AI News : OpenAI Is Three Steps Ahead, Microsofts Agents, Teslas New Autonomy, Driverless Cars...

So there was actually so much news.

This week in AI that I nearly missed some of it. So I want to try to do my best to cover absolutely everything, but let's not waste any time. One of the craziest stories that sets the precedent for the scale of AI that is coming was the interview with OpenAI CPO Kevin Whale on the future of AI.

In this long winded interview he gives.

A detailed explanation of OpenAI and I thought it was super insightful.

Now what he says here is that.

He says that the O1 reasoning model is only at the GPT2 level, so it will improve very quickly. And they are at the very beginning of scaling the inference time compute paradigm. And by the time competitors manage to catch up to 01, they will be three steps ahead. Now I could probably say that this.

Is the biggest news this week because.

I think people are underestimating the significance of the statement. Most of you guys don't know what GPT2 is, but think about it like this.

Do you remember the jump from GPT3 to GPT4?

Now this graphic is quite likely not accurate. I can't count the numbers, but I do know that GPT4 has 1.8 trillion and GPT5 has 170 billion. The point being is that the level of scale, okay, and remember this diagram, okay, and GPT2 was even smaller than this.

Okay? So it was.

If you think about GPT2 as 01 and we're going to get a GPT3 model, which is going to probably be an O2 size model. Imagine what happens when we get to an O4 size model. That's the kind of thing that simply blows my mind. And considering the fact that we're just on the cusp of this paradigm, I'm wondering how many companies are going to try to catch up to OpenAI. And the more important question is where are OpenAI going to be once those companies manage to catch up?

I assume that at some point other labs will figure out the techniques, the ways that we built O1. Some of this stuff diffuses over time and it's our job that by the time they do that, we're already three steps ahead. And we really do feel like with O1 in particular the techniques we're kind of at the, and I mentioned it's a different paradigm of scaling. It's not this massive pre training run, it's inference time scaling. We think we're at the GPT2 sort of phase of that and we see a lot of low hanging fruit that we can go pluck to make these models better very quickly. So by the time people do catch up, we're going to try and be three more steps ahead. I assume that at some point now.

Some of you guys are going to.

Say that that's hype, but I think that they are largely right, considering that this is an entirely new paradigm, one that even led Ilya Sutskova to start his own company. But Sam Altman has been teasing many individuals on Twitt and that is in reference to this tweet where he says, well, ChatGPT's second birthday is next month. What should we give it for a birthday present? Now, next month will be November and if that is going to be ChatGPT's birthday, then we could potentially be getting an update from this company. Now, I think that GPT5 might be likely considering the fact that OpenAI do have a lot of products and many different companies are gearing up to release newer models in this last quarter. Some of you guys might not know, but there are some companies that are slowly prepping because they know other companies are also prepping to release big models. Now, if you look back at this tweet from Sam Altman earlier this year, well, not earlier this year, earlier last year, you'll see that he actually spoke about the things that that were going to happen in 2024. Now with this, I think this was truly fascinating because a lot of the stuff here did really come. Of course, besides AGI, pretty much everything really did come. We got video, we got personalization, better browsing, sign in with OpenAI that didn't really come. I'm not sure if it is. We got better reasoning, better GPT, higher rate limits. The only thing that we really didn't get okay was of course GPT5. Now GPT5, do I think they were going to release it? I'm not entirely sure if they even need to release GPT5 at this point. I think what OpenAI will do is.

They will see what their competition does.

And as per usual, what OpenAI might.

Just do is it might just wait for their model to see if it.

Beats the benchmarks and then update that model and then just show that on the arena because that's what they've done before. They've previously just done these small time updates to the models where they just edge out the other models. Now if you're thinking about what on.

Earth other models are going to be coming, one of the models that could.

Be coming is of course Claude 3.5 opus.

Now the legendary Jimmy Apples has basically stated that look, we won't be getting Claude 3.5 opus, but that doesn't mean we're not going to be getting a.

Different kind of model.

You can see here he says further.

Info, not Opus, an API related agent.

