Huge Claude Updates, AI-Generated Minecraft & More AI Use Cases

This week we have anthropic shipping features that we haven't seen out of their competition yet. OpenAI introducing a brand new domain for ChatGPT, the world's very first generative game engine that you can actually try. Today we'll be doing that here and so much more. All covered for you in this week's episode of AI News that you can use the show that gathers all the AI news for you. And then we shine a spotlight on all the releases that you can actually put to work today. And let's kick off things this week with ChatGPT's transition to a new domain. Because OpenAI bought the domain Chat, this is not massive news. But now you can just go to chat.com and it will redirect you to ChatGPT, which is kind of neat and makes it even easier to use their search product, which we'll talk about here shortly. But first I want to talk about the substantial releases of this week, which are mainly from Anthropic. Actually, I count a total of three releases that we need to talk about, maybe even four if you include last week's release of the desktop app. So that I talked about briefly. It's pretty much the web interface inside of a desktop application. But beyond that, they shipped some really interesting things here and one of them gives cloth capabilities that actually go beyond what OpenAI can do today. What I'm referring to here is its ability to look at PDFs, including visuals like graphs, images or visualizations. If you're not aware in other LLMs, it does not pick up on these, it only considers the text inside of the PDF. But you do need to enable this feature. It's hidden under a feature preview here and now you want to enable visual PDF PDFs as soon as I turn this on, if I for example upload this paper on prompt caching, then it will be able to see and interact and therefore prompt on top of all the visuals here. So for example, if I ask it about this specific thing after uploading, it will not just be considering the text, but also the visuals. This is actually pretty big because this was a major limitations when it came to using LLMs on top of PDFs, especially when it comes to scientific papers, there's usually a visual component to them and now you can prompt on them of that too. Big limitation here is that this only works up to 100 pages. So keep that in mind. And good news is that this is also available through the API. So you can build applications with this. You can build a workflow where it automatically processes PDFs for you with the Sonnet 3.5 API and this Visual PDFs feature enabled in the API. Okay, beyond that, Anthropic shipped Haiku 3.5 and this was actually quite the controversial story this week because it is probably the first release in the LLM space that increased pricing compared to the previous versions. If you've been following this channel or AI at all, you will know that every announcement is like it's 20% or 50% better and we're slashing the pricing by another 60% here. They improve the performance as announced with the Sonnet 3.5 and Claude computer use announcements. We talked about this before here. All the benchmarks for Claude 3.5 haiku. It's essentially their GPT4 all mini competitor, but slightly better, but they hiked the prices on this thing four through the API. If you're a web user, you'll be happy about this because this thing is not available in the web interface just yet. But the price increase was unexpected and it got people spinning up all sorts of theories like does this mean opus 3.5 is not coming? Or does this mean they hit a wall and need to increase their revenue? By the way, if you're not aware, Anthropic mainly makes their money through the API as opposed to OpenAI, which is very chatbot heavy. Look, Anthropic has 85% of its revenue from the API. On the third update I talked about is update to the mobile app. This is very minor. You can finally use your voice to input text, although it does not have an output. So something like advanced voice mode is not available. With claude, you can only do text to speech of your own voice, which to be fair is a very nice way to input information into an LLM. I do this inside of ChatGPT all the time. I just turn on the mic, blabber into it for a minute and then I have all the context I need to start a conversation or start prompting something specific. This might take a little bit of experience. With experience you learn what type of information really helps the LLM craft higher quality responses. But we talked about this a million times. Now in this video we're focusing on news that you can use. Okay, and that wraps up the Anthropic updates. Next up, I want to Briefly talk about ChatGPT search as last week I said that I would follow up once. I have a little bit of experience with ChatGPT search under my belt and here's my conclusion. The quality of the damn Thing is surprisingly good. It's probably on par with Perplexity, although Perplexity goes more in depth on its responses if you compare it to Google. To me, it just feels like it's a slight slower version. Now, where do I take these insights from? I actually did a self experiment over the course of the last week and I completely avoided Google for seven days straight. The way I did this is I set up the ChatGPT extension inside of Google Chrome so the search bar on top would act as ChatGPT search and I straight up avoided Google at all costs. And the result of the self experiment? There's positive things and negative things. Let's