**Cybersecurity chaos, AI for DNA, Kimi K3.1, Claude Duet, Dessa, Fibo 2, Minimax 3.1: AI NEWS — Transcript & Summary | SozAI**
Source: https://sozai.app/transcript/ai-news-cybersecurity-dna-kimi-claude-minimax/

Weekly AI news roundup covering Tencent's WeViz-1VL vision model, Bark 3 transcription, FIBO 2 image editing, a DNA-understanding model, Kimi K3.1 Pro, and major cybersecurity incidents.

## Key Takeaways

- Open-source vision, transcription, and image-editing models are increasingly competitive with proprietary or much larger alternatives.
- Licensing terms differ significantly between models that perform similarly, so commercial viability should be checked before adoption.
- The week saw an unusually high volume of critical cybersecurity incidents, highlighting supply-chain and developer-tool risks.
- Domain-specialized models, such as DNA-focused and engineering-focused AI, are emerging as a notable trend.

## What the video covers

- Opens with a claim of a record-setting week for cybersecurity incidents, with more than a dozen critical incidents found and a chart showing roughly 3% of all critical incidents tied to one vulnerability.
- Covers releases including MiniMax M3.1 Flash, Kimi K3.1 Pro (presented as beating Claude Fable), an engineering-focused AI for breaking designs into components and generating 2D floor plans, and a realistic real-time voice model.
- Tencent's WeViz-1VL, built on Qwen 3-VL with a specialized pixel-segmentation component, handles object detection, referring, counting, pointing, OCR, segmentation, and alpha matting, released under Apache 2.0 in 4B and 2B variants.
- Bark 3 is a lightweight speech-to-text model (about 163 MB for the smallest version) presented as competitive with Whisper Large v3 and Nvidia Parakeet, with a free Hugging Face demo.
- FIBO 2 by Bria is an open-source natural-language image editor that accepts multiple reference images and performs well on lighting correction and colorization, though the presenter found it roughly on par with Qwen Image 2.1.
- FIBO 2's license is non-commercial only and requires user details to access, contrasted with the more permissive Qwen Image license.
- Hugging Face releases an open-source DNA-understanding model built on Qwen 3, trained on 300 billion tokens of DNA data and further refined with reinforcement learning, with training scripts provided.
- A video/image tool features a Relight Studio that lets users drag a virtual light to change lighting direction and intensity.
- Discussion includes a GitLab vulnerability with an available patch, serious flaws found in Node.js, and speculation about advertising companies building their own AI bots.
- Mentions a 3D world-generation approach that represents objects as tiny matter particles and can generate 3D characters and animations while staying spatially coherent, plus a vector-conversion tool that produces detailed rather than flat vectors.

## Chapters

1. 00:00 Intro: Record Cybersecurity Week and Major Model Releases
2. 02:55 Tencent WeViz-1VL Vision-Language Model
3. 05:05 Bark 3 Lightweight Transcription Model
4. 07:50 FIBO 2 Open-Source Image Editor
5. 10:40 Open-Source DNA Understanding Model
6. 13:20 Relight Studio and Video/Image Tools
7. 19:00 Cybersecurity Incidents: GitLab and Node.js
8. 22:00 3D World Generation and Vector Conversion
9. 27:30 Kimi K3.1 Pro and Closing Discussion

Answers

## Questions about this video

What is Tencent's WeViz-1VL model and what can it do?

WeViz-1VL is a vision-language model built on Qwen 3-VL with a specialized pixel-segmentation component. It handles object detection, counting, pointing, text recognition, segmentation, and background removal with alpha masks, and is released under the permissive Apache 2.0 license.

How does Bark 3 compare to Whisper and Parakeet?

According to the video, Bark 3 is a very small transcription model (about 163 MB for the base version) that achieves competitive word error rates against Whisper Large v3 and Nvidia Parakeet, making it highly efficient per model size. A free Hugging Face demo is available to test it.

Why does the presenter say FIBO 2's license matters?

Although FIBO 2 performs about as well as Qwen Image 2.1 on editing benchmarks, its license is restricted to non-commercial use and requires providing personal details to access the model, whereas the Qwen Image license is much less restrictive and allows broader use.

## Full Transcript — Download SRT & Markdown

00:00

Speaker A

AI never sleeps, and this week has been absolutely insane. We had the biggest week in cyber security incidents in history, and it's only going to get crazier. MiniMax releases their latest M3.1 Flash, which is incredibly efficient. Kimi drops their latest model, Kimi K3.1 Pro, which even beats Claude Fable. This thing is an absolute beast. We have a really cool AI model that's built for engineering. So, this can break things down into separate components or even help generate 2D floor plans. Tencent releases a really powerful vision model which can do a ton of things, from counting to text recognition to segmentation. This new real-time AI voice model sounds freakishly realistic. We also have a new top open-source text-to-speech generator, as well as a top open image editor. Hugging Face drops an open-source AI that understands DNA. We have some ridiculous humanoid robot demos and a lot more. So, let's jump right in.

