AI NEWS : Google Catches Up To OpenAI, 2025 Predictions

So there has been many AI announcements that were missed and overshadowed by OpenAI's complete drama of a week. So I'm going to give you guys the quick rundown on many other things you miss. So without further ado, let's take a look at exactly what's going on here. So one of the things that has actually happened and I can't believe I missed this, well, I didn't miss it, so I just didn't have time to make a video on this. But what actually happened was that Google Gemini actually upgraded to Gemini 1.5 Pro significantly. And I think this was by far one of the biggest model upgrades that we got with incredible stats across the board. And I think this was overshadowed completely. So basically they released Gemini 1.5 Pro 0.02 and Gemini 1.5 Flash 002, which is basically a second iteration of their Gemini 1.5 Pro and another iteration of their Gemini 1.5 Flash. It is a little bit confusing because we have 1.5 Pro0.02, so it's like Gemini 1.52, which is a little bit weird. But the long story short is that they just updated these models just like how ChatGPT gets updates from time to time. Google Gemini are doing the same. Now. I think the craziest thing about this is that if you use Gemini on a day to day basis, the long story short is that I'm going to just put this in a full screen is that with the current benchmarks, what we basically have is basically have Gemini Flash currently is basically performing at the rate of previous Gemini Pro. And if you aren't familiar with Gemini Flash, Gemini 1.5 Flash is essentially a model that is basically a faster and lightweight version that accomplishes the same thing. So we can see that the Gemini 1.5 plash on the MMLU Pro compared to gemini, there's a 2% difference on the coding benchmarks, there's like a 3% difference. On math, it actually is better than that. On the hidden math, it's better than that. On the reasoning, it's a lot better. So this Gemini 1.5 Flash version is basically a lot better than the version of Gemini 1.5 Pro that you were currently used to using. So this is going to be one of those things that is quite surprising considering the fact that we have a situation on our hands. We now have this entire model that can do image recognition, video recognition, long context recognition at the rate that it used to, but just for a fraction of the cent. So for those of you who have been building with Gemini 1.5 Pro. This is going to be a remarkable achievement because this is going to be something that reduces your cost. So for those of you who wanted to experiment with models, this is something really good because as you know, models are continuingly getting cheaper and cheaper and this is just something that's really great. Now, of course we do have the fact that these models are quite similar, but let's not discount the fact that of course, Gemini 1.5 Pro, the new version, is actually catching up to OpenAI. You can see that on the MMR upro. This is 75%, a really nice jump on the coding natural to code benchmark. This is 85%, another nice percentage jump. And then surprisingly, one of the biggest jumps that we got was on the math benchmark. So for those of you that think Google's models are not good at math, now you can see they're currently at 86. And then of course, hidden math 52. The reasoning, the GPQA, this is up at 59.1%. So I can't wait for the arena votes to roll around in. I haven't checked them just yet, but this is a remarkable improvement. And I think it goes to show that Google's culture around being a company that doesn't really ship as much as it used to is changing quite a lot. They are now shipping a lot more useful stuff. And of course, for those of you who are using Google Gemini Pro for long context evaluations, you can see that their long context evaluations in terms of the, you know, diagnostics, has gone up to 82%. That isn't the needle in the haystack thing. This is just something that's completely different. So this is going to be something that's remarkably impressive considering how good Google are at what they do. Now, if you're actually wondering what Google CEO is up to, the CEO Sundar Pichai actually says that Google are scaling up their compute infrastructure and working on 1 GW data centers while exploring options for powering them, including small modular nuclear reactors. So these guys are seriously increasing their ability to generate power for their models and for their compute stack. Because in the future, compute is going to be something that is so valuable that companies that have proprietary compute and companies that have as much compute as possible are going to be some of the most valuable companies in the world, I think. So take a look, because we don't often get clips of Sundar Pichai speaking other than in areas where it's like Google I O's event where it's usually prerecorded. So this is something that's pretty cool.

In the short term, it is challenging. In the medium to long term, I'm optimistic because I think it's also bringing a lot of capital investment to, you know, developing new sources of energy. Right. We invested very early in wind and solar because we saw the opportunity there. And today many of our biggest data centers operate at around 90% carbon free basis. Right. And we recently in our Nevada data center, we are running it now with using geothermal. Right. And so we brought a new source and we want others to do that too. I see the amount of money going into SMRs, you know, this small modular reactors for nuclear energy. And so when I look at the capital and the innovation going in, I'm optimistic in the medium to long term.

