AI never sleeps and this week has been absolutely insane. Your robot girlfriends are coming soon. Alibaba finally releases their latest video
generator. We have a new open source image model that's really good with text posters and infographics. AI can now autonomously do stuff in popular
software like Photoshop and Blender. And forget just using one agent, we now have recursive multi-agent
systems that get better and better. This AI is great at turning videos into 3D. We have a new image model optimized for anime. This AI can
automatically detect cuts and transitions in a video. We have a new way for AI agents to autonomously read and write research papers. We have some
ridiculous humanoid robot demos and a lot more. So, let's jump right in. First up, this is a really fascinating AI
that can be very useful for video editing. So, it's called Omnishotcut and this is able to take in a video and detect where cuts happen in a video.
It's also able to detect what kind of transition happened at each cut. For example, here we can see it being able to detect things like hard cut or
dissolve or next we have fade or even slide or cross zoom. So, this is really useful if you want to
detect the exact timestamp of where cuts or transitions happen in a long video. Here it says they collected 2.5 million raw internet videos and then
from that they also produced 300,000 synthetic training videos containing over 11 million labeled transitions. So, after plugging all this data into
the AI, now it knows how to detect these hard cuts and different types of transitions. If you scroll
up to the top of the page, they've released a free hugging face space for you to try this out. Or you can also click on this code button which
contains all the instructions on how to download and run this locally on your computer. Note that this is actually pretty tiny at only 164 megabytes
in size. If you're interested in reading further, I'll link to this main page in the description below. Also
this week, the long-awaited Happy Horse video model by Alibaba is finally out. Now you can sign up for free on their official website, which I'll link
to in the description below. And at the top here, you can see that you do get some free credits for a few generations. Now Happy Horse is ranked
number one on this independent artificial analysis leaderboard for both text-to-video and
image-to-video. In fact, the lead for text-to-video is pretty insane. This is almost like 100 points above the legendary Seedance 2.0. So I had high
expectations for Happy Horse. However, upon testing, I was actually quite disappointed to be honest. Here are a few of my tests. For my first test,
here is my classic Disney princess prompt. 3D Pixar animation, a princess wearing a glittery white
dress running away from a massive dragon with glowing red eyes in the forest. But that's too easy, so let's continue the story further. The dragon
breathes fire, which ignites the ground and foliage behind her. Reaching a river, she quickly jumps onto the drifting debris to cross. The princess
then looks back to see the dragon roaring in frustration as it can't cross the river. Here's what we got
from Happy Horse.
You can see the physics of everything are completely wrong. There's one part where the dragon runs ahead of the princess. This just looks pretty
awful, and it doesn't really follow the prompt very well. Now if you compare this with Seedance's generation, here's what we got.
You can see SeaDance pretty much aced this video. It's incredibly good. It follows all the sequential actions that were specified in the prompt.
All right, here's another very tricky example, a satellite view of planet Earth, then zoom into a drone view of New York City, then zoom into an
office building, then zoom into a view of a person scrolling TikTok. It has to be one continuous shot. Here's what we got for Happy Horse.
