WEBVTT

00:00:02.080 --> 00:00:04.240
Maybe it's just win five on open five.

00:01:06.880 --> 00:01:10.000
You have interfered with our affairs for the last time.

00:01:36.080 --> 00:01:39.200
Oh here we go again They take it

00:02:56.560 --> 00:02:59.680
You have been compared with our bear for the last time.

00:08:48.880 --> 00:08:49.840
Viper sucker.

00:08:50.640 --> 00:08:53.040
Can you turn this not be quite so viper?

00:11:49.520 --> 00:11:51.920
Medical one six six zero or double hardware.

00:14:58.160 --> 00:14:59.600
Should I turn him on?

00:15:01.680 --> 00:15:03.520
Yeah, man's air doing a thing now.

00:15:24.160 --> 00:15:25.200
All on, all on, all on.

00:15:56.800 --> 00:15:58.480
Yeah, I'll eat anything now.

00:16:04.160 --> 00:16:05.440
All on all on all on.

00:17:16.160 --> 00:17:17.280
Alright, alright, alright, alright.

00:17:20.080 --> 00:17:22.160
Google releases their latest video model, Gemini Omni 1.

00:17:22.320 --> 00:17:22.400
2.

00:17:22.480 --> 00:17:25.360
OpenAI drops their latest and best image model, and this one can even make animated gifs.

00:17:25.440 --> 00:17:27.680
NVIDIA open sources a very powerful AI for weather simulation.

00:17:27.760 --> 00:17:29.680
We have some new research that shows how we can use AI to improve other

00:17:30.240 --> 00:17:30.960
Models without further training.

00:17:31.040 --> 00:17:33.200
This could potentially open up an entirely new way to improve AI models.

00:17:33.280 --> 00:17:35.280
This open source AI can recreate any hairstyle in 3D.

00:17:35.360 --> 00:17:37.280
We also have a very powerful open model for upscaling videos.

00:17:37.360 --> 00:17:39.440
The best open model 3K three point one is finally released this week.

00:17:39.520 --> 00:17:41.680
Kansen also drops their best model I 4, which is very impressive.

00:17:41.760 --> 00:17:44.240
We have a ton of 3D model generators and a lot more, so let's jump right in.

00:17:44.400 --> 00:17:48.240
First up, if you ever need to generate 3D hair for whatever reason, this week we have a really cool AI called hair deck.

00:17:48.320 --> 00:17:51.200
You just need to input a single image of a character, and it'll generate their complete hairstyle in 3D.

00:17:51.280 --> 00:17:52.160
Here are some examples for your reference.

00:17:52.240 --> 00:17:54.400
As you can see, it can reconstruct different hairstyles from a ton of different characters.

00:17:54.480 --> 00:17:59.840
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

00:18:00.160 --> 00:18:02.240
Generates the joints of the hair so that it can move naturally in animations.

00:18:02.320 --> 00:18:04.400
So when a character turns their head, you can see the hair moving very naturally.

00:18:04.480 --> 00:18:06.720
It can also preserve their hair colors, including hairstyles with multiple colors.

00:18:06.800 --> 00:18:11.360
And if you look at these benchmarks, hair deck significantly outperforms other hair generating approaches in both geometry accuracy and generation speed.

00:18:11.440 --> 00:18:13.040
The awesome thing is they've released the code and modeled to this.

00:18:13.120 --> 00:18:17.040
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.

00:18:17.120 --> 00:18:19.200
Not only does this work with Nvidia GPUs, but also Apple Silicon.

00:18:19.280 --> 00:18:21.760
And if you check out the Hugging Face folder, note that the total size is only around 2.

00:18:21.920 --> 00:18:22.640
45 gigabytes.

00:18:22.720 --> 00:18:24.400
So this should be able to run on most consumer GPUs.

00:18:24.480 --> 00:18:27.200
And on their GitHub, they've even released the training scripts to this, so you can customize it further.

00:18:27.280 --> 00:18:29.360
Note that this is under the Apache 2 license, which is very permissive.

00:18:29.440 --> 00:18:29.840
You can even use this.

00:18:30.160 --> 00:18:30.640
commercial purposes.

00:18:30.720 --> 00:18:32.880
If you're interested in trying this out, I'll link to this main page in the description below.

00:18:32.960 --> 00:18:35.200
Also this week we have a new AI video upscaler called Relay VSR.

00:18:35.280 --> 00:18:39.280
This can take any low resolution video and upscale it to four times its original resolution, with sharper details and better visual quality.

00:18:39.360 --> 00:18:40.320
Here are some examples for your reference.

00:18:40.400 --> 00:18:42.960
On the left is the original low resolution video and on the right is the upscaled results.

00:18:43.040 --> 00:18:45.520
As you can see it can recover or add much finer details across different types of videos.

00:18:45.600 --> 00:18:46.640
Now what's interesting is how this works.

00:18:46.720 --> 00:18:51.040
Instead of using a massive generative model to process every single frame it only uses the large model to enhance certain keyframes.

00:18:51.120 --> 00:18:54.000
Then a much smaller model uses these keyframes as references to enhance all the frames in between.

00:18:54.160 --> 00:18:55.440
This makes the process much faster and more efficient.

00:18:55.600 --> 00:18:59.520
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.

00:19:00.160 --> 00:19:00.960
thing is they've released this already.

00:19:01.040 --> 00:19:05.440
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.

00:19:05.600 --> 00:19:07.920
Note that this is based on Alibaba's one and the total size is around 10.

00:19:08.080 --> 00:19:08.560
2 gigabytes.

00:19:08.640 --> 00:19:10.560
If you're interested I'll link to this main page in the description below.

00:19:10.720 --> 00:19:15.760
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.

00:19:15.840 --> 00:19:18.080
Now this is based on Alibaba's Z image which is already really good.

00:19:18.160 --> 00:19:22.080
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.

00:19:22.240 --> 00:19:27.920
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.

00:19:28.000 --> 00:19:29.120
Now how this works is quite interesting.

00:19:29.200 --> 00:19:29.840
It uses a special training

00:19:30.080 --> 00:19:32.320
method that analyzes both the generated image and how well it matches the text props.

00:19:32.400 --> 00:19:35.920
This helps the model preserve important details like objects, styles and typography even with very few generation steps.

00:19:36.000 --> 00:19:42.560
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.

00:19:42.640 --> 00:19:46.800
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.

00:19:46.880 --> 00:19:48.640
So this should be able to fit on most consumer GPUs.

00:19:48.720 --> 00:19:50.960
If you're interested in trying this out I'll link to this main page in the description below.

00:19:51.040 --> 00:19:55.120
Also this week NVIO releases a new AI called Strata and this is designed to simulate the atmosphere across the entire planet.

00:19:55.200 --> 00:19:59.840
So this takes information about the current atmospheric conditions including temperature, wind, humidity and pressure and predicts how these conditions will change over time.

00:20:00.000 --> 00:20:01.440
Now, what's really impressive is the level of detail.

00:20:01.520 --> 00:20:04.640
It can simulate the atmosphere across the entire Earth at a resolution of approximately five kilometers.

00:20:04.720 --> 00:20:05.600
So it's incredibly detailed.

00:20:05.680 --> 00:20:08.560
This means it can model weather patterns and atmospheric processes much better than conventional methods.

00:20:08.640 --> 00:20:13.360
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.

00:20:13.520 --> 00:20:14.480
Now, how this works is quite interesting.

00:20:14.560 --> 00:20:16.880
It uses a transformer architecture that divides the planet into smaller sections.

00:20:16.960 --> 00:20:19.920
Each section gets processed separately, and the results are then combined to simulate the global atmosphere.

00:20:20.000 --> 00:20:24.480
And despite the enormous scale of the simulation, the model itself has only about 332 million parameters, not even a billion.

00:20:24.560 --> 00:20:27.840
The awesome thing is NVIDIA has released the model on Hugging Face, and note that the total size is only around 1.

00:20:28.000 --> 00:20:28.560
6 gigabytes.

00:20:28.640 --> 00:20:29.360
So it's surprisingly small.

00:20:29.440 --> 00:20:29.840
Plus, they also

00:20:30.320 --> 00:20:30.960
The training recipes for this.

00:20:31.040 --> 00:20:33.200
And note that it's under the Apache 2 license, which is very permissive.

00:20:33.280 --> 00:20:34.480
You can even use this for commercial purposes.

00:20:34.560 --> 00:20:38.960
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.

00:20:39.040 --> 00:20:40.560
If you're interested, I'll link to this in the description below.

00:20:40.640 --> 00:20:42.400
Also, this week we have a new AI called Flow HMR.

00:20:42.480 --> 00:20:45.280
This can take a regular video of a person and reconstruct their entire body movement in 3D.

00:20:45.360 --> 00:20:49.200
The cool thing is, unlike traditional motion capture systems, this doesn't require any special cameras, motion capture suits, or sensors.

00:20:49.280 --> 00:20:50.080
It just takes in a regular video.

00:20:50.160 --> 00:20:50.960
Here are some examples in action.

00:20:51.040 --> 00:20:54.000
As you can see, it can accurately capture all sorts of movements like kicking a soccer ball or doing push-ups.

00:20:54.080 --> 00:20:56.000
What's especially interesting is that it also considers the laws of physics.

00:20:56.080 --> 00:20:59.360
So instead of just approximating the person's movements, it tries to reconstruct motions that are physically realistic.

00:20:59.440 --> 00:20:59.840
Now, this has some

00:21:00.080 --> 00:21:00.640
Pretty cool applications.

00:21:00.720 --> 00:21:05.200
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.

00:21:05.280 --> 00:21:09.920
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.

00:21:10.000 --> 00:21:13.840
As you can see from these benchmarks, FlowHMR outperforms other similar methods in terms of physical realism and motion accuracy.