Now, Google are not the only people currently focusing on building huge data centers for powering AGI and artificial superintelligence. We can see that OpenAI pitched the white House on unprecedented data center buildout. So it says startup calls for more US data centers as large as 5 gigawatts. That's five times what Google are demanding. And there's something really funny about this that I'm going to show you in a moment but following a recent meeting at the White House which was attended by OpenAI CEO Sam Altman and other tech leaders, the startup shared a document with government officials outlining economic and national security benefits of building a 5 gigawatt data center in various US states. And based on an analysis the company engaged with outside experts on to put that in context, 5 gigawatts is roughly the equivalent of 5 nuclear reactors or enough to power 3 million homes. So yeah, this is absolutely insane. This is actually pretty good because of course, you know, it does boost new jobs, boost GDP and of course ensuring that the US can maintain its lead in AI development. So I mean the things that I've heard Sam Altman speak about, he's actually said that, you know, compute is going to be one of the most valuable resources in the future. That's why I'm saying that companies manage to build out the infrastructure now, they're going to be benefiting from having all of this rather than having to outsource it. But this is pretty crazy. So it says OpenAI is actively working to strengthen AI infrastructure in the United States, which we believe is critical to keeping America at the forefront of global innovation and boosting reindustrialization across the country. See, I think this is needed because whilst most people are currently thinking about AI in its Current state. I still think that we underestimate how much power AI uses. And, and if we actually think about the future in terms of, okay, the future is going to be mainly powered by AI. We know that currently, yes, we can look at systems like ChatGPT that score like 300 million users every single month. When we look at the future, the future, we know that there's gonna be billions of people demanding useful AI. And if you don't have the infrastructure to power that, then you're just gonna be left behind. So I think that's why the build out is there, because whilst yes, we're currently in this AI bubble and we think, okay, AI is every day, you know, everyone knows about AI. Not really. Like, you'd be surprised to, you know, how early it still is. I think within the next two to three years, it's probably going to get a little bit more obvious what AI is going to be capable of doing once we get agents or things like that. But I think it's always better to be prepared for the demand. Also, I wanted to show you guys this as well. So we can see here the cost of tokens falling. And this is crazy because this is actually going to bring me on to my next point. But yeah, the cost of these tokens are actually. In fact, you know what, let me actually show you guys this other thing first, because before we talk about the cost of tokens falling, I want to show you guys that the New York Times Times actually got ahead of the documents from OpenAI's recent funding round and it shows that 350 million people use Chat GPT in August. They've got huge, you know, user growth after they had the anonymous login thing, which is basically where they let people use the platform anonymously. And the subscription apparently is going up to, you know, it's going to go up by $2 by the end of the year. So it could be the Chat GPT is $22 by the end of the year. And in 2029, ChatGPT is largely going to be $44 a month, which is rather interesting. Now, I personally have thought that ChatGPT is really cheap for what it is. And I'm someone who, you know, when you run a business or you run a YouTube channel or whatever you do, I think that you can't underestimate the value of having a second opinion that you can continually jot ideas back and forth with. I think something like this is just completely underestimated. But for the average person, I don't think they see the value in what ChatGPT is.

I think unless you're dealing with certain problems on a day to day basis, I don't think it's probably worth the $20 a month. But I wouldn't be surprised that like I said before, and OpenAI has discussed this, that the future models like Orion and 01 Preview are going to be going up to not $2,000 a month, which they have discussed, but they're going to be going up quite substantially. Because I think what the ChatGPT subscription is going to be, what they're probably going to have is they're probably going to have something where they've got Sora as like an add on, probably got like, you know, the image model area as an add on, like the most advanced reasoning model as an add on. And they're probably going to have agents as an add on, probably all for like $50 a month, $60 a month. And independently they're probably all going to be like maybe $20 a month each if you just want to have individual specific services. Because One thing that OpenAI has is they've got a lot of individual things and I think they could, you know, definitely make a whole load of money out of each one of those by just applying them to different areas. Now you can additionally, as I was saying before, the price of OpenAI's tokens are dropping substantially, which does mean that this company is going to be having more and more profit over the long term. I know many people are thinking that, oh, this company, they're losing money, they're hemorrhaging cash. But companies like Uber, companies like Google lost a lot of money in the early days, so this is not uncommon by any means. So fear mongering, yes, OpenAI has got terrible management in its current state, but the cashing isn't really an issue at all.

