WW3 prep
Maticus · 2026-11-08 · 1:31:58
https://openvibe.video/v/6220
Transcript (AI)

[0:00:02] Maybe it's just win five on open five.
[0:01:06] You have interfered with our affairs for the last time.
[0:01:36] Oh here we go again They take it
[0:02:56] You have been compared with our bear for the last time.
[0:08:48] Viper sucker.
[0:08:50] Can you turn this not be quite so viper?
[0:11:49] Medical one six six zero or double hardware.
[0:14:58] Should I turn him on?
[0:15:01] Yeah, man's air doing a thing now.
[0:15:24] All on, all on, all on.
[0:15:56] Yeah, I'll eat anything now.
[0:16:04] All on all on all on.
[0:17:16] Alright, alright, alright, alright.
[0:17:20] Google releases their latest video model, Gemini Omni 1.
[0:17:22] 2.
[0:17:22] OpenAI drops their latest and best image model, and this one can even make animated gifs.
[0:17:25] NVIDIA open sources a very powerful AI for weather simulation.
[0:17:27] We have some new research that shows how we can use AI to improve other
[0:17:30] Models without further training.
[0:17:31] This could potentially open up an entirely new way to improve AI models.
[0:17:33] This open source AI can recreate any hairstyle in 3D.
[0:17:35] We also have a very powerful open model for upscaling videos.
[0:17:37] The best open model 3K three point one is finally released this week.
[0:17:39] Kansen also drops their best model I 4, which is very impressive.
[0:17:41] We have a ton of 3D model generators and a lot more, so let's jump right in.
[0:17:44] First up, if you ever need to generate 3D hair for whatever reason, this week we have a really cool AI called hair deck.
[0:17:48] You just need to input a single image of a character, and it'll generate their complete hairstyle in 3D.
[0:17:51] Here are some examples for your reference.
[0:17:52] As you can see, it can reconstruct different hairstyles from a ton of different characters.
[0:17:54] Now, this hair is basically made up of 3D bundles or strands of hair, and what's particularly impressive is that this AI doesn't just make solid, rigid hair, but it also
[0:18:00] Generates the joints of the hair so that it can move naturally in animations.
[0:18:02] So when a character turns their head, you can see the hair moving very naturally.
[0:18:04] It can also preserve their hair colors, including hairstyles with multiple colors.
[0:18:06] And if you look at these benchmarks, hair deck significantly outperforms other hair generating approaches in both geometry accuracy and generation speed.
[0:18:11] The awesome thing is they've released the code and modeled to this.
[0:18:13] So at the top of the page, if you click on this code button and scroll down a bit, here you'll find the instructions on how to download and run this locally.
[0:18:17] Not only does this work with Nvidia GPUs, but also Apple Silicon.
[0:18:19] And if you check out the Hugging Face folder, note that the total size is only around 2.
[0:18:21] 45 gigabytes.
[0:18:22] So this should be able to run on most consumer GPUs.
[0:18:24] And on their GitHub, they've even released the training scripts to this, so you can customize it further.
[0:18:27] Note that this is under the Apache 2 license, which is very permissive.
[0:18:29] You can even use this.
[0:18:30] commercial purposes.
[0:18:30] If you're interested in trying this out, I'll link to this main page in the description below.
[0:18:32] Also this week we have a new AI video upscaler called Relay VSR.
[0:18:35] This can take any low resolution video and upscale it to four times its original resolution, with sharper details and better visual quality.
[0:18:39] Here are some examples for your reference.
[0:18:40] On the left is the original low resolution video and on the right is the upscaled results.
[0:18:43] As you can see it can recover or add much finer details across different types of videos.
[0:18:45] Now what's interesting is how this works.
[0:18:46] Instead of using a massive generative model to process every single frame it only uses the large model to enhance certain keyframes.
[0:18:51] Then a much smaller model uses these keyframes as references to enhance all the frames in between.
[0:18:54] This makes the process much faster and more efficient.
[0:18:55] In fact if you compare this with other similar video upscalers note that this new one on the far right is a lot more detailed and accurate.
[0:19:00] thing is they've released this already.
[0:19:01] So at the top of the page, if you click on this code button and scroll down a bit, here are all the instructions on how to download and run this locally on your computer.
[0:19:05] Note that this is based on Alibaba's one and the total size is around 10.
[0:19:08] 2 gigabytes.
[0:19:08] If you're interested I'll link to this main page in the description below.
[0:19:10] Also this week we have a new image model called Nitro Fusion 2 image general just with the fusion steps which makes it extremely fast.
[0:19:15] Now this is based on Alibaba's Z image which is already really good.
[0:19:18] It can generate some really photo realistic images as well as other artistic styles or even images with text or a ton of different elements.
[0:19:22] Now the original Z image turbo requires like eight steps to generate an image but this one can still make a pretty high quality image with just a single generation step which is very impressive.
[0:19:28] Now how this works is quite interesting.
[0:19:29] It uses a special training
[0:19:30] method that analyzes both the generated image and how well it matches the text props.
[0:19:32] This helps the model preserve important details like objects, styles and typography even with very few generation steps.
[0:19:36] The awesome thing is they've released the model and code to this so at the top of the page if you click on this GitHub link and scroll down a bit here it contains all the instructions on how to download and run this locally in Comp UI or diffusers.
[0:19:42] On their hugging face you can see that the full BF16 model is like twelve gigabytes in size, but they also release some more compressed versions.
[0:19:46] So this should be able to fit on most consumer GPUs.
[0:19:48] If you're interested in trying this out I'll link to this main page in the description below.
[0:19:51] Also this week NVIO releases a new AI called Strata and this is designed to simulate the atmosphere across the entire planet.
[0:19:55] So this takes information about the current atmospheric conditions including temperature, wind, humidity and pressure and predicts how these conditions will change over time.
[0:20:00] Now, what's really impressive is the level of detail.
[0:20:01] It can simulate the atmosphere across the entire Earth at a resolution of approximately five kilometers.
[0:20:04] So it's incredibly detailed.
[0:20:05] This means it can model weather patterns and atmospheric processes much better than conventional methods.
[0:20:08] In fact, this can predict the next atmospheric state in 10 minute increments and then keep each prediction back into itself to simulate multiple days into the future.
[0:20:13] Now, how this works is quite interesting.
[0:20:14] It uses a transformer architecture that divides the planet into smaller sections.
[0:20:16] Each section gets processed separately, and the results are then combined to simulate the global atmosphere.
[0:20:20] And despite the enormous scale of the simulation, the model itself has only about 332 million parameters, not even a billion.
[0:20:24] The awesome thing is NVIDIA has released the model on Hugging Face, and note that the total size is only around 1.
[0:20:28] 6 gigabytes.
[0:20:28] So it's surprisingly small.
[0:20:29] Plus, they also
[0:20:30] The training recipes for this.
[0:20:31] And note that it's under the Apache 2 license, which is very permissive.
[0:20:33] You can even use this for commercial purposes.
[0:20:34] So if you're looking for a lightweight but powerful open source AI for whether forecasting or climate research, this would be a great option.
[0:20:39] If you're interested, I'll link to this in the description below.
[0:20:40] Also, this week we have a new AI called Flow HMR.
[0:20:42] This can take a regular video of a person and reconstruct their entire body movement in 3D.
[0:20:45] The cool thing is, unlike traditional motion capture systems, this doesn't require any special cameras, motion capture suits, or sensors.
[0:20:49] It just takes in a regular video.
[0:20:50] Here are some examples in action.
[0:20:51] As you can see, it can accurately capture all sorts of movements like kicking a soccer ball or doing push-ups.
[0:20:54] What's especially interesting is that it also considers the laws of physics.
[0:20:56] So instead of just approximating the person's movements, it tries to reconstruct motions that are physically realistic.
[0:20:59] Now, this has some
[0:21:00] Pretty cool applications.
[0:21:00] For example, they can transfer these human movements onto a humanoid robot like the Unitree G1 and simulate how the robot would perform those same actions.
[0:21:05] You can also use it for 3D character animation, and you can even combine it with an AI video generator to turn text prompts into 3D human movements.
