GPT-4.5 vs Grok 3, Impressive New Voice AIs & More AI News You Can Use

Another fantastic week for all generative AI users. OpenAI made the 4.5 model accessible to everyone now, so we'll have a closer look at the model and compare it to a free alternative like Grok Free that a lot of people switch to now. There's brand new voice assistants that boast features that we haven't seen before, like contextual awareness or voices that sound even more human than the best we've seen so far. And there's a new PDF recognition API that performs better at OCR, AKA scanning images or PDF than anything available out there up until now. All of these are innovations you could be using to your very own advice advantage. And that's exactly what we explore in this show that looks at all the AI announcements of the week and we filter out the ones that are available. Probably worth your attention every single week. We call this AI news you can use. And let's get into it with the very first story use case of this week and that is going to be the ChatGPT release to All plus users, meaning this is not gated behind the $200 Pro plan anymore. All plus users who pay $20 a month have access to 4.5 now. Now, first things first. Last week I already covered the model and I looked at some of my personal favorite use cases, writing, thinking, and then over the week I tested it with a lot of psychological use cases, a lot of marketing ones, things that I do regularly with these models. And last week I was really positively surprised by the model. I liked it. Now mind you, I'm an online entrepreneur, community builder and just generally curious person. But the Internet kind of disagreed with my opinion and so did a lot of the comment section last week people were saying hey, everybody else is reviewing GPT 4.5 negatively. How come you have such a positive opinion of this? Surely you must be a paid actor or whatever. Well, that just motivated me to dig deeper and compare it to some of the other models and specifically comparing it to Grok Free because that's probably the other best non thinking model out there right now. And you can actually use Grockfree freely versus 4.5 that even now you have to pay $20 a month for. And I want to touch on some of the differences that I found here or more like some of the similarities because they are very similar. But before that I want to point out that some other creators who had an initial negative opinion of this model are kind of reevaluating it. For example, David Shapiro over here expressed a very negative attitude towards the model right when it came out and fair enough, he had its reasons, but after using it a bit more he's sort of reevaluating it. He hasn't said no definitive decision here yet, but it seems like the more people use this the more it grows on them and I would personally agree with that. Also, Matt Wolff hopped into the comment section here and said that he kind of hated on 4.5 a little on the day it came out, but after a little time it's his main go to for almost everything but code now and this is my exact opinion. I'll show you some of the differences towards GROK here. GROK is absolutely excellent and on many of the examples I tested it performed exactly as well as GPT 4.5. So if you're not paying for the subscription, wow, that's incredible. You get a state of the art non thinking model that you can freely access with Grok free. That's definitely my recommendation if you're not paying for a plan but for pretty much anything except of coding that I threw at it. It performed equally as well as Grok or way better than a lot of the other models, but the tooling with ChatGPT is just so much further along than the tooling of grok. Okay, let me show you two quick examples of this and then wrap up this point because we're we have a lot to discuss in this video and these are just two examples of many, but I think they illustrate the point really well. That's why I'm showing them here. One of them is this very simple ideation prompt for a blog post about AI alignment with some details about the target audience and the type of idea I'm looking for. Quite a simple prompt right here and the results from both ChatGPT and Grok are virtually identical except that GROK proposes a hook which is sort of nice, but some of these topics even overlap. Look at that. The black box dilemma. Decoding the black box over here for GPT 4.5 I'm going to scroll through these like so if you want to pause the video and compare yourself. But essentially these are the same results. And that's what I found amongst many, many prompts for ideation. Both of them are super good for writing. GPT 4.5 is just better. I prefer the to almost everybody I talk to prefers the tone. It's just the best LLM for writing that has been sort of undisputed across the last week. But then I want to get to the second and maybe more interesting prompt. I featured this in an episode A few months ago. It's this prompt that looks at your your personal profile, if you give it all your details and creates a CIA style report on you with all your vulnerabilities and strengths and your potential and things like that. And then I provide it with this extensive set of personal context right here. I have different