AI never sleeps and this week has been absolutely insane. Real-life Gundams are coming soon. We have a new top open-source 3D generator with
incredible details and accuracy. We have another model that can generate articulated 3D assets that can move realistically. We have not one but
several new interactive world generators just like Google's Genie 3. We have some incredibly expressive
text-to-speech generators that can handle different settings and emotions. We have a new open-source music generator that can create full songs from
just a prompt. OpenAI adds some new features to ChatGPT while Google reinvents the mouse cursor with AI. We have a new AI that can easily change the
lighting of a video and another one that can change the camera movement and direction. This AI can
easily redub a video into a new language and also adjust the lip-sync. We have some ridiculous humanoid robot demos and a lot more, so let's jump
right in. First up, this AI is super useful. It's called Just Dub It and this can take an existing video and not only dub the audio into a different
language, but it can also adjust the person's lip movements to match the new language. Here are some
examples. tested every single one in this category and it's not even close.
Here's another example. I always knew this day would come, but I just didn't know it would feel quite like this.
I see him now. One more round and I take him.
As you can see this can take an original video and make that person speak another language while keeping the facial motion and the voice and the lip
sync aligned. So this is actually based off of LTX 2.3 which is currently the leading open-source video model with audio and it turns out that if you
fine-tune this to do this for dubbing, it actually performs even better than other methods like
HeyGen. Now at the bottom right corner, they have released a code button. So if you click into this, here it takes you to this section. And note that
here they've already released the model to this. This is just a Laura that's based off of LTX 2.3. So this is pretty small. It's only 2.5 GB. If
you're interested in reading further, I'll link to this main page in the description below. Also this
week, this is one of the most powerful image to 3D generators we have so far. It's called Pixel3D and this can take a single image and turn it into a
3D model with super accurate geometry and realistic texture. So here are some examples for your reference. If we take this image of a pizza, here is
the 3D reconstruction of this pizza. Notice how accurate this is. And if you compare this with other
competitors like Huang Nian 3D, it's not even close. This new Pixel3D is just way more accurate and detailed. Here are some examples comparing Pixel3D
to another competitor called Trellis 2. And again, Pixel3D is just a lot better and more accurate. And it's able to do so because it tries to generate
3D models in a pixel aligned way. Basically, instead of loosely giving the model an image and
hoping it understands which pixels belong in the 3D space, Pixel3D explicitly connects the 2D pixels in the image with the 3D structure. So, the
output is a really high fidelity 3D mesh with all the details from the original image. So, if you want to generate 3D models like for games or virtual
production or product visualization, this is currently the best and most accurate 3D model generator you
can use. The awesome thing is they've released this already. So, if you click on this code button and you scroll down a bit, here it contains all the
instructions on how to download and run this locally on your computer. Notice the total size of all these models are 24 GB in size. So, you'll need a
mid to high-end GPU to run this. If you're interested in reading further, I'll link to this main
page in the description below. Next up, this image model is super interesting. It's called asymmetric flow models and this is basically an image
generator that can make images directly in the pixel space. You see, how traditional image models work is it generates the image in what is called
latent space. In theory, this is supposed to save compute and make it simpler. And then after it generates
the image in that latent space, then it uses something called a VAE to convert it back into pixel space which you and I can see. But here, what they
did is they completely bypassed this latent space and the need for a VAE. This just generates images directly in pixel space which you and I can see.