00:21:13.920 --> 00:21:19.120
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.

00:21:19.200 --> 00:21:21.200
Note that this Flow HMR latest is the one released this week.

00:21:21.280 --> 00:21:24.640
This is the one that incorporates the laws of physics into its generation so that the movements are more accurate and realistic.

00:21:24.720 --> 00:21:27.040
There's also an online hugging face demo so you can try it out directly in your browser.

00:21:27.120 --> 00:21:28.800
If you're interested, I'll link to this main page in the description below.

00:21:28.880 --> 00:21:30.000
Also, this week we have a new AI called

00:21:30.080 --> 00:21:30.800
Relift from NVIDIA.

00:21:30.960 --> 00:21:34.880
This takes the regular video and tracks the movement of individual points in 3D space, even when objects disappear behind other objects.

00:21:34.960 --> 00:21:36.000
So here's a demo of it in action.

00:21:36.080 --> 00:21:40.080
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.

00:21:40.160 --> 00:21:41.600
Here are some additional examples for your reference.

00:21:41.680 --> 00:21:43.440
It works with a ton of different environments, motions, and characters.

00:21:43.520 --> 00:21:44.160
Now here's how it works.

00:21:44.240 --> 00:21:46.880
It first uses existing AI models to track points in the video and estimate their depth.

00:21:46.960 --> 00:21:49.920
Then it uses another AI model to correct any errors and predict how these points move when they're hidden.

00:21:50.000 --> 00:21:54.640
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.

00:21:54.720 --> 00:21:56.400
The awesome thing is they've released the code and model weights to this.

00:21:56.480 --> 00:21:59.760
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.

00:22:00.160 --> 00:22:02.560
They've also released the training scripts, so you can train or customize the model yourself.

00:22:02.640 --> 00:22:04.560
If you're interested, I'll link to this main page in the description below.

00:22:04.640 --> 00:22:07.040
Also, this week, Google drops their latest video model, Gemini on the 1.

00:22:07.120 --> 00:22:07.440
2 flash.

00:22:07.520 --> 00:22:09.680
Now, similar to the previous version, this can generate videos up to 15 seconds long.

00:22:09.760 --> 00:22:12.400
And not only does it take in text prompts, but you can also use images, video, or audio as input.

00:22:12.480 --> 00:22:13.360
Here are some examples for your reference.

00:22:13.440 --> 00:22:17.520
As you can see, it can generate cinematic scenes with smooth camera movements, realistic motion, and much better consistency throughout the video.

00:22:17.600 --> 00:22:18.960
One major improvement is reference consistency.

00:22:19.040 --> 00:22:24.320
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.

00:22:24.400 --> 00:22:26.000
It also supports video editing just with natural language.

00:22:26.160 --> 00:22:29.280
So you can simply prompt it to change the background, replace an object, or adjust the lighting of an existing video.

00:22:29.360 --> 00:22:29.920
It can even take motion.

00:22:30.080 --> 00:22:31.600
From one video and apply it to a completely different subject.

00:22:31.680 --> 00:22:35.280
And Google has reduced the price by 30%, so it now costs just 7 cents per second for 720p video.

00:22:35.360 --> 00:22:39.120
Here it says that this new Gemini Ani is available through AI Studio, their API, Google Flow, and various other places.

00:22:39.200 --> 00:22:40.960
If you're interested, I'll link to this main page in the description below.

00:22:41.040 --> 00:22:43.920
Also, this week, OpenAI drops their latest and best image model, GPT Images 2.

00:22:44.000 --> 00:22:44.080
6.

00:22:44.240 --> 00:22:46.480
Now, the previous model was already state-of-the-art, but this one is slightly better.

00:22:46.560 --> 00:22:47.920
It has all the capabilities of the previous version.

00:22:48.000 --> 00:22:50.720
For example, it supports image editing with natural language and reference images as input.

00:22:50.800 --> 00:22:52.720
But a really cool new feature is that it can now create animated GIFs.

00:22:52.800 --> 00:22:53.920
So here are some examples for reference.

00:22:54.000 --> 00:22:55.040
Another major improvement is speed.

00:22:55.120 --> 00:22:59.840
So OpenAI says image generation is now 50% faster than before, and there are actually two different versions available through their API.

00:23:00.000 --> 00:23:01.520
The first is Flare, which is optimized for speed.

00:23:01.600 --> 00:23:03.120
This can generate images in as little as six seconds.

00:23:03.200 --> 00:23:06.960
The second is Sunburst, which prioritizes higher quality images and more consistent details across multiple edits.

00:23:07.040 --> 00:23:12.160
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.

00:23:12.320 --> 00:23:13.440
The nice thing is this is available already.

00:23:13.600 --> 00:23:15.200
At the bottom of the page here, it says that images 2.

00:23:15.360 --> 00:23:17.760
6 is available in ChatGPT and codecs for all subscription tiers.

00:23:17.840 --> 00:23:20.640
Also, this week we have some fascinating research from MIT called weightspace programming.

00:23:20.720 --> 00:23:24.080
Now, to understand this, you first need to know that most modern AI models are basically made of neural networks like this.

00:23:24.240 --> 00:23:27.600
This is basically a huge network of dials and knobs that controls how much data flows through to the next layer.

00:23:27.680 --> 00:23:29.040
These dials and knobs are called the model weights.

00:23:29.120 --> 00:23:29.840
Now, what makes these weights so

00:23:30.160 --> 00:23:32.480
Important is that they essentially encode everything the model has learned during training.

00:23:32.560 --> 00:23:38.160
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.

00:23:38.240 --> 00:23:42.080
So in a sense, these weights kinda represent the model's learned knowledge and abilities, or what we might call its intelligence.

00:23:42.160 --> 00:23:46.400
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.

00:23:46.480 --> 00:23:50.800
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?

00:23:50.880 --> 00:23:52.400
Well, that's exactly what this new research paper is about.

00:23:52.560 --> 00:23:54.960
There are actually a ton of really interesting insights from this, so let's go over the main ones.

00:23:55.040 --> 00:23:58.320
First, they gave GPT6 Astra the weights of four smaller models, each trained to play a different game.

00:23:58.400 --> 00:23:59.760
Without telling it which game they were trained on, Astra

00:24:00.080 --> 00:24:02.160
Correctly identified all four games, including Pong and Pac-Man.

00:24:02.240 --> 00:24:05.840
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.

00:24:05.920 --> 00:24:08.720
In another experiment, they had a car controlled by a small model that kept getting stuck on a ramp.

00:24:08.800 --> 00:24:10.720
So they gave Astra the weights to this model and a recording of the failure.

00:24:10.800 --> 00:24:12.240
Astra then proceeded to change just 3.

00:24:12.400 --> 00:24:15.120
5% of the weights, and the car was then able to successfully complete the course in every test.

00:24:15.200 --> 00:24:18.560
So essentially it figured out which parameters were causing the problem and changed them to fix the car's behavior.

00:24:18.640 --> 00:24:21.200
Now perhaps the most impressive experiment involved Alibaba's Quen 3.

00:24:21.280 --> 00:24:23.040
5, a small model with only 800 million parameters.

00:24:23.200 --> 00:24:24.960
Again, they asked Astra to improve its math capabilities.

00:24:25.040 --> 00:24:29.840
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

00:24:30.160 --> 00:24:31.680
Math 500 benchmark score from just 7.

00:24:31.840 --> 00:24:33.520
9% to 48%, which is crazy.

00:24:33.600 --> 00:24:35.600
Its instruction following performance also improved dramatically.

00:24:35.680 --> 00:24:36.480
Now think about what this means.

00:24:36.640 --> 00:24:43.360
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.

00:24:43.440 --> 00:24:45.760
This could potentially open up an entirely new way to improve AI models.

00:24:45.840 --> 00:24:49.760
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.

00:24:49.840 --> 00:24:50.480
So let's clarify this.

00:24:50.560 --> 00:24:52.160
It turns out the original Quen model had a known problem.

00:24:52.240 --> 00:24:55.840
It would frequently get stuck in repetitive reasoning loops and run out of tokens before it could produce a final answer.

00:24:55.920 --> 00:24:57.920
So Astra didn't necessarily teach it any new math abilities.

00:24:58.000 --> 00:24:59.760
Instead, it just fixed this existing problem, which allowed Quen

00:25:00.080 --> 00:25:02.000
To complete more questions that it was already capable of solving.

00:25:02.160 --> 00:25:08.240
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.

00:25:08.320 --> 00:25:10.800
So the huge improvement from this paper might not be as revolutionary as it seems.

00:25:10.880 --> 00:25:14.000
Another criticism is that Astro wasn't just looking at hundreds of millions of weights and magically figuring everything out.

00:25:14.080 --> 00:25:17.600
For the coin experiment, it actually had access to Python tools and could repeatedly run the model on practice questions.

00:25:17.680 --> 00:25:20.000
It used these results to identify the problem and test different modifications.

00:25:20.160 --> 00:25:24.960
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.

00:25:25.040 --> 00:25:29.920
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.

00:25:30.080 --> 00:25:31.600
Astra couldn't actually improve its performance further.

00:25:31.680 --> 00:25:35.360
So while this research is impressive, it doesn't mean AI can now magically understand and upgrade the intelligence of any model.

00:25:35.440 --> 00:25:39.440
But still, this paper is quite interesting, and this could lead to new ways to use AI to debug or improve other AI models.

00:25:39.520 --> 00:25:41.520
And the nice thing is they've released the experiment logs on GitHub.

00:25:41.600 --> 00:25:43.040
So you can inspect the results and try to model yourself.

00:25:43.120 --> 00:25:45.280
If you're interested, I'll links to this main page in the description below.

00:25:45.440 --> 00:25:48.160
If you want to create viral videos at scale, definitely check out Higgsfield, the sponsor of this video.