[0:21:10] As you can see from these benchmarks, FlowHMR outperforms other similar methods in terms of physical realism and motion accuracy.
[0:21:13] The awesome thing is they've released both the models and code to this, so at the top of the page, if you click on this code button and scroll down a bit, here are the instructions for running this locally.
[0:21:19] Note that this Flow HMR latest is the one released this week.
[0:21:21] This is the one that incorporates the laws of physics into its generation so that the movements are more accurate and realistic.
[0:21:24] There's also an online hugging face demo so you can try it out directly in your browser.
[0:21:27] If you're interested, I'll link to this main page in the description below.
[0:21:28] Also, this week we have a new AI called
[0:21:30] Relift from NVIDIA.
[0:21:30] This takes the regular video and tracks the movement of individual points in 3D space, even when objects disappear behind other objects.
[0:21:34] So here's a demo of it in action.
[0:21:36] As you can see, it can reconstruct the motion of objects in 3D, and you can even rotate the camera to view these movements from different angles.
[0:21:40] Here are some additional examples for your reference.
[0:21:41] It works with a ton of different environments, motions, and characters.
[0:21:43] Now here's how it works.
[0:21:44] It first uses existing AI models to track points in the video and estimate their depth.
[0:21:46] Then it uses another AI model to correct any errors and predict how these points move when they're hidden.
[0:21:50] If you compare this with similar competitor models, you can see that Relift, which is on the far right, on average performs much better at tracking the movements of the subject.
[0:21:54] The awesome thing is they've released the code and model weights to this.
[0:21:56] So if you click on this GitHub button and scroll down a bit, it contains all the instructions on how to run this locally on your computer.
[0:22:00] They've also released the training scripts, so you can train or customize the model yourself.
[0:22:02] If you're interested, I'll link to this main page in the description below.
[0:22:04] Also, this week, Google drops their latest video model, Gemini on the 1.
[0:22:07] 2 flash.
[0:22:07] Now, similar to the previous version, this can generate videos up to 15 seconds long.
[0:22:09] And not only does it take in text prompts, but you can also use images, video, or audio as input.
[0:22:12] Here are some examples for your reference.
[0:22:13] As you can see, it can generate cinematic scenes with smooth camera movements, realistic motion, and much better consistency throughout the video.
[0:22:17] One major improvement is reference consistency.
[0:22:19] For example, you can give it images of a person and several outfits, and it'll generate an entire virtual try on video while preserving the person's appearance and the clothing details.
[0:22:24] It also supports video editing just with natural language.
[0:22:26] So you can simply prompt it to change the background, replace an object, or adjust the lighting of an existing video.
[0:22:29] It can even take motion.
[0:22:30] From one video and apply it to a completely different subject.
[0:22:31] And Google has reduced the price by 30%, so it now costs just 7 cents per second for 720p video.
[0:22:35] Here it says that this new Gemini Ani is available through AI Studio, their API, Google Flow, and various other places.
[0:22:39] If you're interested, I'll link to this main page in the description below.
[0:22:41] Also, this week, OpenAI drops their latest and best image model, GPT Images 2.
[0:22:44] 6.
[0:22:44] Now, the previous model was already state-of-the-art, but this one is slightly better.
[0:22:46] It has all the capabilities of the previous version.
[0:22:48] For example, it supports image editing with natural language and reference images as input.
[0:22:50] But a really cool new feature is that it can now create animated GIFs.
[0:22:52] So here are some examples for reference.
[0:22:54] Another major improvement is speed.
[0:22:55] So OpenAI says image generation is now 50% faster than before, and there are actually two different versions available through their API.
[0:23:00] The first is Flare, which is optimized for speed.
[0:23:01] This can generate images in as little as six seconds.
[0:23:03] The second is Sunburst, which prioritizes higher quality images and more consistent details across multiple edits.
[0:23:07] You can also now do image markup for precise edits and also create and export SVGs, which would be great for logo design or scalable graphics.
[0:23:12] The nice thing is this is available already.
[0:23:13] At the bottom of the page here, it says that images 2.
[0:23:15] 6 is available in ChatGPT and codecs for all subscription tiers.
[0:23:17] Also, this week we have some fascinating research from MIT called weightspace programming.
[0:23:20] Now, to understand this, you first need to know that most modern AI models are basically made of neural networks like this.
[0:23:24] This is basically a huge network of dials and knobs that controls how much data flows through to the next layer.
[0:23:27] These dials and knobs are called the model weights.
[0:23:29] Now, what makes these weights so
[0:23:30] Important is that they essentially encode everything the model has learned during training.
[0:23:32] For example, when you train an AI on billions of examples, it gradually adjusts these weights until it learns to recognize patterns, understand language, solve math problems, and much more.
[0:23:38] So in a sense, these weights kinda represent the model's learned knowledge and abilities, or what we might call its intelligence.
[0:23:42] And normally, if you want to make an AI smarter, you need to train or fine-tune it, which involves repeatedly adjusting these weights using some math.
[0:23:46] But what if another AI could just directly examine these weights, figure out what they do, and modify them to improve the model without training?
[0:23:50] Well, that's exactly what this new research paper is about.
[0:23:52] There are actually a ton of really interesting insights from this, so let's go over the main ones.
[0:23:55] First, they gave GPT6 Astra the weights of four smaller models, each trained to play a different game.
[0:23:58] Without telling it which game they were trained on, Astra
[0:24:00] Correctly identified all four games, including Pong and Pac-Man.
[0:24:02] So it seems like just by looking at the weight of a model, GPT6 Astra could figure out that it was trained on a certain type of game.
[0:24:05] In another experiment, they had a car controlled by a small model that kept getting stuck on a ramp.
[0:24:08] So they gave Astra the weights to this model and a recording of the failure.
[0:24:10] Astra then proceeded to change just 3.
[0:24:12] 5% of the weights, and the car was then able to successfully complete the course in every test.
[0:24:15] So essentially it figured out which parameters were causing the problem and changed them to fix the car's behavior.
[0:24:18] Now perhaps the most impressive experiment involved Alibaba's Quen 3.
[0:24:21] 5, a small model with only 800 million parameters.
[0:24:23] Again, they asked Astra to improve its math capabilities.
[0:24:25] So Astra analyzed the model, identified a problem in its internal reasoning, and modified only a small percent of its weights, and remarkably, this boosted
[0:24:30] Math 500 benchmark score from just 7.
[0:24:31] 9% to 48%, which is crazy.
[0:24:33] Its instruction following performance also improved dramatically.
[0:24:35] Now think about what this means.
[0:24:36] Instead of retraining hundreds of millions of parameters, which takes a ton of time and compute, you can just get a bigger model like Astra to inspect the weights and just change a tiny fraction of them to improve the model's performance.
[0:24:43] This could potentially open up an entirely new way to improve AI models.
[0:24:45] Now, while this paper is legit and its results have been reproduced by others, there's also quite some hype around this, which is misleading.
[0:24:49] So let's clarify this.
[0:24:50] It turns out the original Quen model had a known problem.
[0:24:52] It would frequently get stuck in repetitive reasoning loops and run out of tokens before it could produce a final answer.
[0:24:55] So Astra didn't necessarily teach it any new math abilities.
[0:24:58] Instead, it just fixed this existing problem, which allowed Quen
[0:25:00] To complete more questions that it was already capable of solving.
[0:25:02] In fact, one independent analysis found that the original unmodified coin model could already achieve around 44% on the Math 500 benchmark, just by using different settings.
[0:25:08] So the huge improvement from this paper might not be as revolutionary as it seems.
[0:25:10] Another criticism is that Astro wasn't just looking at hundreds of millions of weights and magically figuring everything out.
[0:25:14] For the coin experiment, it actually had access to Python tools and could repeatedly run the model on practice questions.
[0:25:17] It used these results to identify the problem and test different modifications.
[0:25:20] So rather than directly understanding all the knowledge encoded within the model's weights, it was partly analyzing the model's behavior and experimenting with changes.