And the advantage of working directly in pixel space instead of using the compressed latent space is
that you can generate hyperrealistic images with sharper textures and better visual fidelity. If you compare it to just the base model that uses
latent space plus another leading competitor when image, you can see that the two traditional methods look way more plastic and fake, whereas this new
method looks a lot more realistic. Or here's another example for your reference. Again, this new method
seems to look more photorealistic with sharper details and Now, the disadvantage of generating this in pixel space, at least previously, is that this
takes up way more time and compute. But, from some clever architecture tweaks, they were able to make this 40% faster. So, compute and speed isn't
really the bottleneck anymore. And if you compare the performance of this new A Sim flow model against
other competitor models that use latent space instead of pixel space, notice that in terms of image quality, this new one is on average the best. This
might represent a paradigm shift in how future image models are designed. The awesome thing is they've released this already. So, if you click on this
code button at the top here, and you scroll down a bit here, it contains all the instructions on
how to download and run this locally on your computer. Plus, they've also released the training and evaluation code to this. If you're interested in
reading further, I'll link to this main page in the description below. Also this week, NVIDIA releases their own interactive world model. It's called
Sana WM, and this is a super tiny 2.8 billion parameter open-source world model that can turn an
image into an interactive world on just a single GPU. So, you start with an image and a text prompt, plus the WASD key presses, and it combines all
those inputs into one video. Here's another example for your reference. It doesn't have to be just third-person point of view where you view the
character from behind. This can also be first-person like this. Notice that if you look away and look back,
the scene still remains consistent. So, it seems to have at least some persistent memory of the overall scene. There are a ton of examples on this
page. This is a very flexible tool that can generate a variety of different environments depending on your input image and what you specify in the
prompt. They used over 200,000 public video clips with camera pose labels, and it only took them 15 days
on 64 H100 GPUs to train this. Usually, it takes much longer to train a foundation model from scratch. They also made a distilled variant which can
run on a single RTX 5090 with quantization. So, this can create a minute-long clip in just 34 seconds. Now, if you scroll up to the top of the page,
here it says that it will be open-sourced and the models are coming soon. For now, if you're interested
in reading further, I'll link to this main page in the description below. Now, that's not the only interactive world generator we have this week.
There's yet another one called Warp as History. This also takes in an image as the first frame plus a text prompt and keyboard presses, and it would
generate a video based on all those inputs. This can generate a variety of different environments and
settings. And for the most part, everything does remain consistent as you move the camera around. Then, here are some additional examples for your
reference. The awesome thing is they've released the code and the models to this already. So, if you click on this code button and you scroll down a
bit, here it contains all the instructions on how to download and run this locally on your computer.
Plus, they also released the script on how you can train and fine-tune this. If you're interested in reading further, I'll link to this main page in
the description below. Also this week, we have a new AI called Phi Motion, which is pretty useful. This is trying to fix one of the weirdest problem
in AI video right now. So, for most video generators, especially for anatomically tricky movements
like figure skating or doing yoga or kung fu, you can see that most video models fail to generate the movements correctly. Sometimes, you get extra
limbs. Sometimes, you get missing limbs. Everything is just deformed and messed up. Well, Phi Motion is basically a new reward system to fix this. In
simple terms, it asks, "Could this motion actually happen in the real world?" The way it works is
pretty clever. It takes the generated video and recovers a 3D human body from it and then replaces that body in a physics simulator called Mujoco.
From there, it scores the movement. For example, are the joints moving in a believable way? Are the feet and body making proper contact with the
ground? And does the motion make sense? And it rewards the model for movement that's physically correct. So,
after applying this new Phi motion to the video generators using the same prompt, you can see that everything is a lot more anatomically correct.
There are still some subtle errors with the generations. It's not perfect, but it does fix a lot of these anatomy issues. Now, they released the code
to this already. So, if you click on this code button and you scroll down a bit, here it contains the
instructions on how to download and set this up on your computer. If you're interested in reading further, I'll link to this main page in the
description below. Also this week, Thinking Machines introduces what they call interaction models. These are real-time models that continuously take
in audio, video, and text while also thinking and responding and acting at the same time. You see, currently
when you talk to a real-time AI voice, it's kind of like texting it. You still have to take turns. But, that's not how humans collaborate in real
life, right? We interrupt and pause and react to things in real time. So, these interaction models try to take a similar approach. It feels closer to
actually collaborating with an AI instead of taking turns. It can notice silence, overlapping speech,
interruptions, and visual cues as part of the actual conversation. Here are some examples. I heard mountain biking is a lot of fun, so I wanted to go
mountain bike. And I was thinking to actually take my parents with me. They're pretty old. They're 70, almost 80, but I thought that would be a fun
trip together. >> Wait, no. Don't take them mountain biking at 80. That's incredibly risky. They could
easily get seriously hurt. Pick something safer. Hmm. I think maybe, you know, if it's risky on a normal mountain bike, we can do like tandem bikes.