00:25:48.240 --> 00:25:51.680
They just launched Higgsfield Content Factory, which is kind of like an entire video production team right at your fingertips.

00:25:51.760 --> 00:25:56.160
You can get AI to automatically write scripts, generate visuals, add voiceovers, and edit everything into a complete video, all from a single idea.

00:25:56.240 --> 00:25:58.640
I'm especially impressed by how easy it is to create videos based on popular trends.

00:25:58.720 --> 00:25:59.760
You can choose from over fifty ready-made

00:26:00.000 --> 00:26:02.320
Presets or simply paste a link to a viral TikTok or Instagram video as a reference.

00:26:02.400 --> 00:26:06.000
Higgsfield will automatically follow its style, structure, and pacing, but create a completely new video based on your topic.

00:26:06.080 --> 00:26:09.440
As you can see, it generates an entire story complete with characters, visuals, narration, and scene transitions.

00:26:09.520 --> 00:26:12.800
You can create videos anywhere from 15 seconds to five minutes long, and even generate up to four episodes at once.

00:26:12.880 --> 00:26:15.440
Plus, you can access Content Factory directly inside Claude through Higgsfield MCP.

00:26:15.520 --> 00:26:20.400
Whether you want to create YouTube shorts, TikToks, Instagram reels, educational videos, or marketing content, Higgsfield Content Factory makes the entire process incredibly easy.

00:26:20.480 --> 00:26:22.080
Try Higgsfield today using the link in the description below.

00:26:22.160 --> 00:26:23.760
Also, this week we have a new AI called Lucid Lira.

00:26:23.920 --> 00:26:26.000
This can turn a single image into an explorable 3D scene.

00:26:26.080 --> 00:26:29.920
But what's especially interesting is that it can also recognize individual objects within that scene, so you can select, edit, or extract them.

00:26:30.080 --> 00:26:30.320
Afterwards.

00:26:30.400 --> 00:26:31.440
Here are some examples for your reference.

00:26:31.520 --> 00:26:35.440
As you can see, it can generate detailed 3D environments from regular images and even identify individual objects within it.

00:26:35.520 --> 00:26:37.360
And this works across a variety of different environments and objects.

00:26:37.440 --> 00:26:40.640
For example, you can select an object and move it around, resize it, duplicate it, or remove it entirely.

00:26:40.720 --> 00:26:45.120
It can also convert these scenes into environments for robots to navigate and interact with, which could be useful for robotic simulations.

00:26:45.200 --> 00:26:46.240
Now, how this works is quite interesting.

00:26:46.320 --> 00:26:49.520
It first uses a video generation model to predict what the scene would look like from different camera angles.

00:26:49.600 --> 00:26:52.720
Then another AI model combines this information into a 3D scene and identifies the objects within it.

00:26:52.800 --> 00:26:57.040
And if you look at the comparisons, this outperforms other similar approaches in terms of recognizing and separating objects in 3D scenes.

00:26:57.120 --> 00:26:59.360
Now at the top of the page it says the code is coming soon, so stay tuned for that.

00:26:59.440 --> 00:26:59.920
For now, if you

00:27:00.160 --> 00:27:02.160
interested in reading further I'll link to this main page in the description below.

00:27:02.240 --> 00:27:04.160
Also this week we have a new image model called PH3 image.

00:27:04.240 --> 00:27:08.000
The interesting part is they took the best open video model, Minimax H3, and they basically turned it into an image generator and editor.

00:27:08.080 --> 00:27:08.880
And it turns out it's really good.

00:27:08.960 --> 00:27:14.240
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.

00:27:14.320 --> 00:27:18.080
As you can see it's especially good at photorealistic images, cinematic scenes, portraits, anime and different artistic styles.

00:27:18.160 --> 00:27:21.360
The developers basically modified the architecture of Minimax to specialize in generating images instead of video.

00:27:21.440 --> 00:27:23.680
They also reduced the number of generation steps from 50 to just eight.

00:27:23.760 --> 00:27:25.280
So it's about five times faster than the original model.

00:27:25.360 --> 00:27:28.080
It can generate images up to 16 megapixels and they've also added an AI prompt enhancer.

00:27:28.160 --> 00:27:30.000
So even if you enter a really simple prompt it can automatically expand it

00:27:30.080 --> 00:27:30.720
Into something more detailed.

00:27:30.800 --> 00:27:34.160
The awesome thing is they've released the model weights on Hugging Face, and you can run it with either diffusers or Comp UI.

00:27:34.240 --> 00:27:38.000
There are multiple versions, including the original BF16, 8-bit, and 4-bit versions, so you can download the one that fits your hardware.

00:27:38.080 --> 00:27:40.320
If you're interested in reading further, I'll link to this main page in the description below.

00:27:40.400 --> 00:27:42.880
Also, this week we have a new open source robot toolkit called Open RoboKit.

00:27:43.040 --> 00:27:46.320
This provides a collection of tools to help you control robots, simulate their movements, and train them on different tasks.

00:27:46.400 --> 00:27:47.360
Now there are four main libraries here.

00:27:47.440 --> 00:27:49.760
First there's RoboKit, which helps robots calculate their movements and plan their actions.

00:27:49.840 --> 00:27:53.600
For example, here you can see robots spinning pens, rotating cubes, opening notebooks, climbing obstacles, and even performing jumps.

00:27:53.680 --> 00:27:59.280
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.

00:27:59.360 --> 00:27:59.840
Then there's grasp.

00:28:00.160 --> 00:28:02.000
Which helps robots figure out how to pick up and manipulate different objects.

00:28:02.080 --> 00:28:04.800
It can also train robotic policies and deploy them in simulations for real-world robots.

00:28:04.880 --> 00:28:08.400
And finally, there's RoboSkills, which provides different workloads for AI agents to perform various robotics tasks.

00:28:08.480 --> 00:28:13.120
For example, you can get an AI agents to generate robot movements, reconstruct simulations from videos, or autonomously improve its robot control through experiments.

00:28:13.280 --> 00:28:15.120
The nice thing is they've released the code to all four toolkits.

00:28:15.200 --> 00:28:18.960
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.

00:28:19.040 --> 00:28:21.040
This is a fantastic resource if you're doing robotics, so definitely check it out.

00:28:21.120 --> 00:28:22.480
I'll link to this main page in the description below.

00:28:22.560 --> 00:28:25.360
Also, this week we have a really cool real-time multimodal voice called Moss Live.

00:28:25.440 --> 00:28:27.120
And yes, this is the team behind the awesome Moss image as well.

00:28:27.200 --> 00:28:30.000
The thing is not only is this a real-time voice you can have conversations with, but it's multimodal, so you can

00:28:35.360 --> 00:28:38.480
Honestly, it's my first time here, so I guess you're my tour guide, just planning a nice afternoon.

00:28:38.880 --> 00:28:40.880
Alright, let's head east from Citroen Road to the Boom.

00:28:40.960 --> 00:28:44.000
Then stroll north along the waterfront toward an ancient road, and join the views and the historic buildings.

00:28:44.080 --> 00:28:45.760
We'll find somewhere nearby for dinner before you head to the airport.

00:28:47.600 --> 00:28:52.240
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.

00:28:52.320 --> 00:28:56.000
It can also distinguish your voice from other people or background voices so it doesn't accidentally respond to unrelated stuff.

00:28:56.080 --> 00:28:58.800
It can participate in a meeting and even help you look up information and provide answers at the right times.

00:28:58.880 --> 00:28:59.760
Or here's another example where you

00:29:00.160 --> 00:29:02.320
Use this for travel, you can show your camera to Moss and ask it questions about the blitz.

00:29:02.400 --> 00:29:03.360
Now, how this works is quite interesting.

00:29:03.440 --> 00:29:05.120
The system uses separate foreground and background components.

00:29:05.200 --> 00:29:08.160
The foreground handles real time conversations, while the background takes care of complex reasoning and tool calls.

00:29:08.240 --> 00:29:10.080
This means it can keep talking to you even while it works on another task.

00:29:10.160 --> 00:29:14.480
And if you look at these benchmarks for real time voice conversations, interruptions, and tool use, it performs very well against other leading models.

00:29:14.560 --> 00:29:16.800
You can currently try it through their online demo or access it through their API.

00:29:16.880 --> 00:29:20.480
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.

00:29:20.640 --> 00:29:23.040
Also, this week we have a new open source image generator called Pyramid JIT.

00:29:23.200 --> 00:29:26.640
Now, what's really impressive is that they trained this model from scratch with just $21,000 in compute costs.

00:29:26.720 --> 00:29:27.600
Here are some examples for your reference.

00:29:27.680 --> 00:29:29.920
As you can see, it can generate a ton of different images like realistic portraits, landscape.

00:29:30.240 --> 00:29:31.040
And even stylized illustrations.

00:29:31.120 --> 00:29:35.280
It's a relatively small model with about two billion parameters, but the image quality is surprisingly good for its size and training budget.

00:29:35.360 --> 00:29:36.400
Now, the design of this is quite interesting.

00:29:36.480 --> 00:29:41.280
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.

00:29:41.360 --> 00:29:44.080
It also uses a special architecture that reduces the amount of computation needed during training.

00:29:44.160 --> 00:29:47.840
But they say that this is just an experimental model and it still struggles with complicated scenes, multiple people, and accurate text generation.

00:29:47.920 --> 00:29:49.520
The awesome thing is they've released the model and code choos.

00:29:49.680 --> 00:29:54.000
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.

00:29:54.080 --> 00:29:56.240
And then on their hugging face, note that the total size of this model is 8.

00:29:56.400 --> 00:29:58.880
74 gigabytes, so this should be able to run on most consumer GPUs.

