AI News : So 2027 AI Is Going To Be HUGE, Sam Altman reveals Key Milestones In AI, Googles New Model

With AI rapidly expanding, especially in the model release area, there are a few breakthroughs in today's video that I'm going to talk about, as well as some predictions from the CEOs from the top AI labs that shape how we view what's to come. So one of the first stories for this week in artificial intelligence, or I should say this day in artificial intelligence, with the rate at the update to coming, is Mentabot, a GitHub native state of the art coding agent that scored the 38% on the software engineering bench lite, where the previous state of the art was 33%. Now, this is absolutely crazy because this, I think, is open source, but this is absolutely incredible because it managed to beat Devon, Alibaba, Factory and IBM Research, and they've been using a new cognitive architecture that solves issues in its structured workflow. So it first gathers the context, then it plans and edits. Then of course, you destructure the plan into individual edits and then you apply those edits. And then you can see right here, of course, you test and review, you see if the plan is faulty, you go back to the planning stage. If you need more context, you gather more context. And then of course, finally you submit. So this, you know, small architecture, I say small, but this is actually really, really, you know, fascinating and smart. You can see that they've managed to beat these state of the art. Now, the reason that I think this is so crazy is because it was only a couple of days ago that we got this and this, and this isn't to steal Mentabot's limelight, but it was just crazy that I remember making a video on this where it was literally factory AI, where I was like, wow, these guys managed to get 31.67% on the software engineering bench light using their new infrastructure. And already we can see that this is once again surpassed, which goes to show just how crazy AI development actually is. So this team essentially had months and months of trial and error. And you can see right now this is a product that you can actually use and you were able to, you know, effectively use this to solve these issues, which is pretty, pretty insane. So I'm wondering just how much more software development we are going to be getting over the next years in terms of the future, because it seems like literally every two weeks we are getting, you know, even though it might just be 2% if you compound that over the next couple of years, it seems that the rate of improvement is going to be absolutely incredible. Now, what's absolutely even Crazier is that Google dropped Gemma too, which is, which are two models, one with 27 billion parameters, another with 9 billion parameters. And it's pretty incredible because Google's Gemma 2 actually beats llama 3. 70 billion parameters, COEN 72 billion parameters and command R in the chatbot arena, which is pretty incredible. And I say that, but I don't think you guys are truly understanding that a 27 billion parameter model is beating llama 3, quento and command r in the chatbot arena. Now, whilst, yes, this isn't on the actual, you know, objective benchmarks such as the mmlu, I still think it's rather impressive that, you know, users may not realize what kind of model it is that they are speaking to. And I don't really like this image because I have to turn my head to the left to be able to read it. But what it does show is it does show the rankings of the 27 billion parameter model. You can see that right here, this is the 27 billion parameter model, Google's Gemma 2 and then of course the 9 billion parameter model right here. And it's pretty incredible because it manages to surpass all of these ones right here. Of course, like I said before, it's just on the subjective evaluations, this isn't the hard benchmarks. I'll actually show you the hard benchmarks. They've actually got the hard benchmarks here. But you can see that it actually manages to be on par with llama 3. But of course, with the, you know, the 8 billion versus the 9 billion parameter models, it actually does do a little bit better than llama 3, which is, like I said, really surprising considering that when Llama 3 was released it was kind of this weird state of the art model in terms of, for its size, it just performed the best in terms of how, you know, condensed they've managed to get the information. So this is something that actually did kind of surprise me because whilst yes, we did actually get the information on Gemma 2 before, I didn't think it was going to be this small and this good. So this goes to show that Google are actually shipping and like I said in my previous video, Sundar Pichai and along with Demis Hasabis, they've managed to, I guess, you know, maybe not yet turn Google around, but what we've seen is that they've actually managed to start shipping models. Now. What's really cool about this is that this is actually open source. So this is an open model that has outsized performance, unmatched efficiency and cost. Savings and, and is blazingly fast across anything that you want to try it with. So this is something that is really, really quick, it's something that you can easily use and has broad framework compatibility. So I think when we get to things like Gemma 3 and Gemma 4, we're going to see some remarkable improvements in terms of what they're able to do. So I'm actually really excited for this. Now what's also cool from Google is that they've managed to get Gemini 1.5 Pro to a 2 million context 3 length window and code execution capabilities. And Gemma 2 are available today. So you can see here at I O we announced the longest ever context window of 2 million tokens in Google's Gemini 1.5 Pro behind a waitlist. But today we're opening up Access to the 2 million token context window on Gemini 1.5 Pro for All developers. As the context windows grow, so does the potential for the input cost. To help developers reduce the cost, we're launched context caching in the Gemini API for Gemini 1.5 Pro and and 1.5 Plash. So this is pretty, pretty insane because we now have 2 million context length. And what's crazier is that we haven't even seen other models like OpenAI manage to catch up to this. However, if you did pay attention to GPT4O's secret paper, you'll know that they do actually have a model and I think it is GPT4O with an a million context token length. Because on the secret article I'm actually going to show you guys this. Now, if you actually scroll down to some of the exploration of capabilities and if, if we scroll down to the video one, I'm not sure exactly where it is, but if we just scroll down to lecture summarization, you can see right here that, you can actually see right here that OpenAI's GPT4O is able to take in a 45 minute video. Now I would presume that this is around 1 million tokens because this is pretty much the same exact length that Google did their demo video with. So you can see here that whilst yes, OpenAI actually haven't shipped this model yet, it seems that we're, you know, moving towards an increased length in terms of context length across all major models. However, it seems that Google might have realized what OpenAI have done. And then of course, once again, I'm not sure what they've done, but Demis Hassabis actually did talk about how Google have actually made some kind of breakthrough in order to be get to. In order to get to 2 million context length, they actually spoke about how, you know, they kind of made a breakthrough, and of course, that was what led them to this, you know, length that they got to before everyone else did. So it seems like Google are really, really starting to ship. And I think we can probably expect the most from Google over the coming months in terms of what they're going to be able to give us. And I would think that Google can potentially take the lead in the next year and a half, as long as they continue shipping and of course, managing to prioritize their AI stuff, because they do have an entire suite of AI products that I personally do believe as long as they manage to deploy them safely, they can completely control the AI market. Now, there is also something that was rather fascinating. A Reddit user, Rozzy, the creator, got a sneak peek of ChatGPT's upgraded voice feature that apparently sounds way better and actually generates background effects while narrating. This actually means that, like I said before, you know, the model that we're going to get is a truly capable one, which means in order for OpenAI to roll out this model, they're going to have to do some serious infrastructure scaling, which is why we haven't got access to this model yet. Right?