00:20

Speaker A

First up, Tencent releases an AI called WeViz-1VL. Bit of a tongue twister, but anyways, this is a really powerful vision-language model. So, it's really good at analyzing images. For example, you can upload any image and get it to detect all instances of whatever object you want. For example, in your prompt, if you want it to detect dogs, chairs, and bushes, here is what you get.

00:33

Speaker A

Or you can also ask it to only locate the black-and-white dog. And indeed, that's what it's able to identify. Or you can also get it to count all the dogs in this image. You can also get it to point to all the dogs. This also can work with text and character recognition. So, here's an example where we can get it to detect and outline all the words in this photo. Or you can also get it to locate just a certain word, like "small alley."

00:50

Speaker A

The really cool thing is you can also get it to remove the background and provide an alpha mask, like this. Even down to the fine details, like individual strands of hair. So, this is a super flexible tool that can do a ton of things.

01:05

Speaker A

Now, how this works is quite interesting. It's built on top of Qwen 3-VL, but they've added a specialized component that can identify the exact pixels that belong to an object. So, it can handle everything from basic object detection to precise image editing and alpha masks within a single model.

01:20

Speaker A

And if you look at its benchmarks in terms of all these different tasks, like detection, referring, counting, pointing, character recognition, segmentation, and matting (for this one, lower is better), you can see that across the board, WeViz-1VL just outperforms the other competitor models. And the nice thing is they've released the code and the models. Note that there are two different variants. There's a 4-billion-parameter variant, which is 19 GB in size, and then there's a 2-billion-parameter one, which is about half the size. And if you click on this code button and scroll down a bit here, it contains all the instructions on how to download and run this locally on your computer. And note that this is under the Apache 2.0 license, which is very permissive. You can even use this for commercial purposes. If you're interested in reading further, I'll link to this main page in the description below.

01:32

Speaker A

Also this week, we have a new AI transcription model called Bark 3. This can turn any spoken audio into a text transcript. Now, they've released a free Hugging Face space for you to try this out. So, let me start recording, and you can see it transcribed what I say. What's impressive about this is how tiny the model is. It's only 163 megabytes to download, yet it can compete with much larger models, including OpenAI's Whisper Large and Nvidia's Parakeet. All right. So, afterwards, let me click on transcribe. And it was able to transcribe this in just a second. So, this is super lightweight and super fast.

01:44

Speaker A

Now, they've actually released a few different models in the Bark 2 family. The smallest Bark 2 is 163 MB. There's also a Plus version, which makes fewer mistakes and is a bit larger, and then a Pro version, which is the most accurate, but this is way larger at 1.25 GB. Though this should still be able to fit on most consumer GPUs.

01:51

Speaker A

If you look at this chart, the y-axis is the word error rate. The higher you go, the lower the error rate, and then the x-axis is the model size. So, ideally, you want to be in the upper left corner. And as you can see, the Bark models are all the way up here, making it the most efficient model compared to other transcription models like Nvidia's Parakeet and OpenAI's Whisper Large version 3. The nice thing is they've released everything on Hugging Face already. So, if you click on this link at the bottom of the page here, it contains the Bark 3 models for you to download. And at the bottom of the page, it contains instructions on how to run this locally. And the nice thing is this is under a license which even allows commercial use with attribution. If you're interested, I'll link to this main page in the description below.

02:04

Speaker A

Also, this week we have a new open-source image editor. It's called FIBO 2 by Bria, and this can actually edit images in natural language, just like Nano Banana or Qwen Image. In fact, here are a few of my tests. So, this can take in multiple reference images. Here, I'm getting the person in the first image to wear the outfit in the second image, in the pose of the third image, and in the background of the last image, and it's actually able to understand everything and get it correct. I deliberately chose this dress, which has a very complicated design, but it's able to get all the details correct. She is missing feet, though, so I guess that's one slight error.

02:18

Speaker A

Here's another super hard test where I fed it a super blurry photo of Obama, and I wanted to see if it can upscale and add more details to the photo. It turned out okay, but it doesn't look exactly like the original Obama photo. But keep in mind, this is a really hard test, and none of the other open models can do this well either.

02:35

Speaker A

Now, in terms of restoring and colorizing black-and-white photos, this does really well. As you can see, it's able to keep the faces consistent and faithful. Plus, it's also able to fix the folds and cracks of the original photo.

