Kimi K3, dancing waifus, robot UFC, song to MIDI, GPT Red, hoverboards: AI NEWS

[00:00:00]AI never sleeps and this week has been absolutely insane. Open- source AI is no longer a few months behind. It has actually caught up to Frontier. We can finally do AI video generation locally on just a mobile phone. Speaking of phones, we can also now use one of the best consumer open- source models, Quen 3.6 and Jam Packets run locally on just a phone. Google releases an open source model that allows you to precisely control heads and faces. We have an open- source AI that can make any character dance to any music you input.

[00:00:34]And speaking of music, this AI can turn a full song into the notes for each separate instrument. It's like you're reverse engineering the entire song.

[00:00:42]OpenAI reveals a new model that's designed to attack other AIs. We have a new ultraast image upscaler and another one that's really good at background removal, even for tricky situations. We have some ridiculous humanoid robot demos, including a humanoid cage fight and a lot more. So, let's jump right in.

[00:01:00]First up, Nvidia releases a really useful AI called RD. This is an AI that can generate realistic 3D human movements in real time while letting you change the action or control the character's path. So, here's how it works. You can input a text prompt or specify a movement or even add optional controls like paths, waypoints, key poses, or joint positions. And from that, the AI can generate a continuous 3D human animation from your exact inputs, as you can see from these demos.

[00:01:31]And the cool thing is this responds in real time. You can include some pretty specific details in your prompt, and it's still able to understand and handle all these movements. Plus, you can keep streaming this for a long time. In fact, they say that you can stream this potentially forever. So, this is like infinite streaming. It can keep generating the animation until you press stop. Now, instead of just generating a human animation like this, you can also use this to control a humanoid robot, specifically the Uni Tree G1 in this simulation as well. So, this can be useful for like games, animations, virtual characters, and of course, training robots. The interesting part is that RD plans the motion in two stages.

[00:02:11]It first works out the character's overall position and movement through the scene, then fills in detailed body motion. Think of it like planning the path first, then animating the arms and legs and body afterwards. The awesome thing is they've released the code and 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. And it's actually fairly tiny. So, each model is only like 765 megabytes. If you're interested in reading further, I'll link to this main page in the description below. Also, this week, this is pretty exciting. You can actually now generate videos locally on just a phone. So this is called mobile one and this allows you to run Alibaba's one directly on a mobile device. In fact, here are some examples of the original one 2.25 billion parameter model versus this mobile one which is generated on just a phone. So, this is able to take in a text prompt and output a 5-second video at 480p resolution and 16 frames per second, which is already not bad considering that, you know, phones are not meant to run video generators, right? You often need quite a powerful Nvidia CUDA GPU in order to generate videos. But here, it's able to generate a 5-second video just on a phone in around 20 seconds. Now, as you would expect, this requires some really heavy optimization and compression in order to make this work on a phone. So, here's what they did. It basically generates the video in small chunks instead of processing every frame together. They also had to do some really heavy pruning. So, for example, they removed some less useful attention heads. They reduced generation to just three steps instead of the average 20 to 30 steps.

[00:03:55]And they also used a very memory efficient video decoder. But afterwards, what you get is something that just works on your phone. 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 on your phone. Also, if you click on their hugging face folder, notice that the total size of everything is only 9.9 GB. So, it's fairly tiny. If you're interested in reading further, I'll link to this main page in the description below. Also this week, the fastest open-source upscaler just gotten better. So, Nvidia just released a new update to their PID upscaler, and this is now version 1.5. Now, I already featured version one in a previous video, so see this if you're interested in learning more. But version 1.5 offers even better details and color fidelity.

[00:04:44]For example, here is a comparison between version one, which is what you see here, and then here is version 1.5.

[00:04:51]You can see the details are a lot clearer and the colors are also better.