00:29:58.960 --> 00:29:59.840
Also note that the model is released on the

00:30:00.000 --> 00:30:01.360
Under the Apache 2 license, which is very permissive.

00:30:01.440 --> 00:30:02.560
You can even use this for commercial purposes.

00:30:02.640 --> 00:30:04.320
If you're interested, I'll link to this page in the description below.

00:30:04.400 --> 00:30:06.480
Also, this week we have a new 3D generator called Cola 3D.

00:30:06.560 --> 00:30:10.640
This can turn a single image into an entire 3D scene, and it can separate each object into its own individual 3D mesh.

00:30:10.720 --> 00:30:11.680
Here are some examples for reference.

00:30:11.760 --> 00:30:12.800
Now, this supports three different modes.

00:30:12.880 --> 00:30:17.120
You can let the AI automatically identify and generate all the objects, or you can specify individual objects using points or segmentation masks.

00:30:17.280 --> 00:30:21.440
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.

00:30:21.520 --> 00:30:24.160
This helps preserve the correct positions, sizes, and spatial relationships between objects.

00:30:24.240 --> 00:30:29.280
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.

00:30:29.360 --> 00:30:29.840
In addition, they've

00:30:30.240 --> 00:30:32.000
The training code and data set to this as well, which is fantastic.

00:30:32.080 --> 00:30:34.000
If you're interested, I'll link to this main page in the description below.

00:30:34.080 --> 00:30:36.240
Also, this week we have a super tiny open vision model called Mirage.

00:30:36.400 --> 00:30:38.560
This can analyze any image and identify objects and read text in the image.

00:30:38.640 --> 00:30:42.160
And what's really impressive is that the entire model is only about 22 megabytes, not even a gigabyte in size.

00:30:42.240 --> 00:30:42.800
So here are some demos.

00:30:42.880 --> 00:30:46.400
You can upload any image or use your camera, and the AI will automatically identify different objects and outlines around it.

00:30:46.480 --> 00:30:50.800
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.

00:30:50.880 --> 00:30:54.320
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.

00:30:54.400 --> 00:30:56.240
So you could use this to classify objects or count items.

00:30:56.320 --> 00:30:57.280
Now, how this works is quite interesting.

00:30:57.360 --> 00:30:59.760
It basically uses one shared vision encoder to understand the image and then

00:31:00.080 --> 00:31:02.320
Several specialized components extract the objects, text, and other information.

00:31:02.400 --> 00:31:04.880
This allows it to handle multiple vision tasks with just one tiny model.

00:31:04.960 --> 00:31:08.880
Now, the model has only like 51 million parameters, and with two bit quantization, they've compressed it down to just 22 megabytes.

00:31:08.960 --> 00:31:11.040
So this can process an image in like a tenth of a second on a smartphone.

00:31:11.120 --> 00:31:13.600
And of course you can also run this locally on your laptops, wearables, or other consumer devices.

00:31:13.680 --> 00:31:15.360
Now, the awesome thing is they've released the model and code to this.

00:31:15.440 --> 00:31:18.640
So over here, if you click on the GitHub link, it contains all the instructions on how to run this locally on your computer.

00:31:18.720 --> 00:31:20.560
If you're interested, I'll link to this main page in the description below.

00:31:20.640 --> 00:31:22.240
Also, this week we have a new AI called DreamX.

00:31:22.400 --> 00:31:25.680
This can take the movements from a video of any character and transfer those movements to a completely different character.

00:31:25.760 --> 00:31:26.800
So here are some examples for your reference.

00:31:26.880 --> 00:31:29.840
As you can see, it can take a video of a human hand manipulating objects and use that motion to generate a video.

00:31:30.160 --> 00:31:31.360
Of a robot performing similar actions.

00:31:31.440 --> 00:31:34.160
What's even more interesting is that it can translate those movements into actual robot actions.

00:31:34.240 --> 00:31:36.880
So instead of just generating a video, the robot can physically do this action in real life.

00:31:37.040 --> 00:31:37.920
Now, how this works is quite interesting.

00:31:38.080 --> 00:31:40.880
DreamX uses a video diffusion model to extract motion information from videos.

00:31:40.960 --> 00:31:44.400
It separates the movements from the appearance of the original subject, then applies those movements to another subject.

00:31:44.480 --> 00:31:47.920
And as you can see from the benchmarks, DreamX outperforms some similar methods for motion understanding and transfer.

00:31:48.080 --> 00:31:50.640
Now at the top of their project page, it says the code is coming soon, so stay tuned for that.

00:31:50.720 --> 00:31:52.720
For now, if you're interested, I'll link to this main page in the description below.

00:31:52.800 --> 00:31:55.280
Also, this week we have a new Robotics model called Zero Real Agent Policy, or ZAP.

00:31:55.360 --> 00:31:58.880
This can take camera footage and natural language instructions and then directly control a robot to perform any specified tasks.

00:31:58.960 --> 00:31:59.760
What's especially impressive is

00:32:00.080 --> 00:32:02.400
That it was trained entirely in simulation without any real-world training data.

00:32:02.480 --> 00:32:03.600
Here are some examples for your reference.

00:32:03.680 --> 00:32:07.920
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.

00:32:08.000 --> 00:32:10.240
You can even get it to hold this pen and draw an X on this tic-tac-toe grid.

00:32:10.320 --> 00:32:11.760
And these are real robots, not simulated videos.

00:32:11.920 --> 00:32:12.640
Now here's how they trained us.

00:32:12.720 --> 00:32:17.920
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.

00:32:18.080 --> 00:32:19.920
It initially learned just 40 pretty simple simulated tasks.

00:32:20.000 --> 00:32:23.600
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.

00:32:23.680 --> 00:32:29.840
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.

00:32:30.080 --> 00:32:30.800
And at a fraction of the cost.

00:32:30.960 --> 00:32:32.880
Now, at the top of the page, they've only released a technical paper for now.

00:32:32.960 --> 00:32:34.400
There's no mention whether they will release the model itself.

00:32:34.480 --> 00:32:36.720
But if you're interested in reading further, I'll link to this page in the description below.

00:32:36.800 --> 00:32:38.720
Also, this week, Tencent drops their latest model, Hi4.

00:32:38.880 --> 00:32:40.400
One of the biggest new features is visual understanding.

00:32:40.480 --> 00:32:42.960
The model can now analyze images and solve problems that involve visual reasoning.

00:32:43.040 --> 00:32:47.120
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.

00:32:47.200 --> 00:32:50.960
Like other Frontier models, this has computer use capabilities, so it can interact directly with other software or interfaces.

00:32:51.040 --> 00:32:54.480
Or here's another example where we can get it to create a really detailed 3D environment in FreeJS.

00:32:54.560 --> 00:32:55.360
It's also great at knowledge work.

00:32:55.440 --> 00:32:58.080
So here's an example where we use it to do some financial modeling and equity research.

00:32:58.160 --> 00:32:59.840
Or here's a really cool example where we can get it to reconstruct

00:33:00.000 --> 00:33:02.400
The 3D mouse embryo from 7 million cells in just one and a half hours.

00:33:02.480 --> 00:33:06.720
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.

00:33:06.800 --> 00:33:08.080
However, the model weights haven't been released yet.

00:33:08.160 --> 00:33:10.160
They plan to release the model in the coming days, so stay tuned for that.

00:33:10.240 --> 00:33:14.400
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.

00:33:14.480 --> 00:33:16.560
If you're interested in trying this out, I'll link to this main page in the description below.

00:33:16.720 --> 00:33:18.720
Also, this week we have a new open model called Hydrogen 1.

00:33:18.880 --> 00:33:18.960
0.

00:33:19.040 --> 00:33:22.720
This is an AI model specifically designed for scientific research, including chemistry, material science, and biology.

00:33:22.800 --> 00:33:25.440
Now, this is a massive model with 753 billion total parameters.

00:33:25.520 --> 00:33:26.480
It's based on GLM 5.

00:33:26.560 --> 00:33:29.920
3, but they further trained it on scientific literature, laboratory records, and actual research experiments.

00:33:30.000 --> 00:33:31.680
Here are some examples of the incredible things it can do.

00:33:31.840 --> 00:33:35.360
For example, they got it to write and execute a molecular simulation of polyethylene in a plastic bag.

00:33:35.440 --> 00:33:37.920
The model even identified an error in its code, corrected it, and continued the simulation.

00:33:38.000 --> 00:33:41.760
In another example, it analyzed an NMR spectrum to help identify the chemical structures of possible reaction products.

00:33:41.840 --> 00:33:44.880
It can also design laboratory experiments and generate instructions for robotic lab equipment.

00:33:44.960 --> 00:33:49.920
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.

00:33:50.080 --> 00:33:54.880
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.

00:33:54.960 --> 00:33:59.440
If you look at various benchmarks for scientific reasoning and experiment planning, hydrogen performs exceptionally well compared to other frontier models.

00:33:59.520 --> 00:33:59.760
In some cases,

00:34:00.080 --> 00:34:01.120
Even beating in GPT6 Astra.

00:34:01.200 --> 00:34:04.240
The awesome thing is over here near the bottom of the page, they've released the models for you to download on Hogging Face.

00:34:04.400 --> 00:34:05.520
Note that they released three different versions.

00:34:05.600 --> 00:34:09.120
There's the full model plus an FP4 version for NVIDIA GPUs, as well as an MX version for AMD GPUs.

00:34:09.200 --> 00:34:11.520
Note that even for the FP4 one, it's around 464 gigabytes.

00:34:11.600 --> 00:34:13.280
So you'll need to stack multiple GPUs in order to run this.

00:34:13.360 --> 00:34:15.200
But I mean, this is one of the best open models for scientific research.

00:34:15.280 --> 00:34:16.880
In some cases, even outperforming GPT6 Astra.

