AI News : Gpt4o - Mini CRUSHES Claude, Sam Altman's Aggressive New plans , 3 Years Left Until AGI

So there were actually a few big AI releases, including one from OpenAI. So let's get into some of the news that you might have missed. One of the first stories that I was actually quite Surprised by was GPT4O mini advancing the cost effective intelligence for AI. OpenAI actually released GPT4O mini, which is basically a replacement for GPT 3.5. So you can see that yesterday they essentially announced GPT4O mini, which IS their most cost efficient small model. They actually expect the GPT4O mini. It will significantly expand the range of applications by making AI intelligence much more affordable. One of the most surprising things about GPT4O mini was that it scores 82% on the MMLU and currently outperforms GPT4 on chat preferences in the chatbot arena or the LMSYS leaderboard. The price of this system is actually relatively cheap considering the fact that the cost of intelligence has been decreasing rapidly. It's priced at $0.15 per input tokens and $0.60 per million output tokens and an order of magnitude more affordable than previous frontier models and 60% cheaper than GPT 3.5 Turbo. Now, the reason that I think this is such a game changing statement and most people won't realize the significance of such a release, is the fact that when small cost efficient models are released very cheaply, this allows the AI ecosystem to deploy more products cost effectively. One of the biggest problems preventing many people from deploying GPT applications and from many people actually using ChatGPT or other AI systems is the fact that it's just too expensive. If you've ever tried to build something with AI, the problem is, is that the cost per output just racks up an insane bill that you really can't afford to justify having your software or whatever AI system it is. But with GPT4O mini, you're actually not compromising on the intelligence of your model or the cost of your model. So you're getting a reduction in price and an increase in the reliability and of course the intelligence of the model. Which means that now if there's some kind of software development or some kind of application that you previously couldn't find a reason to justify spending that amount of money or you actually losing money, this might not be the case. And I think this shows us an indication of the further trend in the industry of what's to come. Now, what they also talk about is how GPT4O mini enables a broad range of tasks with low cost latency and applications that chain or paralyze multiple model calls for example calling multiple APIs, passing a large volume of text, for example a full code base or conversation history, or to interact with customers through fast real time responses, for example customer chatbots. Now what's really cool about this is that it currently supports text and vision in the API and it does say that image, video and audio inputs and outputs are going to be coming in the future. That I think is going to be the real game changer for this model. Now something that was actually really interesting was if we take a look at the model evaluation scores, now what they've done is they haven't actually compared this to other frontier models, they've actually compared this to other models that are actually smaller in size, basically comparing them to the respective similar models from the other AI companies. So what we can see here, GPT4O mini is compared to Gemini Flash, Claude Haiku and of course the predecessor GPT 3.5 Turbo. We can see that it is a stark jump in terms of all of the capabilities, because GPT 3.5 just simply didn't compare to Gemini Flash and Claude Haiku. But overall, what we can see here, in a rather surprising turn of events, is that GPT4O mini clearly surpasses the previous state of the art models in terms of their cost effectiveness and their size. On the mmlu, it surpasses Gemini Flash and Claude Haiku on the drop benchmarks, on the gqa, on the mgsm, on the math benchmark. Surprisingly, we can see that there is a huge improvement in terms of what you get for cost effectiveness compared to these other models. In the human evar, that's also the case. And in the MMMU and in the Math Vista we can see that once again there are also surprising benchmarks and Gemini Flash does actually manage to beat GPT4OH mini there. The point being that it's clear that OpenAI probably did use a different method to train this model because the effectiveness for the size of the model, although they didn't actually state it, shows us that we're going to be getting a lot more intelligence out of a lot faster systems in the future. And I do wonder how this is going to look in the next two years with respect to what intelligence level we can get per cost per input. Now, if you do want to, you can actually use GPT4OH mini today. But for example, just to show you the relative speed of this, I said can you list me 20 food items? That was insanely quick and I'm going to do a video later on the actual applications of this Because I think most people miss what you could actually do with such a model. But this is going to be something that I'll be diving into later because it is really effective with what you can do. When a model can respond that quickly with a level of intelligence that's rather effective.