[00:04:55]Or here's another example. Here is version one and then here is version 1.5. You can see everything is just a lot sharper and more detailed and the colors are also slightly better. And then here's another example. The best part is you can use this with different image models including Quinn image, Flux, and Z image. And like I showed in my previous video, this is incredibly fast. This can turn a 512 x 512 latent image into a 2K image in just a few seconds on a consumer GPU. Now, if you're interested in learning more, at the top of the page, they have released a code button. And if you scroll down a bit, here is how you can download and run P 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, this AI is potentially really useful for making music. So this is called audio to MIDI by Mirell and like the name implies you can upload a full song including vocals and multiple instruments and it can generate separate MIDI tracks for each instrument. So MIDI track basically looks like this. These are basically the notes that are played for each instrument. So it's like you're dissecting or reverse engineering the song. In fact, you can try this out for free online. So, let me click on this ML studio and then after you sign up for free, let's click on audio to MIDI. And here's where I can upload a song. So, let me upload a short section from this song.

[00:06:23]>> Was it worth it?

[00:06:29]Tell me.

[00:06:32]>> So, let me drag that over here. And then afterwards we need to click on agree and continue and then click on identify automatically. And let's see what we get. All right. So it has automatically detected voice, drums, synth strings, synth pad. It could not detect bass, but we can also manually add a bass if we wanted to. Let's just leave it at the default and then click generate MIDI. So afterwards, it's proceeding to generate the MIDI notes for all these instruments. How cool is that? And afterwards, let's press play.

[00:07:18]So, not too bad. As you can see, it was first able to separate the voice plus all these different instruments, including the synth pad and the drums.

[00:07:26]And it's able to generate notes for each instrument. This is already really impressive. Now, it's not completely accurate. I think it's like 80% there.

[00:07:34]But then you can plug each track into a DAW and then customize and fix certain things yourself. So if you want to take inspiration from an existing song or if you want to easily extract the notes of all the instruments in any song, this is a very powerful AI tool to use. The awesome thing is this is already open source. So if you click on this GitHub repo, it contains all the instructions on how to download and run this locally on your computer. Note that they released three different variants of the model. There's a small, medium, and large one, but even the large one is fairly tiny at only 1.4 billion parameters. And this large one is only 5.5 GB in size. So, all three models should be able to fit on most consumer GPUs. Here it says that if you don't have a GPU, then the small one could even potentially fit on CPU. One thing to keep in mind is that the model weights are released under the Creative Commons non-commercial license, so you won't be able to use this for commercial purposes. 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 release from the Alibaba Juan team. So, they just released Juan Dancer. Like the name implies, this can basically generate videos of characters dancing, and it can generate really long and synchronized videos. Now, the interesting thing here is that this doesn't take in the existing movements from a reference video like other dance animation tools.

[00:08:56]Instead, this actually takes in music and it can generate a video of any character dancing to the music. So, here's just one example. And it can dance to a variety of different music genres. For example, here is Chinese classical dance.

[00:09:20]And then here's a Latin example.

[00:09:25]And the best part is this can generate really long videos and the character remains consistent the whole time.

[00:09:31]Whereas for other AI dancing or character animation tools, usually they can only remain consistent for a few seconds. So I mean with this tool, you can easily create like virtual dancers, music videos or related content just from an image of a character and a song.

[00:09:46]You don't even need to choreograph or coordinate the dance yourself. Now, instead of just getting this model to generate a dance randomly, you can also insert key frames to control exactly how the character dances. As you can see in these examples, the awesome thing is, as with the previous one releases, this is also open- source. So, 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. However, before you get too excited, note that the total size of everything is 85 GB.

[00:10:17]So, you'll need a really high-end GPU to run this. But, of course, because this is open source, I'm sure there's going to be more quantized versions or maybe GGUFS in the future that can fit on more consumer hardware. If you're interested in reading further, I'll link to this main page in the description below.

[00:10:33]Also, this week, Google releases something called G&M, which stands for generative anthropic, I mean anthropometric model, and ecosystem. And this is basically an AI that lets you design faces and heads with extreme precision.