I got you, Rossi. So picture this. There's this small town, right? Real quiet, everybody knows everybody kind of vibe. Now, there's this old abandoned house at the end of the street that everyone says is on it. Kids dare each other to go there, but most just run away before they even get close. One night, a storm rolls in. Thunder, lightning, the whole shebang. A couple of teens trying to be brave or just plain stupid decide to check out the house during the storm. They sneak in through a broken window and start exploring with nothing but a flashlight and their phones for light. As they creep through the dusty halls, they hear a noise from upstairs. It sounds like footsteps, but slow and dragging. They follow the sound. Like I said, maybe a bit stupid. And it leads them to a room at the end of the hall. The door is slightly ajar, creaking as it sways. They push the door open, and in the middle of the room, there's this old dusty rocking chair just rocking back and forth on its own. No wind, no strings, just moving. The kids are freaked out, but also kind of mesmerized, you know?

Now, I do think the model is trying to generate sound effects here, but one thing that I did actually remember when looking at the exploration of model capabilities when generating sound effects, they actually didn't sound that good. I'm not sure if OpenAI just struggled to do that, but considering the fact that this model is not just combining other models together, it's actually natively modeling multimodal, meaning that, you know, from one input to another input, that's essentially how the model works. It's not like combining a vision system with another system. It's just entirely multimodal in, multimodal out. So I'm guessing that's why the sound effects area isn't as good as the other areas. Now, Anthropic CEO Dario Amodei says by 2027, remember that key date that I spoke about quite a bit about in many of my recent videos? I'll be doing an entire video on this actual date. But he said that AI models will cost up to a hundred billion dollars to train and will be better than most humans at most things. Which means that these next training runs that they're looking at are going to cost $100 billion, which does mean that they're going to have to receive a significant amount of investment from some other people. And that also means that this, you know, this $100 billion is probably going to be going straight towards Nvidia's bottom line, which is very, very interesting considering the fact that many people believe that Nvidia is, isn't going to be a company that is around for much longer, considering many other players are getting into the chip space. But that's not the point of this. The point of this is that we have a situation on our hands. The next training runs after the 10 billion training runs are going to be $100 billion.