Uses a computer and generates clicks. Not sure to be disappointed or still excited. Let's see. So I think this is going to be truly fascinating if this does occur because most people are going to be expecting a reasoning level model from Anthropic, but now it seems that we might just be getting a kind of model that is a little bit different. Many companies are starting to realize that next year will largely be the year for agents and of course in doing so, a lot of companies want to.

Have that limelight for themselves.

Currently, OpenAI hasn't taken the cake for the best autonomous AI agent system. Currently there's no one, so it's free for the taking. So if Anthropic can successfully demonstrate a working AI agent that is usable via the API, I think that's going to be really incredible because they did talk about something like this before. So I'm going to be really intrigued to see exactly what Anthropic are going to be cooking up because Claude 3.5 Sonic is still one of the best models out there that I use on a day to day basis. Now the crazy thing is that OpenAI's agents, like I said before, nobody had taken the cake of course, until today when we had Satya Nadella actually announce Copilot agents. He says today we announce new autonomous agent capabilities across Copilot Studio and Dynamics 365 to help scale the impact of every individual team and business function. Now this is something I did cover in an extended video, but this was something that I wouldn't say it caught me by surprise because I did know that, you know, Microsoft were working on this, but I didn't think that they would do the announcement today and talk about a public release next month. Now basically what this is is that this is where you have agents that are really effective at picking up the slack that humans wouldn't really be most effective at. So if you realize what humans are good at, humans aren't really good at, you know, responding for long periods of times to tedious tasks that can be done by pretty much anyone. And it's not really a good use of human cognition. So these autonomous agen basically pick up the slack and help teams run more efficiently. And this was something that was rather surprising considering the fact that we did think that this was going to come later in 2025. Now, of course, these autonomous agents, as you guys can see on this web page right here, these are made possible because of the fact that we now have the O1 reasoning model. If you remember, the O1 reasoning model wasn't just a model that was reasoning for math, science, and of course physics and any other difficult subject. This was a model that was truly designed to be the backbone for agents. One of the problems that we've had with GPT4O was that the model just didn't generate good reasoning steps and would often make mistakes. But now with a model like O1 preview, and of course the O1, we know that those models can be used as the actual infrastructure that we use for autonomous agents. You can see here it talks about autonomous triggers. So if there's something in your business, these AI autonomous agents can see that, and then like a click of a button, they can immediately respond to that and then go off and do a bunch of tasks. One of the craziest things about this is that they've got this dynamic agent plan. Each business process can have different paths, since agents create dynamic plans on the fly to handle and complete tasks. And you can view exactly what these agents doing, the kind of paths they're going to take, and exactly why this agent chose a particular method, its design process, yada yada yada. And of course, you can see the activity overview for everything that your agent has done. So if you'd like to see where your agent's made a mistake, what channels it has done, if the status is complete, and if it's going to do something else, this is where you can see the entire activity overview.

I personally think that this is going.

To be pretty crazy considering the fact that we're not yet seeing this at scale and it's currently in private use. But the early results from this are pretty promising. And what's crazy, and I spoke about this, but you can see here that Microsoft says, at Microsoft, we are using copilot and agents. So even Microsoft themselves have said that, that they are using AI agents. Which means that, look, if these guys are using it, you know, one of the top companies in the world, then it's definitely going to have an impact downstream when they're released more wildly. Now, that wasn't the only thing that Satya Nadella spoke about. One of the things that he actually spoke about was the fact that AI development is being optimized by OpenAI's 01 model and has entered a recursive phase. We're using AI to build AI tools.

To build better AI.

If you aren't familiar with why this is such broad statement, it's because there has been this really scary concept known as recursive self improvement. It's basically the idea that once we get an AI smart enough, it's going to be smart enough to improve its own training process, which can then improve the overall smartness of the model, leading to an infinite cycle where this model's intelligence just absolutely goes through an insane graph. So this is human progress, but this is the graph that many people are suspecting will essentially happen. Once you get to this tripwire phase where we set off this artificial superintelligence, think of it like this. Let's say you were a robot and you were smart enough to improve your training process. Now instead of, let's say a month for you to train your own model, train your brain, your memory and everything, it takes you 14 days now, because you're now improving at a rate of every 14 days, you can improve the cycle once again. And let's say you shorten it down to seven days, so now you're even smarter. And seven days in, you are now improving yourself at double the rate. So before you're improving once a month, now you're improving every two weeks, now you're improving every seven days and you're even smarter. And because you're even smarter, you can find a way to get that improvement down even more. And the thing is, is that this can keep happening. You can keep improving yourself recursively to the point where the intelligence is, has no limits. And I think this is one of the things that, whilst yes, it does sound crazy that a model could potentially do this, you have to remember that certain people out there, like Elias, that's trying to create artificial superintelligence, that's literally their goal right now. And that guy is a genius. So I wouldn't bet against them to create artificial superintelligence. And it's going to be really hard to sort of tame that kind of thing anyways.