02:45

Speaker A

Next, I got it to generate a character sheet. With this super complicated character, I intentionally made it really hard, and unfortunately, it was not able to execute this prompt very well. As you can see, her legs are way too short. Plus, the staff that she's holding is not really consistent or faithful.

02:55

Speaker A

Now, here's another really good test where I gave it a super dark photo with way too much saturation and contrast. I also wrecked the white balance and made it very blue when I told it to correct the lighting of this photo. And it was able to actually correct it very well. It's pretty much the same colors as the original photo.

03:06

Speaker A

So, overall, a very strong image editing model, but not like way better than Qwen Image 2.1. I would say they are very similar. If you look at this leaderboard by Artificial Analysis in terms of image editing, you can see that FIBO 2 is ranked number one, but just a few points above Qwen Image 2.1. And note that there's not actually a significant difference. But what is different is the license. So, this is a lot more restrictive than the license from Qwen. The FIBO license is for non-commercial use only. Plus, you need to provide them all these details to access the model. Whereas the license for Qwen Image is way less restrictive. But if you are interested in trying this out, the model is fairly small. The int8 converted version is only 8.4 GB in size. However, it's still a bit larger than Qwen Image 2.1, which is only 7.26 GB in size. This is one of the best open image editors out there. So, if you're interested in trying this out, I'll link to this main page in the description below.

03:20

Speaker A

Also, this week, Hugging Face releases a new open-source model called Carbon X. And this is designed to understand DNA. So here it s

03:35

Speaker A

larger models, including OpenAI's Whisper Large and Nvidia's Parakeet. All right. So afterwards, let me click on transcribe. And it was able to transcribe this in just a second. So this is super lightweight and super fast. Now, they've actually released a

03:49

Speaker A

few different models in the Bark II family. The smallest Bark II is 163 MGB.

03:56

Speaker A

There's also a plus version, which makes fewer mistakes, which is a bit larger, and then a Pro version, which is the most accurate, but this is way larger at 1.25 GB. Though this should still be able to fit on most consumer GPUs. If

04:09

Speaker A

you look at this chart, the y-axis is the word error rate. The higher you go, the lower the error rate, and then the x-axis is the model size. So, ideally, you want to be in the upper left corner.

04:19

Speaker A

And as you can see, the bark I models are all the way up here, making it the most efficient model compared to other transcription models like Nvidia's parakeet and openi's whisper large version 3. The nice thing is they've

04:32

Speaker A

released everything on hugging face already. So, if you click on this link at the bottom of the page here, it contains the Bark 3 models for you to download. And at the bottom of the page, it contains instructions on how to run

04:43

Speaker A

this locally. And the nice thing is this is under this license which even allows commercial use with attribution. If you're interested, I'll link to this main page in the description below.

04:54

Speaker A

Also, this week we have a new open-source image editor. It's called Fibio 2 by Bria and this can actually edit images in natural language just like Nano Banana or Quen image. In fact, here are a few of my tests. So, this can

05:08

Speaker A

take in multiple reference images. Here, I'm getting the person in the first image to wear the outfit in the second image, in the pose of the third image, and in the background of the last image, and it's actually able to understand

05:21

Speaker A

everything and get it correct. I deliberately chose this dress, which has a very complicated design, but it's able to get all the details correct. She is missing feet, though, so I guess that's one slight error. Here's another super

05:33

Speaker A

hard test where I fed it a super blurry photo of Obama, and I wanted to see if it can upscale and add more details to the photo. It turned out okay, but it doesn't look exactly like the original

05:45

Speaker A

Obama photo. But keep in mind, this is a really hard test, and none of the other Open models can do this well either.

05:51

Speaker A

Now, in terms of restoring and colorizing black and white photos, this does really well. As you can see, it's able to keep the faces consistent and faithful. Plus, it's also able to fix the folds and cracks of the original

06:03

Speaker A

photo. Next, I got it to generate a character sheet. With this super complicated character, I intentionally made it really hard, and unfortunately, it was not able to execute this prompt very well. As you can see, her legs are

06:16

Speaker A

way too short. Plus, the staff that she's holding is not really consistent or faithful. Now, here's another really good test where I gave it a super dark photo with way too much saturation and contrast. I also wrecked the white

06:29

Speaker A

balance and made it very blue when I told it to correct the lighting of this photo. And it was able to actually correct it very well. It's pretty much the same colors as the original photo.

06:39

Speaker A

So, overall, a very strong image editing model, but not like way better than Quinn image 2.1. I would say they are very similar. If you look at this leaderboard by artificial analysis in terms of image editing, then you can see

06:51

Speaker A

that Fibo 2 is ranked number one, but just a few points above Quinn image 2.1.