[0:25:25] And more importantly, when they tried to improve another model, specifically Google's Gemma, which didn't suffer from these repetitive reasoning loops, then it turns out that.
[0:25:30] Astra couldn't actually improve its performance further.
[0:25:31] So while this research is impressive, it doesn't mean AI can now magically understand and upgrade the intelligence of any model.
[0:25:35] But still, this paper is quite interesting, and this could lead to new ways to use AI to debug or improve other AI models.
[0:25:39] And the nice thing is they've released the experiment logs on GitHub.
[0:25:41] So you can inspect the results and try to model yourself.
[0:25:43] If you're interested, I'll links to this main page in the description below.
[0:25:45] If you want to create viral videos at scale, definitely check out Higgsfield, the sponsor of this video.
[0:25:48] They just launched Higgsfield Content Factory, which is kind of like an entire video production team right at your fingertips.
[0:25:51] You can get AI to automatically write scripts, generate visuals, add voiceovers, and edit everything into a complete video, all from a single idea.
[0:25:56] I'm especially impressed by how easy it is to create videos based on popular trends.
[0:25:58] You can choose from over fifty ready-made
[0:26:00] Presets or simply paste a link to a viral TikTok or Instagram video as a reference.
[0:26:02] Higgsfield will automatically follow its style, structure, and pacing, but create a completely new video based on your topic.
[0:26:06] As you can see, it generates an entire story complete with characters, visuals, narration, and scene transitions.
[0:26:09] You can create videos anywhere from 15 seconds to five minutes long, and even generate up to four episodes at once.
[0:26:12] Plus, you can access Content Factory directly inside Claude through Higgsfield MCP.
[0:26:15] Whether you want to create YouTube shorts, TikToks, Instagram reels, educational videos, or marketing content, Higgsfield Content Factory makes the entire process incredibly easy.
[0:26:20] Try Higgsfield today using the link in the description below.
[0:26:22] Also, this week we have a new AI called Lucid Lira.
[0:26:23] This can turn a single image into an explorable 3D scene.
[0:26:26] But what's especially interesting is that it can also recognize individual objects within that scene, so you can select, edit, or extract them.
[0:26:30] Afterwards.
[0:26:30] Here are some examples for your reference.
[0:26:31] As you can see, it can generate detailed 3D environments from regular images and even identify individual objects within it.
[0:26:35] And this works across a variety of different environments and objects.
[0:26:37] For example, you can select an object and move it around, resize it, duplicate it, or remove it entirely.
[0:26:40] It can also convert these scenes into environments for robots to navigate and interact with, which could be useful for robotic simulations.
[0:26:45] Now, how this works is quite interesting.
[0:26:46] It first uses a video generation model to predict what the scene would look like from different camera angles.
[0:26:49] Then another AI model combines this information into a 3D scene and identifies the objects within it.
[0:26:52] And if you look at the comparisons, this outperforms other similar approaches in terms of recognizing and separating objects in 3D scenes.
[0:26:57] Now at the top of the page it says the code is coming soon, so stay tuned for that.
[0:26:59] For now, if you
[0:27:00] interested in reading further I'll link to this main page in the description below.
[0:27:02] Also this week we have a new image model called PH3 image.
[0:27:04] The interesting part is they took the best open video model, Minimax H3, and they basically turned it into an image generator and editor.
[0:27:08] And it turns out it's really good.
[0:27:08] Not only can it generate images from text prompts but it can also edit existing images with natural language and even combine up to eight reference images into one scene.
[0:27:14] As you can see it's especially good at photorealistic images, cinematic scenes, portraits, anime and different artistic styles.
[0:27:18] The developers basically modified the architecture of Minimax to specialize in generating images instead of video.
[0:27:21] They also reduced the number of generation steps from 50 to just eight.
[0:27:23] So it's about five times faster than the original model.
[0:27:25] It can generate images up to 16 megapixels and they've also added an AI prompt enhancer.
[0:27:28] So even if you enter a really simple prompt it can automatically expand it
[0:27:30] Into something more detailed.
[0:27:30] The awesome thing is they've released the model weights on Hugging Face, and you can run it with either diffusers or Comp UI.
[0:27:34] There are multiple versions, including the original BF16, 8-bit, and 4-bit versions, so you can download the one that fits your hardware.
[0:27:38] If you're interested in reading further, I'll link to this main page in the description below.
[0:27:40] Also, this week we have a new open source robot toolkit called Open RoboKit.
[0:27:43] This provides a collection of tools to help you control robots, simulate their movements, and train them on different tasks.
[0:27:46] Now there are four main libraries here.
[0:27:47] First there's RoboKit, which helps robots calculate their movements and plan their actions.
[0:27:49] For example, here you can see robots spinning pens, rotating cubes, opening notebooks, climbing obstacles, and even performing jumps.
[0:27:53] Next, there's retargeting kit, and this can take human movements from videos, motion capture, or VR devices, and translate them into movements for robotic hands or full humanoid robots.
[0:27:59] Then there's grasp.
[0:28:00] Which helps robots figure out how to pick up and manipulate different objects.
[0:28:02] It can also train robotic policies and deploy them in simulations for real-world robots.
[0:28:04] And finally, there's RoboSkills, which provides different workloads for AI agents to perform various robotics tasks.
[0:28:08] For example, you can get an AI agents to generate robot movements, reconstruct simulations from videos, or autonomously improve its robot control through experiments.
[0:28:13] The nice thing is they've released the code to all four toolkits.
[0:28:15] If you click on this GitHub icon at the top right corner and scroll down a bit, you can find all the instructions on how to install and use them.
[0:28:19] This is a fantastic resource if you're doing robotics, so definitely check it out.
[0:28:21] I'll link to this main page in the description below.
[0:28:22] Also, this week we have a really cool real-time multimodal voice called Moss Live.
[0:28:25] And yes, this is the team behind the awesome Moss image as well.
[0:28:27] The thing is not only is this a real-time voice you can have conversations with, but it's multimodal, so you can
[0:28:35] Honestly, it's my first time here, so I guess you're my tour guide, just planning a nice afternoon.
[0:28:38] Alright, let's head east from Citroen Road to the Boom.
[0:28:40] Then stroll north along the waterfront toward an ancient road, and join the views and the historic buildings.
[0:28:44] We'll find somewhere nearby for dinner before you head to the airport.
[0:28:47] It can listen and speak at the same time, so you can interrupt it halfway through a sentence, and it'll understand whether you're asking a new question or simply acknowledging what it's saying.
[0:28:52] It can also distinguish your voice from other people or background voices so it doesn't accidentally respond to unrelated stuff.
[0:28:56] It can participate in a meeting and even help you look up information and provide answers at the right times.
[0:28:58] Or here's another example where you
[0:29:00] Use this for travel, you can show your camera to Moss and ask it questions about the blitz.
[0:29:02] Now, how this works is quite interesting.
[0:29:03] The system uses separate foreground and background components.
[0:29:05] The foreground handles real time conversations, while the background takes care of complex reasoning and tool calls.
[0:29:08] This means it can keep talking to you even while it works on another task.
[0:29:10] And if you look at these benchmarks for real time voice conversations, interruptions, and tool use, it performs very well against other leading models.
[0:29:14] You can currently try it through their online demo or access it through their API.
[0:29:16] There's no mention of releasing the model weights, but anyway, if you're interested in trying this out, I'll link to this main page in the description below.
[0:29:20] Also, this week we have a new open source image generator called Pyramid JIT.
[0:29:23] Now, what's really impressive is that they trained this model from scratch with just $21,000 in compute costs.
[0:29:26] Here are some examples for your reference.
[0:29:27] As you can see, it can generate a ton of different images like realistic portraits, landscape.
[0:29:30] And even stylized illustrations.
[0:29:31] It's a relatively small model with about two billion parameters, but the image quality is surprisingly good for its size and training budget.
[0:29:35] Now, the design of this is quite interesting.
[0:29:36] Instead of compressing images into a separate and latent space like many other image generators, it works directly in pixel space, which you and I can see.
[0:29:41] It also uses a special architecture that reduces the amount of computation needed during training.