[00:10:48]So this takes adjustable parameters such as a person's identity, facial expression, head rotation, eye direction, and position, and you can drag all of these settings with sliders, and it basically outputs a dense 3D head mesh with these exact facial and head details. If you've been playing around with AI image generation, you might find that most characters generated with AI kind of look the same. they kind of have the same faces, especially with the older models. So, if you really want to generate a unique looking character with a very specific looking face, then GNM by Google might help you achieve your desired results. So, this even includes really precise settings for internal structures like the eyeballs, teeth, gums, and the tongue. And all these controls are separate, so you can change only that specific feature without affecting anything else in their expression or face. Now, currently this G&M head model contains 253 identity controls, including like 170 controls for the head and even 80 controls for teeth components. And then in terms of expression, it has 383 controls, including 100 settings for each eye, plus 150 settings for the lower face and even 32 settings for the tongue, which is pretty crazy. The awesome thing is they've released this under the Apache 2 license, so you can pretty much do whatever you want with it, including commercial usage. And here it contains all the instructions on how to download and run this locally on your computer.

[00:12:16]If you're interested in reading further, I'll link to this main page in the description below. Also, this week, we have a super powerful open-source background remover called Lucida. This is able to remove backgrounds of images while preserving really difficult details like glass, camouflage, text, shadows, glow, or other tricky stuff.

[00:12:36]Here are some examples. You can choose to remove all the pedals and only keep the face. And Lucida is able to handle this very well. Whereas for other background removal models, you can see it's unable to detect this because the appearance of the face and the background are very similar. And same with transparency. If we remove this background, note that this vase should also be slightly transparent. So, Lucida is able to handle this, but ideoggram as you can see was not able to add this extra transparency. And then same with text. Let's say we only want to isolate this text, but not the shadow. Lucida is able to handle this very well. Whereas for audiogram, you can see that it still seems to preserve the shadow of this text. This also works very well for hair and other tricky stuff. So, let's say we want to remove this background. Well, as you can see, Lucida is able to remove more of the details near the hair, whereas this other competitor is not able to remove these details. The awesome thing is they released the model already. So, on this page, if you scroll down a bit, it contains all the instructions on how to download and run this locally on your computer. They also released details on how it was trained and also the data set. So, this is completely open- source and it's under the MIT license, which has very minimal restrictions. And if you look at the size of this, this is fairly tiny at less than 900 megabytes. So you should be able to fit this on most consumer GPUs. They also have a free demo on hugging face. So if you click on this link, you can try out Lucida over here for free. For example, let's enter this image and then click on submit. And it should proceed to remove the background of this image. And notice that it's fairly fast. This only took like 6 seconds. And as you can see, it's also able to make, you know, the water and the glass a bit transparent, which is correct. If you're interested in reading further, I'll link to this main page in the description below. Also, this week, we have a really cool AI called motion for motion. This is an AI that can transfer the movement from one video over to a completely different video without requiring any matching skeletons or any control net or anything like that. Each video could have different characters with different body proportions. So, here are some additional examples. This works for a ton of different environments. Here's a really tricky example where, you know, we transfer the movements of this fox to a giraffe, which has completely different proportions, but it's still able to handle this very well. And of course, this can also handle dancing as you can see here. Now, how this works is instead of using skeletons or segmentation or control net, what this does is it tracks points across the source video to create a motion flow map as you can see here. And it matches those points to the target and injects their new positions into a video model in order to create the new movement. And so that's how it's able to transfer the movement of one video onto another video, even if the character looks completely different. Now, currently, they've only released a technical paper.

[00:15:29]They do have a code button, but it doesn't really take you anywhere.

[00:15:32]Hopefully, they will open source this.

[00:15:34]But for now, if you're interested in reading further, I'll link to this main page in the description below. AI agents can do a ridiculous amount of work. But did you know you can also turn them into a full content creation team? That's exactly what Higsfield, the sponsor of this video, allows you to do with Higsfield MCP. You can connect Claude, ChatgPT, OpenClaw or other agent systems and turn them into a complete creative production studio. Just describe what you want and the agent can help create images, videos, ads, landing pages, campaign ideas, and entire creative workflows. The reason this works so well is that the quality of the AI generated content often depends on the quality of the prompt. And the frontier models are incredibly good at understanding your intent, improving rough ideas, and planning out the right creative process.