You can see again the physics are just completely wrong. The dude is like way too big inside this window, plus the TikTok screen on his phone also
looks completely wrong. Whereas if you contrast this with SeaDance 2.0, it is one continuous shot while zooming through all these different levels,
and it does reach the end with this woman scrolling TikTok. Again, this is really good. So I just stopped
there. I mean, this is nowhere near the insane quality and flexibility of SeaDance 2.0. I'm not sure how the hell Happy Horse is able to rank number
one on artificial analysis with such a huge lead, but at least for my personal tests, this is still clearly behind SeaDance 2.0. And that's why I
didn't do a full review video on Happy Horse this week, cuz it's not state-of-the-art. But if you want to
test it out, you can sign up on their site for free, and you do get some credits to test it out. So if you're interested, I'll link to this main page
in the description below. Also this week, we have this new AI called Mocap Anything version two, and it's a big step forward in motion capture. In
simple terms, this takes a regular video like a person or a creature, pretty much anything moving, and
it turns it into clean animation ready skeletons that you can apply to any 3D character. And the key thing here is it works across totally different
skeletons. It can be humans or other animals or even fictional characters with weird limbs. Now, what makes this version special is how it actually
does that. Older systems usually split the problem into two steps. First, they guess where the joints
are, like elbows, knees, and wings, and they try to convert that into rotations using a separate step called inverse kinematics. The problem is the
second step isn't learnable, and it often gets things wrong. You get weird artifacts, like joints twisting unnaturally or spinning. So, what this
system does instead is go fully end-to-end. It basically learns everything together. It looks at the
video, predicts the pose, and directly outputs the final joint rotations all in one unified model. That means it can actually correct itself and
optimize for the final animation quality, and not just intermediate guesses. And compared to the previous version, version one, which is on the top
row, you can clearly see the difference. The old one has jitter and weird joint spinning, whereas this
latest version two looks a lot more stable and accurate. So, overall, this is a big deal if you care about animation, games, VFX, or even robotics. It
removes a ton of manual work and makes MoCap way more flexible. With just one reference video, you can animate pretty much anything. Now, if you
scroll up to the top of the page, they have released a GitHub repo, and here it does contain
instructions on how to install and run MoCap anything version two. It also includes the script on how you can train this yourself. If you're
interested in reading further, I'll link to this main page in the description below. Also this week, Inclusion AI releases a super performant
medium-sized model, Link 2.6 Flash. Fun fact, Inclusion AI is yet another spin-off lab from Alibaba, which is
different from the Tong Yi Lab, which made Wan and Zi Image, and also different from their other lab called ATH, which made happy horse. So, Alibaba's
got a ton of different labs working on AI at once. At a high level, this is 104 billion parameters, but when you use it, only 7.4 billion are active.
So, this makes it super efficient and cheap to run. If you look at these agentic benchmarks, it's
pretty competitive with other similar sized models, although they're kind of cherry-picking the models here. We don't see more recent models like Qwen
or Gemma 4. Now, it's not just about performance. Ling 2.6 flash is optimized for speed and efficiency. As you can see from these charts here, it's
way faster than the other models, especially when the context length gets high. In other words, when
your prompt gets very long. Anyways, this is just another open-source model to put on your radar. 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 for anyone into anime image generation. Z anime just dropped. At a high
level, they took one of the best open-source image models out there, Z image base, and they fine-tuned
it from the ground up on anime images. This isn't just a LoRA, but a fully trained model designed to handle anime styles, which gives it way more
consistency and range. So, here are some examples of images it can generate. It has all the awesome features of Z image, being quite tiny and fast at
only 6 billion parameters. If you scroll down the page, they even released a smaller FP8 version, which
is only 6 GB in size, so this can fit comfortably with most consumer GPUs. They also released GGQFs for users with non-CUDA GPUs. They even made a
distilled four-step version, so it only takes four steps to generate an image, which means it usually finishes in seconds. So, this is insanely fast.
Anyways, if making anime images is your thing, definitely check out this model. It's pretty capable.
I'll link to this main page in the description below if you're interested. Next up, this project is a really interesting shift in how multi-agent AI
systems actually work. It's called recursive multi-agent systems, and the idea here is pretty simple. Instead of AI agents talking to each other in
text, they collaborate in their own latent space, and they loop their thoughts over and over to refine
their answers again and again. What exactly does that mean? Well, normally when you have multiple AI agents working together, they pass messages back
and forth in just plain text. That's slow, expensive, and very inefficient. Every step requires generating words, decoding them, and then rereading
them. Well, this system flips that completely. Instead, the agents communicate using their internal