One of the coolest things I have seen seen recently is with O1 coming to GPD or other to GitHub Copilot, you can sort of use AI to do the next level of optimization. I think what you have on the slides, you know, behind me is the auto encoder we use for GitHub copilot is being optimized by O. So think about the recursiveness of it, which is we're using AI to build AI tools to Build better AI. So that's sort of, you know, so it's, it's just a new frontier.

Now this is major AI news, but for those of you who are interested in videos, we've got the artificial analysis releasing their video arena, the first crowdsources comparison for text to video models. So it talks about how text to video is accelerating pretty rapidly and honestly guys, it is advancing more rapidly than I can even keep up with. I haven't even done a video on Minimax Haylou AI, which is pretty insane. And I haven't done an entire video on Cling's AI stuff because I just simply haven't had the time. These models are just being released at a rate to which even I can't keep up. And essentially what this is, is this is a, you know, arena where you can actually look and compare which models are better. I think this is one of the best things to do because when you look at how we've previously analyz statements between large language models, we could always see immediately which one responded the best. Now with this, you can see here that we do have a situation where we can do the same now with videos. So if you're wondering which video is going to be the best for you in terms of whatever use case you do have, you can use the text to video arena and input your prompt and see which one works the best.

So one of the examples I wanted.

To show you all here is, was.

This comparison done by AI tech Realm on Twitter. So thank you very much for that because I asked them to just essentially.

Use the same driving image for various different video platforms.

Now I think this is one of the fairest comparisons that you can see because all of the models are given the same driving image, the same prompt.

So we can see how these different.

Video models have their internal world models in terms of video representation. And I think things like this are really important to actually see where certain models excel. For me personally, I don't use text.

To video that much, but I do.

Find it useful when I can compare different AI tools because it allows me to understand different models for different use cases. For example, I use Gemini in certain use cases where it beats everything else. And I use claw 3.5 sonnet in.

Use cases where it beats everything else. So I think if you're trying to.

Get more out of AI, doing constant comparisons in things like Chat arena or.

Whatever is something that you can really.

Utilize to ensure that your use case isn't just bound to one model. Because some people would think that, oh.

Chatgpt struggles with this.

No other AI is ever going to get it when that's not really the case. Trust me, you'd be surprised what is.

Out there now if you're actually trying.

To think about the future. There was a fascinating talk with Jensen Kwan, the CEO of Nvidia, and he says that a number of industries soon will go from zero to billion dollar markets. And these markets are autonomous vehicles, humanoid robotics, quantum computing and digital biology. But he talks about this in 10 to 15 year timeframes.

How many $0 billion markets are there left out there?

There's a whole bunch. One of my favorite ones right now.

Av, I'm not going to give up on that.

We're now, we're now just R and D, budget wise, probably 6, $7 billion into autonomous vehicles. And we will absolutely ship great cars with Mercedes and JR and a few others after that. So, number one, number two, humanoid robotics, which is a cousin of the autonomous people. There'll be billions of these shipped. Zero today, billions someday. And then we're working on the next generation of computing, quantum classical computing. And so we have a new market.

My question to you, the audience, is which of these do you think is going to be the billion dollar market? First, is it going to be autonomous vehicles, humanoid robots, quantum computing? My best bet is on autonomous vehicles because that has a ton of different applications.

And once you get that to work.

You can easily deploy that to any major city. And those companies are going to be saving so much money on drivers. And if you've seen the experience from.

Waymo, it looks remarkable.