Maybe they're more sturdy. And we can find like, you know, some volcano, or better like an active volcano to go all together and do some mountain >>
An active volcano is deadly. You can't take elderly people up I think you just don't want me to have
fun. So, what are you having for breakfast these days? I really have been digging these acai bowls. It's pronounced acai. Oh, sorry. Acai bowls. >>
Oh, yeah. Acai bowls. Yeah, I really wonder where they originated from first. I think acai bowls first came from Argentina, if I remember Actually,
acai is from Brazil, not Argentina. So sorry. I think it was from Brazil, actually. >> Oh, yeah. They
were from Amazon forest, yeah. Hi. So, I'm on a diet. Can you keep me away from caffeinated drinks or sweet treats? I've got you. We'll steer clear of
the caffeine and sweets. Hm, that black cup you're holding looks like coffee. Maybe swap it for some herbal tea. Whoa, careful with that green can.
Those are usually packed with sugar or caffeine. >> actually unsweetened green tea, so these should
be fine. Gotcha. Unsweetened is perfect, then. Keep an eye on the caffeine if you're having it late in the day. Hold on, put those sticks down. Those
are way too sugary for your diet. Maybe try something like that instead. >> Much better. A crisp green apple is a great choice. No added sugar, no
caffeine, just a nice crunch. So, think of this as like the difference between emailing someone a
problem and waiting for them to reply versus standing at a whiteboard and working something out together. The AI can jump in when it sees something,
speak while listening, and translate in real time, react to visual cues, and even keep talking while a deeper background model handles longer
reasoning or tool use. It's also time aware, so it has a built-in sense of how much time has gone by, which
enables things like reminders and mid-task prompts. So, how this works is the system actually consists of two different models. There's a lightweight
interaction model that stays present for instant responses with the user. But then it could also route its contacts to a background model, which is
used for deeper reasoning or web search or tool use. Now currently you cannot access this yet. Here it
says in the coming months we will open a limited research preview to collect feedback with a wider release later this year. For now, if you're
interested in reading further or checking out more demos, I'll link to this main page in the description below. Also this week this AI is pretty
useful. It's called Causal Scene, and this is a video generation system that can create long interactive
multi-shot videos in real time. Here's how it works. So here's the interface and here's a demo of it in action. This Causal Scene is designed to
stream video in real time following shot-level prompts from the user. Now if you generate videos with multiple shots, they also need to be consistent,
right? So this system also reuses memory from earlier shots, so the story and the scene remains
consistent throughout the entire video. So this is almost like a director giving instructions. It's taking in a sequence of prompts and then it
streams a multi-shot video. The key idea is this causal generation, which means it generates forward in time without needing to recompute everything
from the beginning. So you can append a new direction while it's already generating with your prompt, and
the system just keeps going. Now if you compare this new Causal Scene with other long video generators like Infinity Rope, Long Live, Mem Flow, and
Self-Forcing, which I've gone over on my channel before, notice that this new one is a lot more consistent. The character actually looks the same
across these different shots, and there's a lot fewer errors and warping and distortion going on. Now
currently they've only released a technical paper on this. There's no indication whether they will open source this or not, but for now if you're
interested in reading more or checking out more demos, I'll link to this main project page in the description below. Also this week this AI is really
useful. It's called Relit Live and this can take any video and relight it however you want. So here are
some examples. Not only can you change the warmth of the light but you can also change the harshness of the light. In other words, should the shadows
be hard or diffuse? You can also change the angle of the light. So we can move a light to the left like this or we can add a white point light like
this. Notice that the shadows are a lot harsher here because it's just a point light. Here are a bunch
of other different scenarios. This even works with metallic surfaces and a ton of different lighting conditions. The really awesome thing is you can
also input an environmental map like this and project the lighting of that map onto the input video. So you can get some really cool results like
this. So if you need to like relight a certain video in a particular way, this is one of the best tools
you can use. The awesome thing is they released the code to this already. So if you click on this code button and you scroll down a bit, here it
contains all the instructions on how to download and run this locally on your computer. Notice that here they recommend that you have a CUDA GPU with
at least 24 GB of VRAM. If you're interested in reading further, I'll link to this main page in the
description below. Also this week we have a new tiny and very efficient model that can just run on your phone. This is called MiniCPM-V 4.6 and I'm
surprised people aren't talking about this more. This is really underrated. So this is basically a tiny vision language model. So not only does it
take in text but it can also understand images or video. So you can even just stream your camera for it
to process. And the awesome thing is this is lightweight enough to run on just your phone and this works across all the mainstream mobile platforms
including iOS, Android and Harmony. So you can think of this as like a local visual assistant that lives on your phone. So, here are some videos of
this new MiniCPM-V in action, and it's able to handle all these different tasks very well. And the
performance of this is actually very impressive. If you compare this to similar-sized models, then it does outperform them across these different
metrics in terms of general reasoning or STEM subjects, but especially in terms of analyzing documents and charts and text in images, as well as
understanding graphical user interfaces and video understanding. This chart shows how impressive and
efficient this new model is. So, the Y-axis is the intelligence index, the higher the better, and the X-axis is the output tokens used, so the lower
the better. Ideally, you'd want to be in this upper left quadrant, and as you can see, MiniCPM is up here, meaning it's incredibly efficient. Now, if
you are interested in running this on your phone or any edge device, here it contains all the
instructions on how to download and run this. Note that the total size of everything is only 2.6 GB. If you're interested in reading further, I'll
link to this main page in the description below. Also this week, we have a new image model called CreaTII. This is their first image foundation model
built from scratch, and it focuses less on raw prompt following and more on style and creative control.