00:34:16.960 --> 00:34:19.200
If you're interested in trying this out, I'll link to this main page in the description below.

00:34:19.280 --> 00:34:21.520
Also, this week OpenAI drops their latest model, GPT6 Cyber.

00:34:21.600 --> 00:34:23.920
Now, as the name implies, this is a specialized model for cybersecurity.

00:34:24.000 --> 00:34:28.160
It's designed to autonomously find vulnerabilities in software and help developers fix them, and some of its results are quite impressive.

00:34:28.320 --> 00:34:29.840
For example, OpenAI used this model to identify

00:34:30.160 --> 00:34:34.720
81 previously unknown i severity vulnerabilities in Firefox, even after extensive security testing with previous AI models.

00:34:34.800 --> 00:34:37.520
It also helped develop working exploits for over 900 vulnerabilities in Linux.

00:34:37.600 --> 00:34:38.080
That's pretty crazy.

00:34:38.160 --> 00:34:43.840
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.

00:34:43.920 --> 00:34:45.120
Now, we plebs won't have access to this.

00:34:45.200 --> 00:34:46.640
So here it says they're introducing access

00:34:59.600 --> 00:34:59.600
to the money.

00:35:01.120 --> 00:35:03.520
Hey, I've kinda well here waiting.

00:35:12.000 --> 00:35:13.600
No, that needs to be fixed.

00:35:13.680 --> 00:35:15.760
I don't know that needs to be fixed.

00:35:18.480 --> 00:35:19.120
Hey.

00:35:26.240 --> 00:35:27.040
Still here.

00:35:27.200 --> 00:35:28.480
What's on your mind?

00:35:30.160 --> 00:35:32.160
wait a second.

00:35:32.960 --> 00:35:34.000
I'll tell you.

00:35:36.240 --> 00:35:38.880
how about that?

00:35:39.120 --> 00:35:40.320
okay.

00:35:40.560 --> 00:35:50.640
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.

00:35:53.600 --> 00:35:55.280
Go ahead, I'm listening.

00:35:55.440 --> 00:35:57.120
Describe what you're seeing.

00:35:58.080 --> 00:35:58.400
Okay.

00:36:00.240 --> 00:36:01.920
Yeah, what's happening?

00:36:02.160 --> 00:36:02.560
Yeah.

00:36:03.040 --> 00:36:19.840
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.

00:36:20.160 --> 00:36:22.800
And we have gain and exposure.

00:36:23.360 --> 00:36:29.760
And I accidentally hit default and now the camera is super dark and the gain and the exposure

00:36:30.320 --> 00:36:37.360
Aren't really working on the clickies there, nor are they working on the key commands that are attached.

00:36:37.680 --> 00:36:40.640
So we need to look that specifically.

00:36:41.600 --> 00:36:44.240
Camera defaults nuking the picture.

00:36:45.200 --> 00:36:48.400
Not just that, but the actual whoa.

00:36:51.360 --> 00:36:52.960
Sounds like there's more to it.

00:36:53.120 --> 00:36:54.160
What else is going on?

00:36:54.800 --> 00:36:59.760
I mean yeah, the more to it is is that the gain and the exposure aren't working.

00:37:02.000 --> 00:37:03.840
That's the same thread I'm already on.

00:37:04.080 --> 00:37:07.920
Is it that they move but the image doesn't change or do they not move at all?

00:37:08.320 --> 00:37:18.720
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.

00:37:18.960 --> 00:37:20.800
But I mean it's wonderful garbage.

00:37:20.960 --> 00:37:22.640
But anyway, I'm getting off track.

00:37:22.880 --> 00:37:28.080
The the but the gain and exposure are not adjusting like the they move.

00:37:28.240 --> 00:37:29.760
Oh wait, now they don't move.

00:37:30.160 --> 00:37:50.240
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

00:37:50.240 --> 00:37:59.840
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

00:38:00.160 --> 00:38:04.880
I think I might just set you loose on this one in a second, okay?

00:38:05.440 --> 00:38:06.320
Sure thing.

00:38:06.480 --> 00:38:08.000
I'm ready when you are.

00:38:09.760 --> 00:38:10.400
Alright.

00:38:10.640 --> 00:38:13.200
Yeah, this is a thing that needs to happen.

00:38:13.280 --> 00:38:15.360
So I guess make it happen.

00:38:15.760 --> 00:38:16.720
On it.

00:38:18.320 --> 00:38:23.920
Should I close Actor Core or just you're just gonna do your own thing on your own computer?

00:38:24.000 --> 00:38:25.360
How's this gonna work?

00:38:28.000 --> 00:38:28.960
Leave it open.

00:38:29.120 --> 00:38:29.680
I'm working off.

00:38:30.160 --> 00:38:31.440
the same files on your machine.

00:38:31.520 --> 00:38:35.120
So as long as nothing's locked, I'll read and patch right alongside it.

00:38:35.600 --> 00:38:36.160
Sweet.

00:38:38.720 --> 00:38:40.000
Alright, thank you.

00:38:40.080 --> 00:38:41.200
hit me up.

00:38:41.600 --> 00:38:44.560
Ping me on my wrist when it's ready.

00:38:46.640 --> 00:38:47.440
Will do.

00:38:47.600 --> 00:38:49.200
Wrist ping when it's done.

00:38:55.280 --> 00:38:58.080
Fucking superhero in my computer, dude.

00:38:59.520 --> 00:38:59.840
Alright.

00:39:05.680 --> 00:39:09.200
yeah, this whole video is freaking crazy.

00:39:11.200 --> 00:39:13.600
Hair deck, I need a hair deck.

00:39:14.320 --> 00:39:15.600
Is it ready?

00:39:20.000 --> 00:39:21.040
Well shit.

00:39:43.920 --> 00:39:45.520
Okay, I got the ping.

00:39:45.680 --> 00:39:53.360
I got the ping, so should I just shut it down and and you can edit some code or what's up?

00:39:54.000 --> 00:39:55.840
The ping means the fix is in.

00:39:55.920 --> 00:39:57.040
You can close it whenever.

00:39:57.200 --> 00:39:58.080
Nothing's locked.

00:39:58.320 --> 00:39:59.840
Want me to walk you through what changed?

00:40:00.320 --> 00:40:02.480
Oh you actually the fix is in.

00:40:02.640 --> 00:40:03.040
Excellent.

00:40:03.120 --> 00:40:05.680
Okay, I'll restart it now and give it a test.

00:40:06.320 --> 00:40:07.040
Go for it.

00:40:07.200 --> 00:40:07.360
Yeah.

00:40:07.680 --> 00:40:11.680
Pay special attention to whether gain and exposure actually shift the image this time.

00:40:11.840 --> 00:40:15.360
Yeah, yeah, that's I'm gonna be paying very special attention to that.

00:40:16.480 --> 00:40:17.440
That's the one.

00:40:17.600 --> 00:40:19.680
I'm right here if the numbers still lie.

00:40:33.680 --> 00:40:34.560
We'll see here.

00:40:34.640 --> 00:40:35.440
What we got?

00:40:35.680 --> 00:40:37.120
Some more music.

00:40:39.040 --> 00:40:40.240
It sounds easy.

00:40:43.840 --> 00:40:45.920
Okay, so this quick.

00:40:49.280 --> 00:40:50.080
Okay, well.

00:40:51.280 --> 00:40:52.240
Important pressure.

00:41:05.280 --> 00:41:07.120
Wait a minute, what?

00:41:07.680 --> 00:41:15.520
can we can we patch it and see if it I mean what's the what's the situation?

00:41:53.360 --> 00:41:55.280
Oh, that's confusing.

00:42:03.520 --> 00:42:06.000
I'm gonna have to talk to Chief about this.

00:42:08.080 --> 00:42:14.640
Chief, can you take can you take a look at Core, Actor Core Engineer Bot?

00:42:14.880 --> 00:42:20.160
I'm trying to get I have an issue and Core's not cooperating.

00:42:20.320 --> 00:42:22.320
I'm gonna need your help here, I think.

00:42:22.880 --> 00:42:24.640
Core's giving you trouble?

00:42:25.360 --> 00:42:27.600
Digging into what Core's been doing.

00:42:32.800 --> 00:42:36.080
That's the read-only pass on core's definition in recent records.

00:42:36.240 --> 00:42:38.080
What's it actually refusing to do?

00:42:42.400 --> 00:42:48.720
Well, I wanted to fix After Core and apply it I mean cre pass bug.

00:42:51.920 --> 00:42:57.360
Core Engineer is your After Core engineering bot, and you're after it fixing After Core and getting that bug through.

00:42:57.520 --> 00:42:59.760
My look at its latest thread should show where it's stuck.

00:43:00.480 --> 00:43:00.800
Yeah.

00:43:01.040 --> 00:43:04.080
Now have a look at that and let me know what you think.

00:43:07.520 --> 00:43:09.440
Pulling up Cores thread now.

00:43:10.640 --> 00:43:13.120
Maybe you need to give it a skill, I think.

00:43:14.080 --> 00:43:15.760
A skill could be the miss Yeah.

00:43:16.800 --> 00:43:18.880
That's worth testing once I see what it said.

00:43:19.040 --> 00:43:20.720
What's the skill you had in mind?

00:43:21.120 --> 00:43:22.080
I d I didn't.

00:43:22.160 --> 00:43:23.200
It was just a thought.

00:43:23.520 --> 00:43:29.520
You can go ahead and get back to me when you have some great news for me.

00:43:30.320 --> 00:43:31.040
We'll do

00:43:36.480 --> 00:43:38.880
which only allows approved security researchers to use these advanced models.

00:43:38.960 --> 00:43:40.160
But if you're interested in reading further, I'll

00:43:50.320 --> 00:43:50.960
be able to do that.