[00:16:21]Instead of spending hours experimenting with prompts yourself, GPT or Claude can act as your creative director, while Higsfield takes care of generating the actual content. For example, Higsfield MCP lets you connect Higsfield directly inside codecs or claude code. The LLM can come up with the concept, write the creative brief, generate the images or videos, and even save the finished files directly in your project folder. You don't have to constantly switch tabs, rewrite prompts, download files, or move assets between different platforms. You could use it to produce launch videos, create UGC ads, recreate viral formats in your own brand style, run a faceless content channel, or build an entire marketing campaign from a single chat.

[00:17:02]And Hakesfield supercomputer takes this even further by giving you a full creative team in one place. You can choose any agent to handle the thinking, then use tools like Seed Dance, Cinema Studio, and Marketing Studio to turn the idea into a finished result. Whether you're producing ads, social videos, or any other kind of content, Hicksfield can supercharge your creative workflow.

[00:17:23]Try it today using the link in the description below. Also this week, we have a new family of models that can run locally on just your phone, and this is called Bonsai 27B. This is based off of Quen 3.627B, the medium-sized, open- source model from Alibaba. And this is one of the most popular and capable models that can actually fit on consumer devices. So, this has gotten like over 5 million downloads just this month. The problem is, if you click on their main repo and you look at the total size of this, this is 56 GB. So, you'll need like a high-end GPU to run this. But here, what Bonsai did is they basically took this massive Quinn 3.6 model and compressed it and made it efficient enough to run on just a mobile phone. Now, here's the problem they faced. Like I showed you, the full model is like over 50 GB in size. And even if you compress it down to 4bit, it's 18 GB. And this will be still too large to run on a phone or even for most laptops. But here they compressed it even further and they released two different variants of this.

[00:18:25]The first one is called Turner Bonsai 27B which is just 5.9 GB while the more aggressively compressed 1 bit version is only 3.9 GB. This is small enough to fit within like an iPhone 17 Pro which is pretty insane to think about. I mean who would have thought you could actually jam-pack a medium-sized 27 billion parameter model which is like over 50 GB to be able to run on just a phone. You see, here's the reason why this is able to compress the model so much. Normally, AI model weights, these are like the parameters of the neural network.

[00:18:59]They're stored as very precise numbers.

[00:19:01]So, something like 0.7342 or some really long decimal. But turnary weights replace each of these detailed numbers with only minus1, 0, or + one.

[00:19:11]You can think of it as like instead of remembering the exact strength or value of every connection, turnary weights only remembers whether that connection should push the signal down, do nothing, or push it up. And as you can imagine, if you don't need to save some really long decimal number for each value, then you can shrink the size down by a lot.

[00:19:29]So this actually becomes dramatically smaller and cheaper to run. But of course, it can also reduce the accuracy and quality because a lot of this detail is discarded. But the really impressive thing is that here it says this Turnary Bonsai 27B retains 95% of the full precision baseline or basically the quality of the original model and then the smaller 1-bit variant retains 90%.

[00:19:51]If you look at these benchmark scores then as you can see the full Quen 3.6 model scores 85% on average across all these benchmarks and then Turner Bonsai scores 80.5% which is not too bad. It's not like a huge drop in quality even though they were able to compress the size by a massive amount. And so now with this Turner Bonsai, you can have the intelligence of Quen 3.6 completely offline on just your laptop or phone.

[00:20:19]The awesome thing is they released this already. So in the right column, if you click on this GitHub button, it contains all the instructions on how to download and run this locally on your mobile device or laptop. And you know this is the power of open source because Alibaba released the model weights to Quinn 3.6 the community can build on top of that and now we are able to fit Quen on just a mobile phone making this intelligence potentially accessible to well anyone with a strong enough smartphone. 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 number one open-source AI model. And get this it has caught up to the frontier on various benchmarks. This performs just as good or even better than GPT and the best Claude. Now, I already did a full review video right when it came out with a ton of really impressive tests. I also went over its specs and benchmarks and pricing and how and where to use it. So, I'm not going to repeat too much there. See this video if you're interested in learning more.