[0:29:44] But they say that this is just an experimental model and it still struggles with complicated scenes, multiple people, and accurate text generation.
[0:29:47] The awesome thing is they've released the model and code choos.
[0:29:49] So at the top, if you click on this model code button and scroll down a bit, it contains all the instructions on how to download and run this locally on your computer.
[0:29:54] And then on their hugging face, note that the total size of this model is 8.
[0:29:56] 74 gigabytes, so this should be able to run on most consumer GPUs.
[0:29:58] Also note that the model is released on the
[0:30:00] Under the Apache 2 license, which is very permissive.
[0:30:01] You can even use this for commercial purposes.
[0:30:02] If you're interested, I'll link to this page in the description below.
[0:30:04] Also, this week we have a new 3D generator called Cola 3D.
[0:30:06] This can turn a single image into an entire 3D scene, and it can separate each object into its own individual 3D mesh.
[0:30:10] Here are some examples for reference.
[0:30:11] Now, this supports three different modes.
[0:30:12] You can let the AI automatically identify and generate all the objects, or you can specify individual objects using points or segmentation masks.
[0:30:17] What's interesting is that instead of generating each object separately and then trying to piece them together, this model processes the entire scene at once.
[0:30:21] This helps preserve the correct positions, sizes, and spatial relationships between objects.
[0:30:24] The awesome thing is they've released the code to this, so at the top of the page, if you click on this code button and scroll down a bit, here it contains all the instructions on how to run this locally on your computer.
[0:30:29] In addition, they've
[0:30:30] The training code and data set to this as well, which is fantastic.
[0:30:32] If you're interested, I'll link to this main page in the description below.
[0:30:34] Also, this week we have a super tiny open vision model called Mirage.
[0:30:36] This can analyze any image and identify objects and read text in the image.
[0:30:38] And what's really impressive is that the entire model is only about 22 megabytes, not even a gigabyte in size.
[0:30:42] So here are some demos.
[0:30:42] You can upload any image or use your camera, and the AI will automatically identify different objects and outlines around it.
[0:30:46] It can also read any text in the image and understand spatial relationships such as which objects are attached to each other or which objects are on a table.
[0:30:50] Another interesting feature is that you can give it a set of choices and it'll automatically make a decision based on what it seeks.
[0:30:54] So you could use this to classify objects or count items.
[0:30:56] Now, how this works is quite interesting.
[0:30:57] It basically uses one shared vision encoder to understand the image and then
[0:31:00] Several specialized components extract the objects, text, and other information.
[0:31:02] This allows it to handle multiple vision tasks with just one tiny model.
[0:31:04] Now, the model has only like 51 million parameters, and with two bit quantization, they've compressed it down to just 22 megabytes.
[0:31:08] So this can process an image in like a tenth of a second on a smartphone.
[0:31:11] And of course you can also run this locally on your laptops, wearables, or other consumer devices.
[0:31:13] Now, the awesome thing is they've released the model and code to this.
[0:31:15] So over here, if you click on the GitHub link, it contains all the instructions on how to run this locally on your computer.
[0:31:18] If you're interested, I'll link to this main page in the description below.
[0:31:20] Also, this week we have a new AI called DreamX.
[0:31:22] This can take the movements from a video of any character and transfer those movements to a completely different character.
[0:31:25] So here are some examples for your reference.
[0:31:26] As you can see, it can take a video of a human hand manipulating objects and use that motion to generate a video.
[0:31:30] Of a robot performing similar actions.
[0:31:31] What's even more interesting is that it can translate those movements into actual robot actions.
[0:31:34] So instead of just generating a video, the robot can physically do this action in real life.
[0:31:37] Now, how this works is quite interesting.
[0:31:38] DreamX uses a video diffusion model to extract motion information from videos.
[0:31:40] It separates the movements from the appearance of the original subject, then applies those movements to another subject.
[0:31:44] And as you can see from the benchmarks, DreamX outperforms some similar methods for motion understanding and transfer.
[0:31:48] Now at the top of their project page, it says the code is coming soon, so stay tuned for that.
[0:31:50] For now, if you're interested, I'll link to this main page in the description below.
[0:31:52] Also, this week we have a new Robotics model called Zero Real Agent Policy, or ZAP.
[0:31:55] This can take camera footage and natural language instructions and then directly control a robot to perform any specified tasks.
[0:31:58] What's especially impressive is
[0:32:00] That it was trained entirely in simulation without any real-world training data.
[0:32:02] Here are some examples for your reference.
[0:32:03] As you can see, you can control this robotic arm to move this test tube, or here it's getting it to move this chess piece according to the diagram.
[0:32:08] You can even get it to hold this pen and draw an X on this tic-tac-toe grid.
[0:32:10] And these are real robots, not simulated videos.
[0:32:11] Now here's how they trained us.
[0:32:12] They started with a vision language model that already understands objects and instructions, but then they taught it how to move a robot using reinforcement learning in a virtual environment.
[0:32:18] It initially learned just 40 pretty simple simulated tasks.
[0:32:20] But the cool part is when they applied this to a real robot, it turns out that it could perform new tasks it had never seen before.
[0:32:23] And the impressive thing is this model has just 27 billion parameters, but in real-world tests, it even matched the success rate of the much larger GPT-6 Astra, while being significantly faster.
[0:32:30] And at a fraction of the cost.
[0:32:30] Now, at the top of the page, they've only released a technical paper for now.
[0:32:32] There's no mention whether they will release the model itself.
[0:32:34] But if you're interested in reading further, I'll link to this page in the description below.
[0:32:36] Also, this week, Tencent drops their latest model, Hi4.
[0:32:38] One of the biggest new features is visual understanding.
[0:32:40] The model can now analyze images and solve problems that involve visual reasoning.
[0:32:43] So, for example, you can get it to analyze some satellite images, or you can get it to use Blender to reconstruct this entire home.
[0:32:47] Like other Frontier models, this has computer use capabilities, so it can interact directly with other software or interfaces.
[0:32:51] Or here's another example where we can get it to create a really detailed 3D environment in FreeJS.
[0:32:54] It's also great at knowledge work.
[0:32:55] So here's an example where we use it to do some financial modeling and equity research.
[0:32:58] Or here's a really cool example where we can get it to reconstruct
[0:33:00] The 3D mouse embryo from 7 million cells in just one and a half hours.
[0:33:02] And if you look at these benchmarks across agent decoding, 3D modeling, and knowledge work, it does seem like this is one of the best open models out there.
[0:33:06] However, the model weights haven't been released yet.
[0:33:08] They plan to release the model in the coming days, so stay tuned for that.
[0:33:10] For now, you can try it through Tencent Cloud or Work Buddy, which is like their free harness, very similar to Codex and Claude Code.
[0:33:14] If you're interested in trying this out, I'll link to this main page in the description below.
[0:33:16] Also, this week we have a new open model called Hydrogen 1.
[0:33:18] 0.
[0:33:19] This is an AI model specifically designed for scientific research, including chemistry, material science, and biology.
[0:33:22] Now, this is a massive model with 753 billion total parameters.
[0:33:25] It's based on GLM 5.
[0:33:26] 3, but they further trained it on scientific literature, laboratory records, and actual research experiments.
[0:33:30] Here are some examples of the incredible things it can do.
[0:33:31] For example, they got it to write and execute a molecular simulation of polyethylene in a plastic bag.
[0:33:35] The model even identified an error in its code, corrected it, and continued the simulation.
[0:33:38] In another example, it analyzed an NMR spectrum to help identify the chemical structures of possible reaction products.
[0:33:41] It can also design laboratory experiments and generate instructions for robotic lab equipment.
[0:33:44] For example, here it created a protocol for a chemical experiment, tested the instructions in a virtual environment, and automatically corrected an error before execution.
[0:33:50] What's interesting is that they've trained it on complete research trajectories, including failed experiments, so it can learn how to make better scientific decisions over time.
[0:33:54] If you look at various benchmarks for scientific reasoning and experiment planning, hydrogen performs exceptionally well compared to other frontier models.
[0:33:59] In some cases,
[0:34:00] Even beating in GPT6 Astra.