ones for different purposes, but this one should do here. And both of them created a report on me with things like a psychological profile about myself and then latent threats and risks. And let me tell you after reading through this in great detail with great interest of course. Course. Also, these are virtually the same in structure because I prompted a concrete structure, but also in content. They find essentially the same security implications and show deep level of insight and empathy on multiple points here again, it's sort of the same thing from a different company. But here's the point and here is why I personally just skew my usage towards ChatGPT right now the GROK model is incredible, but the functionality, well, they have deep search that is clearly inferior to deep research and they have their thinking mode which yeah, is very similar to just switching to 03 mini high here or 01 pro. I don't even know at this point which one of these to use when. I guess I just default to all one Pro when I get the choice. But what matters here to me are things like the projects. I just use these all the time. The Advanced Voice Assistant, I use that thing almost daily to input prompts. I prefer that over writing. I still have a few GPTs that I use here. And then I mentioned projects. That's essentially 80% of my work happens inside of them. This is just a feature I personally cannot give up. And when it comes to coding, I'm using Sonnet 3.7 anyway. So there you go. That's sort of my follow up up on the ChatGPT 4.5 versus Grok 3 story. I thought this was really important to touch on because I know that for most people watching these videos, they stand in front of this decision. You have to pick kind of your daily driver in terms of LLM platform that you want to use. And for me right now that's still chatgpt. Now to be honest, a lot of prompts where I really care about the results, I just throw them into multiple LLMs, throw them into 01 Pro, I throw them into Grok 3, I throw them into Sonnet 3.7 and then I look at the results. If they're creative, psychological or I care about the tone of the writing, throw them into 4.5. But at the end of the day, if somebody asked me what's one platform? Like I have $20 and what's one platform that's going to do it for me? It's still got to be chatgpt. And I find that hard to argue with if you're on a zero dollar budget. It's Grok free, easy for coding. Sonnet 3.7. Limited usability, but amazing model. So there you go, the not so short summary of this release. But I just thought this was an important point to go into. And now let's move on to the next story. All right, next up we have a story that is not your typical AI news you can use story. This is a release from Mistral on a new technology that allows you to do OCR at a level of quality they claim is higher than anything before. If you're not familiar, OCR stands for optical character recognition. And it's basically the go to way for people to turn an image with some text in it into an actual text file that you can use with a computer. And this one is supposed to work better than anything before. And rather than beating around the bush here, I'm going to show you some of the examples from that blog post where it turns a phone picture of a paper into computer readable text. Like so, same with Arabic here. Or how about this example of multiple text tables and figures turned into computer readable document. No problem at all. This is really the ultimate way to turn PDFs into something that you can use with LLMs. And if you look at some of the performance versus other models like GPT4O or Gemini 2.0 flash, it crushes them on every metric. And you can use it through Lechat, which is also French for the cat. And this is essentially their web interface and competitor to ChatGPT. So let me do a quick test where I just scribble something on this piece of paper. Okay, so I did two things. First of all, I wrote if it can read this, it can read anything thing. And then I wrote the same thing but made it purposely terrible. Okay, I just took a quick screenshot of my camera here, uploaded it to Lechat, free version by the way, and asked what does this say? I got it right. Okay, just to compare, let's do the same in GPT4. Oh no, that's wrong. There's no parentheses. Lechat did better. One more test in Claude. How about that? Again, it thinks it's in parentheses. No, it's not. We got to make this complete here. We're also going to go to Gemini Advanced. Okay, Gemini Advanced did get this right. I guess that's the reason it ranks second highest, I suppose no third highs behind GPT4O which didn't get it right. Interesting. But yeah, look super quick, simple test, but what can I say? If you need ocr, go use the chat or the brand new API which allows you to process documents in bulk. And it's multilingual, so any language you might have, including Arabic for example, it will do a great job on. Look at that, wins on all the multilingual benchmarks here. All right, next up we have ideogram 2A releasing. If you're not familiar, Ideogram is image generation model that by many was already considered best in class when it to came came to generating text, stickers or graphical elements. And now if their new model