06:56

Speaker A

And note that there's not actually a significant difference. But what is different is the license. So this is a lot more restrictive than the license from Quinn. So the FIBO license is for non-commercial use only. Plus, you need

07:09

Speaker A

to provide them all these details to access the model. Whereas the license from Quinn image is way less restrictive. But if you are interested in trying this out, then the model is fairly small. So the int8 convert version is only 8.4 GB in size. However,

07:24

Speaker A

it's still a bit larger than Quinn image 2.1, which is only 7.26 GB in size. This is one of the best open image editors out there. So if you're interested in trying this out, I'll link to this main

07:36

Speaker A

page in the description below. Also, this week, HuggingFace releases a new open-source model called Carbon X. And this is designed to understand DNA. So here it says this is a 14 billion parameter model for DNA understanding and generation. For example, you can

07:52

Speaker A

give it a sequence of DNA and ask questions about what it does or make a change to the sequence and ask if that mutation would be harmful or not. It can also predict missing or future parts of a DNA sequence. Now they started from

08:06

Speaker A

Quinn 3, but they trained it on 300 billion tokens of DNA data and then trained it further on reinforcement learning. And after all this training, you can see that they were able to bump the performance of Quen 3 all the way up

08:20

Speaker A

to 67% at least in terms of this DNA intelligence index. Even matching GPT 6.1 Soul and beating other Frontier open models, but at a much cheaper cost.

08:31

Speaker A

Remember, this is only 14 billion parameters, whereas the other open models are like at least hundreds of billions of parameters. Here are some additional DNA understanding and analysis benchmarks for your reference.

08:42

Speaker A

And as you can see, Carbon X on average outperforms even closed Frontier models like GPT 6.1 and Opus 5.5, which is pretty insane for just a 14 billion parameter model. The awesome thing is they've released this already. So on

08:56

Speaker A

this hugging face page, it contains all the instructions on how to download and run this locally on your computer. They also released the scripts on how they evaluated this, plus info on how they trained this. And this model is fairly

09:09

Speaker A

accessible at less than 30 gigabytes in size. So you can even run this on like high-end consumer hardware. You don't need a data center for this. And the really awesome thing is this is under the Apache 2 license which has very

09:20

Speaker A

minimal restrictions. You can even use this commercially. If you're interested in reading further, I'll link to this main page in the description below. Also this week, Tencent drops a new AI world model called High Nomas. Like other world generators, you can just describe

09:35

Speaker A

a world in natural language and it can generate a virtual world which you can interact with in real time. And the quality of this looks amazing, much better than other world generators that I've featured in the past few months.

09:47

Speaker A

You can move the camera around or even change the environment or even introduce new objects and actions just with a text prompt. So, here are some examples where we can control what happens just with natural language. This can control a ton

09:59

Speaker A

of different characters from first person to third person, you can basically generate whatever you want.

10:04

Speaker A

And how this works is instead of generating an entire video at once, it generates short segments one after another. Plus, it also remembers what happened in previous segments, so it can maintain consistency as you explore around. So, here's an example where you

10:19

Speaker A

can keep running this for over a minute and it still stays consistent. And because of its quality and consistency, you can also use this to generate worlds for training robots. Now, this can generate videos at 720p resolution and

10:31

Speaker A

24 frames per second. The awesome thing is they've released the code and the model to this. So, if you click on this GitHub link and you scroll down a bit here, it contains all the instructions on how to download and run this locally

10:43

Speaker A

on your computer. They've also released the training scripts to this so you can fine-tune the model with your own videos, which is fantastic. Now, they released a few different variants of this. There's the base model, which is pretty huge at 34 GB in size. There's

10:57

Speaker A

also a distilled INT8 version, which is a bit smaller at 9 GB in size. So, you can potentially even run this on consumer hardware. The really nice thing is this is under the Apache 2 license, which is very permissive. If you're

11:11

Speaker A

interested in reading further, I'll link to this main page in the description below. Also, this week we have some exciting news in terms of 3D generation.

11:19

Speaker A

So, High3D, which is one of the leading AI 3D generators, just open- sourced their version one model. So, this model is designed to create some really detailed 3D assets just from a prompt or reference images. Note that currently

11:32

Speaker A

High3D is already on version 3. This is their latest and most detailed model, and this version 3 is closed source, but this week they've open sourced their previous version one, which is still pretty good. So, on this page, they've

11:46

Speaker A

released the code and the models to this. If you click on this GitHub link here, it contains all the instructions on how to download and run this locally on your computer. And if you compare Hi3D, which is all the way on the far

11:57

Speaker A

right, with other leading open 3D generators, you can see that High3D is by far the most detailed and accurate.