[00:21:18]But here are a few more updates since that video came out. So, Kim K3 was also added to some other leaderboards after my video was published. One of them is this debate leaderboard. This measures how well AI models perform in multi-turn debates across a wide range of topics.

[00:21:34]So the model needs to be really knowledgeable across different domains.

[00:21:37]Plus, it needs to be factually accurate.

[00:21:39]And it also needs to stay responsive and defensible over several rounds. And as you can see, Kimmy K3 is second place, just slightly below Claude Fable 5, but it even beats GPT 5.6 and Opus 4.8.

[00:21:54]Really impressive performance for an openw weights model. And then here's another crazy benchmark called kernel bench. This measures how well an AI is at optimizing kernels or basically really low-level GPU code. This is really complex and specialized code that determines how a GPU runs calculations that ultimately power AI models and other stuff. And as you can see for kernelbench CUDA, Kimik 3 even beats Claude Opus 4.8. And then for Colonel Bench Hard, it is slightly below Claude Fable 5, but it does beat Claude Opus 4.8. 8. I'm not sure why GPT 5.6 isn't on this list. In fact, one really neat detail that I didn't mention in my previous video is that they also got Kimik 3 to design a chip to serve a smaller model built on its own architecture. So, it ran for like 48 hours autonomously and it was able to actually design the chip and optimize it so that it met speed and size requirements. And the result was a very small chip that could run fast enough in simulation. And this chip includes special hardware for low precision math, so it can handle the model's calculations effectively. Now, this is just a proof of concept, but I hope you can see how this ability can be super helpful, especially for Chinese labs that are tight on compute and they don't have access to the best Nvidia GPUs.

[00:23:10]They need to take whatever existing hardware they have and squeeze every last drop of juice out of it. And with Kimmy K3, because it's so good at optimizing kernels or this low-level GPU code, it can serve a crucial role in designing and optimizing this step.

[00:23:25]Here's another update. So, the official Deep Suite leaderboard also added Kimmy K3. And note that it scored 69% which is actually 1.5% above what the Kimmy team originally reported. So, they actually under reported this score. Now again, Deep Suite is one of the most accurate measures of long horizon software engineering tasks. And as you can see, Kimik 3 is among the best models out there. Notice that all these confidence intervals overlap. So there's actually no significant difference between these models. It's as good as GPT 5.6 and Claude Fable. Now, this leaderboard looks at how good an AI is at solving New York Times connections puzzles. It's just a fun leaderboard, but as you can see, Kimmy K3 scores number three, just slightly below GPT 5.6. six, but it beats Claude Fable. And interestingly for this benchmark, somehow Gemini 3.1 Pro is still number one. Also, interestingly, if you look at this leaderboard called LiveBench, then you can see that Kimik 3 ranks all the way down here slightly below Opus 4.8 and even below GPT 5.5. It is extremely impressive in terms of reasoning. But for other categories like agentic coding or mathematics data analysis, at least according to this leaderboard, it is relatively weaker compared to these other models. Anyway, those are just some additional reference points for Kim K3. Again, see this video for a full review. I'll link to it in the description below. Also, this week, Real Steel is already here. Apparently, we had the first ever humanoid robot MMA event held in Shenzhen, China. And this featured the full-sized T800 humanoid robot from Engine AI. As you can see here, they're trading punches and flying kicks in this cage. In fact, the first flying kick from this robot here kind of looks like Conor McGregor, if you know what I'm talking about. Note that these are currently teleoperated by someone with a remote control in the background.