That brings me to my next point, which is something that I'm pretty sure most people did miss. But essentially, if you are familiar with.

Waymo, which is really good at autonomous.

Driving, there is this small company in the UK that is a leading developer of embodied intelligence for AI vehicle autonomous vehicles. And basically, they managed to do 75 minutes of uninterrupted autonomous driving in busy central London.

Now, I'm someone who drives in London.

From time to time, and I can.

Tell you you do need to keep.

Your eyes peeled constantly because there's people.

Weaving in and out of traffic. And if you're from America, you. What you might not understand about the.

UK is that the thing with this kind of driving, although it might look unimpressive, you have to understand that oftentimes some people just run into the road. You see, jaywalking isn't a crime in the uk. People can cross as they please. So that means that at any time you could actually run into Someone. Now, fortunately that's never happened to me, but I've seen a number of close calls. So getting an AI, robot machine driver, whatever you want to call it, to work this well in a dense city is something that is a remarkable achievement. And it goes to show that many companies across the board are edging forward in terms of their success. There was also this, which is the truth terminal. This was basically where an AI discovered a religion, turned it into a meme coin and managed to take that meme coin market cap to over a billion dollars or 400 million, 400 million dollar market cap.

It was pretty crazy.

It's pretty crazy story. I think it's still developing. But there was a recent update on this story where the LLM made a mistake and they thought that it was a person that was typing behind the screen and so the coin dropped 50%. Now, typos do happen in LLMs. They are pretty rare. But this is something that I think was pretty funny considering the fact that most people don't understand how these ars work. But I think the AI agent trend is going to be one that is a lot more prominent in the future when we have various different personalities and characters that are just on social media saying different things. And the crazy thing about it is that they're going to be able to interact with people 24 7. So this kind of AI agent that manages to make the initial user a million dollars where the wallet was just continually increasing in size due to it holding a lot of the goat token, I think we're going to start to see that trend happen a lot more because this is something that is not revolutionary. But the weirdest narratives play out in crypto.

And I think, as Elon Musk said.

The most unusual outcome is the most unlikely.

I'm not giving a direct quote, but.

He said something along those lines.

But it's pretty crazy. We had LLMs talking each other in.

The infinite backrooms and developing some crazy theories about life. I mean it was just genuinely one of the insane stories.

Now I made a full video on.

It which I will link in the description. But this was something that is just completely surprising. Now my question to you is, what do you think about this? Because this is coming sooner than you realize. And Microsoft's Mustafa Suleyman says that his team is crafting AI companions who will see and remember everything we do, which will constitute an intimate relationship with AI. Now, I think our relationship with AI is going to change. I think we're going to go from chatbots to these systems that are Constantly on, constantly monitoring us and we're constantly going to be able to say, hey, is this okay? Is that working? Yada, yada yada. And just literally, like for example, right now I'd be able to say, hey Chat gbt. How do you think I'm doing recording this video? And it's going to be able to respond. Now, of course, I don't think it's going to be something that's forced upon.

Us, but I do think that over.

Time a lot of people will opt for that lifestyle because it might just be easier rather than saying, hey chat gbt, you know, opening up the app and saying, hey, can you note this down? They're probably just going to say, hey chatgpt, you're listening, note this down. And I think that will be most certainly embedded into some future phone as a native app on device that works really well.

Don't even think of this as an application. I think of this as fundamentally a relationship. You know, my team and I are now in the business of engineering personality. We're crafting a lasting, meaningful, trusted relationship that is the new platform, as far as I see it, because it's not just about voice and text or language. You know, it's really going to be about vision. You know, your, your companion is going to see everything that you see in your browser and on your desktop in real time, understanding both the text and all of the images and able to talk to you about it as fluently as I'm talking to you about it now. And so trying to craft that into an application and a business model too soon almost sort of misses.

We actually also had the OpenAI 01 team talk a lot about what they, you know, I don't want to say they think it's coming next.

It's them. They're going to, you know, tell you.

Exactly what's coming next. And basically it was a good video because it got to show us, you.

Know, the real use cases of the O1 model.

And at the end of the video, they showed us what's actually next. So they talked about how, you know, we're going to be getting some model releases.