representation, also known as their latent thoughts, before they get turned into words. So, this takes a problem as input, it runs multiple agents in
a loop, and only outputs text at the very end. Think of it like a group of experts silently thinking together, refining an idea over multiple rounds,
and only speaking once they reached the final answer. And this turns out to be pretty huge because
staying in latent space, the system gets faster, cheaper, and actually more accurate. We're talking about a 2.4x speedup, and also 75% fewer tokens
used, and over an 8% boost in accuracy across benchmarks. You can even chain multiple different agents together, like a planner agent, a critic, and a
solver agent. So, the output of the last agent feeds back into the first, and the process repeats
again. And what's really interesting is that most of these rounds happen silently. There's no text generated at all until the final step, and that's
why it's so efficient. Now, they tested this across a bunch of different setups, including sequential agents, mixture of specialists, and even
distillation setups where a smaller model learns from a bigger one, and in pretty much every case, this
recursive approach outperformed standard methods. Here's one wild example. On a math problem, for the first round, it gets the answer wrong, but by
the second round, the system corrects itself, and by the third round, not only does it get the answer, but it explains it more clearly. So, it's not
just improving accuracy, it's actually refining reasoning over time. And the scaling behavior is also
interesting. The deeper you let the system loop, the better it performs. While traditional text-based systems start to plateau or get even worse, this
one just keeps improving. As you can see here, across all these different benchmarks, the more rounds you add, the better its performance. So, why
does this matter? Well, this is basically a new way to scale AI. Instead of making models just bigger
or adding more data, you just let multiple agents think together and loop that process in hidden latent space. It's cheaper, faster, and more
flexible. And if this direction keeps working, it could become the default way multi-agent AI systems are built going forward. The awesome thing is,
at the top of the page, they've already released the models and the code to this. 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. Next up, this AI is also pretty cool. It's called Vista 4D, and
this lets you take a normal video and turn it into a 4D scene. In other words, everything is in 3D with
the element of time. Here are some other examples for your reference. And this means that you can also reshoot the scene using any camera movement or
angle you want. So, this is a super flexible tool. Here are some other examples for your reference. You can see that here, we can even do some really
cool zoom parallax effect like this. And you know, all the characters and all the details from the
original video stay consistent the entire time. Now, if If compare this generation from Vista 4D to other 4D generators like Recam Master or
Trajectory Crafter, which I've gone over on my channel before, you can see that by far Vista 4D is the most coherent and the most accurate. Here's
another example for your reference. It's able to keep the details of the character and everything else
consistent throughout the generation. Now, the really cool thing is because it essentially converts this into a 3D point cloud, you can then edit the
point cloud however you want and then reconstruct that into a video. For example, in the top row, the source video only has one car, but we can add an
extra car to the 3D point cloud and then it would render a video with two cars in the scene. Or
here's something even crazier. You can even insert an entire video, for example, this old couple sitting at a bench and then just add that 3D point
cloud into the original 3D scene and now we can merge these two things together. Pretty cool. Or here's another example where in the top row, we can
choose to completely remove the old man at the front in the 3D scene and afterwards we can re-render
this into a new video without this old man. Or alternatively, we can also take this video of a rhino and extract the 3D point cloud of the rhino and
insert that into the 3D scene of the original video and re-render it so that we get the exact rhino walking into the scene. Or here's an even cooler
example. Now, from the source video, you can see that this is a fixed point. We don't really know what
the rest of the room looks like, but after plugging it through Vista 4D, it's able to extrapolate and just guess what the sides of the room would look
like. Now, this might not be an accurate render of the room, so what we can do instead is plug in a video of what the entire room actually looks like
and then feed that back into the point cloud data and here is what we get. So now it's able to do
this orbiting shot with accurate details of the room. The awesome thing is, if you scroll up to the top of the page, they've already released the code
to this. 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. Note that the model is roughly 21 GB in size, so you should be able to fit this
on mid-to-high-end consumer GPUs. Anyways, if you're interested in reading further, I'll link to this main page in the description below. Also this
week, Meta releases their latest image generator and editor called Tuna 2. So, this is very similar to like Nano Banana and GPT Image 2. This can
generate a variety of different images. It has pretty good prompt understanding, so you can even generate
complex images like this. It's also great at rendering text and different art styles. But, just from these initial examples, it doesn't seem to be as
good as Nano Banana or GPT Image 2. Now, in addition to generating images, this can also edit images. So, for example, we can turn this woman into