[0:34:01] The awesome thing is over here near the bottom of the page, they've released the models for you to download on Hogging Face.
[0:34:04] Note that they released three different versions.
[0:34:05] There's the full model plus an FP4 version for NVIDIA GPUs, as well as an MX version for AMD GPUs.
[0:34:09] Note that even for the FP4 one, it's around 464 gigabytes.
[0:34:11] So you'll need to stack multiple GPUs in order to run this.
[0:34:13] But I mean, this is one of the best open models for scientific research.
[0:34:15] In some cases, even outperforming GPT6 Astra.
[0:34:16] If you're interested in trying this out, I'll link to this main page in the description below.
[0:34:19] Also, this week OpenAI drops their latest model, GPT6 Cyber.
[0:34:21] Now, as the name implies, this is a specialized model for cybersecurity.
[0:34:24] It's designed to autonomously find vulnerabilities in software and help developers fix them, and some of its results are quite impressive.
[0:34:28] For example, OpenAI used this model to identify
[0:34:30] 81 previously unknown i severity vulnerabilities in Firefox, even after extensive security testing with previous AI models.
[0:34:34] It also helped develop working exploits for over 900 vulnerabilities in Linux.
[0:34:37] That's pretty crazy.
[0:34:38] Now, if you look at these cybersecurity benchmarks, GPT6 Cyber significantly outperforms their previous cyber model, and even GPT6 Astra on most of these specialized tasks.
[0:34:43] Now, we plebs won't have access to this.
[0:34:45] So here it says they're introducing access
[0:34:59] to the money.
[0:35:01] Hey, I've kinda well here waiting.
[0:35:12] No, that needs to be fixed.
[0:35:13] I don't know that needs to be fixed.
[0:35:18] Hey.
[0:35:26] Still here.
[0:35:27] What's on your mind?
[0:35:30] wait a second.
[0:35:32] I'll tell you.
[0:35:36] how about that?
[0:35:39] okay.
[0:35:40] well I've kind of rediscovered a bug in Actor Core, and I'd like to tell you about it and see if we can get a fix in.
[0:35:53] Go ahead, I'm listening.
[0:35:55] Describe what you're seeing.
[0:35:58] Okay.
[0:36:00] Yeah, what's happening?
[0:36:02] Yeah.
[0:36:03] So in ActorCore, Mission Control, under Little Buddy, in the overview sub tab of the little buddy page of ActorCorps, there are controls for the camera.
[0:36:20] And we have gain and exposure.
[0:36:23] And I accidentally hit default and now the camera is super dark and the gain and the exposure
[0:36:30] Aren't really working on the clickies there, nor are they working on the key commands that are attached.
[0:36:37] So we need to look that specifically.
[0:36:41] Camera defaults nuking the picture.
[0:36:45] Not just that, but the actual whoa.
[0:36:51] Sounds like there's more to it.
[0:36:53] What else is going on?
[0:36:54] I mean yeah, the more to it is is that the gain and the exposure aren't working.
[0:37:02] That's the same thread I'm already on.
[0:37:04] Is it that they move but the image doesn't change or do they not move at all?
[0:37:08] Well after clicking default, then I try to make it because the default was way too dark because that camera on Cosmo is kind of garbage.
[0:37:18] But I mean it's wonderful garbage.
[0:37:20] But anyway, I'm getting off track.
[0:37:22] The the but the gain and exposure are not adjusting like the they move.
[0:37:28] Oh wait, now they don't move.
[0:37:30] I saw you oh wait yeah now they move when I click the buttons in the UI sometimes But the image doesn't change same thing with exposure I it the values change but the image stays dark Values moving with no image change points to the driver
[0:37:50] call not going through Alright well That's noted for the pack I'm not go on I'm not sure how we're doing on usage let me take a look
[0:38:00] I think I might just set you loose on this one in a second, okay?
[0:38:05] Sure thing.
[0:38:06] I'm ready when you are.
[0:38:09] Alright.
[0:38:10] Yeah, this is a thing that needs to happen.
[0:38:13] So I guess make it happen.
[0:38:15] On it.
[0:38:18] Should I close Actor Core or just you're just gonna do your own thing on your own computer?
[0:38:24] How's this gonna work?
[0:38:28] Leave it open.
[0:38:29] I'm working off.
[0:38:30] the same files on your machine.
[0:38:31] So as long as nothing's locked, I'll read and patch right alongside it.
[0:38:35] Sweet.
[0:38:38] Alright, thank you.
[0:38:40] hit me up.
[0:38:41] Ping me on my wrist when it's ready.
[0:38:46] Will do.
[0:38:47] Wrist ping when it's done.
[0:38:55] Fucking superhero in my computer, dude.
[0:38:59] Alright.
[0:39:05] yeah, this whole video is freaking crazy.
[0:39:11] Hair deck, I need a hair deck.
[0:39:14] Is it ready?
[0:39:20] Well shit.
[0:39:43] Okay, I got the ping.
[0:39:45] I got the ping, so should I just shut it down and and you can edit some code or what's up?
[0:39:54] The ping means the fix is in.
[0:39:55] You can close it whenever.
[0:39:57] Nothing's locked.
[0:39:58] Want me to walk you through what changed?
[0:40:00] Oh you actually the fix is in.
[0:40:02] Excellent.
[0:40:03] Okay, I'll restart it now and give it a test.
[0:40:06] Go for it.
[0:40:07] Yeah.
[0:40:07] Pay special attention to whether gain and exposure actually shift the image this time.
[0:40:11] Yeah, yeah, that's I'm gonna be paying very special attention to that.
[0:40:16] That's the one.
[0:40:17] I'm right here if the numbers still lie.
[0:40:33] We'll see here.
[0:40:34] What we got?
[0:40:35] Some more music.
[0:40:39] It sounds easy.
[0:40:43] Okay, so this quick.
[0:40:49] Okay, well.
[0:40:51] Important pressure.
[0:41:05] Wait a minute, what?
[0:41:07] can we can we patch it and see if it I mean what's the what's the situation?
[0:41:53] Oh, that's confusing.
[0:42:03] I'm gonna have to talk to Chief about this.
[0:42:08] Chief, can you take can you take a look at Core, Actor Core Engineer Bot?
[0:42:14] I'm trying to get I have an issue and Core's not cooperating.
[0:42:20] I'm gonna need your help here, I think.
[0:42:22] Core's giving you trouble?
[0:42:25] Digging into what Core's been doing.
[0:42:32] That's the read-only pass on core's definition in recent records.
[0:42:36] What's it actually refusing to do?
[0:42:42] Well, I wanted to fix After Core and apply it I mean cre pass bug.
[0:42:51] Core Engineer is your After Core engineering bot, and you're after it fixing After Core and getting that bug through.
[0:42:57] My look at its latest thread should show where it's stuck.
[0:43:00] Yeah.
[0:43:01] Now have a look at that and let me know what you think.
[0:43:07] Pulling up Cores thread now.
[0:43:10] Maybe you need to give it a skill, I think.
[0:43:14] A skill could be the miss Yeah.
[0:43:16] That's worth testing once I see what it said.
[0:43:19] What's the skill you had in mind?
[0:43:21] I d I didn't.
[0:43:22] It was just a thought.
[0:43:23] You can go ahead and get back to me when you have some great news for me.
[0:43:30] We'll do
[0:43:36] which only allows approved security researchers to use these advanced models.
[0:43:38] But if you're interested in reading further, I'll
[0:43:50] be able to do that.
[0:47:01] Hey, I'm having a little bit of an issue with Active Bud, I'd like you to help me touch the bug.
[0:47:09] Sure.
[0:47:10] What's the bug doing?
[0:47:15] Well, when I go to the little buddy section in the overview sub tab of the little buddy page in ActRecord.
[0:47:30] The camera controls gain and exposure.
[0:47:36] They change when I click them.
[0:47:37] They change when I push the key commands, but the image remains dark.
[0:47:43] So they're not affecting Cosmos camera.
[0:47:53] Hang on, I'll get in text.
[0:47:55] Sure.