[00:25:17]And for most of the fight, they do kind of miss. You can see that they're just punching or kicking in the wrong direction. But eventually, one of the robots head gets knocked loose and it could keep fighting for a while, but it eventually falls to the ground and loses balance. Now, while it looks pretty basic and comical right now with the robots missing and just randomly hitting in all directions, keep in mind this is the worst it's going to get. Robotic technology is accelerating like crazy.

[00:25:43]And these humanoid robots are only going to get more responsive, powerful, and accurate. In fact, we might not even need tea operation anymore in the future. Eventually, we could just design robots that can autonomously fight each other. Let me know in the comments what you think of all of this. First of all, would you be interested in watching such events? And if we were to train robots that are able to, you know, pull off these attacks autonomously, are you concerned that this could get out of hand? What if these robots escaped or were let loose? Or what if they were programmed to attack humans instead of other robots? Let me know in the comments what you think of this. Oh, by the way, the fight isn't the only cool part about this event. So apparently these robots were actually flown to the arena using this hoverboard designed by a company called Coolfly. And this specific hoverboard is called the Coolfly Urban, which is designed to be like an EV tool device. In other words, an electronic vertical takeoff and landing aircraft. And this one is designed to transport a robot or of course humans. So, you know, in the future, maybe instead of driving a car to a certain place and getting stuck on the highway for hours, especially during rush hour, maybe you can just grab one of these hoverboards and fly to your destination way faster. In other humanoid robot news, forget about robots with two legs. Now, we have apparently the world's first centaur robot. So, another Chinese company called Run Robotics just unveiled a hybrid robot with a humanoid upper body and a lower body with four wheeled legs. There's not much info on it right now, but this is intended for intense industrial settings as you can see from this video. So things like nuclear plants or steel mills, oil fields, mining, and other intense settings. It can reportedly carry up to 210 kg, which is a much higher load compared with other leading humanoid robots. And you know, this design is pretty interesting. Who says we need to design robots like humans with only two legs? With this design, not only can it roll around more efficiently, but with four legs, it's also able to maintain balance a lot more effortlessly. And it can also use these legs to climb steps or get through uneven terrain. Plus, with this platform on its back, you could kind of think of it like a pickup truck where it could potentially carry really heavy stuff on its back. Or maybe a human could even ride this like a horse. Anyway, a pretty interesting design from this lesserknown robotics company. Also this week we have a demo from another lesserknown robotics company called Sunday robotics. They just unveiled a demo of their act 2 preview robotics model. So this is like the brain controlling the robot.

[00:28:20]Apparently this runs completely offline.

[00:28:22]You don't need to connect it to the cloud or anything. And this is trained to do really well on household tasks.

[00:28:28]It's able to do tasks in homes it has never seen before. Zero shot with a 99% success rate which is really impressive.

[00:28:36]Specifically, it achieved a 99.1 success rate in laundry folding. Apparently, for shorts, tops, and polos, it's able to achieve 100% success rate. It's even able to autonomously handle situations it has never seen before in its training data. For example, picking up clothes off the ground and handling clothes of different sizes from baby sizes up to 8 XL shirts. Now, folding clothes is just one of the many tasks it can do. Here they say that the same robotic model is now learning to vacuum, organize things, and even zip clothing, and turn clothes inside out. So, this could potentially be a very capable household robot that can do a ton of chores with a high success rate. Another robot company to put on your radar. Also, this week, Google DeepMind introduces a new AI called Genception. This basically turns a standard video model into a generalpurpose system that can understand videos and perform many different vision tasks. For example, instead of just building one model for depth, another for object tracking, another for camera motion, this genion basically contains all these abilities in just one unified model. So for example, you can take an input video and use it to predict the normal or the depth, the camera pose, segmentation, density pose, and even the 3D key points of any objects in the video. And as you can see, compared to other specialist models, this Genception is even better at predicting all these different visual attributes. Here's another example where you can easily take one input video and use it to predict the depth, normal estimation, segmentation, etc., etc. And you know the cool part is you can then merge all these predictions together to even reconstruct a 4D scene and you can also ask it to identify things in the scene because again it has a unified visual understanding of everything. So this research suggests that video models might not be only useful for creating clips but they can also be repurposed into really powerful visual understanding systems. Now for now they've only released a technical paper.

