AI is BOOMING! Google CRUSHES it, Open AI Overhauls Chat Memory, Open Source models & MORE!

This week has been absolutely colossal with AI news. I mean, so many different releases from big names in the tech space, little open source tidbits, cool demos, you name it, we got it today. These are the kinds of weeks that I like to see in the AI world, so let's just jump right in. First up, quick little recap. Llama 4 dropped from Meta. I did a whole video on it. It's a little bit controversial. Two new models that were released, Llama 4 Scout and Llama 4 Maverick, both have 16 million active parameters, but Scout only has 16 experts, while Maverick comprises of 128 experts. Neither models, even the smallest one Scout, run on consumer grade hardware. These are meant for corporations, these are meant for businesses to run. And they're still open source, but they don't have the greatest license in the world. TLDR is that this release is pretty meh, not as exciting as previous Llama releases and overshadowed by the rest of what we're going to talk about in this video, more or less. Hey, one thing though is that Pliny has already jailbroken Meta's llama 4. It seems like no AI model is truly bound by its guidelines. As long as Pliney's around, we've got a new image generation model hitting the scene. This is High Dream AI. It's got three variants for a full quality, a dev quality or a fast quality. Now this thing is MIT licensed. However the text encoder is llama 3 based, so that still maintains the llama 3 license, but other than that, yeah, this sucker's fully open source. It's a brand new image gen model. It looks like you can test them all out on hugging face for completely free, so that's going to be linked down below. Now for the High Dream model, it does require pretty significant amount of VRAM to run, but the benchmarks, if they are to be believed, and I don't recommend you take them at face value, do look promising at the very least beating out Dall E3SDXL flux even. However, not by much. Not by much. Keep in mind in mind that those hugging face links are using a quant, so it doesn't fully represent the biggest, baddest version of the model. But I ran this example and it's a difficult prompt and it kind of nails it. It does a great job incorporating all of the text. The character is really coherent overall, not too much AI mush. The background is so so. But yeah, I am pretty impressed with this model. Actually we've got another image generation model as well. But this one is more for image sty and it's supposed to kind of mimic or be similar to the image stylization control that we see in GPT4 Omni. It's actually built on the Flux Dot 1 model. Pretty great base to start off with, but a lot of people do say that, you know, it's not super great at actually taking the original character whatever image you upload and transferring the style over. Or at least it's not as good as GPT4.0, but I still think it does a pretty decent job considering it's a diffusion based model and not an auto regressive one. I think this example is a little bit better. You'll probably want to dial in your settings a little bit while you're using it. And yes, you can try this one for free On Hugging face here we've got a quick Ghibli ify of myself. You know, I think it does a halfway decent job. It's not perfect, it's not going to be as good as GPT4O, it's just not going to have the same level of adherence to detail. But it's a faster solution nonetheless. And it's also fully open source, which is great to see. This is Apache 2.0 licensed. You can see they've got the GitHub for the code here and a paper app as well. So this is pretty dope actually. Next up, Google had a ton to launch this week. They had a really fat week. My last video covers the bulk of it, but I want to go over some stuff I missed and some new comments I have. So one of the bigger releases was Firebase Studio. It's essentially an AI vibe coding style platform, aims to automate a lot of typical coding processes. Uses Gemini under the hood, but not 2.5 Pro, which is the very good coding model Google has. It's using a little bit of a weaker model and that that has led to some less than satisfactory examples. I've seen a lot of comments from folks in my own YouTube video saying that Firebase Studio isn't really holding its own too well. And even this example from Flavio here says he tried generating an app and it just didn't really go well. Mintroy responds though, saying that he's got it working a little bit better somewhat, but also had some environment issues at launch. So while Firebase Studio looks super promising, it seems to be in the pretty early stages and they've got some stuff to iron out with this one for sure. So wanted to let you guys know on that. Anyways, yes, there are a ton of other updates. Like I said, chubby here brings us some extra ones that I didn't cover in my last video, like updated image in 3 image generation text to Music is coming soon as well. They're working