We're all very hungry. Maybe that's how I put it. You know, we could have, Yanis and I could have stayed and tried pushing agents, you know, at DeepMind. And as I said, I think, you know, the reason we decided to do it in our own way is because we, we think we can move quickly and much faster against this goal. Some of this urgency is driven by a, a real belief that we are, we are three or so years away from something that resembles a digital AGI. And by that, that's a, that's what I've been referring to as a universal agent. Something has both the, this kind of breadth and depth of knowledge, and that means we're actually on a very accelerated timeline. Yeah, you're, you know, a few months in, you're kind of 5% away from, from hitting that timeline. And maybe some of this is also driven by how quickly AlphaGo went from experts in the field doubting this is possible. Yeah, things kind of decades human level, or like expert human level, Go play was decades away. And how effectively they were able to solve that problem within months, I think we're seeing a similar kind of acceleration happen.

So that was Misha Laskin, someone who is previously working at Google DeepMind. They went off to start their own company and, and you can see that he's speaking about the AI timeline. And I think it's always refreshing to kind of hear what some of the industry insiders actually think about the timelines, rather than a few skeptics on podcasts that are just talking about AI very briefly. And when I say that, I'm talking about news anchors and other things, and of course, I do think that they are important for the discussion, getting it out into the wider community. But what I don't think people understand is that if the timelines are as what we're looking at here, which is, you know, three years, currently it's 2024, this would be 2027. This would mark, I think around now, the sixth or seventh person that hails 2027 to 2028 as the time when we're going to be getting a truly capable system that can actually act and do things within the environment. Now, I think that that is a startling revelation because it shows us at the rate at which AI is expected to increase. And the funny thing is that in the past people have actually been wrong at how quick these kind of systems have advanced. So we could even argue that perhaps even these indications are somewhat conservative. Now, what that means is that within the next five years, I do think that there is going to be so many different breakthroughs on so many different levels and we are going to be getting more and more capable systems out there. But I do wonder if people are ever truly going to realize that what is coming, because the capabilities of these systems are going to have some really profound implications beyond just fancy chatbots. And whilst those chatbots are being developed, and whilst obviously other generally intelligent systems are being developed, we're also going to see a lot more money flowing into AI around the space as well. Now, as for those of you who don't think that maybe AI is going to be advancing as quickly as former Googlers think, we can see here that this is going to put even further pressure on on AI development, as Trump's allies are working on a plan to create Manhattan Projects for AI military tech. So essentially, Trump's allies are reportedly drafting plans for an executive order on AI. Basically, if you don't know, Donald Trump is running to become president again and it is very likely that he will win this election. And right now all eyes are on his plans for the largest developments with artificial intelligence, because of course he has some unique ideas on where things are going to go. According to the Washington Post, several people close to this former president are writing up an AI executive order to advance US interest in the technology. It will include the creation of industry led agencies that will study AI models and protect them from foreign powers, the report said. The framework also contains a Make America first in AI which aligns with the 2016 Trump administration comments to strengthening American leadership in the field. This order demands that major tech companies communicate the risks of their AI models with the federal government and it also limits the government's use of AI systems in high risk situations and includes programs to potentially study harmful AI in healthcare practices. Now, I do think this is going to happen anyways, as I've said before, when these elected leaders start to truly understand the magnitude of what the potential of AI is. I don't think there's going to be any question whether or not there are going to be Manhattan projects popping up in various parts of the world, not just America. And I think that the research landscape for AI could most definitely change as we get to more and more powerful systems. As in by 2030, maybe by 2035, the level of research that we're seeing published in certain areas might actually grind to a halt. This doesn't mean the level of innovation. It just means what we're publicly seeing and hearing from some of the frontier labs. What's also interesting is that it seems that OpenAI is continuing to talk to Broadcom about developing a new AI chip. One of the major winners of the AI boom was Nvidia. They've gained a trillion dollars in market cap and profited heavily from the rise of artificial intelligence. Considering the fact that many people need these graphics cards to train their models now, we do know that Sam Altman was doing this last year. Last year he decided to start a new company that could develop and produce a new AI chip and help set up chip factories and make them and data centers to house them. But now his plans have slightly changed what he's been doing. He's actually been hiring former members of a Google unit that actually produces Google's AI chip, the Tensor Processing Unit, and has actually sought to develop an AI server chip, according to three people who have been involved in the conversation. And he's also been talking to chip designers, including Broadcom, about working on the chip, according to certain people. Now, what's crazy about this is that it does say here that a new server chip that would rival the kind made by Nvidia is a long shot that would take years to come to fruition. And in trying to develop a chip, OpenAI risks upsetting Nvidia, OpenAI's most important chip supplier. But it could provide OpenAI with potential leverage in future pricing negotiations with the company. Now, me personally, I'm not sure why OpenAI is trying to do this. Maybe they're just trying to do this because they realize that one day they'll have superintelligent AI, and then super intelligent AI could in house, push that chip fabrication company forward and they could have the best chips on the planet. I think that might be the play here. Although it does sound a little bit crazy and although it does seem like that would be a very long term plan, I'm pretty sure Sam Altman is a very long term thinker. And the reason that I do state that is because these AI chips are actually really hard to design. There's a reason Nvidia is the market leader. It's because they're the very best and have the entire infrastructure, the integrations, the partnerships, the trade routes. Simply everything is in their favor to be benefiting from the AI boom. You can see here that some People don't even believe this, he says. But Altman's factory expansions ambitions struck numerous executives as improbable because it would require a lot of capital and specialized labor. Altman was too aggressive for me to believe, TSMC's CEO Wei said in a news conference last month when asked about the prospect of new factories. Like I said before, this is no walk in the park even for the very experienced entrepreneurs now. OpenAI's co founder actually had an interesting statement on the future of evolving AI.

