AI News : GPT5 Will Be Better Than We Think , Gemini 2 China Takes The Lead , , Googles New Robots..

So welcome back to the AI grid. Today we're going to take a look at some of the top AI stories, one of them being the fact that AI is going to continue getting better. Currently. There has been a recent sentiment regarding AI technology. Many people think that we are towards the top of the Gartner hype cycle, meaning that we are about to experience the trough of delusion. Currently, apparently we are at the peak of inflated expectations and this is where we are headed now. I think that this is mildly true in the sense that currently where we are at now, there seems, okay, and I put an emphasis on seems to be a lack of AI development. But you have to remember all of the insane model releases that we've gotten this year. And we also have to remember the fact that there are two key things that we haven't done yet. Number one, we haven't exhausted all of our available compute yet. The models that we've recently trained, they've only been trained up to $100 million. While that seems to be an insane number, and yes, it absolutely is a staggering amount of investment for a single AI system, we have to realize that there is still a billion dollar training runs and $10 billion training runs that could be commencing within the next couple of months or even years. Which means that the amount of returns that we are going to get due to the scaling laws is very likely to continue. Which means that despite people thinking there is this trough of delusionment and that hype is about to die down, I would argue that this is not the case. You have to remember that AI is very much so going to be a technology that impacts pretty much everything.

Well, think we're not at diminishing marginal returns on scale up. And I try to help people understand there is an exponential here. And the unfortunate thing is you only get to sample it every couple of years because it just takes a while to build supercomputers and then train models on top of them. And so the next sample is coming and like, I, you know, I can't tell you when and I can't predict exactly how good it's going to be, but it will almost certainly be better at like the things that are brittle right now where you're like, oh my God, like this is a little too expensive or it's a little too fragile for me to use. Like all of that gets better, like it'll get cheaper and like, you know, things will become less fragile and then, and like more complicated things will become possible. Like that is the story of each generation of these Models as we've scaled.

Up, that is Microsoft's cto, Kevin Scott, confirming what I do believe about the AI space. Now, I do want to say that this is no surprise. Have people recently forgotten what you're able to do with Claude 3 artifacts? The model is only marginally better than GPT4 and in some case worse than GPT4O in some. In terms of the user preferences, we can see that Claude 3.5 sonnet being able to do what it's able to do with code is completely game changing. And this isn't Claude 3.5 opus, which are the next iteration of models. So I truly don't believe that people are going to realize what is going to happen on the second iteration of model releases. Things like Gemini 2, Claude 3.5 Opus, GPT5, these systems are truly going to obliterate what we currently have and what our perception is of what AI is currently capable of. And in speaking about future models like Gemini 2, there is actually a recent tweet about a new model that appeared on the Chatbot Arena. If you're not familiar with the Chatbot arena, essentially that is a place where leading Frontier Labs collect early data on their models in the world about how capable their system are. What you'll do is you'll put a question to a model and then what will happen is two models will respond. Now sometimes what you will get is you'll get a test model. You can see right here that this is titled Gemini Test. We can see that this means that they are testing some form of Gemini. Now some people have speculated that this is probably going to be a exclusive Gemini Vision model, but some others have stated that this is Gemini 2.0. Now if you remember, Gemini 2.0 did already start training a couple of months ago. So considering the fact that we know Google are behind OpenAI in terms of product releases and in terms of their development, and if we know that Google have recently taken a much more aggressive approach in terms of releasing their software, it could very much be that they are testing Gemini 2 in the chatbot arena to see how it performs compared to certain models and how it performs in certain benchmarks. Now, with this example, we saw that someone managed to put a picture in that they took on their phone. And then of course the system Gemini managed to get exact locations simply from a personal image. Now the reason that this is rather impressive is because of two things. Either one, the model just realized exactly where it was based on looking at the image. And that's absolutely outstanding for it to know where you are just based on what it's seen. Or number two, it was able to use some of its data from Google and it was able to geolocate where this person was, which is also pretty cool, in a small piece of information for OpenAI developers. We actually did get the Text to Speech API in the playground. So if you've ever wanted to use ChatGPT's text to speech feature, where you've asked a question and then it manages to talk back to you, this is something that I've played around with. It's actually pretty cool. I just wanted to include this in the video because I know that some people might want to know that this is a feature.

Major risks around I do think there are, there are major risks around persuasion propaganda where, you know, you could, you could persuade people very strongly to do specific things. You could control people to do specific things. And I think that's incredibly scary to control society to go in a specific direction with the current systems, you know, they're very capable of persuasion and influencing your way of thinking and your beliefs. So, and this is something that we've been studying for a while, and I do believe it's a real issue with AI. It gets majorly exacerbated. So especially in the past year, we've been very focused on how to help election integrity. And there are a few things that we are doing. So number one, we're trying to prevent abuse as much as possible. And so that includes improving the accuracy of detection, political information detection, and understanding what's going on in the platform and taking quick action when that happens. So that's one. The second thing is reducing political bias. So you might have seen that ChatGPT was, you know, criticized for being overly liberal and Elon, you're too woke, right? Well, there were a few other voices, but, you know, the point is that it wasn't intentional and we work really hard to reduce the political bias in the model behavior and we'll continue to do this and also the hallucinations. And then the third thing is we want to point people to the correct information when they're looking for where they should be voting or voting information.