I've said this a few times, but, you know, back in 10 years ago when all of this was kind of science fiction, I used to talk about AGI a lot. I now have a different perspective where I don't think of it as one point in time. I just think we're on this smooth exponential. The models are getting better and better over time. There's no one point where it's like, oh, the models weren't generally intelligent, and now. And now they are. I just think, you know, like, like a human child learning and developing. They're getting better and better, smarter and smarter, more and more knowledgeable. And I don't think there will be any single point of note. But I think there's a phenomenon happening where over time, these models are getting better, better and better than even the best humans. I do think that if we continue to increase the scale, the amount of funding for the models if it goes to, say, 10 billion.

So now a model would cost what.

100 million right now? 100 million. There are models in training today that are more like a billion, right? I think if we go to 10 or 100 billion, and I think that will happen in 2025, 2026, maybe 2027, and the algorithmic improvements continue apace, and the chip improvements continue apace, then I think there is in my mind a good chance that by that time we'll be able to get models that are.

I mean, this is something that, you know, Leopold Aschenbrenner actually spoke about. If we actually take into account every single thing that contributes to an improvement in AI models, not just the ordinance of magnitude in terms of the scale, but the algorithm improvements, the inference improvements, the, the chip improvements, and everything that we can manage to do, as well as the agentic workloads, frameworks that we can use with these models. It is going to be no surprise when these models are truly capable in 2027 after the $100 billion supposedly training runs now, I'm not sure who's going to be funding that or where you're even going to get money to train a model for $100 billion because that amount of cash is just not even liquid at most companies, unless you're Apple. But that is something that is going to be pretty incredible. And, and I'm wondering how much we're going to be getting out of these next models that are a billion dollars or even just $10 billion, which is just a staggering amount. Now, Dario Amade also said something about scientists accelerating discoveries in biology and curing diseases. And this is something that we've actually seen already with the likes of Google's models. So I wouldn't be surprised if in the future we get specialized models that are literally just focused on breakthroughs regarding science.

Let's say those models get to the point where, you know, they're kind of, you know, graduate level or strong professional level. Think of biology and drug discovery. Think of a model that is as strong as, you know, a Nobel prize winning scientist or, you know, the head of the, you know, the head of, head of drug discovery at a major pharmaceutical company. I look at all the things that have been invented. You know, if I look back at biology, you know, crispr, the ability to like edit genes. If I look at, you know, car t therapies which have cured certain kinds of cancers, there's probably dozens of discoveries like that lying around. And if we had a million copies of An AI system that are as knowledgeable and as creative about the field as all of those scientists that invented those things, then I think the rate of those discoveries could really proliferate and you know, some of our really, really long standing diseases, you know, could be addressed or even cured.

I do think that this is going to be something that does happen. AI is actually able to accelerate the rate of discovery. Whilst yes, this is something that is quite futuristic and hard. If we do take a look at what AI models are able to do and if we're able to just look at fundamentally what they are, I do think that, you know, these discoveries are going to be made sometime in the future. If we do look at what we were able to do with of course, AlphaFold, that was able to accelerate the discovery of those proteins by, by not a million years, but a huge amount of time. And I do think in the future we're going to be getting models that are going to do that exact thing. They're going to be able to experiment a million different ways and a million different ways that we couldn't even think about. So it's going to be something that is truly fascinating as discoveries continue to happen. And I think we just need to get to that inflection point where we get a model that can continuously do that. All we need to do is just put all our compute into that and then of course those discoveries will be made. What's even weirder was that Sam Altman in a recent live interview actually spoke about this same phenomenon, although he actually got a lot of flack for this. But I'm going to explain why he shouldn't have.

I think unlike other sort of, we're aware even if today we're like, okay, ChatGPT is, you know, once I use it, I'm not scared of it. There is a sense of super understandable anxiety about where, what does it mean if these tools keep getting more capable, they've been getting capable at and there's tons of wonderful things and we could talk about those all day, but there is this, what is the future going to look like? Like even if we solve every safety problem, even if we solve every, you know, misuse problem, even if we figure out the perfect regulatory regime, like what are, what are our lives going to be like when it's not just like the computer understands us do these things, but we can say like hey computer, like discover all of and it can go off and do that. What does it mean when we can say like hey, start and run a great company. It can go off and do that. So that's a big change.