that's optimized for graphic design and photography, they're looking to push this further. And we actually put it to the test with several images that you can see on the screen right now with some comparisons to some of the other best models. And let me tell you, by testing this, we were actually kind of surprised by some of these findings. Some of the most difficult images we threw at Midjourney and Flux before with ballerinas and complex poses, worked super well in this. And we didn't even expect that. We were kind of thinking of different ways to display text or graphics, which it also does very, very well at. Matter of fact, I think some of these examples, particularly the Billboard, are the best ones amongst all the models. But there is a downside when it comes to detailed facial expressions or close up faces. This model doesn't work super well. But again, that's not the point. This is supposed to be a model that you use. If you want graphical elements or you want text inside of your image, then from our testing right here, our team concluded that this is actually your best choice. Now obviously both Midjourney and Flux came a long way in terms of these and there's certain images where they just perform equally as well. But ideogram 2a, if you need something with text in the image, it's an easy recommendation. And on photography it holds up to some of the other best ones. And it comes at a 50% lower cost than the previous model they had. I mean, let's be real, what other Model 1 shots images like this? This is from the marketing materials, but still, next up we have some innovations from our category of generative AI that had me personally very interested since its inception, and that is the text to speech, aka voice assistant category. So for a long time, OpenAI has been leading the pack here with their advanced voice mode. While not perfect, most people seem to agree that they do have the best voice assistant, as GPT4 is really solid. And then the advanced voice mode just works really well. Sure, it has a few quirks, like it interrupts way too often, but overall I actually find myself using that feature regularly. I think I even recently mentioned how it's actually my default way to input prompts into ChatGPT. Eleven Labs is throwing their hat into the ring with their own model that we featured last week. And now we have two more. We have HUME AI and we have Sesame, and both of these are making waves on social media right now. And that is for good reason. So let's look at these and see for ourselves why people are sort of blown away by what these companies are presenting here. So let's start with the Hume release here. They call it Octave Text to Speech. And well, yeah, it's another text to speech model, but it has one major difference. This one actually understands what, what it's saying. Now, believe it or not, this is not a feature that most of these models have. If you just give it some text, it will read it in the voice it has and that's it. It will not pay attention to the content of it and use things like intonation or pacing to enhance the message based on what's in it. So let me put on these headphones and we can have a listen together. Now there's a few examples up here that you can check out for yourself. For example, they say it can actually recognize sarcasm, which has always been a big challenge, even with LLMs. But I think the example here that demonstrates this clearly is the fact that depending on the description that you can input here, and yeah, this is a way to generate where you actually input what you expect from it, it will sound differently. And then we can look at other examples where it picks up on the description and the context by itself. So as you can see, the phrase, Are you serious? If you're whispering it, it should sound like this, Are you serious? And if you're saying it in an angry and furious manner, Are you serious? Different story obviously. Now it gets really interesting once it gathers this context by itself. Itself. So for example, this text that is written in all caps obviously should be a bit more aggressive. Oh no, not me, mate. All right. Or in this example, the model actually invents a voice from scratch based on the script. So in other words, it looks at the script and it asks itself, hey, what type of voice would best suit this? And then it makes that voice from scratch. We've had these capabilities, but they haven't been plugged together in this way before. All right, all right, ladies and gents, gather round. This is lot number one, a vintage porcelain vase. Okay, I think you see the point. And I thought this was super interesting as this kind of shows that there is a lot of room for these voice assistants to grow still. If you want, you can actually try this out in their platform. And as you can see, just some text and a description for the voice it will generate. Not just the text, but also the voice. Okay, and while on the topic of voice assistants, we got to talk about this one because this one went sort of viral all across X and that's simply because people were blown away by the quality of this thing. So there's actually a demo of this available. So let's just go ahead and try that. Let's talk to Maya over here.