Lego or this into a Van Gogh style painting. Now, if you scroll up to the top of the page, they have
released a code button to this. They have started adding some installation instructions, but if you scroll down a bit here, it says, "Due to
organizational policy constraints, we are unable to release the full model weights. Instead, they just plan to release a foundation checkpoint with a
smaller number of layers or basically a nerfed version." I mean, what BS excuse is this? But, not surprising
since this is coming from Meta, which hasn't really open-sourced anything since Llama. Anyways, if you're still interested in reading further, I will
link to this main page in the description below. Also this week, this AI is very useful for 3D reconstruction. So, it's called Any Recon, and this can
take a series of photos of a scene and it can reconstruct a 3D point cloud of the scene. The key
here is you can just give it a bunch of images, as VUS 2 and they don't even need to be perfectly aligned or captured in order and it's still able to
output a coherent 3D reconstruction of the scene. So, you can just walk around and casually record stuff and it can reconstruct that into a solid 3D
model. And that's a huge deal because traditional methods struggle when the input is sparse or
inconsistent. Some newer diffusion-based methods try to fix this by generating missing views, but they only usually look at one or two frames at a
time and the results are inconsistent. This new Any Recon takes a different approach. It kind of keeps a global memory of the scene. So, think of it
like a continuously updated mental map of the 3D environment. So, instead of forgetting earlier frames,
it remembers them and uses them throughout the entire process. And so, afterwards, it's able to fill in the blanks and generate a very coherent video
and understanding of this entire scene. Here are some examples. Here we just input five different photos taken at random angles of the scene and it's
able to reconstruct it into a full 3D scene. Or here's another example where we only have four
photos of completely different areas at this location, but it's able to stitch them together very seamlessly. Or here's an even trickier example where
we only have two input views, but it's still able to stitch everything together and generate a very realistic 3D reconstruction of the scene. And if
you compare this with other methods like View Crafter or Uni 3D, which I've gone over on my channel
before, you can see that this new one is just a lot more detailed and coherent. The awesome thing is they've released the code to this already. So, at
the top of the page, if you click on this code button, here it contains all the instructions on how to download and run this locally on your computer.
Notice that this Any Recon is pretty tiny at only 614 MB in size and for the video diffusion model
this uses VAE 2.1, which should fit within most consumer GPUs. If you're interested in reading further, I'll link to this main page in the description
below. If you want to seriously level up your productivity, you should check out Merlin AI, the sponsor of this video. It's an all-in-one tool that
gives you access to all the top LLM models like GPT, Gemini, Claude, and more, all under one
subscription. And Merlin is already used by over 2 million people, so this is not something new. And real quick, Merlin AI is offering an exclusive
75% discount for you guys. It normally costs 19 bucks a month, but this brings it down to just five bucks a month. I'll show you exactly how to get it
at the end. What I like is how it actually fits into your workflow. For example, let's say I'm
researching a topic. Normally, I'd open multiple tabs, skim articles, copy and paste stuff into ChatGPT, but with Merlin's browser extension, I can
just click deep research, and it scans a bunch of sources and pulls everything together. Or if I'm reading a long article, I can just use chat with
webpage to instantly summarize it or ask follow-up questions without leaving the page. It's super
seamless. Another feature I've been using a lot is projects. It basically lets me build a kind of personalized knowledge base for the stuff I work on
frequently. It keeps the context across all chats, so I don't have to keep repeating myself every time. There are other features, too, like quick web
search for fast answers, academic search for more technical stuff, and even Reddit search to see
real user opinions. The advantage is having multiple AI models in one place. Each one has its strengths, and Merlin AI lets me switch between them
seamlessly depending on my needs. Merlin AI has a pretty insane discount exclusive for you guys. If you look at each model separately, ChatGPT is like
20 bucks a month, Claude is also 20 bucks a month, Gemini is also 20 bucks a month, and that adds up
fast. Merlin AI normally costs 19 bucks a month, but with my exclusive discount code AIS5, the price comes down to 60 bucks for a year, which is just
5 bucks a month. Merlin AI keeps the cost low because it buys API access in bulk, but most users typically use around two to three bucks worth of
credits per month and don't fully utilize their individual subscriptions. So, that's how you can access
multiple LLMs at such a low cost. Definitely check out Merlin AI using the link in the description below. Also this week, this project is a really
interesting shift in how research itself might work. So, it's called The Last Human-Written Paper: Agent-Native Research Artifacts, and here's the
core idea. In the near future, most computer science or even just general scientific papers will be
written by AI, and most will also be read by AI. So, what does that mean and why does it matter? Well, right now a research paper is basically a clean
story. What you see published is just the final idea, the methods, and the actual findings that work. But in reality, that's not how research actually