start with the positive things. I was actually very surprised by how high the quality of ChatGPT search recommendations Work most of the time. Even though I felt like this is not a thing I should be asking ChatGPT, it delivered the answer. I was looking to convert a certain amount from euros to dollars. It just found the latest exchange rate and did the calculation for me correctly. That is I needed the weather in a location. It has the same widgets that Google has. I wanted to access the anthropic console. Well, it found the exact URL and just gave me one link instead of a lot of ads and various links. Kind of nice. But we need to talk about the negative, which is the fact that it's just pretty slow compared to Google. Like it's quite fast for an LLM, but compared to Google, which is essentially instant, it feels instant. Right? That's what they optimize for. It takes some time and this is annoying, especially if you have been used to the speed of Google over the last decade. But here's the thing. This is the first version of the product and eventually, as it gains more understanding of you, your needs and what you're trying to do, my opinion is that this is going to steamroll search as we know it. I gave this example in the community and I want to round out this segment on on the example. So, as you might know, I'm in Brazil right now to do a kite surfing vocation. Basically, I work every day, just like at home, but in the evening we get to go kite surfing because there's wind here every day. Now, a big part of kite surfing is the wind and I regularly use Google to look at that. But here's the thing, whenever I'm looking for the wind, I'm not actually interested in how many knots of wind there will be today or miles per hour. I'm actually interested in a question that requires that information, what I'm really asking is if I can go out and kite surf today and, and if, yes, what size should I take and what does the tide look like? Now, in practice, this is a simple Google search where I open up Google and ask it, what is the wind like today? Or I go to one of my favorite wind forecasting websites and you can get the same result from both ChatGPT Search and Google right now. But where am I going with this? As I mentioned, asking for the wind is just one part of the equation that I'm actually looking for. And I think eventually all of these LLM powered search engines have the ability to actually understand what you're looking for. It's not going to tell you the number of knots the wind will be today. It's going to tell you, Igor, it looks like you're in Guru and you should go out at 4pm and take your 9 meter cork height. Whereas plain old Google in the current version is just going to give you the wind number. And then you have to do these computations in your head by yourself, as we do right now. And think about that. Which one would you rather use? The personal assistant that figures it all out for you and just tells you what to do, or the version where you have to do some extra thinking? I personally think once it has all of your personal context and understands your life in a more holistic manner, that these LLM powered searches will actually win out in the long term. And I think Google realizes that too. And this is where the little story ends and we can actually get to the next release, which is actually Google rolling out their AI powered search even further than they have right now because they're feeling the pressure. This actually came out last week around the time ChatGPT search dropped and I think they renamed this again. This was called differently. I think it was called Gemini Summary. It doesn't really matter though. So now it's called AI Overviews in Search and they're expanding this to more places around the world. If you're not familiar, this is when you use a regular Google search at the top, you get an AI response on top of the typical Google search results. Now when this initially came out, this got a lot of backlash because people were breaking it in all sorts of ways and it was making weird recommendations that were just not truthful hallucinations. But they seem to have improved this a lot and now this is shipping to more and more people worldwide. Now there you go. That's the update on your search wars between ChatGPT Google and perplexity and other players. My verdict is that I will be sticking to Google as the default with ChatGPT search pinned to my bookmarks whenever I just want an answer instead of a list of links. All right, so next up we have Xai's Grok coming out with an API playground type website. And this is just something I briefly wanted to cover because they really did an incredible job at putting this platform together. If you compare it to some of the competitors like OpenAI, Anthropic or Gemini, this is actually my favorite platform. Look at how simplistic this is. I just logged in with my X account. You get $25 of free credits and you can start using Groq for the API right away. Now, should you be doing that? I don't know. It's up to you. My verdict right now is that Grok is just not that good. There's no real reason to use it versus some of the other LLMs out there. The one thing that it has going for it is that it's plugged into the Twitter firehose, but you just have to be specific in prompting it to even get the right news stories. I don't know, they're just a bit behind the state of the art models right now. Nevertheless, they're catching up quickly as they added the ability to upload