A really important concept to understand is that AI technology has been the greatest cost appreciating technology ever invented. So you know, when we launched our models, they're pretty expensive to use. They were $36 per 1 million token. Today our models are basically for a leading model is 25 cents. So we're almost getting to the point where like AI is becoming free. Like this technology is depreciating so much that this will really change how you'll be able to apply AI no matter what you're doing. This creates whole powerful different examples. You know, a quick way that we do it at OpenAI, we get thousands of people filling out forms every single day. These are really hard informations. For us to understand, like what information we should actually listen to and read to versus what we should ignore. Most people don't do these or add kind of text fields on forms because no one actually reads them. We're able to use actually natural language processing in our automation where it reads every single form, can actually call out things that are interesting for us that we know are going to be things that we want to talk about, grade every single response and then automatically route that to the right person in the company for who we should talk about. So we completely automated every single way that we collect information coming into the company.

Then of course we actually did have Alpha Chip. So Alpha chip was incredible. From helping design state of the art CPUs, building AI models to CPUs and data centers, basically AI transformed the way that they designed microchips. You can see right here that in 2019, the first of its kind, AI using reinforcement learning to design the physical layout of a computer chips. And this can take human engineers months.

Of work and they manage to do.

It with reinforcement learning. Then in 2020, Alpha Chip begins powering the next generation of Google TPUs, their custom designed AI accelerators optimized for training large AI models. And then of course alpha chip code is open source, helping the research community unlock its potential. Then of course we have Google introducing Trillium and designed by Alpha chips. And this is a state of the art TPU which is a specialized chip which are essential to building AI models. Now all of this might be a little bit confusing, but basically I'm going to show you guys how Demisus Harbors puts it. Basically puts it as this feedback loop where you train a state of the art chip design model which is Alpha chip, then use it to design better AI chips, then use those better chips to train better models which design better chips. And this is part of the reason that their TPU stack is so good. So that's currently the cycle that they have. Every iteration cycle cycle gets faster because you get AI that can design better chips. And because it couldn't design better chips, the next cycle just becomes a little bit faster. So overall we have, I wouldn't say recursive self improvement kind of, but just a slower version of that, which is really cool to see from Google. Like trust me guys, they are not to be underestimated at all. Then we also got the X CEO of Google giving some rather fascinating predictions about AI for the next year, which is 2025. So take a listen, does what you wanted.

It actually writes the code and there's Evidence that this, in my lifetime will probably replace most programmers, which is bad news for programmers, but good news, good news for the rest of you because programmers are kind of obnoxious and expensive. Speaking as a computer scientist programmer. So those three, which I think are occurring this year, set up a very interesting set of questions. And I'm just going to assert that, that those three are going to occur. They're going to happen, you're going to use them, they'll be powerful, you'll like them.

So I think this talk is rather fascinating because it shows us quite a lot. And once again, I mean, it seems like it's a tough time to be a programmer. You know, you've got this burning question of, you know, should I stay in this industry? Should I pivot? What do I do? I think. I think it's quite a tough time because all the top leaders are constantly saying that this is going to replace you. And we also have Jensen Huang actually saying this as well. He's actually speaking about, you know, AI won't replace you, but someone using an AI will. And I think that's true for the most part, but in some cases that might not be true.

To anybody who is concerned about an AI taking their job, you should probably worry about someone who uses AI taking your job.

Yeah.

And so that reminds you to get going and to go learn this new tool and to make this new tool, you know, your advantage. And surely as a company, we're going to do that.

Yeah.

And as a country, we should do that.

Now, lastly, we also do have this, which is Alibaba presenting Mimo Controllable Character Video Synthesis with Spatial Decomposed Modeling. Basically, you can input any kind of image you want and then you can use that image to replace any kind of video. We first saw this with viggle AI. They used, I'm guessing, an earlier iteration of this. But you can see that now with this kind of technology, it just looks remarkably more effective. I'm not sure where you're going to use this, maybe for previs. I think this would be quite useful in certain scenarios. But you can see right here that this is novel 3D motion control, which looks remarkably effective. I mean, this kind of stuff is just slowly churning on in the background of many different, you know, companies. And what I mean by that is, I mean that you won't necessarily hear about this kind of news, but it is something that is continually getting better with every single update. So if you guys did enjoy today's video, don't forget to, like, let me know what you thought about the stories. And I will see you guys in the next amazing video.