00:47:01.600 --> 00:47:07.120
Hey, I'm having a little bit of an issue with Active Bud, I'd like you to help me touch the bug.

00:47:09.920 --> 00:47:10.320
Sure.

00:47:10.960 --> 00:47:12.240
What's the bug doing?

00:47:15.760 --> 00:47:28.880
Well, when I go to the little buddy section in the overview sub tab of the little buddy page in ActRecord.

00:47:30.880 --> 00:47:35.200
The camera controls gain and exposure.

00:47:36.400 --> 00:47:37.440
They change when I click them.

00:47:37.760 --> 00:47:42.000
They change when I push the key commands, but the image remains dark.

00:47:43.440 --> 00:47:45.920
So they're not affecting Cosmos camera.

00:47:53.680 --> 00:47:55.120
Hang on, I'll get in text.

00:47:55.920 --> 00:47:56.240
Sure.

00:48:01.840 --> 00:48:07.120
So yeah, go ahead and it yeah it's it's yeah right there

00:48:08.800 --> 00:48:16.400
camera controls aren't affecting the camera but they appear to change in the software not sure what's going on there

00:52:02.480 --> 00:52:04.000
To this release page in the description below.

00:52:04.160 --> 00:52:07.280
Also, this week, Meta is introducing real-time voice conversations and expressive avatars to views.

00:52:07.360 --> 00:52:09.680
If you haven't heard of Muse, it's basically a personal AI assistant from Meta.

00:52:09.840 --> 00:52:15.040
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.

00:52:15.120 --> 00:52:18.880
For example, you can ask it to help plan your next application, brainstorm ideas, or even rehearse for a job interview.

00:52:18.960 --> 00:52:22.480
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.

00:52:22.560 --> 00:52:26.080
They also released an avatar which can react to you with realistic facial expressions and gestures that match the conversation.

00:52:26.240 --> 00:52:29.280
Now, here it says this is now rolling up in the US through the Muse app on mobile, web and Mac.

00:52:29.440 --> 00:52:29.920
So you can just download.

00:52:30.160 --> 00:52:31.840
Or upgrade your use app to try out this new voice feature.

00:52:31.920 --> 00:52:33.520
Also, this week we have a new AI called Mobius.

00:52:33.680 --> 00:52:37.600
This can generate realistic human movements from text descriptions and it can also understand, complete, or edit existing motion sequences.

00:52:37.680 --> 00:52:38.800
Here are some examples for your reference.

00:52:38.880 --> 00:52:41.520
You can simply describe an action and it'll generate a 3D character that performs that movement.

00:52:41.600 --> 00:52:43.520
But what's interesting is that this model can also do the reverse.

00:52:43.600 --> 00:52:46.080
You can give it a motion sequence and it can describe what the character is doing in natural language.

00:52:46.240 --> 00:52:47.200
It also supports motion completion.

00:52:47.280 --> 00:52:50.240
So if you provide only part of a movement, it can automatically generate the missing parts to complete the action.

00:52:50.320 --> 00:52:51.760
You can even provide text instructions to guide the results.

00:52:51.920 --> 00:52:52.880
And finally, it supports motion editing.

00:52:52.960 --> 00:52:55.040
So you can modify existing animation through natural language instructions.

00:52:55.120 --> 00:52:56.080
Now, how this works is quite interesting.

00:52:56.160 --> 00:52:58.640
It uses a single AI architecture that can process both language and human motion.

00:52:58.800 --> 00:52:59.840
This allows the same model to understand

00:53:00.080 --> 00:53:00.800
Movements and generate new ones.

00:53:00.880 --> 00:53:04.880
And if you look at these benchmarks for motion generation and understanding, it actually outperforms many competing models across various metrics.

00:53:04.960 --> 00:53:08.160
Now, at the top of the page, if you click on this GitHub link, here it says the code is under internal review.

00:53:08.240 --> 00:53:09.760
So hopefully this gets approved and open source.

00:53:09.840 --> 00:53:12.480
For now, if you're interested in reading further, I'll link to this main page in the description below.

00:53:12.560 --> 00:53:14.080
Also, this week the model waits to the best open model.

00:53:14.240 --> 00:53:14.880
Kimmy K3.

00:53:15.040 --> 00:53:16.000
1 are finally released as prompt.

00:53:16.080 --> 00:53:17.680
Now I already mentioned this when they first announced it a few weeks ago.

00:53:17.760 --> 00:53:23.280
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.

00:53:23.360 --> 00:53:25.840
And as promised, this week they released the full model on Hugging Face, which is great.

00:53:25.920 --> 00:53:27.600
Now, this is one of the biggest open models out there, with 2.

00:53:27.760 --> 00:53:29.840
8 trillion total parameters, and it's a mixture of experts, so only

00:53:30.080 --> 00:53:31.280
104 billion parameters are active at a time.

00:53:31.360 --> 00:53:33.520
As you can see from the hugging face repo, the total size of this is massive at 1.

00:53:33.600 --> 00:53:34.320
56 terabytes.

00:53:34.400 --> 00:53:37.440
Further down this page, it contains all the instructions on how to download and run this locally on various platforms.

00:53:37.520 --> 00:53:40.000
Now, while this is huge, that's the size you'd expect from a model with Frontier Intelligence.

00:53:40.080 --> 00:53:42.320
It's aimed more at enterprises with the hardware to run this rather than consumers.

00:53:42.400 --> 00:53:45.120
But it's already amazing that they're willing to release this model for free for anyone to download and use.

00:53:45.200 --> 00:53:48.480
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.

00:53:48.560 --> 00:53:50.080
If you're interested, I'll link to this page in the description below.

00:53:50.160 --> 00:53:52.480
Also, this week, Baidu drops their latest open language model, earning 5.

00:53:52.640 --> 00:53:53.040
5 flash.

00:53:53.120 --> 00:53:58.080
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.

00:53:58.160 --> 00:53:59.680
Like most Frontier models, it also supports a context.

00:54:00.240 --> 00:54:01.040
Of up to 1 million tokens.

00:54:01.120 --> 00:54:02.640
So you can fit a ton of stuff into your prompt at once.

00:54:02.720 --> 00:54:07.600
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.

00:54:07.680 --> 00:54:11.040
The awesome thing is Baidu has released the model weights on Hugging Face, and they've actually released two versions.

00:54:11.120 --> 00:54:15.040
The FPH model is about 288 gigabytes, while the full BF16 version is around 573GB.

00:54:15.280 --> 00:54:18.400
If you scroll down on their Hugging Face page, you'll also find instructions on how to run the model locally on your computer.

00:54:18.480 --> 00:54:22.880
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.

00:54:23.040 --> 00:54:25.040
Anyway, if you're interested, I'll link to this main page in the description below.

00:54:25.120 --> 00:54:27.680
Also, this week we have another multimodal model from Antgroup called Ling3.

00:54:27.840 --> 00:54:28.240
2 Flash.

00:54:28.320 --> 00:54:30.000
Not only can it understand text, but it also supports image and

00:54:30.000 --> 00:54:30.480
And video inputs.

00:54:30.560 --> 00:54:32.720
So this makes it a lot better at things like 3D and spatial understanding.

00:54:32.800 --> 00:54:35.200
Here are some examples where we can input any image for it to reconstruct in 3D.

00:54:35.280 --> 00:54:37.680
Or here's another example where we can get to create this animated website of London.

00:54:37.760 --> 00:54:39.280
It also has quite powerful agentic capabilities.

00:54:39.360 --> 00:54:43.600
So, for example, you can get to coordinate a team of agents to turn financial data into an analysis report and slideshow like this.

00:54:43.760 --> 00:54:48.960
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.

00:54:49.040 --> 00:54:52.320
And if you look at these benchmarks, it's quite competitive with the leading open models out there, including GLN 5.

00:54:52.400 --> 00:54:54.960
3 Flash and Deep Seek, so it's definitely among the best multimodal open models out there.

00:54:55.040 --> 00:54:59.840
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

00:55:00.080 --> 00:55:00.960
Model weights to this in the near future.

00:55:01.040 --> 00:55:03.200
Anyway, if you're interested, I'll link to this main page in the description below.

00:55:03.280 --> 00:55:05.600
Also, this week we have another open model called Ripple65B.

00:55:05.680 --> 00:55:12.080
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.

00:55:12.160 --> 00:55:16.080
Now, this uses a dense architecture, so unlike a mixture of experts' models, it activates all 65 billion parameters when you use it.

00:55:16.160 --> 00:55:20.400
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.

00:55:20.480 --> 00:55:23.280
Now, interestingly, they originally developed this model to power its AI chip design platform.

00:55:23.360 --> 00:55:26.640
The idea is to use AI to design better chips, then use those chips to train more powerful AI models.

00:55:26.720 --> 00:55:29.760
And according to the team, the entire research and development process for Ripple was only $32 million.

00:55:30.240 --> 00:55:32.160
Which is way cheaper than the training costs of many Frontier labs.

00:55:32.240 --> 00:55:34.560
The awesome thing is they've released the model on Hugging Face for you to run this locally.

00:55:34.640 --> 00:55:37.040
So if you click into this Hugging Face page, note that they've released two versions.

00:55:37.120 --> 00:55:42.400
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.

00:55:42.640 --> 00:55:44.400
And this is under the Apache 2 license, which is very permissive.

00:55:44.480 --> 00:55:46.000
If you're interested, I'll link to this page in the description below.

00:55:46.080 --> 00:55:49.280
Also this week, Chinese food delivery company Mage1 drops their latest AI model, Longcat 2.

00:55:49.360 --> 00:55:49.440
5.

00:55:49.600 --> 00:55:53.840
And this is a really powerful multimodal model that's especially good at coding, visual understanding, and doing complex tasks autonomously.

00:55:53.920 --> 00:55:55.200
Now this is quite a massive model at 1.