on voice cloning apparently too. And Gemini 2.5 Flash is now live. Lots of ships from Google this week and they've really done a great job. Honestly, most of it was really exciting and really cool. Another one that wasn't covered in that last post is a new TPU built for inferencing AI inferencing. They're calling it Ironwood. This is their 6th gen TPU. 192 gigabytes of RAM per chip and it's got 4.5 faster data access. A lot of these big Companies, these big AI companies are looking for alternative methods to Nvidia GPUs, cheaper ways to do AI inferencing and it looks like this might be a pretty solid competitor. Google also has officially dropped VO2 publicly. You can actually use it right in Gemini, which I've been testing out, but no image uploading to Gemini I noticed, so that's a little disappointing. But you know, VO2 is a great model regardless, so it's been a ton of fun to mess around with. In the API they have inpainting and out painting features which is pretty awesome. Apparently there's also the ability to do camera presets like panning to the right which is really cool. Obviously this is more for like professional team level work where they're trying to create a commercial or a create a story, you know, larger groups of people managing these systems at a higher level. And of course first and last frames are also on here through the API. It's cool to see they're bringing VO2 to the public through an API. I mean that is really great to hear considering how good VO2 is of a model. I mean it's a top charter. Most models don't compete with it, but here's one that kind of does. Gen 4 Turbo is now available. If you remember last week we talked about the official Gen 4 full release, but now they have a Turbo model that is five times faster and half the cost of the original Gen 4 obviously does not have the best quality or prompt coherence in comparison to Gen4, but it does allow for rapid ideation and the quality is honestly still pretty darn good considering it's a Turbo model. Here's a dramatic lightning strike Slow pan, an example that I think it Handled pretty nicely. The model struggles with large motion. If it's not humans, you can definitely see that in the animation examples. I don't think RunwayML has ever been the best AI video generator for animation. You can see it also struggles with more complex prompt. Here's me trying to catch a lemon and then squeeze the juice into my mouth in like a really impressive way. But it just generates all kinds of really strange videos relating to me and lemons. You can see I opened my mouth and then the lemon juice just falls out. It's. It's a little scary, honestly, but. But I think the most impressive part is how consistent it kept me throughout all of these shots. For being a turbo model. For being a turbo model. Moving on. We've got more AI video generation news. This is a new paper. One minute video generation with test time training. I did a full deep dive video on this where we looked at some of the examples up close. And essentially this thing generates Tom and Jerry cartoons. Like full Tom and Jerry cartoons. They're one minute long, but you could possibly extend that to five and 10 minutes. This paper is really promising. The stories are coherent. They actually interact in ways that make sense. Jerry in one of them unplugs the computer and the computer goes fuzzy for Tom and he gets angry and starts looking to see what's wrong. Just like a classic Tom and Jerry episode. It's really, really incredible. So, yeah, there's. There's a full video that I recommend you guys check out. Honestly, an underrated paper that we saw this week. Now, that's not all that's going on in the AI video space. Higsfield AI also has an update. This is a brand with a brand new AI video generation model centered around different camera techniques and camera work. But now they're adding even more. Now you can combine multiple motion controls at once in a single shot using Higgs field AI, including moves that aren't even possible with real cameras. That's pretty crazy. They're also dropping 10 new motion controls built for speed, tension and cinematic impact. Overall, if you really care about camera control above all else with your AI video generations, you need to check out Higgs field AI because it is pretty much what they are targeting and specializing in. They've got plenty of examples though. This is a crash zoom in and an explosion. Here we've got building explosion in the background. Look at the hat fly off of her head. That is pretty epic. A crane plus a crash zoom in. It doesn't look like the crash zoom in worked all too well, but the crane worked a little bit better. So yeah guys, if camera work means the world to you for AI generated video, this is a no brainer. But hey, maybe actors are your thing. And other video generation news. LTX Studio has now finally added actors. You can train custom AI characters using your own reference images and keep faces, outputs and styles consistent