[0:48:01] So yeah, go ahead and it yeah it's it's yeah right there
[0:48:08] camera controls aren't affecting the camera but they appear to change in the software not sure what's going on there
[0:52:02] To this release page in the description below.
[0:52:04] Also, this week, Meta is introducing real-time voice conversations and expressive avatars to views.
[0:52:07] If you haven't heard of Muse, it's basically a personal AI assistant from Meta.
[0:52:09] It's very similar to OpenAI Docs, Glockbot for OpenClock, but now instead of typing, you can just talk to Muse naturally with your voice, just like regular conversation.
[0:52:15] For example, you can ask it to help plan your next application, brainstorm ideas, or even rehearse for a job interview.
[0:52:18] It can also work on tasks while you're talking to it so you don't have to like pause your current conversation or wait for it to finish.
[0:52:22] They also released an avatar which can react to you with realistic facial expressions and gestures that match the conversation.
[0:52:26] Now, here it says this is now rolling up in the US through the Muse app on mobile, web and Mac.
[0:52:29] So you can just download.
[0:52:30] Or upgrade your use app to try out this new voice feature.
[0:52:31] Also, this week we have a new AI called Mobius.
[0:52:33] This can generate realistic human movements from text descriptions and it can also understand, complete, or edit existing motion sequences.
[0:52:37] Here are some examples for your reference.
[0:52:38] You can simply describe an action and it'll generate a 3D character that performs that movement.
[0:52:41] But what's interesting is that this model can also do the reverse.
[0:52:43] You can give it a motion sequence and it can describe what the character is doing in natural language.
[0:52:46] It also supports motion completion.
[0:52:47] So if you provide only part of a movement, it can automatically generate the missing parts to complete the action.
[0:52:50] You can even provide text instructions to guide the results.
[0:52:51] And finally, it supports motion editing.
[0:52:52] So you can modify existing animation through natural language instructions.
[0:52:55] Now, how this works is quite interesting.
[0:52:56] It uses a single AI architecture that can process both language and human motion.
[0:52:58] This allows the same model to understand
[0:53:00] Movements and generate new ones.
[0:53:00] And if you look at these benchmarks for motion generation and understanding, it actually outperforms many competing models across various metrics.
[0:53:04] Now, at the top of the page, if you click on this GitHub link, here it says the code is under internal review.
[0:53:08] So hopefully this gets approved and open source.
[0:53:09] For now, if you're interested in reading further, I'll link to this main page in the description below.
[0:53:12] Also, this week the model waits to the best open model.
[0:53:14] Kimmy K3.
[0:53:15] 1 are finally released as prompt.
[0:53:16] Now I already mentioned this when they first announced it a few weeks ago.
[0:53:17] As you can see from these benchmarks across agent encoding, general knowledge, and scientific research, it is the best performing open model out there, and it even comes close to some of the top closed models.
[0:53:23] And as promised, this week they released the full model on Hugging Face, which is great.
[0:53:25] Now, this is one of the biggest open models out there, with 2.
[0:53:27] 8 trillion total parameters, and it's a mixture of experts, so only
[0:53:30] 104 billion parameters are active at a time.
[0:53:31] As you can see from the hugging face repo, the total size of this is massive at 1.
[0:53:33] 56 terabytes.
[0:53:34] Further down this page, it contains all the instructions on how to download and run this locally on various platforms.
[0:53:37] Now, while this is huge, that's the size you'd expect from a model with Frontier Intelligence.
[0:53:40] It's aimed more at enterprises with the hardware to run this rather than consumers.
[0:53:42] But it's already amazing that they're willing to release this model for free for anyone to download and use.
[0:53:45] I mean, it probably costs millions of dollars and a ton of effort to train this, so props to the Kimmy team for making this open.
[0:53:48] If you're interested, I'll link to this page in the description below.
[0:53:50] Also, this week, Baidu drops their latest open language model, earning 5.
[0:53:52] 5 flash.
[0:53:53] Now, this is quite large with 280 billion total parameters, but it's a mixture of experts, so only 14 billion parameters are active when you use it.
[0:53:58] Like most Frontier models, it also supports a context.
[0:54:00] Of up to 1 million tokens.
[0:54:01] So you can fit a ton of stuff into your prompt at once.
[0:54:02] And if you look at these benchmarks, it performs extremely well against other open models of similar sizes, particularly in reasoning, coding, and general knowledge.
[0:54:07] The awesome thing is Baidu has released the model weights on Hugging Face, and they've actually released two versions.
[0:54:11] The FPH model is about 288 gigabytes, while the full BF16 version is around 573GB.
[0:54:15] If you scroll down on their Hugging Face page, you'll also find instructions on how to run the model locally on your computer.
[0:54:18] Just keep in mind that this is a base model, so it's primarily intended for further training rather than being a helpful assistant right out of the box.
[0:54:23] Anyway, if you're interested, I'll link to this main page in the description below.
[0:54:25] Also, this week we have another multimodal model from Antgroup called Ling3.
[0:54:27] 2 Flash.
[0:54:28] Not only can it understand text, but it also supports image and
[0:54:30] And video inputs.
[0:54:30] So this makes it a lot better at things like 3D and spatial understanding.
[0:54:32] Here are some examples where we can input any image for it to reconstruct in 3D.
[0:54:35] Or here's another example where we can get to create this animated website of London.
[0:54:37] It also has quite powerful agentic capabilities.
[0:54:39] So, for example, you can get to coordinate a team of agents to turn financial data into an analysis report and slideshow like this.
[0:54:43] Now, this is quite a large model with 560 billion total parameters, but it's a mixture of experts, so only 25 billion parameters are active when you use it.
[0:54:49] And if you look at these benchmarks, it's quite competitive with the leading open models out there, including GLN 5.
[0:54:52] 3 Flash and Deep Seek, so it's definitely among the best multimodal open models out there.
[0:54:55] Now, currently you can try it out through Ant Group's developer platform, but they have open sourced these bling models in the past, so I'm hopeful that they will also release the
[0:55:00] Model weights to this in the near future.
[0:55:01] Anyway, if you're interested, I'll link to this main page in the description below.
[0:55:03] Also, this week we have another open model called Ripple65B.
[0:55:05] This is a 65 billion parameter model trained from scratch, and they claim this is a new state-of-the-art open model in terms of agent decoding, at least if you compare it to other open models from the US.
[0:55:12] Now, this uses a dense architecture, so unlike a mixture of experts' models, it activates all 65 billion parameters when you use it.
[0:55:16] Like most Frontier models, it also supports a context window of up to 1 million tokens, so you can fit a ton of stuff into your prompt at once.
[0:55:20] Now, interestingly, they originally developed this model to power its AI chip design platform.
[0:55:23] The idea is to use AI to design better chips, then use those chips to train more powerful AI models.
[0:55:26] And according to the team, the entire research and development process for Ripple was only $32 million.
[0:55:30] Which is way cheaper than the training costs of many Frontier labs.
[0:55:32] The awesome thing is they've released the model on Hugging Face for you to run this locally.
[0:55:34] So if you click into this Hugging Face page, note that they've released two versions.
[0:55:37] There's the full BF60 model, which is 130 gigabytes in size, but they also released a smaller FP4 version for NVIDIA GPUs, which is around 57GB.
[0:55:42] And this is under the Apache 2 license, which is very permissive.
[0:55:44] If you're interested, I'll link to this page in the description below.
[0:55:46] Also this week, Chinese food delivery company Mage1 drops their latest AI model, Longcat 2.
[0:55:49] 5.
[0:55:49] And this is a really powerful multimodal model that's especially good at coding, visual understanding, and doing complex tasks autonomously.
[0:55:53] Now this is quite a massive model at 1.
[0:55:55] 6 trillion total parameters, but uses a mixture of expert's architectures, so only around 48 billion parameters are active when you use it.
[0:56:00] Here are some examples of the incredible things it can do.
[0:56:01] For instance, they got it to create this interactive fluid simulation website complete with realistic physics and visual effects.
[0:56:05] And beyond coding, it can also perform professional office tasks.