[00:30:41]It says the code is to be announced. So, it looks like they are planning to release this as well, which is great.

[00:30:47]For now, if you're interested in reading further, I'll link to this main page in the description below. Also, this week, OpenAI unveils their internal model, which is called GPT Red. And this is essentially an automated safety red teamer. It's designed to find vulnerabilities and attack other models so that it can train them to be more robust. And here's why it's so important. You see, people are going to use AI agents more and more to do things like browse websites, scrape data, read emails, access files, and take actions on your behalf. But this also creates a huge security problem called prompt injection. This is where someone can hide malicious code within a web page or email or attachment. So when the AI agent reads such a file and sees these hidden instructions, it could end up leaking private data or changing account settings or taking some harmful action as you can see from the response from GPT 5.1. Well, human security researchers already test for these vulnerabilities, but that process is really slow and difficult to scale. So what OpenAI did is they made a model specifically designed to find these vulnerabilities and attack existing models and they called it GBT red. This can generate and test a huge number of attacks automatically. Plus it doesn't stop there. Once it sends an attack, it watches how the target model responds and then it also adjusts its strategy until it finds something else that works. So it's constantly like self-improving to find more and more vulnerabilities. The interesting part is that OpenAI trained it through self-play where GPT Red continuously attacks some AI models while these models simultaneously learn to resist these attacks. So basically, as the defenders become harder to break, GPT Red is forced to invent even more sophisticated attacks. And by the end of training, GPT Red could break nearly every model it was tested against, even models up to GPT 5.5. As you can see from this attack success rate, if you scale up compute, then GPT red was able to successfully attack models around like 84% of the time, whereas the human baseline is only 13%. Now, the important part is what happens next. So, they used these attacks discovered by GPT Red to train their current best model GPT 5.6 so that it can learn how to resist these malicious attacks. As a result, GPT 5.6 six produced six times fewer failures compared to the last model 5.5 from just a few months ago. So obviously GPT Red is quite dangerous. This is a model that intentionally looks for ways to attack existing models. So this is going to remain internal. We're not going to have access to GPT Red. But the big idea here is that AI models can now help improve not only the intelligence of future models but also their security. If you're interested in reading further, I'll link to this main page in the description below. Also this week, this virtual girlfriend guy is back. So, if you don't recognize him, basically in the past few weeks, I featured this tool called One Streamer, which allows you to talk and interact with any character in real time. Well, this week they released a better version 0.3. You see, the previous versions could already have live video conversations with synchronized speech and lip movement, eye contact, or even nodding and different facial expressions, but it can only do like portrait shots. But this new version 0.3 goes further by allowing the character to freely interact with its surroundings instead of just sitting there and talking. So here are some examples.

[00:34:15]>> Make a peace sign with your fingers.

[00:34:18]>> All right. Peace sign. It is.

[00:34:20]>> Put your hands on your hips.

[00:34:23]>> Hands on hips. What's the next command?

[00:34:26]>> Fix your hair.

[00:34:28]>> My hair. Is it messy? Is that better?

[00:34:32]>> Look to the right.

[00:34:34]looking to the right.

[00:34:36]>> Look down at the floor.

[00:34:38]>> Okay.

[00:34:39]>> And the key here is to divide the video into two parts. The world and the event stream. The world contains everything that should remain consistent like the character's appearance, the room, the visual style, the background, etc. And then the event stream contains everything that's changing over time, including the speech and body movements, camera motions, and other events. And then during a live session, the user's camera, microphone, and text also become part of this event stream. For now, they've only released a technical paper.

[00:35:08]Hopefully, they will open source this in the future. But for now, if you are interested in reading further, I'll link to this main page in the description below. Also, this week, Thinking Machines dropped a massive open-source model called Inkling. First of all, if you're not familiar with Thinking Machines, this is founded by previous OpenAI CTO Mera Morati. They've been building in stealth for the past few months, but finally they released this new open waste model called Inkling.