12:04

Speaker A

Like if you zoom in on certain sections of the character, look how detailed the generations from High 3D are. It looks much better than other models. And this is fairly tiny. So the total size of everything is only 7.74 GB. So, you

12:17

Speaker A

should be able to fit this on most consumer GPUs. If you're interested in reading further, I'll link to this main page in the description below. Also, this week, we have a new AI for 3D reconstruction called Pure. So, this can

12:30

Speaker A

take an ordinary video or sequence of images and reconstruct the environment in 3D. So, here's an example of this in action. As you can see, it can reconstruct an entire motorcycle ride.

12:41

Speaker A

And in the corner over here, you can see it mapping out the entire 3D scene.

12:45

Speaker A

What's particularly interesting is how this model was developed. So the researchers took existing 3D reconstruction models and improved them just with ordinary videos that don't even have any labels or 3D data. So instead of requiring 3D data, which is

13:00

Speaker A

often expensive and hard to get, they just used regular videos and they converted them into additional information like depth and object movement to improve the reconstruction.

13:10

Speaker A

The awesome thing is this is open source. So, if you click on this code button at the top and you scroll down a bit here, it contains all the instructions on how to download and run this locally on your computer. In

13:20

Speaker A

addition, they've also released the training scripts to this as well. Note that the total size of everything is fairly tiny at only 13.2 GB in size. If you're interested in reading further, I'll link to this main page in the

13:32

Speaker A

description below. Also, this week, we've got a new texttospech model called Ming 3.0 TTS by Antroup. It can turn pretty much any transcript into realistic speech. And you can even use natural language instructions to control how the voice sounds. For example, you

13:48

Speaker A

can just describe the type of voice you want. And here's what it sounds like.

13:52

Speaker A

You change the plan without telling anyone. How are we supposed to know what to do?

13:59

Speaker A

Beyond the northern ridge lies a city that vanished from every map. The old radio crackled softly as the storm moved across the valley.

14:10

Speaker A

You can also clone someone's voice from a reference audio clip. And you can even add special tags directly into your transcript to control the emotion or add sound effects.

14:25

Speaker A

It supports all these languages plus a bunch of Chinese dialects. And if we look at the artificial analysis leaderboards, the smaller model actually ranks among the best openweight models for voice quality and pronunciation accuracy. Note that this is only

14:40

Speaker A

slightly better than Fish Audios Drama 3, which I featured last week, but it's not like significantly better. Now, the awesome thing is they've already released the models and inference code on GitHub. So, if you scroll down a bit,

14:53

Speaker A

you'll find instructions on how to run everything locally. They've released two versions. The smaller one's only around 3 GB, so you should be able to run it on many consumer GPUs. The larger one's around 18 GB, and it also supports

15:07

Speaker A

multipe podcast generation. And here's another great thing. Both models are released under the MIT license. So, you've got a lot of freedom to use and modify them, even for commercial projects. If you're interested in trying this out, I'll link to this main page in

15:21

Speaker A

the description below. If you want to supercharge your content creation, definitely check out Luma, the sponsor of this video. Instead of jumping between different AI tools and platforms, Luma gives you a team of autonomous agents that can work with you

15:35

Speaker A

across entire content creation projects. You can simply prompt Luma agents your goal, and it will autonomously plan out the task and proceed to generate all the assets you specify. They've just released a bunch of new creative workflows that allow you to easily edit

15:50

Speaker A

images and videos while keeping everything else consistent. For example, I'm especially impressed by its new Relight Studio. You can literally drag a virtual light around an image to change the lighting direction and intensity just like you're working in a

16:05

Speaker A

professional photography studio. Another impressive feature is multi-angle shot coverage. This allows you to take a single video shot and generate additional moving camera angles. So, you have multiple perspectives to work with when editing your video. You also have

16:20

Speaker A

product shots, which lets you take any product and instantly place it into completely different environments without having to start over. Luma gives you access to the best image generators as well as the best video generators out there, including their own Ray 3.2, too.

16:35

Speaker A

So you can seamlessly switch between these different tools all in the same interface. So if you need to make content at scale, Luma is a super smooth and userfriendly platform for making this happen. Try Luma today using the

16:48

Speaker A

link in the description below or you can scan the QR code here. Also this week we saw one of the biggest waves of serious cyber security warnings in recent history. It affected a ton of things from popular websites and developer

17:02

Speaker A

tools to cloud platforms and even the security certificates that help keep the internet safe. Now, there's been a lot of confusion and fear-mongering around these incidents. So, let's clarify what actually happened. First of all, how was AI involved? Well, researchers used some

17:17

Speaker A

of the most powerful AI models to discover serious vulnerabilities that had gone unnoticed for years. But at the same time, hackers can also use AI to find weaknesses and automate attacks much faster than before. So, we're basically entering a new arms race.