across every shot using LTX Studio. If you are a big LTX Studio fan, let me know if this feature is something that you've been dying for. I remember one of the biggest holdbacks for me regarding LTX Studio was the fact that the characters were not consistent enough for me. I might have to check this out. You can see they include a little demo clip here with this red haired girl. Her hair is super consistent, although it is cut in, yeah, that shot right there. But most of the time it's pretty much the same and it really does keep her eyes, her freckles, all her facial features intact for the most part. Let's move towards audio generation. Eleven Labs has a couple of updates for us. First of all, they have their new MCP server. This gives Claude and Cursor access to the entire 11Labs AI audio platform and allows you to essentially ping it or access it with simple text prompts. Nice and easy, real easy. Use case to kind of give you an idea of what this can provide for you. If you use cloud or Cursor, you can spin up voice agents to perform outbound calls for you like ordering a pizza or other things. They've also got a couple of more examples included down here. Simple text to speech, speech to text, designing custom AI voices or you know, conversational dynamic voice agents. Now eleven Labs has also upgraded their professional voice cloning, allowing you to produce high quality voiceovers that sound more like you than ever before. Eleven Labs has been a top text to speech competitor for the longest time since like AI Audio's inception. Pretty much. I've used their professional voice cloning before and it is very good and very popular. To see this brand new version of professional voice come out that makes it easier to create a near perfect match of your voice leaves me pretty excited. I want to test this out. Seems like a pretty swanky upgrade. 11 labs. And by the way, I know I'm like flying through these AI news pieces because we have so much to cover this week, but if you do want me to take a closer look at anything, let me know in the comments and give me a reason as to exactly what you're interested in, such as professional voice cloning for example. All right, now I've got some cool Minecraft demos to show you. This is voxel diffusion in Minecraft. Check this out. This comes from the stable diffusion Reddit. You can see it starts out as a bunch of randomly generated noise of blocks and it eventually diffuses into an actual 3D house structure. So that's something to keep in mind. I mean, we're traditionally used to diffusion models generating 2D video or 2D images, but no, they can work in 3D as well. This is voxel diffusion actually diffusing a 3D structure from noise. How cool is this? And these examples, they all seem to just be generating houses, but I would love to try to generate other structures, really unique things, things like statues. This is also really cool. This is an AI assistant that dynamically can play Minecraft with you. If you start building a house, for example, it'll figure out what you're doing and then jump into help using the same blocks, trying to follow and match your patterns, which is really crazy. Apparently this assistant wasn't trained with RLHF. Instead it's powered by assistant games. It's a better path forward for building AI agents. Unlike RLHF, Assistance games explicitly treats user assistant interaction as a two player game where the user knows their goal, but the assistant doesn't. So this model is designed to communicate about the goal from the user and actually collaborate to achieve the end goal. This is pretty crazy too. It actually displays emergent helpful behaviors like active learning and learning from corrections. The assistant literally in this case, builds the walls one block too high to see if the user will start breaking it. To say, hey, no want the roof to be that tall. The assistant will then learn and actually start to break the blocks. Pretty crazy. The Grock 3 API after all this waiting is finally here. And the pricing is actually not as bad as I thought. They've got a bunch of different models. Grok 3 beta, Grok 3 fast beta, Grok 3 mini beta, and then Grok 3 mini fast beta. The most expensive model here is Grok 3 fast beta. Must be running on different GPUs. Text input is is 5 bucks. Text completion is 25 bucks per 1 million tokens. The regular Grok 3 beta is 3 dollars per million token input and 15 bucks per million token output. So these two models up here are pretty expensive. Not as competitive in pricing, but down here, if we go to Grok 3 mini, this sucker has $0.30 per million input tokens and $0.50 per million output tokens. That is a lot more competitive. Now, interestingly enough, Mini Fast Beta is like way more expensive for output tokens, but only twice as expensive for the input. I don't really know. The good thing is all of these have the same context length of 131,000 tokens, which is pretty hefty now. Epoch AI also ran their own independent evals