[00:35:34]Now, this is a mixture of experts models. So, think of it as like a team of specialized AIs working together. And the total size is quite huge at almost a trillion parameters, but when you use it, only 41 billion parameters are active. So, it's quite efficient. And like most Frontier models, this has a context window of up to a million tokens. Here they say that it was pre-trained on 45 trillion tokens of text, images, audio, and video. So a massive training data set. And because it's trained on all these different types of media, this makes it super powerful, especially for multimodal capabilities. For example, here it says Inkling can reason natively over text, images, and audio. And if you look at these benchmarks in terms of reasoning, agentic coding, factuality, and also audio and vision understanding, you can see that inkling, which is the blue surface area, is actually not too bad.

[00:36:23]If we compare this to like one of the leading open- source models, GLM 5.2, you can see it's lacking in both vision and audio capabilities. Same with GPT 5.6 Soul. It doesn't really have audio.

[00:36:34]And also the same with Claude Fable 5.

[00:36:37]If you look at its performance in this design arena, then it's also pretty decent. Here are some PDFs that it has coded up. You can see this looks very professional and the layout is also very nice. Or here's a multiplayer snake game that it has coded up. Now, in addition to the massive full Inkling model, which is almost a trillion parameters, they also are sharing a preview of Inkling small, which is much tinier at only 276 billion parameters, 12 of which are active. And of course, this runs a bit faster. This can fit on lower memory, but with some sacrifice in quality, especially for agentic coding. Now, if you look at this independent leaderboard by artificial analysis, then Inkling is all the way over here, even behind Deepseek V4, which was released like months ago, and way behind the current leading openw weights models, GLM 5.2 and Kimik 3. And if you look at speed, Inkling is also not the fastest. So, it's much slower than Mini MaxM3 and also GLM 5.2, which is all the way over here, which is more than double the speed. So, not too impressive in my honest opinion, but I do like that it has multimodal capabilities, especially audio and visual understanding. If you scroll down here, they've already released the models to this. So, if you click on this hugging face link, note that the full model is 1.9 tab in size, so you're not going to be able to fit this on consumer devices. They also have an NVFP4 version, which is much smaller at 592 GB in size. Plus, Unsloth has already released GGUFS for Inklings. So, the most quantized one-bit version is only 270 GB in size. Although, note that for Q1 compression, you're going to sacrifice a lot more quality. If you're interested in reading further, I'll link to this main page in the description below. Also, this week, Nvidia releases a new family of models called Neotron 3 Embed. And this is actually quite interesting. You see, these models aren't designed to answer your questions directly, but instead these are used to help AI systems find or retrieve information. So, for example, this takes in text or code or documents or search queries as input and it outputs numerical representations that can be used to find similar information. So, this is really useful for like search algorithms or rag systems or coding agents especially in terms of retrieving information. Now, this is fairly tiny.

[00:38:53]The full model is 8 billion parameters.

[00:38:54]They also released a 1 billion parameter variant. But as you can see in terms of its retrieval accuracy across these benchmarks, it even outperforms other similarsized embedding models like Quen 3 and Gemma. It's also extremely costefficient. So for this chart, the X-axis is the cost and the Y-axis is the retrieval accuracy. So ideally, you want to be in this left corner and indeed that is where all three of these Neotron 3 embed models are located. So these are the most costefficient models given their size. Again, while this is not the usual AI model which you can chat with and answer questions directly, this embedding model is really useful for search engines and rag systems. The awesome thing is they've released everything already. So, if you click on this hugging face link, note that the 8 billion parameter model is 16 GB in size, which should be able to fit on most mid to high-end GPUs. And then the 1 billion version is even smaller at only 2.3 GB in size. Both models are also multilingual, so it's able to support these languages. And if you scroll down a bit here, it contains all the instructions on how to run this locally on your computer. 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.

[00:40:20]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 uptodate 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.

[00:40:42]Thanks for watching, and I'll see you in the next one.