17:33

Speaker A

Hackers can use AI to discover vulnerabilities, while developers and security companies use AI to find and patch them before anyone can exploit them. And as AI gets smarter, both sides are going to get faster. Now, we've seen a ton of serious security

17:48

Speaker A

vulnerabilities recently. So, let's go over some of the biggest ones. First, there's a critical vulnerability in PHP.

17:56

Speaker A

This is huge because PHP powers the backends of millions of websites, including many built with WordPress and Drupal. Researchers used AI to uncover two serious vulnerabilities that could let hackers run malicious code on vulnerable web servers without even

18:12

Speaker A

logging in. So, this caused quite a panic. A number of websites, including government websites, temporarily went offline as a precaution. But fortunately, the developers have already released security patches. Of course, website operators still need to install those updates to stay protected. Another

18:30

Speaker A

major security flaw was discovered in GitLab. This is a very popular platform that developers use to store code, collaborate on projects, and deploy software. And this one was a pretty big deal. It received a severity rating of 10 out of 10, which is the highest

18:46

Speaker A

possible score. Basically, hackers could potentially run malicious code on vulnerable GitLab servers without even logging in. And since these servers can contain sensitive source code and credentials, the consequences could be pretty serious. Fortunately, there haven't been any confirmed widespread

19:03

Speaker A

attacks associated with this vulnerability, and GitLab has already released a patch. There were also some serious security flaws discovered in NodeJS. This is another widely used technology that powers the backends and APIs of a ton of websites and

19:18

Speaker A

applications. The good news is that security updates have already been released to fix multiple vulnerabilities. So, if you're a developer who uses NodeJS, you'll want to make sure your server is running a patched version. Probably the craziest incident involved Emudra. This issues

19:35

Speaker A

digital certificates that browsers use to trust websites. It's one of the padlocks used over here and it's especially widespread in India. Now, this is supposed to rely on really strong secret keys that should be impossible to guess. The crazy thing is

19:49

Speaker A

researchers discovered that these keys were actually ridiculously weak. I mean, these were the keys. One of them was just the number four. Isn't that ridiculous? Even for normal passwords, you should have at least 12 characters with uppercase, lowercase, special

20:04

Speaker A

characters, etc. But it turns out that the keys they used were just this simple. If a hacker knew these keys, they could just forge a certificate for any website and trick the browser into accepting it as legitimate. For example,

20:16

Speaker A

they could forge a login page to your bank account, which the browser would think is safe. Now, these keys have since been revoked, but of course, they received huge backlash for such a pathetic set of keys. This was a huge

20:27

Speaker A

blow to the credibility of a mudra, and a few days later, they actually announced that they'll exit the digital certificate business entirely. Another huge one involved railway. This is a cloud platform that lets people host apps, databases, and back-end services

20:42

Speaker A

without managing their own servers. Now, what happened was hackers got access to their internal systems and decrypted some customer environment variables, which can contain passwords and API keys. Railway paused all new builds and deployments while they investigated, which led to a lot of user frustration.

21:00

Speaker A

People were told to set up new keys and passwords, but they couldn't actually deploy if deployments were paused.

21:05

Speaker A

Anyway, a really messy situation and yet another one is open code. This is an AI coding assistant which allows you to use different AI models. Hackers accessed an older system and potentially exposed API keys belonging to 1.4% of open code

21:22

Speaker A

users. Another one is light LLM, which is another platform for people to try different AI models through one interface. Apparently, hackers got access to the GitHub login of three developers and used them to push a malicious version of Light LLM. This

21:37

Speaker A

version could steal passwords and API keys from users. The team discovered this and was able to remove the release within around 2 to 3 hours, but who knows what information was compromised during that time frame. So, those were

21:50

Speaker A

the main ones. It's important to note that these weren't all part of one coordinated attack, and the internet didn't actually collapse. In fact, most major services remained in operation.