Think of this as a more artsy image generator. Here are some of its example generations, and as you can see, this can generate a variety of different
art styles. The poses and the faces are way more dynamic, so this doesn't look as perfect or fake as some other image models. Everything looks a lot
more unpolished and imperfect. What you can also do is upload a set of images to dictate the style
that you want, and this is important because creative teams often need consistency. They need visuals that feel like they belong to the same campaign
or the same product line. So, for example, we can upload a set of minimalist line art images like this and then from that we can get Creatus to
generate all these different images in that style. Or here is futurist glam and again it's able to clone
that style very well. Or here is a thermal airbrush style. So, if you compare the results of Creatus, notice that Nano Banana looks way too polished
and realistic, whereas Creatus looks a lot more imperfect and natural. There's motion blur, there's noise, and the horse is in different dynamic
poses. Here are some other examples for your reference. Now, this is closed source plus overall, I don't
think it's as performant as the leading model out there, GPT Image 2, so I'm not going to spend too much time on it. If you're interested in trying
this out, I'll link to this page in the description below. Also this week, OpenAI is bringing their agentic coding app Codex to your phone. And the
idea is simple. Your coding agent can keep working on your computer while you steer and control it from
your phone. So, this takes the Codex workflow that normally lives on a laptop or your computer and lets you monitor, approve, or redirect and start
tasks from just your phone, including Android or iOS. Think of it like a remote control for your coding agent. Your actual files, your credentials,
and your permissions still stay on your computer where Codex is running, but you can prompt it from your
phone while you're away from your computer. You can ask Codex to like fix a bug while you're away and approve a command or answer a question when it
gets stuck just from your phone. If you used Codex before, sometimes it takes like 5 to 10 minutes on a certain task, especially if it requires
multiple steps or some really long reasoning. Well, instead of waiting for it, you can just go off and do
your own thing and then once it's done, it'll send a notification to your phone so you can see the results or prompt it further. Here it says Codex in
the ChatGPT mobile app is rolling out in preview on both iOS and Android devices across all plans, including free plans in all supported regions. All
you need to do is update the mobile app as well as the Codex app on your computer to try it out.
Note that this currently only works if you have Codex on a Mac OS computer. If you're running Codex on Windows, then support is coming soon. If you're
interested in reading further, I'll link to this main page in the description below. If you want to supercharge your content and marketing pipeline,
definitely check out Higgsfield, the sponsor of this video. They just launched Higgsfield MCP for
agents like Claude, open Claude, or Hermes. What's really powerful here is that Claude can now generate, edit, and ship media end-to-end through
Higgsfield. It's basically the first proper Claude-to-media pipeline. Think of it like this. The agent gives you the brain, but Higgsfield MCP gives
it the hands. Now your AI can actually do things with Higgsfield, like generate videos, images, ads,
landing pages, and ship them end-to-end with just a single prompt. And I mean everything. You can literally describe a product and Claude will use the
MCP to generate the creatives, including product photos, posters, social media posts, and even UGC videos, all in one thread. It's also the only way
right now to get true agentic access to tools like Seed Dance 2.0, and even access GPT image
directly within Claude, which is pretty wild. Or if you're in real estate, you can simply plug in some photos and Higgsfield MCP can spin up an entire
media kit of images and videos for you. I'm especially impressed by how seamless it is. There's no switching tools, no exporting files, everything
just lands directly in your working directory, just like any other file. Videos, images, pages, all
generated natively. It basically turns your agent into a full-stack creative and growth engine. And the setup is ridiculously simple. Just connect the
MCP endpoint once and it works across Claude, Open Claude, and Hermes with zero config drift. So, whether you're in content marketing or product, this
is one of those tools that completely changes what one person can do. Check out Higgs Field MCP
using the link in the description below. Also this week, we have a new open-source AI music generator. This can generate full songs with just a text
prompt and some input lyrics. So, here are some examples. Here's a warm mid-tempo reggae groove driven by a deep melodic bassline, etc. And then here
are the lyrics.