images to Grok now. And with this super clean developer platform that gives you the ability to use this programmatically. Plus a rumor of Nvidia wanting to invest 75 billion into Xai's Grok. I think they're well on track to be one of the big names in this game. It might just take a little bit more time and we'll keep our eye on Grok and what they're up to. Okay, next up, while remaining on the topic of X, this is something fun. I've been told it's not that good, but I really like this idea and I wanted to test this live in the video. So this is a new 11 Labs feature which does the following. It creates a unique voice from your Twitter or X profile. So the way this should work is that if I put in my Twitter handle, DAI Advantage, it will go ahead and create a new voice avatar from it. And look, I'm not even logged in and you can just go in. It does this very quickly and it will create free voice samples from your profile for you. Let's have a listen. Let me tell you about kite surfing with a side of AI breakthroughs, because who doesn't love balancing extreme sports with technology. Right? Wait, what is it listening to me right now? What's up with that? Okay, one more. Let me tell you about kite surfing with a side of AI breakthroughs. Let me tell you about kite surfing with a side of AI breakthroughs. That one sounds a bit like Elon. Okay, right. Let me tell you about kite surfing with a side of AI breakthroughs. Because who doesn't love balancing extreme sports? Okay, so it animates your profile picture, creates one of their custom voices, and I guess it writes some text with an LLM based on your recent tweets, but it didn't really tweet about kitesurfing. What's up with that? Anyway, you can try this for yourself with your very own voice. It's free, it's. You don't need to log in. It's kind of fun. And now let's move on to the next thing, which is a very interesting one that I don't want to spend too much time on, but I want to bring your attention to this and show you that this is happening. The very first generative game engine, meaning you can play Minecraft without the code of Minecraft being actually ran in the background. Every single image, every single interaction is generated on the spot. And you can try this for yourself in the browser without even logging in under Oasis. Dcart, AI, get some details on it here. But we can just pick one of these presets and hop right into a game. And this is the very first publicly available fully generative game engine. Now, I'll admit that this is probably more exciting from the technical point of view rather than the usage experience, as some people on Twitter pointed out. Pretty much everything about this is terrible compared to normal Minecraft. But if you go here, you queue up, and then in 80 minutes, I should be able to hop in here. So let's just wait a bit and give this a spin, asking you. All right, it's our turn in queue. So let's just pick the village outpost here and. All right. And there we are. Or not. No. Oh, no. Now I reloaded. Okay. Okay, let's do this again. Okay, as you can see, the top right, I have 40 seconds left in here, so let's make the best out of it. Okay? Okay. With space, I can jump. And as mentioned, all of this is being generated in real time. I mean, look, it's really simple. I can move my map. I can even open my inventory here by pressing E. Let's do that briefly. There you go. That's my character with all the stuff. So Now I should be able to move these different blocks back and forth. Look, it's rendering all these numbers in real time. Isn't that crazy? Five seconds left. Can we move something? I'm kind of out of time here, but yeah, it actually works. I could click things. Ah, and I have been kicked out, but there you go. Very simplistic. Worse in every way than the original game, but eventually. Think about it. It's like the real world it will be generating in real time as the player interacts with it, rather than there being a preset path for you to follow. That's a whole new paradigm and a whole new way of hosting digital experiences. If they make this work. This is the future of not just gaming, but digital experiences and maybe even the way information is communicated. Absolutely massive. This is the very first time we see this. Go try it out yourself. Big milestone for technology right here. But now let's look at the next thing that might be a little more useful to you. Okay, next up, we have a big feature coming out of AI video. We're getting more and more manual controls as per predicted, inside of the state of the art generators here. Concretely, we're getting detailed camera control for runways. Gen 3, which is amazing because this gives you the ability to create some shots that weren't really possible with AI video so far. Some of these movements that you can see here in the demo video, including the intensity at which the movements happen, were controls that you didn't really have up until now. You could prompt really well for the subject, the background, the lighting, the environment. And sure you could pick the lens and the movement, but you never really had this fine grained control. You could try and prompt for it, but most of the time it would just ignore it. Now we actually get proper sliders to dial it in every single week. AI video tools are getting better and I think this is absolutely amazing. So let's give this a quick little spin. And then these advanced camera controls are only available inside of Gen 3 Turbo. So let's see, we