[0:56:07] For example, you can get it to analyze business data or generate reports.
[0:56:09] Now, one of the biggest upgrades over the previous Longcat model is its native visual understanding.
[0:56:12] So it can now understand and analyze images.
[0:56:14] And if you look at these benchmarks for coding, identifies and visual understanding, you can see that this new model is quite competitive with some of the top models from OpenAI and Anthropic, at least their older generations.
[0:56:20] Now, currently you can try LongCat 2.
[0:56:21] 5 through the official Longcat chat website or access it through their API.
[0:56:24] At the top, the GitHub and hutting face releases are marked as coming soon, so it looks like they are planning to open source this as with their previous models.
[0:56:29] For now, I'll link to the
[0:56:30] Main page in the description below.
[0:56:30] Also, this week, thinking machines unveils their latest model, Inkling2 Small.
[0:56:33] Now, despite its name, this is actually a pretty large model with 594 billion total parameters.
[0:56:36] And this is a mixture of experts, so just 17 billion are active when you use it.
[0:56:38] It's designed to be especially strong at coding, visual understanding, and long autonomous tasks.
[0:56:41] For example, they got it to create an entire interactive city simulator with 3DS, including procedurally generated buildings, vehicles, and infrastructure.
[0:56:46] Or here's another example where they got it to create a 3D animation in Blender.
[0:56:48] So as you can see, it's quite good at operating different software and interfaces.
[0:56:50] What's even more interesting is its ability to carry out autonomous AI research.
[0:56:52] In one experiment, it independently trained and improved another AI model for time series forecasting.
[0:56:55] It designed experiments, evaluated results, and adjusted the training process all on its own.
[0:56:58] Now, if you look at these benchmarks on H
[0:57:00] Decoding, visual reasoning, and knowledge work, it's extremely competitive with other leading open weight models.
[0:57:03] In fact, once they release this, it might very well be the best open model from the US.
[0:57:05] Now, at the bottom here it says you can try English Small right now through their platform called Tinker.
[0:57:09] And they do plan to release the model weights next week under the very permissive Apache 2 license.
[0:57:11] If you're interested, I'll link to this release page in the description below.
[0:57:13] Also, this week we have a new 3D generator called Vox Solution.
[0:57:16] This can generate entire 3D indoor scenes, like apartments, offices, and bedrooms based on just a category or a simple layout.
[0:57:20] Here are some examples for your reference.
[0:57:20] As you can see, it can generate fully furnished rooms with quite detailed geometry, even down to smaller objects like table waves and cabinet dividers.
[0:57:25] Now, how this works is quite interesting.
[0:57:26] It first generates a rough layout of the room in boxes, then it progressively increases the resolution to add
[0:57:30] Now at the top of the page it says the code is coming soon, so stay tuned for that.
[0:57:33] For now, if you're interested in reading further, I'll link to this main page in the description below.
[0:57:35] Also, this week we have a new AI called Solver.
[0:57:37] This can take multiple photos or video frames of a scene and reconstruct its 3D geometry, along with the camera positions.
[0:57:41] You can give it photos from different angles, and you can figure out how everything fits together in 3D.
[0:57:43] What's particularly interesting is that Solver can continuously refine its existing reconstruction.
[0:57:46] You can add more images at any time, even in random order, and it'll update the scene accordingly.
[0:57:49] It also supports real time video streams, so it can progressively reconstruct the scene as new frames arrive.
[0:57:53] Now, how this works is that it uses a loop transformer, which seems to be getting more popular.
[0:57:55] It repeatedly applies the same loop to improve its estimates of the 3D scene and camera positions.
[0:57:58] Now, if you look at these benchmarks for 3D
[0:58:00] Reconstruction, Solver is quite competitive with other state-of-the-art models.
[0:58:02] It's also particularly efficient for continuous video processing.
[0:58:04] Now, at the top, they have released a GitHub for this.
[0:58:05] If you click into it, note that it says the code and models will be released within a week, so stay tuned for that.
[0:58:08] If you're interested in reading further, following to this main page in the description below.
[0:58:11] And that sums up all the highlights in AI this week.
[0:58:13] Let me know in the comments what you think of all of this.
[0:58:14] Which piece of news was your favorite, and which tool are you most looking forward to trying out?
[0:58:17] As always, I will be on the lookout for the top AI news and tools to share with you.
[0:58:20] So if you enjoyed this video, remember to like, share, subscribe, and stay tuned for more content.
[0:58:22] Also, there's just so much happening in the world of AI every week, I can't possibly cover everything on my YouTube channel.
[0:58:30] The link to that will be in the description below.
[0:58:31] Thanks for watching, and I'll see you in the next one.
[0:58:50] Elon sat down with Ron Barron and put some important numbers on the table.
[0:58:53] He said they're tracking annual recurring revenue in excess of $100 billion this December, and that it's quite likely that the figure exceeds $500 billion by December next year.
[0:58:59] A trillion.
[0:59:00] It's no longer out of question.
[0:59:00] He also said they're finally confident enough to catch Ship 42 with a power in about a week, that Starlink already has 11,000 satellites up with version 3 just starting, and that the only way to compete with China's power body is to take digital intelligence into space.
[0:59:10] Now, also, there's that Fallcom D.
[0:59:11] And underneath the Polcom hard radio, Starlink Mobile is a brand for both the satellite and the new terrestrial 5G build, one umbrella for a fully integrated cellular company that is global across Earth, across this planet, all the moment, but not just the US.
[0:59:20] So today I want to know how that Starlink mobile evolution changes the story, how the ARR boss sits against the SpaceX figures and whether Ship 42
[1:00:54] Manages to get out of its testing ring.
[1:00:57] You and I agree very strongly that the
[1:01:00] The rate of innovation in this war is faster than it has ever occurred.
[1:01:03] It's faster than the rate of innovation that I have seen in my many decades in tech.
[1:01:07] So imagine a situation where an agent bot manages to get out of its testing ring by accident, but ultimately triggers a nuclear response because it signals that there's one coming and then the nuclear response happens automatically.
[1:01:20] Well, first place, we're not going to do a pause because the entire incentive system in America is not to pause.
[1:01:26] I mean, we need a solution that's implementable for a pause.
[1:01:30] I don't see it.
[1:01:30] So let's imagine I'm an evil dictator and I wake up one day and I have a perfect AI system and I say, attack this country, tell me when you're done.
[1:01:38] Right?
[1:01:38] That's my command.
[1:01:39] Notice I didn't say anything else.
[1:01:40] The system would be free to do whatever he thought was appropriate, which could include massive war crimes.
[1:01:46] That is Eric's.
[1:01:48] He ran Google for a decade, then chaired the Pentagon's Defense Innovation Board and America's National Security Commission on AI.
[1:01:55] Today, his own drone company builds interceptors that, by its count, have shot down more than four.
[1:02:00] 4,000 Russian drones over Ukraine.
[1:02:02] The economist's editor-in-chief sat down with him in Kyiv.
[1:02:05] Here is his full warning.
[1:02:06] Let's just dive down into some of those specifically.
[1:02:09] So jet shaheeds, you know, the Russians took the Shahed drone, Iranian drone that they now make in huge numbers, and added jet propulsion, which as I understand it means they go faster and higher and much harder to intercept.
[1:02:21] How close is Ukraine to a solution to that?
[1:02:24] There are at least five companies in Ukraine that have come up with solutions that are in test now.
[1:02:30] It should be soon.
[1:02:31] But this was a major event in the war.
[1:02:33] And I didn't fully foresee it, which I hold myself responsible.
[1:02:36] We knew that the Russians were going to take the Shaheds, make the back stronger, add these small jet engines, which you can get from China.
[1:02:44] We thought they were going to use about ten percent of the supply, and we thought that that was overcomable.
[1:02:49] Russia accelerated the use of these to more than half of the Shaheds, but they did one other thing.
[1:02:55] They don't have access to the equivalent of Starlink for Russian use in Ukraine.
[1:02:59] Thank goodness.
[1:03:00] So they built a mesh network out of Chinese parts, and they typically have two Shaheds that are sitting in a going around in a circle between Belarusia and Kyiv, for example.