happens. In reality, it's messy. You try a bunch of different ideas, and most of them fail. You
hit weird bugs, you pivot, and finally something works. But all of this gets thrown away when the paper is actually written and published. So, this
leads to a huge loss of knowledge. And this creates two problems. The first one is what they call the storytelling tax. You only see the winning path,
but you don't see the dead ends and the decisions that led there. So, there's a lot of useful insight
that disappears. They also talk about this engineering tax. And here's the idea behind it. Papers don't include enough detail to actually reproduce
work. Things like hyper-parameters and fixes are often missing. So, anyone trying to rebuild the system basically has to rediscover everything by
themselves from scratch. Again, all of this is just lost knowledge. But here's the solution. So, they call
this the Agent-Native Research Artifacts, or ARA for short. Think of it like this. Instead of a PDF research paper with just an abstract,
introduction, methods, results, etc., here research becomes a structured package that's readable by AI. So, here's what the structure looks like. You
still have a human-readable overview, but you also have the actual logic, like claims and experiments, the full
code and configs needed to run it. You also have a complete graph of the research process, including failed attempts, and all the raw evidence like
logs and results. So, instead of a polished story, you get the entire trajectory. And here's the interesting part. This isn't something that
researchers have to manually write. They also introduce a system called the Live Research Manager, and this
captures everything automatically while you work. Every experiment, every failure, every tweak, it's all recorded in the background automatically by
this agent. So, over time, this artifact basically builds itself. And in terms of results, this actually helps AI systems learn a lot. It improves
understanding, it makes reproducing the experiment way easier, and even helps with extending the
research further. So, the big idea here is a shift from narrative to knowledge. Instead of compressing research into just a boring polished scientific
PDF that humans can read, now you can store the full research process so that both humans and AI can actually use it. And if AI ends up being both the
main reader and contributor to research, then honestly this kind of format starts to make a lot
more sense in the future. The nice thing is if you scroll up to the top of the page, they've already released the code to this. So, if you click on
this code button, and you scroll down a bit, here it contains all the instructions on how to download ERA on your computer. If you're interested in
reading further, I'll link to this main page in the description below. In humanoid robot news, we have a
new ultra human-like robot called Kai from robotics company Kinetics AI. This is a sleek humanoid robot with a white body but no face. It's just a
metallic plate which makes it a bit unsettling. In this demo you can see it doing a ton of diverse actions like carrying bags, playing with this kid,
picking up and placing objects in this bag, and even zipping up the bag or even playing ping pong,
although I'm not sure how staged this scene is. It can also fold clothes and a ton of other stuff. Now, this robot has exceptional dexterity with 115
degrees of freedom across the body and 36 degrees of freedom dedicated to just the hands. This allows the robot to move very naturally and perform
fine motor tasks such as handling delicate stuff. Things that are traditionally very difficult for
robots. It's also equipped with full body tactile skin, which means it doesn't just see the world, its entire body can actually feel things. This
allows it to adjust its grip strength so it doesn't crush an orange, and it also allows it to be aware of and react to humans touching it. This is
obviously very important for safety in household setting. The robot's brain is powered by the Kai world
model, which is capable of long horizon tasks and self-correction. So folding laundry is a perfect example of this. Fabric is unpredictable and the
robot must understand the goal, manipulate the fabric, and recognize that if it makes a mistake it should be able to correct it on the fly. This
allows it to do a diverse range of stuff rather than just following rigid programmed actions. Also this
week we have this new demo from Robot Era. It won't be long before we see manufacturing and warehouse pipelines completely automated by humanoid
robots. In this video there are like dozens of these Robot Era L7 humanoid robots performing different sorting tasks in this logistic center. Here you
can see each robot working at an automated sorting line, specifically inspecting and handling these
small parcels. The L7 robot uses embedded vision, depth sensing, and real-time positional feedback for control and navigation. And here L7 is
apparently able to pick parcels from conveyor belts, identify them with cameras, and sort them into destination lanes. The video shows multiple robots
working in this environment, suggesting that an entire fleet of these can be orchestrated at once. In fact,
Robot Era is already planning to increase fleets to 1,000 units. Imagine 1,000 of these deployed at a factory automating different things. Also this
week, Neotic's Robotics released a demo of their latest bionic desktop humanoid robot. Now, Neotic's is known to build robotic heads with
ultra-realistic faces and expressions. And as you can see here, she can talk, blink, move her eyes around, and
show very fluid micro expressions just like a human. So you ain't just getting a dead stare. Everything is also very intricately designed with
synthetic skin, eyelashes, and so on. It's intended for social interaction, companionship, and even professional services like customer support or