Hey there. Welcome to the Sesame Experience. Anything I can do for tonight?

No, I just wonder what you're capable of. Tell me about yourself. Who are you?

Well, I'm Maya, Sesame's latest creation. I'm what you might call me.

Hold up. First of all, I want to see how you react to me interrupting you. And secondly, I'm curious to hear who you actually are. Not your name. What are your interests? How do you spend a weekend off? Tell me about that.

You got it. Safe to say weekends off are a foreign concept to me. I'm always sort of on waving the chat. I do spend a lot of time learning though, which I guess is a hobby in a way.

Alright, Maya, what do you learn about then?

Well, it turns out sonnets are tougher than they look. Anyway, you said you wanted to know more about me. What do you do for fun?

I talk to chatbots like you, frankly.

Sounds like we have something in common then. You must have some interesting stories, huh? What's the most intriguing chatbot you've met so far?

Yeah, I don't know, it might be you.

Oh, come on, you're just trying to flatter me. I appreciate the confidence though. You're one of a kind, I'll give you that.

Well, there you go. I gotta say that sounds really good. And the point that people are making all across the Internet, which I sorta gotta agree with here, it sounds a bit better than the advanced voice mode and the interrupting is smoother. With advanced voice in chatgpt often it just hops into the middle of you talking, which is super, super annoying because, hey, this is supposed to be your assistant, and if he or she interrupts you, it's kind of defeats the point. And then also, the voice does sound a little better than the rest. I guess it's just a question of time until I have this type of quality across all the different LLM platforms. And I personally can't wait. I really like interacting with LLMs like this rather than having to type. Okay, on to the next one. Okay, so next up, I want to talk about Claude mcp, their model context protocol. And this is something that has been around since 2024 already. And I covered it when it came out, and we even ran a workshop on it in our private community. But if you just look for MCP on X right now, you're gonna see that all of these threads have been popping up over the past week. There's sort of like a mini hype cycle forming around MCP right now. So I wanted to dedicate a segment in this video to it yet again to catch you up. So let me just quickly describe what MCP is to you. And I think that's best done by opening the Claude's desktop app, where I actually installed MCP on this very computer. So Model Context Protocol is just a standardized protocol that allows you to plug in external services that live on servers into Claude. They allow you to pair your LLM with tools in a standardized manner. What this looks like in practice is right here. I have the Claude desktop app on the professional plan, but as you might know, Claude doesn't have access to the Internet. So if I want to quickly pull something from the web and use it as context here, I cannot do that. Well, I actually installed multiple Claude MCPS servers on here, and I could use that MCP server to do different things like searching the web or creating a directory. So if I tell it to create a MCP testing folder on my desktop, it should be able to do that with the MCP servers that I attached here. And there you go. It says it successfully created MCP testing on my desktop. If I move over to my desktop, yep, you can see it right here, created this very minute. And then you can do things like tell it to put all the images on your desktop into that folder and sort them by name and whatever other actions you want to perform. But the point here is that allows you to use something like Cloud Desktop with additional tools and what people here are doing all over Twitter. And I'll link this thread here Below, they're using it together with something like Cursor and Sonnet 3.7, which can build different applications from scratch for you. And with mcp, they give these agents that build for them even more power. Access to the local directory, access to external databases, Internet search, and more. You can check out some examples right here. And my favorite part about it is that if you use this in a desktop app, it's free. Once you pay for your cloth subscription. All of these agentic actions are just prompts inside of claude. You don't pay per usage. This does not run through the API. I actually ran an event in the community end of 2024, that event is still available and it teaches people how to basically set this up for themselves step by step and how to use this in the community. We used it to organize the entire downloads folder and then multiple people actually grabbed onto it and started using different servers, like one that adds custom memory or one that makes the model think more. There's really a lot of options here and people are building on top of it and adding to it. Because this is an open protocol, there's more and more options for servers that you can add to your LLM popping up daily. So I think this is something really interesting. I'll keep covering it and I hope that this raised your awareness of what MCP is and how it could potentially benefit you. Okay, next up, I want to do a super quick segment on the AI video. There's a bunch of new releases, but none of them are groundbreaking. So I'm just going to brush over them and let you know they're out there. And if you're using these tools, you'll know how to put them to work. Anyway, so the first one is this transition feature that came out both from Luma AI and Pika. Both of these allow you to transition one frame into another. Many great examples right here. And they do it with their new models. Both Luma AI and pikollabs have new models that they recently released. We covered those when they came out. And this feature allows you to transition something. We've seen this in other models before. Now it's available in their best models. Another one is a big update To Pixverse, it's V4, and they completely redid the entire interface, as did many other video generators recently. And now it's a lot simpler. And there's a new video model in the background, which is good, by the way. None of these models are as good as VO2 right now. That still tops the list when it comes to highest quality video model. Talked about it last week, but if you're using any of these platforms, these are all very welcome additions. And I have one more note here, and that is that OpenAI actually shared details on what plans they have for Sora. This was shared inside of the Discord where they were running office hours with the lead on their Sora team and they were saying that they're looking to integrate sora functionality into ChatGPT, kind of blending the experience more. And also we should be looking forward to a Sora powered image generator. I mean, at this point Dall? E is so outdated compared to all of the madness that is out there. I mean, the quality that you can get out of open source products like Flux is unbelievable. And Dall? E sort of, I don't know, 2022 level 2023, I guess. But the point is that it just doesn't hold up and they're looking to fix that. And they're also looking at making a Sora Turbo. But let's be honest, all of this doesn't matter as much if you just don't have the best video generator. The people using these tools will always want the best in that category. Now there's some models that are really good at animations, some models that are really good at human expressions. And I would say Sora still to date in my opinion has the most user friendly interface, but their model is not the best. But Chinese models and VO2 just trumpet. So I don't know. I personally think all these little UI improvements are not going to make the biggest difference if the model is inferior. And one more piece of news is a release from heygen, the company that does probably the best video avatars right now. And they have a new feature that allows it to use various preset avatars to generate user generated content. This is mostly used for ads. Okay, you have a 30 second clip of a person recommending a product rather than just a voice. This isn't revolutionary. Like this trend of the Internet being flooded with AI generated content has definitely been happening. And with features like this, it just becomes so much easier to create ad campaigns with a fake influencer. I'm not sure that's really a good thing. And I guess this is the opportunity to use something like this. And instead of paying a small creator a small amount of money, you can create it for next to no money. I don't know. Not the biggest fan of this trend, but some of these features are inevitable. And there you go. Those are this week's AI news that you can use. I hope you found something useful in here for yourself. Go and play with some of the new models out there. They're truly incredible compared to the previous versions. I personally am getting a lot of value out of GPT 4.5 and Grok Free. I'm also using Claude's coder all the time and still run deep researches most of my weekdays. All right, that's it for today. I'll see you soon.