The models that we do get, a lot of them are going to be.

Getting access to tools. And then of course we're also going to get dynamic model selection.

So these are going to be three.

Areas where we actually know what is coming next. Because of, of course, a lot of.

The time there's a lot of speculation.

But having this where we see, okay, what we're going to get is model releases. So of course the full 01 model.

Right now we're on the 01 preview.

And of course the O1 mini.

Of course there might be GPT 4.5.

Or GPT 5, although I do think it will just be GPT 5, although.

We never know, of course.

Access to tools Having these systems have access to tools just basically means they.

Are a whole lot better to use.

So that's something that you can also plan for. Dynamic model selection basically just means your.

Life is going to get a whole lot easier because if they do release more models, you've got a ton of.

Different models to choose from. You've got GPT4 Mini, GPT4.0, GPT4 Canvas.

01 Mini 01 Preview. And then you might have the 01 model and you might have GPT 4.5.

Or you might have GPT5, which is.

Just way too many models for the.

Average person to pick from. And we all know that that can lead to analysis paralysis. So the paradox of choice is something that they're going to remove with where you enter one prompt and then it picks the best model to answer that.

And it's a much more seamless experience.

Let's touch on what's next. We are planning to make several updates to these models over the coming months. We plan to continue developing and releasing models in the new OpenAI 01 series as well as our GPT series. In addition to model updates, we expect to add web browsing, file and image uploading and other features to make them more useful in use cases in ChatGPT. And while even today you are able to switch between models in the same conversation like you saw in the demo, we're working to enable ChatGPT to open automatically. Choose the right model for your given prompt. In terms of model updates, we've already started to test and run evals in OpenAI01, which is the next iteration in this following this research Preview. Compared to O1 Preview, users can expect more powerful reasoning model that is even better at coding and math, as you can see in these evals where O1 represents the orange bar. We look forward to making O1 available to customers as we build, as it will build on the use cases we saw today and likely unlock more use cases with its powerful reasoning capabilities. And now I'll hand it back and.

I can't believe I didn't realize this when I actually made the video, but you can see right here on this blue bar that this is actually GPT4O. So GPT4O in a model update Very soon is going to be getting a decent buff to its competition math and its PhD level science question to GPQA. So that's going to be pretty cool because I can't believe I didn't realize that because I was just so focused on oh one. But you're going to see that those models are actually getting a huge boost now. We also had Nvidia's Nematron model which surpassed GPT4.0 on many different benchmarks. It's basically a fine tuned version of Llama 3.1, the 70B version that manages to surpass the current version of GPT4O.

Which blew everyone's socks off.

I think most people didn't realize that whilst Nvidia is a hardware company, they can still cook up some pretty cool models using a variety of different techniques. And that's exactly what they did with the Nematron model. And this was something that took the AI community by quite the surprise. I did make a full video on that too, of course. Link will be in the description or just go on my channel. It's been uploaded in the last couple of days. I know some of you may have saw this, but I wanted to include this one once again because I think this was one of the most underrated updates by far. Tesla releasing their autonomous update where they showcase that this robot is now a lot more autonomous than you do think. I think the reason that this one was so important is because a lot of people wrote off Tesla due to their recent event having tele operators, when in fact Tesla just wanted to collect data in certain environments. You can see here that these robots.

Can literally charge themselves. They can pick things up, they can walk upstairs, they can use their vision systems.

This update was surprising genuinely, because humanoid robots are just one of the hardest things you do.

In fact, OpenAI said that they were.

Going to work on it, but they were like, you know what? No way, we cannot work on this.

This is too hard.

We're just going to focus on software and companies like Tesla are racing forward with the most human looking robots that I've ever seen that are now even more autonomous. So this is something that, I don't know, it's kind of blowing my mind that the future is actually here and this robot is actually autonomous. As someone that watched the IROBOT movie, you know exactly how creepy these robots can get. But nonetheless, it just goes to show how quickly these teams are moving, especially with even the hardest kind of technological problems. And I know there have been so.

Many memes about the Tesla bot, but let me know if these were priced.

Reasonably and price wasn't a factor. Like, let's say these were, let's say, $500.

Would you get one, or does that fit you?