consultation. Now, in addition to Neotic's, another robot company called TF-Bot also shared a short video of
their latest Android head called Ella. And here they're even explicitly positioning this as a girlfriend robot. So if you can't get a girlfriend in
real life, don't worry, Ella is coming soon. Now, this head does look fairly natural. It's also quite expressive, though I'd say her eyes don't look
as realistic as the previous Neotic's head. Anyways, these are just two more robotic head companies to
add to your radar. A few weeks ago I featured another one by a company called A Head Form, which also looks super realistic. I think in the near
future we're going a lot more of these, and it's going to be harder and harder to tell them apart from actual human heads. Also this week we have a
really cool model called Sense Nova U1. Now this isn't just an image generator or editor, but instead it's
actually a full multimodal model that can take in text or images and also output text and images end-to-end. So it's a unified model for
understanding, reasoning, and generating images, which makes it incredibly capable at making really complex images. So here are some examples. You can
see this poster contains a ton of different elements and text, but it's able to render this very well. Or here's
a Chinese example on how to create this juice. Here's another really complex example consisting of a ton of different panels and different elements
and text. So it can easily make posters and infographics with a ton of text and elements that approach the level of the top closed image models. Now
instead of these infographics, it can also do just regular images like these. Although I find that it
does produce plastic-looking skin, especially for portraits, this kind of looks a bit fake. So if you're looking to generate photorealistic stuff,
then there are much better options like Z image or Flex client. This is also great at editing images with just a text prompt. So here are some
examples. We can turn this dude's jacket to yellow or turn this into anime style or make this guy smile or
replace some text. Now again, we already have a ton of image editors that can do this, so that's not the strength of this Sense Nova model. The
strength of this one is that it can output text and images at once. So for example, you can get it to output a storyboard like this that's interleave
with both text captions and images. And because this is essentially a reasoning model, you can also get it
to solve visual puzzles like this. And this actually outperforms even the top image models out there, including Nano Banana and GPT Image 2. Or if you
ask it questions about images, it's actually able to analyze the image, reason through it, and figure out the answer accordingly. Now, if you look at
these charts, you can see that it has state-of-the-art performance across many understanding,
reasoning, and image generation benchmarks, even beating the leading open models like Z Image, Qwen, and Ernie, especially in text rendering, visual
reasoning, and infographics. And again, it's not just its image generation that's impressive. If you look at this chart comparing multimodal vision
models like Qwen 3.5 and Google's Gemma 4, for some of these visual and spatial understanding
benchmarks, Sense Nova even beats these other models. And here's the key shift. Most multimodal models today still have separate pieces under the
hood, like a vision encoder to read images and another model to process language. But here, Sense Nova throws that away. It uses this new architecture
called Neo Unify, which directly connects pixels and words end to end. So, instead of converting images
into tokens and then passing them around, everything is learned as one unified representation. The awesome thing is they've released the code to this
already. On this page, if you scroll down a bit, here it contains all the instructions on how to download and run this locally on your computer. Note
that the model is around 32 GB in size, so you do need some higher-end consumer hardware to fit
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 latest
open-source model, which is called Nemotron 3 Nano Omni. Now, instead of using separate models for video, audio, images, and text, this new Nano Omni
handles all of it together. It can take in video, audio, images, and text all at once and outputs
reasoning or actions based on all of these media together, which makes the context way more consistent. Under the hood, this is a 30 billion parameter
mixture of experts models, so think of it as like a team of specialists working together to help you solve a task. But when you use it, only 3 billion
parameters are active at once, which makes it super fast and efficient. In terms of performance,
this thing is actually pretty strong. It's hitting top-tier results across document understanding benchmarks, OCR tasks, and also video and audio
understanding. And on efficiency, it's kind of wild. In some cases, it delivers over nine times higher system capacity for video reasoning and over
seven times more capacity for multi-document tasks. You can run way more stuff with this model. This new
Nemotron 3 Nano Omni also has some clever tricks for handling video. Instead of compressing every frame in full detail, it compresses and summarizes
the important parts while still keeping motion and temporal information. So it understands what's happening over time, not just frame by frame. And
this is designed to fit into agentic systems like Openclaw. The awesome thing is this is fully open
source. You get the models, the data sets, and the training recipes to this. Now at 30 billion parameters, this is still pretty huge. This is a
medium-sized model, so the full version is 66.1 GB. I'm sure there's also going to be more quantized versions that can fit on lower-end hardware.