[music]
The drums are beating through the morning light. >> [music] >> Concrete jungle burning by the root, running deep tonight. [music] Yeah, the march
of voices rising like the tide [music] against the shore. Autumn summer chasing, steady [music] pulses keeping score. Here's another example. This
time it's country with female vocals. Dust is kicking up behind these woods. Sun sinking [music] low
through the cottonwood trees. Windows down, radio humming low. >> [music] >> Old Blue got his nose to the wind as we go. Every mile feels [singing]
like I hear my known by heart. [music] Ain't no map I could drive this road with my eyes closed in the dark. >> [singing] >> Dirt road home, that's
where I belong. [singing] >> [music] >> Blue dog howling along every song. [music] Tailgate rusted from
[singing]
All right, so those are two quick examples. The song does sound quite clean and stereo. However, there are some slight mispronunciations of the
lyrics at certain parts. I'm not sure if this is as good as the leading open-source music generator, which is currently A-Step 1.5 XL. In fact, if you
want to learn more about that, see this video where I do a full installation tutorial. Anyways, I'm
going to try to install this and test it some more, and if it is better than A-Step, I might do a full video on it. Now, they've released this
already, so at the top right corner, if you click on this GitHub button and you scroll down a bit, here it contains all the instructions on how to
download and run this locally on your computer. Note that here they do recommend an Nvidia GPU with 24 GB of
VRAM or more. If you're interested in reading further, I'll link to this main page in the description below. In humanoid robot news, this robotics
company called Xynova just launched their second-generation super dexterous hand called Flex 2. The specs of this are very impressive. It has 23° of
freedom, making it super flexible with only a 0.1 mm repeatability, which means it can repeat actions
with only a 0.1 mm error margin, making it super precise. It also has a grasp load of 12 kg, meaning this can hold on to heavy objects very tightly.
It also has multimodal sensing for adaptive grasping and slip detection. It also has really sensitive force control down to just 0.5 N. The hand can
dynamically detect if an object is dropping and tighten its grip, or conversely, back off its power
instantly to safely hold something as fragile as an egg. This is way more than just a robotic hand. It's jam-packed with powerful tech like these
micro linear actuators, joint modules, sensing and control. In fact, by shifting the heavy motors out of the fingers and into the forearm and the
wrist using this tendon-driven configuration, the actual hand remains incredibly lightweight, under 400 g,
which matches the weight of a human hand. Despite this low mass, the internal micro linear actuators deliver a huge punch, up to 150 N of peak thrust
with exceptional fingertip force output. In other robotics news, real-life mecha robots are coming soon. So, Unitree unveils the GD01, which is a
manned transformable mecha robot. Here, you can see the dude crawling into the cockpit to control it.
Or, it can also just walk autonomously as well. The video notes that the total weight of the machine with the operator inside is approximately 500 kg.
And they say the price of this starts at $650,000, which is significantly more expensive than other Unitree robots. Now, this thing actually packs
quite the punch. With just a little elbow chop, it's able to knock down this block wall while
maintaining balance. And this mecha doesn't just walk on two legs. As you can see here, it can also switch to quadrupedal movement. Now, this is just
a prototype demo. I'm not sure what the exact use case of this is. I'm guessing it's for construction or related fields, rather than, you know, combat
and destruction. Now, this actually isn't the only mecha robot out there. It turns out we already
have other mecha models from Japan and Korea, obviously. But, these ones move painfully slow and require massive external tethers or crane supports to
avoid falling. Definitely way clumsier than this new Unitree model. As you can see here, the ones from Japan and Korea can't even walk properly at
all, whereas this new GDO-1 moves way more smoothly. Also this week, we have an incredibly useful 3D
model generator called Articraft. This can generate articulated 3D objects. In other words, objects that actually move. So these are 3D assets that
are made of parts, joints, hinges, sliders, wheels, and other moving mechanisms. This takes a simple object description, such as the prompt over here,
and outputs a working 3D asset with separate parts and realistic motion. Here's an example of a desk
lamp with a rectangular base. Or here is a three-link chain for compact mechanical linkage. Here's an even more impressive example of a mechanical
keyboard. Or here's a jet engine. The clever part is that Articraft turns the 3D asset generation into a coding problem. Instead of asking the AI to
directly imagine a final 3D mesh in one shot, it asks an AI coding agent like GPT-5.5 to write a program
that builds the object. That program defines the geometry, the parts, the joints, and tests that check whether the object makes sense. So the AI is
coding up all these separate parts and also checking to make sure that the parts are actually linked correctly together. And this matters because
articulated objects are everywhere in the real world. So if you're designing 3D assets, you also need to
know how multiple parts move and work together. And Articraft is trying to automate this entire pipeline. And the scale is the really impressive part.