of Grok 3. Now that it does finally have an API, it still stacks up and holds up pretty well. Grock 3 mini with low reasoning is just behind Claude 3.7 sonnet. Somehow Grok 3 actually does worse here, but only ever so slightly. Crazy part is all of these GROK models do beat out GPT 4.5 and the older Claude 3.7 sonnet without any thinking at all. Nothing touches Gemini 2.5 Pro, but I am impressed at how good these models are and it definitely seems like they're continuously updating them and got them pretty well polished for the API release. Finally, let's cap off this AI news roundup with some OpenAI news. We've got a brand new feature in ChatGPT. It is extended memory. It can now reference all of your past chats to provide more personalized responses. It will draw on your preferences and your interests to give you more insightful answers. It understands your personality, the way you ask questions, and it can make some pretty spookily accurate insights just based on the way you communicate with ChatGPT. So this is in addition to the saved memories that were already there, but it's going to build upon this. Referencing past chats to deliver responses that are more relevant and useful to you in the moment. It naturally builds upon what it already knows about you, makes interactions feel smoother and more uniquely tailored. You can always also opt out of this or disable it. But but yeah, in my experience it definitely learns a lot about you and it talks very naturally. It's super confident about what it knows about you from conversations. I did get a little bit of hallucinating when I would ask it about like very specific stuff, but it was quick to correct itself when I pointed it out, which is good to see. But other than that, yeah, it's spooky accurate. It'll tell you things about yourself you didn't even know you did, which is definitely a little bit creepy. So here's a screenshot of some of the craziest examples that I personally have encountered. It gave me a bunch of different insights based on the way I conversate and talk with ChatGPT. And honestly, I wasn't aware of all of these. It gave me a little bit more Nuance, a little perspective, crack into how I might appear to other people or how a very intelligent person or AI model would take notes on me as a person. Knowing more about yourself can a lot of times actually boost your own productivity or creativity in the real world. So try all kinds of different prompts. So that should be rolling out to chat GPT users. Most seem to already have access. I got it, I think last night. But they are preparing three new models for release and we've already been talking about this for quite some time. We know it's coming this month. A lot of people are saying next week, but we've got 04 mini.04 mini high and the full big bad 03. These models should be cutting edge. They should be really competitive. I'm interested to see what OpenAI brings to the table. A little bit of an inside scoop on this too, apparently. And take this with a grain of salt. Testers are concerned that Sama is pushing too hard to rush these models out next week. Again, nothing is confirmed, but we've got quotes here saying that because there's more demand for the technology, Sam Altman wants to push it out faster. Apparently there's a member of staff quoted saying this is a recipe for disaster. But it again, like take this with a grain of salt. You don't know where this information's really coming from. I know that they care very much about safety evaluations and OpenAI typically does pretty decent safety evaluations if that's something you're concerned with. But I'll be real with you, all of these models are just jailbroken immediately once they're released. I mean, follow Pliny on X. He jailbreaks everything. The safeguards don't mean much when you can simply input a prompt into the model from Pliny and have your model jailbroken immediately. Anyways, let's cap this off with something nice. Olama is hanging out with Sam Altman and they say let's go open source. They gave him a little Olama mug. There he is. Olama also got an OpenAI mug. Seems like they're in talks. This could potentially be related to OpenAI saying they want to release a decent large language model, fully open weight and open source pretty soon. Those OpenAI open source hopes are really flowing right now. I really hope OpenAI delivers with this and the license is actually satisfactory. We want to build and accelerate. So yeah, that's going to be it from me for today. Thanks so much for watching. Let me know if there's anything I missed or let me know if there's anything that you want to dive a little bit deeper into, anything that's really interesting or standing out to you. If you want a deeper look at some of the larger topics from today, I've got three videos from earlier this week diving into those respectively. Some really cool stuff to dive into there. I hope you guys enjoy your weekend and I will see you next time. Thanks for watching.