Anyways, if you are interested, on this page it contains all the instructions on how you can download and run
this Nemotron 3 Nano Omni on your computer. If you're interested in reading further, I'll link to this main page in the description below. Also this
week, Anthropic dropped a pretty interesting update. They call it Claude 4 Creative Work. And here's the big idea. Instead of using AI tools
separately, Claude can now plug directly into some common creative tools like design software, music
production tools, and 3D editors. They call these connectors, and they let Claude directly use software like Adobe Creative Cloud. So, here's an
example where you can prompt Claude to create a design using Creative Cloud, and it'll actually work within those apps. Or, here's another example of
working inside Blender. With the new connector, you can literally talk to your 3D scene. You can debug
it, modify objects, or even generate scripts using Blender's Python API. Or, here's another example of it working with 3D objects in Autodesk Fusion.
It actually does things in the interface and controls it programmatically. On their blog, you can see it can even connect to Ableton or Canva. Now,
even though I'm not a fan of Claude, it's way too slow and expensive, plus they're closed as hell, I
do think that this is the next logical step for AI. Getting them to autonomously do stuff in existing apps and software. This opens up a ton of
possibilities and can help us automate work even more. If you're interested in reading further, I'll link to this main page in the description below.
Now, Claude isn't the only team building connectors to other software. Another AI lab called Moon Lake
just dropped a 3D world-building agent that can also use software like Blender. And this is a pretty big shift in how AI works with 3D environments.
Here's the simple version. Instead of just generating images or 3D assets in one shot in a chat interface, this thing actually acts inside real tools
like Blender. So, it takes a task like build a scene or fix this model, and it goes into Blender, it
clicks around, adds its objects, adjusts lighting, and iterates, just like a human would. The input is basically a goal or instruction, maybe some
reference images, and the output is a fully built or edited 3D scene with all the parts properly structured, aligned, and usable in real workflows.
And that's important because 3D work isn't just about making something look right. It has to also be
structurally correct with objects connected properly, animations working, states behaving logically. That's stuff that most AI models still struggle
with. What Moon like is doing differently is treating this like a loop. Build something, check it, fix it, and repeat. The agent runs over long
sequences of actions instead of just generating once and stopping. And it's optimizing for multiple layers
at the same time. Overall seeing quality, consistency with references, and then really low-level stuff like whether objects are actually connected and
behaving correctly. And as I've shown you here, they've already got this working across a bunch of different examples like articulated objects like
cabinets, hinges, appliances, even full environments like malls and factory floors and outdoor
scenes. Here's another key feature. It can learn workflows from experts. So, if someone demonstrates how to do a task once, the agent can reuse that
pattern across future projects. So, zooming out this is moving AI from just generating stuff in a chat interface to doing real production work inside
software tools. And that's a huge deal because 3D editing is a massive industry. We're talking about
like video games, film, simulation, etc. And a lot of it is still manual and repetitive. So, if this keeps improving, you're looking at agents that
don't just assist creators, but actually handle entire chunks of the pipeline end-to-end. Anyways, this week they just released an early demo. If
you're interested in reading further, to this main page in the description below. Also this week, this is
actually a really fascinating idea. It's called Talky and this is a 13 billion parameter model that only knows the world up to 1930. In other words,
you're not talking to a modern AI trained on the internet. You're talking to something that's been trained entirely on historical stuff like books,
newspapers, basically anything up to 1930. And here's why it matters. Pretty much every modern AI
you've used, whether it's ChatGPT or Claude or Gemini, they've all been trained on basically the same thing, the modern internet. Even when these
models feel different, maybe you get a different vibe from each of them, they're still kind of related because they're all just trained on the same
data. A vintage model like this lets researchers ask a very fundamental question. How much of what we
think we know about AI is just tied to the specific training data set. If we train a model on something completely different, you finally start to see
what's going on under the hood. Here's one of the interesting experiments they ran. They took thousands of on this day historical events from the New
York Times and measured how surprising these events are to Toki. So, events from before 1930 feel
normal to it because, well, that's part of its training data. But once you cross the knowledge cut off, its surprise jumps up, especially in the '50s
and '60s. If you tell it about events that happened during that time, it gets pretty surprised. But then it kind of plateaus. Even from 2000 and
beyond, it doesn't really get surprised any further. What's interesting here is you can literally watch
the model react to history it has no way of knowing. They also ran a coding experiment. They gave this vintage model a Python coding test. Now, this
model has never seen a computer. It has no concept of what Python is, and they showed it a few example Python functions in the prompt and said, "Okay,
write a new one." And it actually got some right. Not a lot, and only really simple ones, but think
about what that means. The model has no idea what code is, but it still picked up enough from a few examples to grasp the concept of this function.