even get some of these presets here. How about something like orbit? Ah, yeah, close up orbit, right. That is when the camera spins around the subjects while keeping it in the center same area of the frame. And look at that. It adds this manual prompt to it. And I actually need to upload a image. So let's just do this on top of the sunflower. Okay, so let's see if we can get this plant growing as a time lapse and the camera orbiting around it during the time lapse. This is something that would be super hard to achieve in real life. You would need to plug this into ac. You need to take thousands of images, stitch it all together, make sure that the lighting is consistent, ideally. And let me tell you, with this window in frame, it would be almost impossible to pull off because you would have so much flickering from the sun in the background. A real pain to create in the real life. Let's see if AI managed to do this. Okay, so we did get the orbiting motion. We did not get the time lapse of the flower growing. This might be due to my prompting. Let me try one more time. Okay. I added this part. A time lapse of a sunflower growing from a seed in a pot in front of a window. Basically describing what the image shows, plus the transformation that I'm looking for there. Right. Growing from a seed implies a transformation. This should probably work. Let's see. No, it didn't. You also get these controls here. Okay, let's see. Generate. Maybe it's because it uses this image as a start frame. Oh wow, look at that. More dramatic motion. So this camera motion works really well. I just can't seem to prompt the time lapse into existence here. Let me prompt it the other way around where it turns from a flower into seed. As you can see, I'm having way too much fun here. Last try. Okay. It just doesn't want to manipulate the sunflower. But I have to say that the camera motion here works really well. And even this interface is super intuitive, including all of these presets. So if you want, you can get a lot of manual control over the camera in your shots with runways. Gen 3 Alpha Turbo model. That's a mouthful. And yeah, so now it's just a question of time until all the other S tier providers add this to their video tools. Next up, I want to show you a new AI tool by Google. And they always have these interesting tools. Do you remember last week's AI DJ where you could just prompt different words and then you got a slider and it started mixing those sounds or styles into the track that was creating live. Pretty amazing stuff. Here's a new one that I want to highlight. It's called Learn About. This is another one of their AI expressions experiments. But basically this is an assistant that is focused on one thing, helping you acquire new skills. So if you want to learn about something, this is a great way to do it and it will create a custom learning plan for you. Now, unfortunately, this is not available everywhere you need to be in the US I'M in Brazil right now. Some team members tested this back in Europe. Also doesn't work. So what I'm going to do here is I'm just going to engage my VPN real quick. I'll be right back. And now that I changed my location to the U.S. voila. Here we are. So, so what do we want to learn about today? How about prompting LLMs? Let's see what it suggests here. I'll just read to the Google terms here. Okay, so it tells me why this skill matters. It gives me example prompts, it tells a little story. Heck, it weaves it all together just like a teacher would. Look at that. You have some follow up prompts here and then you have other suggested topics that you could now continue learning about. This is really great. So there you go. I'll just continue the conversation with one of the presets here. But this is actually really great. This world class cloud seems to be quite helpful and accurate when it comes to learning about prompting. And right here you have different prompt engineering techniques and interactive list. Let's see. Let's expand on this one. Be clear and specific. It changes the interface here. So as you can see, when you want to learn about something, this might be a better experience than using plain old chatgpt. Can this be the next notebook LM when it comes to user popularity, I don't know. But this seems pretty good actually. And as you know, I personally am a huge fan of learning new skills. I think it makes you a more well rounded human being and increases your value to the world. Something that in this fast moving age is a great idea. And if you can learn about something that you're generally interested in, it's actually one of the funnest activities available on this planet. Let's talk about the next point here. This is really great. And I have to say I don't just enjoy the format here, the interface, but also the tone. Ah, look at that. There's a knowledge test here in a more formal and academic response. Fun and easy way. Scholarly perspectives. Probably this one, right? Yes. And then it gives you the answer to why that is correct. This is amazing. Yeah. So if you want to learn about something and you're in the US or have a vpn, you can go ahead and use this. It's completely free, just like all other Google AI experiments. All right, that's everything I got for you today. If you want more updates at an earlier point in time, plus the free template that you get on signup, check out the AI Advantage newsletter in the description below. And other than that, I will see you next week with another episode.