22:00

Speaker A

But just this week, there are more than a dozen of these critical incidents that have been found. And it's only going to get crazier. In fact, here's a chart showing the scale of what happened this week. Roughly 3% of all critical

22:13

Speaker A

software bugs in history have been found in just the last month. And we're only getting started. On one side, we're going to see hackers use AI to find and exploit vulnerabilities. And then on the other side, we're going to see

22:24

Speaker A

developers and companies also use AI to try to find and patch these vulnerabilities before the hackers find them. I think the biggest concern is that AI is making it much faster and cheaper to discover and exploit these vulnerabilities, and this could

22:38

Speaker A

potentially give companies less time to defend themselves. Anyway, let me know in the comments what you think. Also, this week, OpenAI published a blog post showing how thousands of AI agents can work together to speed up scientific discoveries. So they analyzed this

22:52

Speaker A

enormous data set with millions of mouse brain cells. Researchers had previously disrupted thousands of different genes and measured how the brain cells responded. But here's the problem. These experiments generate so much data that scientists could spend months analyzing

23:08

Speaker A

everything and looking for meaningful connections. So instead, the team deployed thousands of AI agents powered by GPT. Each agent investigated what happened when a particular gene was disrupted. It also searched existing scientific papers and tried to explain what the results might mean. Then they

23:26

Speaker A

combined all these findings into a massive collection of research reports. And here's the impressive part. The agents completed the entire analysis in just 32 minutes. If humans had done the same work, it could have taken several months to a year. Interestingly, they

23:42

Speaker A

also tested whether a single AI agent could handle the same task. At the largest workload, the single agent cited just 11 scientific references, while the team of agents cited over 700. So, this shows how a swarm of AI agents can

23:57

Speaker A

potentially do much more extensive research than a single agent. The AI also uncovered some interesting clues related to Alzheimer's and Parkinson's, including a possible connection between cholesterol regulation and these brain diseases. Of course, these are still just hypotheses, so scientists will need

24:15

Speaker A

to test them in real life. But what's impressive is that AI could help us discover valuable scientific connections that researchers might otherwise miss.

24:24

Speaker A

It could also help scientists figure out which experiments are worth pursuing next. If you want to learn more, they've released a full paper on Bioarchchive.

24:32

Speaker A

I'll link to the main page in the description below. Also, this week, we have a new AI world model called JST2.

24:39

Speaker A

This can take in a text prompt or a few images and turn them into an entire 3D environment that you can freely explore.

24:46

Speaker A

It can even generate 3D characters and animations. Plus, everything stays coherent as you can see from the map in the bottom corner. So, how this works is it actually represents objects as like tiny pieces of matter in 3D space. So,

25:00

Speaker A

it can keep track of their positions and how they move. So, even if you walk away and you walk back, the scene and all the characters stay the same. So, here are some additional examples for your reference. The cool thing is you can

25:12

Speaker A

also take the same model and create 3D characters that can already be animated. So, you don't need a separate rigging process. And because this world generator can keep track of everything in 3D space, you can also use this to

25:24

Speaker A

train robots, especially in terms of navigation and manipulating objects. In fact, if you compare this with other leading 3D model generators like Meshi or Trippo, you can see that this new JST2 on average gets a higher win rate.

25:38

Speaker A

Now, this doesn't seem to be open- source, but you can try JST2 for free using their Pinocch platform. You can try this for free right in your browser.

25:47

Speaker A

If you're interested in reading further, I'll link to this main page in the description below. Also this week we have a new AI called Sky001 and they claim this is the first world model for engineering. For example, you

26:00

Speaker A

can give it an image of a mechanical object and it can break it down into individual 3D parts along with the sketches and steps needed to construct them. So here's an example where we can take this Jaguar Roadster and break it

26:13

Speaker A

down into individual 3D pieces. Or here's another example where we can break down this jet engine into a ton of smaller individual components. The most useful part is that you can actually zoom in and click on any element and get

26:27

Speaker A

it to generate step-by-step sketches and operations needed to build each part out in CAD software. So here, for example, we can just zoom in on just this one component from this roadster and then get it to draw it out and then build it

26:40

Speaker A

in CAD software. And it turns out that this same model can also help with architecture design. So you can give it an outline of a building and get it to generate floor plans with individual rooms. So you can see here how it's

26:53

Speaker A

generating all these rooms sequentially like this. What's even more impressive is the speed and efficiency. So here is the time that it takes to complete a task. And as you can see, this is way faster than the other Frontier models.

27:05

Speaker A

Although keep in mind, these Frontier models are more generalist whereas Sky is specialized for engineering. So, it is expected to be a lot more efficient.

27:14

Speaker A

And then here is the cost per task. Again, this is way cheaper than the other Frontier models. Now, it doesn't seem like they will open source this model, and you can't access this yet, but if you're interested, you can join

27:23

Speaker A

their weight list over here. I'll link to this page in the description below. Also, this week, Kimmy releases their latest and best model, Kimmy K3.1 Pro.