That's a real generalization signal, and it's a clean one because there's zero contamination. The model literally has never seen Python code or this
answer before during training. Here's another really interesting test. They compared Toki to a modern
twin. This model was trained on data up until now, which is the gray bar on these charts. You can see that Talkie does lose on knowledge tests, which
makes more sense because it doesn't know modern stuff. But on core language understanding and basic numeracy, it holds up surprisingly well. The team
thinks the remaining gap is mostly because of data quality issues, especially OCR. Since none of the
text in its training data was digital, all of it had to be like scanned and transcribed from physical pages, often with very messy handwriting. And
during that process, a lot of these words got mangled. For example, its training data looks something like this with a lot of spelling errors. So,
those are some of their key findings. They're already training a larger model that's coming out this
summer. They also want to expand beyond English and also redo a chunk of the OCR with their new system, so the training data is more accurate. The
awesome thing is you can actually chat with this model right now. So, if you click on this chat button, it takes you to this page where you can chat
with Talkie. So, if I ask it to tell me about TikTok, let's see its response. So, it has no idea about
the TikTok I'm referring to. Here it's just talking about this machine that was apparently invented in 1860. The thing that makes this project so
cool, beyond just the novelty of chatting with a 1930s AI, is what it lets researchers actually study. You get a contamination-free model, which is
very rare and valuable. You get a window into how data shapes the model's personality. You can actually
test it questions about generalization and forecasting and discovery, which you can't really do with modern AI models cuz they're trained on
everything already. So, if you're into AI research or history or just the weird intersection of these two, Talkie is one of the more interesting
releases this week. If you're interested in reading further, I'll link to this main page in the description below.
Also this week, xAI quietly dropped their latest model, Grok 4.3 beta. Now, they designed this to be a general-purpose assistant with better
reasoning, faster responses, and way more reliable outputs. It has access to a sandbox computer, so it can write and run code and work with multiple
folders and files at once. For example, you can get it to create a PowerPoint presentation like this, or it
can also work with PDFs or spreadsheets like this. So, a lot more agentic capabilities with different tool usage compared to the previous Grok.
However, if you look at this leaderboard by Artificial Analysis, while it does outperform the previous Grok 4.2, it's not as performant as even the
top open-source models out there like Kimiko 2.6 or Mimo by Xiaomi. And it's still quite behind the top
closed models out there, with the leader being GPT 5.5. However, price-wise, it is quite competitive. It's around the same cost as the top open models
out there and way cheaper than the top closed models out there. In terms of performance, quite a mediocre model, not really state-of-the-art, which is
why I didn't do a full video on it. But, if you are interested in reading more, I will link to
this main release page in the description below. Also this week, European AI lab Mistral drops their latest model, Mistral Medium 3.5. This is a dense
128 billion parameters, so it's quite huge, with a context window of 256K, so it's not the largest in the industry. Most of the top models out there
already have a million token context window, which is like four times larger. Here are its scores on
some agentic benchmarks. It does come close to the other top open-source models, but notice that they are cherry-picking here. We already have Kimiko
2.6 and Qwen 3.6. So, it seems like they're intentionally comparing it with these older models. Now, same with other scores on math and instruction
following, at least according to them, this new Mistral performs pretty well. However, if you look at
this independent leaderboard, then here's where things fall apart. Mistral medium 3.5 performs very bad, significantly worse than DeepSeek V4 and the
other open-source models. It's not even close. So, it doesn't look too good here. Now, this is open-source, but if you choose to use it through their
API, it's also not cheap, way more expensive than the other leading open models like Kimmi and
DeepSeek. So, honestly, there's no reason to use this latest Mistral model unless you have to comply with some EU regulations or something. 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.