00:55:55.360 --> 00:55:59.920
6 trillion total parameters, but uses a mixture of expert's architectures, so only around 48 billion parameters are active when you use it.

00:56:00.000 --> 00:56:01.520
Here are some examples of the incredible things it can do.

00:56:01.600 --> 00:56:05.440
For instance, they got it to create this interactive fluid simulation website complete with realistic physics and visual effects.

00:56:05.520 --> 00:56:07.440
And beyond coding, it can also perform professional office tasks.

00:56:07.520 --> 00:56:09.600
For example, you can get it to analyze business data or generate reports.

00:56:09.680 --> 00:56:12.640
Now, one of the biggest upgrades over the previous Longcat model is its native visual understanding.

00:56:12.720 --> 00:56:14.000
So it can now understand and analyze images.

00:56:14.080 --> 00:56:20.400
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.

00:56:20.480 --> 00:56:21.600
Now, currently you can try LongCat 2.

00:56:21.760 --> 00:56:24.160
5 through the official Longcat chat website or access it through their API.

00:56:24.240 --> 00:56:29.120
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.

00:56:29.200 --> 00:56:30.000
For now, I'll link to the

00:56:30.160 --> 00:56:30.880
Main page in the description below.

00:56:30.960 --> 00:56:33.200
Also, this week, thinking machines unveils their latest model, Inkling2 Small.

00:56:33.360 --> 00:56:36.240
Now, despite its name, this is actually a pretty large model with 594 billion total parameters.

00:56:36.320 --> 00:56:38.560
And this is a mixture of experts, so just 17 billion are active when you use it.

00:56:38.640 --> 00:56:41.440
It's designed to be especially strong at coding, visual understanding, and long autonomous tasks.

00:56:41.520 --> 00:56:46.160
For example, they got it to create an entire interactive city simulator with 3DS, including procedurally generated buildings, vehicles, and infrastructure.

00:56:46.240 --> 00:56:48.240
Or here's another example where they got it to create a 3D animation in Blender.

00:56:48.320 --> 00:56:50.400
So as you can see, it's quite good at operating different software and interfaces.

00:56:50.480 --> 00:56:52.560
What's even more interesting is its ability to carry out autonomous AI research.

00:56:52.640 --> 00:56:55.760
In one experiment, it independently trained and improved another AI model for time series forecasting.

00:56:55.840 --> 00:56:58.560
It designed experiments, evaluated results, and adjusted the training process all on its own.

00:56:58.640 --> 00:56:59.680
Now, if you look at these benchmarks on H

00:57:00.080 --> 00:57:03.280
Decoding, visual reasoning, and knowledge work, it's extremely competitive with other leading open weight models.

00:57:03.360 --> 00:57:05.760
In fact, once they release this, it might very well be the best open model from the US.

00:57:05.920 --> 00:57:09.120
Now, at the bottom here it says you can try English Small right now through their platform called Tinker.

00:57:09.200 --> 00:57:11.840
And they do plan to release the model weights next week under the very permissive Apache 2 license.

00:57:11.920 --> 00:57:13.760
If you're interested, I'll link to this release page in the description below.

00:57:13.840 --> 00:57:15.920
Also, this week we have a new 3D generator called Vox Solution.

00:57:16.000 --> 00:57:19.920
This can generate entire 3D indoor scenes, like apartments, offices, and bedrooms based on just a category or a simple layout.

00:57:20.000 --> 00:57:20.880
Here are some examples for your reference.

00:57:20.960 --> 00:57:25.360
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.

00:57:25.440 --> 00:57:26.400
Now, how this works is quite interesting.

00:57:26.480 --> 00:57:29.760
It first generates a rough layout of the room in boxes, then it progressively increases the resolution to add

00:57:30.720 --> 00:57:32.960
Now at the top of the page it says the code is coming soon, so stay tuned for that.

00:57:33.040 --> 00:57:35.600
For now, if you're interested in reading further, I'll link to this main page in the description below.

00:57:35.680 --> 00:57:37.200
Also, this week we have a new AI called Solver.

00:57:37.360 --> 00:57:41.040
This can take multiple photos or video frames of a scene and reconstruct its 3D geometry, along with the camera positions.

00:57:41.200 --> 00:57:43.840
You can give it photos from different angles, and you can figure out how everything fits together in 3D.

00:57:43.920 --> 00:57:46.880
What's particularly interesting is that Solver can continuously refine its existing reconstruction.

00:57:46.960 --> 00:57:49.600
You can add more images at any time, even in random order, and it'll update the scene accordingly.

00:57:49.680 --> 00:57:53.040
It also supports real time video streams, so it can progressively reconstruct the scene as new frames arrive.

00:57:53.120 --> 00:57:55.600
Now, how this works is that it uses a loop transformer, which seems to be getting more popular.

00:57:55.680 --> 00:57:58.720
It repeatedly applies the same loop to improve its estimates of the 3D scene and camera positions.

00:57:58.800 --> 00:58:00.000
Now, if you look at these benchmarks for 3D

00:58:00.080 --> 00:58:02.240
Reconstruction, Solver is quite competitive with other state-of-the-art models.

00:58:02.320 --> 00:58:04.320
It's also particularly efficient for continuous video processing.

00:58:04.400 --> 00:58:05.760
Now, at the top, they have released a GitHub for this.

00:58:05.840 --> 00:58:08.720
If you click into it, note that it says the code and models will be released within a week, so stay tuned for that.

00:58:08.800 --> 00:58:11.200
If you're interested in reading further, following to this main page in the description below.

00:58:11.280 --> 00:58:12.960
And that sums up all the highlights in AI this week.

00:58:13.120 --> 00:58:14.320
Let me know in the comments what you think of all of this.

00:58:14.480 --> 00:58:17.200
Which piece of news was your favorite, and which tool are you most looking forward to trying out?

00:58:17.360 --> 00:58:19.920
As always, I will be on the lookout for the top AI news and tools to share with you.

00:58:20.000 --> 00:58:22.720
So if you enjoyed this video, remember to like, share, subscribe, and stay tuned for more content.

00:58:22.880 --> 00:58:26.240
Also, there's just so much happening in the world of AI every week, I can't possibly cover everything on my YouTube channel.

00:58:30.000 --> 00:58:31.120
The link to that will be in the description below.

00:58:31.200 --> 00:58:32.720
Thanks for watching, and I'll see you in the next one.

00:58:50.560 --> 00:58:53.280
Elon sat down with Ron Barron and put some important numbers on the table.

00:58:53.360 --> 00:58:59.280
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.

00:58:59.360 --> 00:58:59.680
A trillion.

00:59:00.160 --> 00:59:00.880
It's no longer out of question.

00:59:00.960 --> 00:59:09.920
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.

00:59:10.080 --> 00:59:11.280
Now, also, there's that Fallcom D.

00:59:11.360 --> 00:59:20.480
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.

00:59:20.560 --> 00:59:25.680
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

01:00:54.480 --> 01:00:57.120
Manages to get out of its testing ring.

01:00:57.680 --> 01:00:59.680
You and I agree very strongly that the

01:01:00.080 --> 01:01:03.120
The rate of innovation in this war is faster than it has ever occurred.

01:01:03.200 --> 01:01:07.360
It's faster than the rate of innovation that I have seen in my many decades in tech.

01:01:07.680 --> 01:01:20.240
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.

01:01:20.480 --> 01:01:26.000
Well, first place, we're not going to do a pause because the entire incentive system in America is not to pause.

01:01:26.240 --> 01:01:29.520
I mean, we need a solution that's implementable for a pause.

01:01:30.000 --> 01:01:30.560
I don't see it.

01:01:30.640 --> 01:01:38.320
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.

01:01:38.400 --> 01:01:38.560
Right?

01:01:38.640 --> 01:01:39.200
That's my command.

01:01:39.360 --> 01:01:40.800
Notice I didn't say anything else.

01:01:40.960 --> 01:01:46.320
The system would be free to do whatever he thought was appropriate, which could include massive war crimes.

01:01:46.640 --> 01:01:47.360
That is Eric's.

01:01:48.320 --> 01:01:54.880
He ran Google for a decade, then chaired the Pentagon's Defense Innovation Board and America's National Security Commission on AI.

01:01:55.040 --> 01:01:59.680
Today, his own drone company builds interceptors that, by its count, have shot down more than four.

01:02:00.160 --> 01:02:02.160
4,000 Russian drones over Ukraine.

01:02:02.240 --> 01:02:05.120
The economist's editor-in-chief sat down with him in Kyiv.

01:02:05.200 --> 01:02:06.560
Here is his full warning.

01:02:06.800 --> 01:02:09.520
Let's just dive down into some of those specifically.

01:02:09.600 --> 01:02:21.200
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.

01:02:21.360 --> 01:02:24.640
How close is Ukraine to a solution to that?

01:02:24.880 --> 01:02:29.920
There are at least five companies in Ukraine that have come up with solutions that are in test now.

01:02:30.240 --> 01:02:31.200
It should be soon.

01:02:31.360 --> 01:02:33.520
But this was a major event in the war.

01:02:33.600 --> 01:02:36.720
And I didn't fully foresee it, which I hold myself responsible.

01:02:36.960 --> 01:02:44.400
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.

01:02:44.560 --> 01:02:49.280
We thought they were going to use about ten percent of the supply, and we thought that that was overcomable.

01:02:49.440 --> 01:02:55.120
Russia accelerated the use of these to more than half of the Shaheds, but they did one other thing.

01:02:55.280 --> 01:02:59.280
They don't have access to the equivalent of Starlink for Russian use in Ukraine.

01:02:59.360 --> 01:02:59.760
Thank goodness.

01:03:00.400 --> 01:03:09.920
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.