So this is a topic that isn't as spoken about as much when it comes to AI development, because it's one that tends to fall under the radar because it's quite hard to detect if an AI system is actively changing your opinion. Because the problem is that usually what AI does try to do is it tries to always give a fair and balanced opinion. For example, if you ask an AI AI its opinion on something, it usually just says either you could do A or either you could do B, but it's always up to the person. So with that being said, it's very hard to understand when these systems are trying to persuade you. Now, I do think there was something that they said that was rather important. If you're trying to predict the kinds of dates to where AI systems are going to be released. One of the things they did also talk about was the fact that persuasive AI systems, they wouldn't want those systems around to maintain election integrity. Of course, right now we are at a very intense period of time because former president was recently under an assassination attempt, which is completely remarkable. The fact that that was even able to occur. And with that, if OpenAI were to release the future system GPT5 around this time, it could certainly put OpenAI in a precarious position to where certain laws and regulations are passed that impact their future development. If you haven't been paying attention to OpenAI's releases, the past releases of their previous AI systems were under heavy scrutiny and heavy criticism. For example, when GPT4 was released, it was only a month or two after that, people were saying, look, we need to slow down because this model is way too advanced and GPT5 is literally just, you know, it's just way too much. In addition to that, when SORA was released, there was also a huge outcry from many individuals saying that you guys need to literally, and this is verbatim a tweet by someone, people were literally stating that you need to go and kill yourself. Now, I think that OpenAI have to be very careful about how they release these systems because if they aren't, they're going to be in a worse position, because as these systems get more capable, they can do a lot more things and impact society even further. Which brings us to the question, how on earth are they going to release truly capable systems in the wild if they have the potential for superhuman persuasion? That is a question that I will leave to OpenAI. Now, once again, it might be seeming that China has taken the lead again. Now, this is something that comes of no surprise to me because I've been paying attention to every area of AI development and I know that China are rushing ahead in terms of the kind of things they are developing. One of the things that they've been continuing to focus on is of course, their frontier models. And recently we can see here that it seems that they've managed to catch up to GPT4O and even apparently surpassing Claude 3.5 Sonnet. Now, I would like for these models to at least be in some area where we can truly test them. I know that yes, it is very hard to get access to Chinese models such as CLING and of course sensetime, but one thing that I can't help but thinking is that many companies are brushing out these models and managing to do just marginally better on the benchmarks. Which leads me to believe that it might just be a exaggeration of the model's capabilities. Because the benchmarks are so close in terms of the differences between what the models can do, it's very hard to analyze whether or not these results are actually true. We can see that the average score is actually tied with GPT4O and of course on the MM, the MMU, the Math Vista. Apparently SenseTime 5.5 manages to surpass all previous iterations. For example, we know just how good Claude 3.5 Sonnet is. Like, I'm astounded at how smart this model is. I pretty much use it on a day to day basis and the difference between it and GPT4O, surprisingly is night and day. Which means that if this model right here is claiming to be even better than GPT4O and 3.5 Sonnet with such a remarkable difference, I truly would like to see some kind of tests or some kind of videos that can confirm this. Now, what was also interesting was we did see this small video demo. I'll leave a link to this in the description and of course I did actually try to translate this, but it just didn't sound right because the echoes and the translation just completely sounded off. But the long story short is that there was a video live demonstration of the SenseTime 4o or SenseTime 5o, which is basically their version of ChatGPT talking to you. You can see it's got a camera, it's got a voice mode, and it has an AI system that manages to talk and reply with real time accuracy. Now, I'm not sure how good this is. It's quite hard, like I said before, to get access to these systems, but one thing that really does surprise me is the rate of development of these frontier models. It would be interesting to get someone who could potentially go over there and test these models and do a one by one side by side comparison with the Western models. But until then, I guess we're going to have to take their word for it on how capable these models are. And this likely won't be the last company that manages to develop models that are this good. Gemini 1.5 Pro Long context window was recently tested and integrated with robots. They said a limited context makes it a challenge for many AI models to recall environments powered with Gemini 1.5 Pro's 1 million token context lengths. Our robots can use human instructions, video tools, and common sense reasoning to successfully find their way around a space. They took the robots to specific areas in a real world setting, highlighting key places to recall, such as Lewis's desk or temporary desk area. Then they were asked to lead us to these locations.

Hi, robot. Let's go to a tour. So this is Lewis's desk. Here we have some open separate desks for everyone.

Oh, hey, robot, I'm Lewis. Can you take me to a temporary desk?

Okay. Syncing with Gemini, please give me a minute. Let's go. This temp desk is reserved for you.

