure that you have the best people and people have done this before and that you really leverage that you have the full understanding of your product all through your organization. Right. And if anything goes wrong, you can redesign, you can fix it. And that's again kind of like back to controlling your own destiny while being vertically integrated and controlling your own supply chain. You actually have you come blame anyone else. You have the power to fix this if something's wrong. It's going to be a lot of work. Yeah, yeah.

In some really good news from Google, we also had Alpha Proteo generates novel proteins for biology and health research. So basically Alpha Proteo is a new AI system created by Google DeepMind. It's designed to make new proteins that can stick to specific target molecules in our body. Now, we this is important because it could help scientists develop new medicines and understand diseases better. Now, why is this important? Well, you know, the proteins are like tiny workers in our bodies and they do all sorts of jobs, from helping cells grow to fighting off diseases. And sometimes we want to create new proteins that can attach to specific targets in the body, like viruses or cancer cells. And this helps us to study these new diseases or to create new treatments. Now, if you're wondering how this works, it's been trained on a huge amount of information about how proteins work and stick together. And when the scientists give it a target molecule, it can design a new protein that will stick to that target. Now, this is special because it is actually faster and more successful than the older methods of designing proteins. The proteins it designs stick more strongly to their target and it can actually create targets and it can create proteins for targets that were difficult to work with before. Now this is great because this could be used for developing new medicines, creating tools to study diseases, making senses to detect specific molecules in the body, and helping crops resist pests. Now, this does have some limitations. For example, it couldn't design proteins for one particularly challenging target. And they're still working on improving that. But I think that this is rather fascinating because the implications are profound. Now, the team at Google are actually being careful about how they share this technology because it is powerful and could potentially be misused. And they are working with experts to ensure it is used responsibly. Now, lastly, I will leave you all with this. This is the leaderboards currently for Seal, which is a company that focuses on the private evaluations. So this is where you models that are currently evaluated on a private data set so there's no contaminations. Models can't be trained on these answers, so it's just raw intelligence being measured. And we can see that in coding. Claude 3.5 Sonnet is ranking first and they've added GPT4O, Gemini Pro and Mistral Large 2. And surprisingly, Mistral Large 2 encoding actually beats out Google Gemini 1.5 Pro. And I've told you guys many times before that Mistral Large 2 is one of the models that is severely underrated. You can see here that in math, GPT4O ranks second and even instruction following GPT4O doesn't rank first. So it's clear that by private metrics, GPT4O has been dethroned. And I am wondering how long is it going to be before OpenAI reclaims their title. I do think that they do have a decent lead in terms of their brand image and the popularity of ChatGPT.