01:03:10.080 --> 01:03:11.200
And those are relays.

01:03:11.360 --> 01:03:17.200
And those relays allow them to see literally 8K video of whatever their target is.

01:03:17.360 --> 01:03:23.520
So right next door, they manage to go right through the central, central window of the SBU headquarters, which is like the CIA.

01:03:23.680 --> 01:03:26.080
And that's a good warning of the power of these things.

01:03:26.240 --> 01:03:29.680
There are reports that there are plans to use up to 80,000 such.

01:03:30.240 --> 01:03:34.400
Jetshaheads against Kyiv and the other towns between now and the end of December.

01:03:34.560 --> 01:03:35.760
This is a major point.

01:03:35.920 --> 01:03:38.560
the government here is extremely focused on this.

01:03:38.720 --> 01:03:40.880
I know of a number of efforts that are likely to work.

01:03:41.200 --> 01:03:47.520
You mentioned that the Russians, and this would be even worse, are close to or potentially close to building their version of Starlink, Rasviet.

01:03:47.840 --> 01:03:48.560
How close are they?

01:03:48.560 --> 01:03:52.880
'Cause it seems to me, and I'm no technician on this, but it seems that that would be a real game changer.

01:03:53.200 --> 01:03:57.680
there's evidence that they are have started launching, there's evidence that it doesn't work yet.

01:03:58.080 --> 01:03:59.760
My own opinion is it's probably

01:04:00.240 --> 01:04:02.480
A year away, because these networks are hard.

01:04:02.640 --> 01:04:11.440
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.

01:04:11.680 --> 01:04:16.160
Ukraine today does not have the ability to use Starlink to hunt inside America.

01:04:16.400 --> 01:04:22.400
So you can imagine a situation both countries have that capability, the level of damage that they could do to each other.

01:04:22.560 --> 01:04:26.960
In the case of Ukraine, one of their highest priorities is hitting the oil derives and so forth and so on.

01:04:27.360 --> 01:04:29.440
They do that without really being able to see what they're doing.

01:04:30.160 --> 01:04:32.800
Imagine how much more the Ukrainians could be against Russia.

01:04:33.280 --> 01:04:35.280
If this actually from the U.

01:04:35.280 --> 01:04:35.280
S.

01:04:35.520 --> 01:04:43.200
This begs a question, which many of the Ukrainian commanders I've spoken to mention why can't we use Starlink in Russian territory?

01:04:43.360 --> 01:04:55.680
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.

01:04:55.920 --> 01:04:59.120
And they say if we can't hit the arrow, why can't we go after the archer?

01:04:59.360 --> 01:04:59.680
And why

01:05:00.160 --> 01:05:04.800
Can't someone persuade Elon Musk to let us for that purpose use Starlink in Russia.

01:05:04.960 --> 01:05:05.920
Do you think that would be a good idea?

01:05:06.240 --> 01:05:11.680
I'm in favor of giving Ukraine the tools to allow them to stand the war by force.

01:05:11.920 --> 01:05:13.760
Literally to stop the Russian aggression.

01:05:14.000 --> 01:05:16.240
The question you're asking is really a political decision.

01:05:16.400 --> 01:05:17.840
It's really a government decision.

01:05:17.920 --> 01:05:20.640
And the US government needs to grapple with this question.

01:05:20.880 --> 01:05:21.760
Put a number on that.

01:05:21.920 --> 01:05:27.520
On March 24th this year, Russia fired 948 drones at Ukraine in 24 hours.

01:05:27.680 --> 01:05:29.840
The biggest attack of the war, according to Ukraine's.

01:05:30.160 --> 01:05:30.640
Air Force.

01:05:30.720 --> 01:05:38.560
The day before, a Russian company called Bureau 1440 launched the first 16 satellites of Rasviet, and 16 more went up in July.

01:05:38.720 --> 01:05:44.800
By August, fewer than four in 10 had reached their working orbit, and partial service is expected in 2027.

01:05:44.880 --> 01:05:48.240
So the year he's guessing at is already on somebody's launch schedule.

01:05:48.400 --> 01:05:49.120
Now listen to this.

01:05:49.280 --> 01:05:50.400
Quick thing before we go on.

01:05:50.560 --> 01:05:52.960
Most of the people who watch this channel aren't subscribed.

01:05:53.120 --> 01:05:56.240
If these videos are worth your time, take a second to hit subscribe.

01:05:56.400 --> 01:05:59.840
It's free, it helps us keep making them, and you can change your mind whenever you like.

01:06:00.400 --> 01:06:01.520
Right, back to Kyiv.

01:06:01.680 --> 01:06:02.880
Let's talk about what's coming next.

01:06:02.960 --> 01:06:07.520
I mean that's all pretty, pretty alarming, and that's these are the challenges Ukraine faces now.

01:06:07.600 --> 01:06:13.200
But coming down the road is even greater technical innovation, and particularly AI.

01:06:13.280 --> 01:06:19.040
And you know, in most contexts, you and I have conversations about what is happening in Silicon Valley, not what is happening in Ukraine.

01:06:19.200 --> 01:06:26.000
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.

01:06:26.160 --> 01:06:29.680
People talk about is AI being used, but as I understand it, what's really

01:06:30.240 --> 01:06:34.640
being used is quite well known technology of machine vision for final targeting.

01:06:34.800 --> 01:06:35.600
Could you explain that?

01:06:35.680 --> 01:06:37.200
Because I think that's an important thing to make clear.

01:06:37.520 --> 01:06:43.280
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.

01:06:43.440 --> 01:06:53.120
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.

01:06:53.280 --> 01:06:56.160
That analogy scales up in terms of autonomy.

01:06:56.320 --> 01:06:59.760
So it's reasonable to expect that every weapon has a

01:07:00.320 --> 01:07:08.480
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.

01:07:08.640 --> 01:07:09.760
All of these are not AI.

01:07:10.080 --> 01:07:12.480
They're just good computer science, good computer engineering.

01:07:12.720 --> 01:07:14.640
Artificial intelligence is systems that learn.

01:07:14.880 --> 01:07:19.840
And artificial intelligence, learn like we sort of learn, you get smarter and smarter and smarter.

01:07:20.000 --> 01:07:21.120
You're self-improving.

01:07:21.200 --> 01:07:23.440
The technical term is recursive self-improvement.

01:07:23.600 --> 01:07:26.720
And we're just beginning to see the use of AI in war.

01:07:26.880 --> 01:07:29.840
There's a lengthy study on Russian use of a Chinese model

01:07:30.320 --> 01:07:31.760
To do some of their decision making.

01:07:32.000 --> 01:07:34.640
These are all from leaked memos, I don't know what the truth is.

01:07:34.880 --> 01:07:37.440
In Ukraine, there's relatively little use of that.

01:07:37.600 --> 01:07:44.080
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.

01:07:44.320 --> 01:07:46.000
Again, it's just at the beginning.

01:07:46.160 --> 01:07:49.440
So speaking as a computer scientist, it's pretty obvious what it should look like.

01:07:49.600 --> 01:07:58.000
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.

01:07:58.160 --> 01:07:59.680
And have these large language models.

01:08:00.160 --> 01:08:01.600
use tools to do reasoning.

01:08:01.760 --> 01:08:04.000
To me it's fantastic they can think at this level.

01:08:04.160 --> 01:08:05.600
They don't think like humans, by the way.

01:08:05.760 --> 01:08:07.520
They're reasoning around an objective.

01:08:07.760 --> 01:08:16.240
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.

01:08:16.400 --> 01:08:19.040
But the more interesting questions are why am I losing?

01:08:19.200 --> 01:08:21.680
Why is this group not as good as the other group?

01:08:21.840 --> 01:08:23.520
You want the system to anticipate that.

01:08:23.600 --> 01:08:26.560
And you want the system to say, General, you have a problem.

01:08:26.720 --> 01:08:28.160
Here is our recommended solution.

01:08:28.320 --> 01:08:29.040
Can we help you?

01:08:29.200 --> 01:08:29.840
That's I think the

01:08:30.480 --> 01:08:30.960
That's where we're going.

01:08:31.280 --> 01:08:33.440
That kind of system already has a track record.

01:08:33.680 --> 01:08:41.200
In April 2024, Plus 972 magazine and local call reported on a tool the Israeli army used in Gaza called Lavender.

01:08:41.360 --> 01:08:51.280
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.

01:08:51.440 --> 01:08:57.200
They said roughly one in ten of the people it marked were known in advance not to be in Hamas's military wing.

01:08:57.360 --> 01:08:59.360
Israel's military disputed the account and described

01:09:00.480 --> 01:09:02.640
As a database for cross-referencing intelligence.

01:09:02.800 --> 01:09:07.600
Either way, there was a human in that loop, and 20 seconds was how long the loop lasted.

01:09:07.760 --> 01:09:08.640
Watch what he says next.

01:09:08.800 --> 01:09:19.520
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.

01:09:19.760 --> 01:09:23.920
Firstly, what does it mean for how autonomous warfare becomes?

01:09:24.080 --> 01:09:29.600
Because now you can have this last mile guidance, but then presumably you can have the AI not only giving the general

01:09:30.320 --> 01:09:34.640
lessons but providing a battle plan and and down the road you know executing them autonomously.

01:09:34.960 --> 01:09:42.640
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.

01:09:42.720 --> 01:09:43.440
Right, that's my command.

01:09:43.680 --> 01:09:45.120
Notice I didn't say anything else.

01:09:45.280 --> 01:09:51.120
The system would be then free to do whatever it thought was appropriate, which could include massive war crimes, right?

01:09:51.280 --> 01:09:52.400
All sorts of terrible things.

01:09:52.560 --> 01:09:59.360
It could destroy civilians, it could destroy the infrastructure, it might attack another country that was uninvolved in order to annoy the first country.