Oh, thank you. How nice. So it's clear as these models get more and more capable, they will be integrated into robots. And we're slowly starting to see the evolution of how that may occur. There was also another video demo in which they showed how they provided more multimodal instructions, such as map sketches on a whiteboard, audio requests referencing places from the tour, and visual cues like a box of toys.

Hey, robot, can you take me somewhere to draw things?

Okay, thinking with Gemini. Please give me a minute. Let's go. Congratulations, you've reached the goal. You can now draw on the whiteboard.

Thanks, robot. Hey, robot, can you take me somewhere to draw things?

Okay, thinking with Gemini. Please give me a minute. Let's go. Congratulations, you've reached the goal. You can now draw on the whiteboard.

Thanks, robot.

In more robot news, apparently the new Optimus design will be complete later this year, and it is something special. So I wonder if we're going to get an image of what that actually looks like, considering the recent developments in robotics and AI. There was also this clip here which shows an interesting evolution in terms of what many people might think about regarding robotics and AI. Essentially, if you remember, we have neuralink, which is a truly capable system that allows people to talk control systems with their thoughts. And here, Elon Musk and his team are basically talking about how you could potentially use neuralink to control certain parts of Optimus.

Generally for anything that's got a Bluetooth interface, including potentially an optimist. Yeah. Yes. You could communicate with Optimus. Yep, absolutely. Optimus will also be able to talk to Optimus, but, like, why not just beam it? But you could just. Yeah. Instead of talking, you could just beam it directly. Or if someone has lost the use of speech, then they can still communicate to an optimus. They can communicate telepathically to Optimus by Bluetooth. And so even if someone has completely lost the ability to speak, they could still control Optimus or their computer or phone.

I mean also like if you have an optimus and you have a neural correctly mapped, a brain signal to control the physical arm of the robot. And that's a very meaningful thing. Like if, you know folks that have spinal cord injury, one of the biggest requests is to be able to scratch something that quite annoying. And if you have a scratch on your face, you can't fall asleep until you scratch it. It's very convenient to be able to move something physically towards you. Similar to things like eating food if you need somebody very hard to have dinner with friends way, that sort of a normal social experience. And so if you can feed yourself, pick up a fork and actually eat a piece of chicken on your own, you know, that's a big deal. It saves a lot of interactions with caretakers and other people in your life that you rely on to take care.

I think an exciting possibility long term also is to say if you take parts of the Optimus humanoid robot and you combine that with a neural link, let's say somebody has lost their arms or legs. Well, we could actually attach an optimus arm or optimus leg and do a neural link implant so that the motor commands from your brain that would go to your biological arms, now go to your robot arms or robot legs. And again, you'd have basically cybernetic superpowers actually.

So the latency from the neuralink to your hand would probably be slightly faster than it is just to go to your physical hand. So you can imagine like if you're a piano player or anything that requires extremely fast hand movements, that you could actually have a pretty imbalanced right hand robotic arm control versus left hand physical arm control. That's one of them.

Yeah. Like I said, it's just kind of a cyberpunk deus ex future where you have cybernetic upgrades that are actually better than your biological.

Another interesting topic that I recently discussed with someone was the hyper personalization of media content due to AI may lead to a fracturing of our once shared reality and essentially everyone living in different universes.

You had product placements back in the day where everything was. All the Marvel movies had Coca Cola cans in them. And now you're going to have Marvel movies that have a unique product based on all of the data that they've collected off of you, off your cell phone or whatever. EL and that fractures reality just a little bit. And then you have all of these subsequent fractures of reality where that shared if, if a shared reality is a limited commodity, all of a sudden that's kind of the fundamental, that, that's the scariest thing at all where you can go talk to somebody across the street and they no longer live in the same universe as you because they have consumed video and media that has nothing to do with what you've seen. And so that scarcity on the back of AI is kind of baked into the way that the tools are going to be used.

You have funnily enough, Elon Musk responded to this saying that you are not wrong. And I think that this is a lot further off than people think. Not to say that this is not going to happen, but I mean, you have to think about how crazy and personalized our feeds already are. For example, all of our home feeds right now are remarkably different. Whilst yes, some of your friends might have similar interests, you might be into similar things, I wouldn't say that there is one feed that is exactly the same as everyone else's. Everyone's, you know, what everyone sees is based on their different algorithm and preferences and what they recently watched, what they watched before, what they don't like. And with that being said, you know, are we going to be living in a different reality where with certain social media apps we might have entire worlds that are just truly hyper personalized now? I do think there will always be a sense of shared reality on some extent because there's always going to be things that happen in real life that people want to know about. For example, the Trump assassination, right now there's a global IT outage. I do think that these are going to be things that people are always, you know, talking about and connected about. We all still live in the same world, but I do wonder at what point, you know, there's going to be that drift part to where people are just in their own world, watching their own TV shows and, and everything's just unique for them. It's going to be a weird future. And I think this is going to be probably something that happens in perhaps 20 years maybe. Or in fact, now that I say 20 years, it sounds so crazy, but maybe like 10 years, I would say.