In case of cells, it's actually possible to get the cancer and cancer is a prevalent phenomenon in the nature. Once then there is an increased number of AIs and some sense, you know, there is a process of mutation which is AIs are modifying its own code, humans are modifying their own code. There is a process of natural selection. And you can say that the AI that literally wants to maximally speak spread will be the one that will exist. The things in the universe that want to replicate are the things that exist here. The main difference is that AI will have just huge power.

You could imagine a stable equilibrium, right, where things aren't allowed to consume all the resources, for example, right? Or am I missing something?

It might be the case that other AIs are kind of defending the system or maybe they are never, they were never trained for defending. And it's very hard to predict the dynamics in multi agent setup. With one AI, you can maybe predict what, what are the possibilities. It would be still, let's say extremely hard. But once you have many of them competing in some sense for resources, very hard to say actually what might be the consequences.

Now I know this might not count as AI news, but genuinely this is something that I believe most people should take a look at because it shows what, what you can do when you actually use AI generated software to produce something interesting. I've looked at tons of different AI videos and this is by far one of the best series I've ever seen. I'm not going to play to you guys the entire clip because I want you to actually check out the original creation, but I will show you guys 30 seconds from an AI generated TV show, which might show you guys how the future of entertainment is going to change with regards to enabling solo creators to produce full fledged TV shows that they otherwise wouldn't be able to access.

On tonight's episode of Unanswered Oddities. Humans. Are they real or just a youngling's fairy tale? Humans have been a great mystery of our time. These creatures of law supposedly ate potatoes and enjoyed sleep. They're thought to be long extinct, if they ever existed at all. But recent sightings all across the world have drummed up the debate once again. Rural Morg, two hours outside of Deedle, a hiker claims to have seen a human.

So, yeah, it was. It was a Tuesday night. I had just gotten home from spitting. I went out for my midnight lurk. You know, I was just being creepy and weird, and that's when I saw it. It was a hump.

So for me, when I actually saw this, I was thoroughly impressed because I've seen a lot of different AI videos, and this was one that actually made me smile and. And I gotta be honest, I was truly engaged at the thought of watching a full episode of this. Now, once things become more streamlined and there's a lot of plug and play, I do think we might get automated TV shows. But for those of you guys who think this was written by an AI, it actually wasn't. It was written by a human, with the production basis actually being AI. So I do think that you guys should check this out. As always, I'll leave a link to everything and hopefully you have a wonderful day.