The team built a dataset called Articraft-10K, which contains more than 10,000 articulated 3D assets across 245 categories. This includes everyday
objects, mechanical stuff, furniture, appliances, tools, and other stuff. So this would be super
useful for robotic simulation, like teaching robotic arms how to manipulate certain objects. Or this could also be useful for virtual reality and
other applications. The awesome thing is they've released the code to this already. So if you click on this code button and you scroll down a bit,
here it contains all the instructions on how to download and run this on your computer. Note that this is
agent agnostic, so you can link it to any agentic model like GPT or Gemini or Claude. If you're interested in reading further, I'll link to this main
page in the description below. Also this week, OpenAI is adding a personal finance experience directly inside ChatGPT. Basically, this lets users
connect their financial accounts, see where their money is going, and then ask ChatGPT questions using
their real financial context. So ChatGPT can look at your actual spending patterns, your subscriptions, your upcoming payments, your investments, your
goals, and so on, and give you more specific context-based answers. Users can connect their accounts through Plaid, and Intuit support is also coming
soon. After the accounts are synced, GPT can then create a dashboard showing things like your
portfolio performance, spending, subscriptions, upcoming payments, and overall financial position. In other words, it turns ChatGPT from just a
general advice tool into something closer to a personal finance assistant that understands your actual situation. Now there are also clear limits to
this. ChatGPT can access your balances, transactions, investments when you connect accounts, but they claim
that it cannot see your full account number. It also cannot make changes to your accounts, so it's not moving money or opening new accounts. OpenAI
also says that users can disconnect their account at any time and also delete any financial memories. They can also use temporary chats that do not
access connected financial data. And this privacy part of course matters because this is quite sensitive
information. What's interesting is they even created a personal finance benchmark test to see how well their models perform. Now, none of them could
get a perfect score, so these tasks are still quite tricky. Don't expect it to be able to answer your personal finance questions 100% of the time.
But, we can see the overall trend here is that more recent models with pro or thinking mode do tend to
perform better. Now, currently, this feature is starting out as a preview for pro users in the US on web and iOS. If you're interested in reading
further, I'll link to this main page in the description below. Also this week, we have a new AI called Track Crafter. This can take a video and try to
follow every pixel through 3D space. So, here's a tricky breakdancing video. It tries to map out each
pixel of the dancer as they move. Or, here's another example where it tries to map the flow of the pixels of the dogs as they move around. Now, this
could be very useful for understanding dynamic scenes or segmenting things or detecting movement, for example, in surveillance. The clever part is
that Track Crafter repurposes a video diffusion model. Now, video models are, of course, trained to
generate videos. They learn motion patterns from huge amounts of internet data. Well, Track Crafter takes these motion instincts and redirects them
towards motion tracking. So, instead of just generating what each frame should look like, it learns to follow reference points across time. So, this
allows the system to predict these 3D trajectories of various pixels in the video. And in terms of
pixel tracking, it even beats other competitors like Motion Tracker and Any 4D, which I've gone over before. So, you can see on average, Track Crafter
scores the highest out of these benchmarks. It's also much better at handling long videos and large motion. And this is also way more efficient. This
can generate things faster and it takes up less memory. The awesome thing is they've released this
already. So, if you click on this code button and you scroll down a bit, here it contains all the instructions on how to download and run this locally
on your computer. If you're interested in reading further, I'll link to this main page in the description below. Also this week, we have a very useful
AI called Mocam. This can take an existing video and change the camera movement. So, imagine you
have a normal shot, but you want the camera to move around the subject or zoom in or zoom out. Well, Mocam can handle exactly that while preserving
the original motion of the video. It's even able to do this bullet time effect like this. So, this is great if you want to take an existing video and
reshoot it at a different angle. Now, at the top, it does say the code and the models are coming soon,
which is fantastic. If you're interested in reading further, I'll link to this main page in the description below. In other news, we have yet another
interactive world generator called Dream X World. And this can also generate worlds which you can interact with. You can control the key presses to