I think people fundamentally misunderstood what Sam Altman was trying to, you know, say here. I think one of the first things is that some people just looked at this and they said, oh my gosh, why would he state that an AI model can solve all of physics that doesn't make any sense, then clicked off. But I think what Sam Altman is trying to state here is that there is a lot about the universe that we simply don't understand. And if we do get to a future, bear in mind a very, very far future to which we can say, look, AI model, this is physics. Whatever your understanding of physics is, we need you to just simply just focus on making breakthroughs and focusing on what we don't understand. Because there's actually quite a lot, there are actually going to be quite a lot of paradigm shifts in the future when we start to discover certain things, things and what they kind of do. Because there's this constant debate about whether or not we know exactly what we know. And our understanding is always kind of changed by one brilliant mind or just a few brilliant minds over the course of a few decades which fundamentally reshape our understanding and worldview of the world. Like, you know, for example, we thought the earth was flat. We didn't know about, you know, the microorganisms that live on our skin and inside of our bodies. There are just all of these kinds of things that I can guarantee once we manage to discover them, whether or not it's going to be simply an AI system continually making game changing discoveries, which I think will be very hard to do considering its physics. And you know, AI is fundamentally software. So unless it can have like a simulation that can one to one map stuff, which probably could happen in the future, I think of course this is going to be something that's extraordinarily difficult. But I do think the fundamental point that we are going to be using AI systems to research and make breakthroughs is going to be something that whilst, yes, now seems crazy, just remember a lot of the things that we even see in today's day and age does seem crazy itself. There was also a Chinese robot toy maker, Liju Robotics, launching a full size humanoid robot, Kaofu, and it's integrated with Huawei's multimodal LLM Panju to understand natural commands, plan tasks and execute with bi manual coordination. I would follow the humanoid hub because it's actually a really effective resource for just humanoid robots. It's something that I actually stay up to date with when I'm looking for certain humanoid robot news that I otherwise wouldn't have found. And, you know, niche news like this, like, like these Chinese robots is always, always interesting because it shows us what China is doing. And if you're wondering why are China putting out so much in terms of the humanoid robots, number one, China are very, very effective at producing factories and building stuff like their, you know, ability to execute when it comes to factories and stuff like this is just absolutely remarkable. So I think what we will see in the future, especially considering the fact that China have did this kind of initiative where they've kind of incentivized these companies to make a lot more humanoid robotics, I do think that this is going to be something that truly, truly gives us the insight to how the future is going to be, because we're likely going to see China expand their horizon, expand their fleet first, and then of course, the US is probably going to execute on another level as well. We also got this from OpenAI and I think a lot of people are getting mad at OpenAI when they do tweet stuff like this, because the problem is, is that OpenAI didn't tweet stuff before. And usually when they did tweet stuff, it was usually like a major, major update. So now when people see tweets and updates and blog posts and it's not GPT 4.5, not GPT 5, a lot of people do get, you know, I guess you'd say a little mad. But they're working diligently to produce the GPT4O voice model. But essentially they state here, finding GPT4's mistakes with GPT4. Critic GPT, a model based on GPT4 writes critiques of ChatGPT's responses to help human trainers to spot mistakes during reinforcement learning with human feedback. Basically, they trained the base model GPT4 called Critic GPT to catch errors in GPT's code output. We found that when people get help from Critic GPT to review ChatGPT code, they outperform those without help 60% of the time. We're beginning the work to integrate Critic GPT like models into our RLHF labeling pipeline. Basically, this is recursive self improvement with AI with humans in the loop, I guess an extended form of recursive self improvement. They do highlight here that as we make advances in reasoning and model behavior, ChatGPT becomes more accurate and its mistakes becomes more subtle. And this can actually make it hard for AI trainers to spot the inaccuracies when they do occur, making the comparison task that powers the powerful RLHF much harder. And this is a fundamental limitation of RLHF, because they make it increasingly difficult to align models as they gradually become more knowledgeable than any person could tolerate that could provide feedback. So basically, they're stating that, look, if you have a dumb model, it becomes easy to provide human feedback because you need only a human that's smarter than a dumb robot in order to do that. But as these models get increasingly smarter, you're going to need increasingly smarter humans, and there's an increasingly smaller pool of humans to actually provide feedback from. And at what point do you just start using the model to improve itself if those subtle mistakes are things that we can't even find? So this actually brings about a bigger question. You know, when are we going to get to that recursive self improvement with RLHF? And of course, I'm not stating that that's how recursive self improvement works. I'm just stating that I think we're going to eventually get to a point where we're probably going to use AI models to evaluate AI models. And that wouldn't be a surprise, but we can see here that this is something that is currently being used and is currently very effective.