[1:03:10] And those are relays.
[1:03:11] And those relays allow them to see literally 8K video of whatever their target is.
[1:03:17] So right next door, they manage to go right through the central, central window of the SBU headquarters, which is like the CIA.
[1:03:23] And that's a good warning of the power of these things.
[1:03:26] There are reports that there are plans to use up to 80,000 such.
[1:03:30] Jetshaheads against Kyiv and the other towns between now and the end of December.
[1:03:34] This is a major point.
[1:03:35] the government here is extremely focused on this.
[1:03:38] I know of a number of efforts that are likely to work.
[1:03:41] You mentioned that the Russians, and this would be even worse, are close to or potentially close to building their version of Starlink, Rasviet.
[1:03:47] How close are they?
[1:03:48] 'Cause it seems to me, and I'm no technician on this, but it seems that that would be a real game changer.
[1:03:53] there's evidence that they are have started launching, there's evidence that it doesn't work yet.
[1:03:58] My own opinion is it's probably
[1:04:00] A year away, because these networks are hard.
[1:04:02] Once that network is in place, it is more or less unblockable, which means the Russians will be able to again hunt anywhere inside of Ukraine.
[1:04:11] Ukraine today does not have the ability to use Starlink to hunt inside America.
[1:04:16] So you can imagine a situation both countries have that capability, the level of damage that they could do to each other.
[1:04:22] In the case of Ukraine, one of their highest priorities is hitting the oil derives and so forth and so on.
[1:04:27] They do that without really being able to see what they're doing.
[1:04:30] Imagine how much more the Ukrainians could be against Russia.
[1:04:33] If this actually from the U.
[1:04:35] S.
[1:04:35] This begs a question, which many of the Ukrainian commanders I've spoken to mention why can't we use Starlink in Russian territory?
[1:04:43] And the it's usually come up in the context of ballistic missiles because the other you mentioned it, the other thing that is terrorizing Ukraine is Russian launches of ballistic missiles, which are dramatically more powerful, huge damage terrorising the cities.
[1:04:55] And they say if we can't hit the arrow, why can't we go after the archer?
[1:04:59] And why
[1:05:00] Can't someone persuade Elon Musk to let us for that purpose use Starlink in Russia.
[1:05:04] Do you think that would be a good idea?
[1:05:06] I'm in favor of giving Ukraine the tools to allow them to stand the war by force.
[1:05:11] Literally to stop the Russian aggression.
[1:05:14] The question you're asking is really a political decision.
[1:05:16] It's really a government decision.
[1:05:17] And the US government needs to grapple with this question.
[1:05:20] Put a number on that.
[1:05:21] On March 24th this year, Russia fired 948 drones at Ukraine in 24 hours.
[1:05:27] The biggest attack of the war, according to Ukraine's.
[1:05:30] Air Force.
[1:05:30] The day before, a Russian company called Bureau 1440 launched the first 16 satellites of Rasviet, and 16 more went up in July.
[1:05:38] By August, fewer than four in 10 had reached their working orbit, and partial service is expected in 2027.
[1:05:44] So the year he's guessing at is already on somebody's launch schedule.
[1:05:48] Now listen to this.
[1:05:49] Quick thing before we go on.
[1:05:50] Most of the people who watch this channel aren't subscribed.
[1:05:53] If these videos are worth your time, take a second to hit subscribe.
[1:05:56] It's free, it helps us keep making them, and you can change your mind whenever you like.
[1:06:00] Right, back to Kyiv.
[1:06:01] Let's talk about what's coming next.
[1:06:02] I mean that's all pretty, pretty alarming, and that's these are the challenges Ukraine faces now.
[1:06:07] But coming down the road is even greater technical innovation, and particularly AI.
[1:06:13] And you know, in most contexts, you and I have conversations about what is happening in Silicon Valley, not what is happening in Ukraine.
[1:06:19] And I wanted to sort of stand back because I think there's a bit of misunderstanding about what's going on with AI and this war right now.
[1:06:26] People talk about is AI being used, but as I understand it, what's really
[1:06:30] being used is quite well known technology of machine vision for final targeting.
[1:06:34] Could you explain that?
[1:06:35] Because I think that's an important thing to make clear.
[1:06:37] So in the progression of war, you start with person with the soldier with the bullet, with the gun and the bullet and the aim.
[1:06:43] At some point it should be relatively straightforward to have the gun be able to see the subject with a laser or what have you identify and then the human pulls the button and it always hits.
[1:06:53] That analogy scales up in terms of autonomy.
[1:06:56] So it's reasonable to expect that every weapon has a
[1:07:00] final autonomous segment where basically it's close enough, it's near the target, it's going to roughly hit the target, you press a button and it hits it precisely.
[1:07:08] All of these are not AI.
[1:07:10] They're just good computer science, good computer engineering.
[1:07:12] Artificial intelligence is systems that learn.
[1:07:14] And artificial intelligence, learn like we sort of learn, you get smarter and smarter and smarter.
[1:07:20] You're self-improving.
[1:07:21] The technical term is recursive self-improvement.
[1:07:23] And we're just beginning to see the use of AI in war.
[1:07:26] There's a lengthy study on Russian use of a Chinese model
[1:07:30] To do some of their decision making.
[1:07:32] These are all from leaked memos, I don't know what the truth is.
[1:07:34] In Ukraine, there's relatively little use of that.
[1:07:37] There is a project which has been discussed by the MOD to try to help automate the planning of war and planning of where people go.
[1:07:44] Again, it's just at the beginning.
[1:07:46] So speaking as a computer scientist, it's pretty obvious what it should look like.
[1:07:49] Take all the information that you have, all the open source, all the proprietary, all the super secret stuff, feed it into one large computer and say what's going on.
[1:07:58] And have these large language models.
[1:08:00] use tools to do reasoning.
[1:08:01] To me it's fantastic they can think at this level.
[1:08:04] They don't think like humans, by the way.
[1:08:05] They're reasoning around an objective.
[1:08:07] It makes sense to me that eventually you'll say, how can I best position my assets to achieve my objective, such as winning the war, defending the city, so forth.
[1:08:16] But the more interesting questions are why am I losing?
[1:08:19] Why is this group not as good as the other group?
[1:08:21] You want the system to anticipate that.
[1:08:23] And you want the system to say, General, you have a problem.
[1:08:26] Here is our recommended solution.
[1:08:28] Can we help you?
[1:08:29] That's I think the
[1:08:30] That's where we're going.
[1:08:31] That kind of system already has a track record.
[1:08:33] In April 2024, Plus 972 magazine and local call reported on a tool the Israeli army used in Gaza called Lavender.
[1:08:41] According to the intelligence officers they quoted, it marked some 37,000 Palestinians as suspected militants, and one officer said he spent about 20 seconds on each target.
[1:08:51] They said roughly one in ten of the people it marked were known in advance not to be in Hamas's military wing.
[1:08:57] Israel's military disputed the account and described
[1:09:00] As a database for cross-referencing intelligence.
[1:09:02] Either way, there was a human in that loop, and 20 seconds was how long the loop lasted.
[1:09:07] Watch what he says next.
[1:09:08] So let's talk about what the consequences will be when you have real AI, when you have this ability to reason, to infer, to learn that we talk about constantly in Silicon Valley coming here.
[1:09:19] Firstly, what does it mean for how autonomous warfare becomes?
[1:09:24] Because now you can have this last mile guidance, but then presumably you can have the AI not only giving the general
[1:09:30] lessons but providing a battle plan and and down the road you know executing them autonomously.
[1:09:34] So let's imagine I'm an evil dictator and I wake up one day and I have a perfect AI system and I say, attack this country, tell me when you're done.
[1:09:42] Right, that's my command.
[1:09:43] Notice I didn't say anything else.
[1:09:45] The system would be then free to do whatever it thought was appropriate, which could include massive war crimes, right?
[1:09:51] All sorts of terrible things.
[1:09:52] It could destroy civilians, it could destroy the infrastructure, it might attack another country that was uninvolved in order to annoy the first country.