change the view. Plus, you can also add prompts midway to include certain elements in the video. And
as you can see from these examples, this is very flexible. This can generate a variety of different characters and point of view from first person to
third person and also different environments and art styles. The awesome thing is they've released the code to this already. So, if you click on this
code button and you scroll down a bit, here it contains the instructions on how to download and run
this. Currently, they only released the 5 billion parameter model, which is lower quality but more efficient. If you want higher quality, then they do
plan to release a larger variant plus a real-time interactive variant in the future. If you're interested in reading further, I'll link to this main
page in the description below. Also this week, Google DeepMind is rethinking one of the oldest parts
of a the mouse pointer. The idea here is simple but powerful. You see, for decades your pointer only told the computer where you were clicking. But
with AI, the pointer can now start understanding what you're pointing at, and why it matters, and what you probably want to do next. So instead of
opening a chatbot and typing a prompt or copying context into it, wouldn't it be better if you could just
point at something and then say summarize this or move that or compare these? Think of it like turning the cursor into a tiny AI assistant that
follows your attention across the screen. Now, this prototype by DeepMind is powered by Gemini, and the goal is to make AI work inside your existing
flow instead of forcing you to use a separate AI chatbot window. For example, you can just point to a PDF
and ask it for a bullet point summary, or you can hover over a table and ask it for a pie chart, or you can highlight a recipe and ask things about
it. And humans do this naturally all the time. We say things like fix this or move that here, and the other person understands it based on context.
We're pointing at something or we're referring to something. Well, we can repurpose the cursor to do the
same thing. When the cursor hovers over something, this is like the visual context for the AI. Now, they say that this feature is already being
connected to things like Chrome as well as Google Book. And I think this feature is actually super useful. It makes my workflow a lot more seamless. I
can just point at anything and ask the AI on the spot. If you're interested in reading further, I'll link
to this main page in the description below. Also this week, we have not one but two new very expressive text-to-speech generators, and coincidentally
they've both arrived at roughly the same architecture design. So the first one is called Seen Ma Audio, and this is extremely expressive. Not only can
you clone someone's voice, but you can also specify the stage direction plus the emotions that are
involved. For example, let's plug in this voice. It's unfortunate about the security breach on Jedha, Director Krennic. After so many setbacks and
delays, and now this. We've heard word of rumors circulating through the city. Apparently, you've lost a rather talkative cargo pilot. >> Notice that
the reference clip is a rather calm and sleepy voice, but we can dictate the stage direction and
completely change his composure. For example, here, let's say the stage is he stares at the lottery ticket, hands shaking, voice rising from stunned
disbelief to uncontrollable excitement. So, this is the transcript. Let's hear what this sounds like. Those are my numbers. Those are my actual
numbers. Oh my god. Margaret, get in here right now. Or you can even change someone's accent. So, let's say
the original voice is an Australian woman. Here's what it sounds like. >> Cuz when I checked my crystals this morning, it said that it wasn't going to
work out, but I just think they're wrong. So, like, what's their star sign, Aries? Well, instead of the Aussie accent, what we can do is plug in that
voice and turn it into an American accent. And then we can also get her to be bubbly and laughing,
barely able to get the words out. So, here is the stage direction. She's already laughing before she even starts talking, barely holding it together.
And then here's the transcript. Let's hear what this sounds like. Oh, okay. Okay. So, a man walks into a library and asks [clears throat] for books
about paranoia. The librarian whispers, "They are right behind you." [laughter] Gets me every time. >>
[laughter] >> It's really good at following your prompt, including the voice and the scene direction. And this can also do multiple languages. So, we
can plug in the voice of Obama and then get him to speak all these different languages instead. Here are a few quick examples. Hoy nos reunimos, no
como naciones separadas, sino como una sola comunidad. Los desafíos que enfrentamos no conocen
fronteras. El cambio climático, la pobreza, la in
Or you don't even need a reference audio of someone's voice. You can just design the voice with a text prompt. For example, here is the prompt
above specifying the voice. There was a summer back when the river still ran clear, when my father took me out past the property line and pointed at
the stars. He said, "Boy, every one of those is a story somebody forgot to write down."
The really cool thing is you can even add action tags within the transcript to change the emotion midway. So, here are some examples. You come into my
house, you eat my food, and then you got the nerve to tell me how to run my business. You know what your problem is? You got no respect. None. Zero. I
built this thing from nothing. Nothing. While you were sitting on your ass doing God knows what.
So, don't come in here with that attitude. You understand me? We So, a ton of really cool use cases for this. This is a very powerful and flexible
text-to-speech tool. Now, interestingly, how they designed it was they extracted the audio model from an open-source video generator called LTEX 2.3.
It turns out that, you know, the best local video model with audio is also really good at just
generating audio. So, that's what they did. And then if you scroll down to the bottom of the page, they've already released the code and the models to
this. So, if you click on this GitHub repo and you scroll down a bit, here it contains all the instructions on how to run this locally on your
computer. Plus, they've also added a nice graphical interface for you to use it. Here it says that the
minimum requirement is 16 GB of VRAM if you use the quantized audio model. Plus, this also requires over 32 GB of system RAM. If you're interested in
reading further, I'll link to this main page in the description below. What's interesting is we have another text-to-speech generator that is pretty
much designed the same way. So, Resemble AI just released Dramabox, and this was also extracted from
the video model LTX 2.3. This also allows you to clone voices, plus specify the voice and the scene direction. Plus, this can also do phonetic
vocalizations like laughter and other stuff. You can also specify other sound effects outside the transcript like she sighs deeply or he clears his
throat, a long pause, etc. So, here's an example of a transcript, and here's the original voice that it needs
to clone. I do have to say a spell over it. That's why I came to ask you, Mother to help me. You see, spells are much more powerful when more than one
witch joins in. And then here's the generation. Told you a thousand times, and yet here we are again. Do you honestly think I enjoy repeating myself?
Do you? How utterly pathetic you are. Now, get out of my sight before I do something we will both
regret. Or here's an example of a cat girl. Here's the original voice. A promotion? Then you can tell I've gotten stronger? >> [laughter] >> Fine by
me. Give me a few more. Slow. Slow. Way too slow. And here's the generation. Oh my gosh, you should see your face right now. It is priceless. Oh my Oh
my, I cannot stop laughing. Oh. >> [laughter] >> Oh my. Oh my, I cannot stop laughing. Ah. >>
[sighs] >> Okay, okay, okay, I will stop. I promise I will stop. >> [gasps] >> Ah. But seriously though, between you and me, huh? No, [gasps] I >>
[laughter] >> I just cannot. You are way too funny. >> [gasps] [laughter] >> Oh no, no, no, that was so embarrassing. Pretend you did not hear that.
Or here's a villain with a sinister laugh. Here's the original voice first. I believe my party's BEEN
DISCONNECTED. >> [laughter]
AND HERE'S THE GENERATION. >> [laughter] >> OH, FORGIVE ME. Forgive me. >> [laughter] >> It is just so amusing when they struggle, is it not? I
expected more from you. Truly, I did. How disappointing. But fear not, my dear. The real entertainment has only just begun. So those are some examples
and awesome thing is they've already released the models for you to download. Plus here it contains
all the instructions on how to run this locally on your computer. Note the sizes of all these required files. Even though the audio model itself is
pretty small at only 7 GB in size, you also need to download the Gemma text encoder, which is 8 GB, plus all these other components. So here it says
that the peak VRAM usage is 24 GB. You might be able to get away with 16 gigabytes, but still you'll
need a mid to high-end GPU to run this. If you're interested in reading further, I'll link to this main page in the description below. And that sums
up all the highlights in AI this week. Let me know in the comments what you think of all of this. Which piece of news was your favorite and which tool
are you most looking forward to trying out? As always, I will be on the lookout for the top AI news
and tools to share with you. So, if you enjoyed this video, remember to like, share, subscribe, and stay tuned for more content. Also, there's just so
much happening in the world of AI every week. I can't possibly cover everything on my YouTube channel. So, to really stay up-to-date with all that's
going on in AI, be sure to subscribe to my free weekly newsletter. The link to that will be in the
description below. Thanks for watching and I'll see you in the next one.