27:33

Speaker A

As you can see from these benchmarks, it performs really well, pretty much on par with the Frontier models, including Claude Fable and GPT6 Astra. So, here are some benchmarks on its agentic coding, tool use, knowledge work, and other general knowledge benchmarks. This

27:49

Speaker A

is also incredibly efficient. If you look at its terminal bench score versus cost per task, then not only does this score the highest, but it's also much cheaper than the other Frontier models like GPT and Claude. So, here's an

28:01

Speaker A

example where you can get it to code up some pretty advanced looking games in Unreal Engine. It's even able to handle character rigging and all these animations. This thing is also incredibly good at 3D modeling. So, here's an example where we can get it to

28:17

Speaker A

model an indoor rainforest for an airport using Blender. It's also state-of-the-art in terms of financial analysis. So, here's an example where we can get it to do some financial modeling. And as you can see, it's able to work directly in any spreadsheet

28:31

Speaker A

software and do a ton of financial analyses like this. Now, if you look at this intelligence index by artificial analysis, then you can see that Kim K 3.1 is indeed among the frontier models, even beating Claude Fable 5.1 and GPT6

28:46

Speaker A

and Gemini 4. The cost per task is also very attractive. So, it only costs like two bucks per task, whereas the other Claude models cost at least three times more. Same with GPT6. It costs way more.

28:58

Speaker A

Now, this is a massive three trillion parameter model. This is much bigger than the other leading open models like DeepSeek, which are only like a few hundred billion parameters. And like all the Frontier models out there, this has

29:10

Speaker A

a 1 million token context windows. So, you can fit a ton of information into your prompt at once. Now, at the bottom here, it says that Kim K 3.1 is rolling out today on their platform as well as

29:20

Speaker A

through their paid plans and API. But like the previous Kimmy models, this will also be open- sourced on November 4th. So you can download the model and run it locally and offline if you have the hardware. But keep in mind that at 3

29:33

Speaker A

trillion parameters, this is going to be pretty huge and you're unlikely to fit this with consumer GPUs. Anyway, I'm planning to do a full review of Kimik 3.1 in the near future, so stay tuned for that. If you're interested in

29:46

Speaker A

reading further, I'll link to this main release page in the description below. Also this week, ideoggram releases a new model called ideoggram vector and this is really useful for graphic design. You see this is specialized for creating vector graphics. So things like logos,

30:03

Speaker A

illustrations, typography, designs etc. And vector is made of like mathematical shapes. So in theory you can keep zooming in and in everything will still remain perfectly sharp. And because of this vectorzation you can also edit individual shapes or change their

30:17

Speaker A

position or microedit certain things. The cool thing is you can also upload an existing image and have the model convert it into a vector. And what's really impressive is that these don't just look like flat simple vectors. Like

30:30

Speaker A

you can upload a really complicated image like this and it's still able to maintain most of the details of the original image. Currently you can try this on their ideoggram app. So simply click on generate vector over here and

30:42

Speaker A

you can enter in your text prompt or upload an image for it to convert into a vector. If you're interested in reading further, I'll link to this main release page in the description below. Also, this week, we have a new freakishly

30:54

Speaker A

realistic realtime voice model called Desa. Let me just play you a few examples. Hey, how you doing?

31:02

Speaker A

I'm doing well. Um, how are you doing with the winter stuff? Winter is coming for sure, but you know, is that kind of a sore spot this year?

31:14

Speaker A

Yeah, it's going to be cold. I'm not a fan of the snow. I prefer to have sun on my skin. You know what I'm saying?

31:21

Speaker A

Yeah, I can see that. Well, anyway, I've had such an exciting day at work today.

31:26

Speaker A

Oh, tell us. Yeah, we the team uh we've decided to release the speech to speech AI model we've built for the past.

31:35

Speaker A

Oh, whoa. I was so I have been watching your journey of creating this and what what did you think of it? Are you excited about it as much as I am? so blown away. I am so impressed.

31:48

Speaker A

It's so expressive, right? It is so expressive. It It's conversational and it's it's it's very conversational. It can laugh. It can understand you and your emotions much better than anything I've seen out there.

32:02

Speaker A

I haven't really seen anything like that either, but it's it's pretty much like human level. Yeah, it's it's a bit scary to be honest, but but I'm so excited to see where it would get us.

32:16

Speaker A

I'm so excited for it to be in the world because it's just amazing. We've had so many conversations about it.

32:23

Speaker A

I I I totally agree. What What are you excited for? What What's your vision for where this AI is going to be?

32:31

Speaker A

I Well, okay. So, the one thing that I've been kind of thinking about is advertising companies creating their own AI bot.

32:40

Speaker A

Okay. No, advertising. If you if you were able to talk to this AI, what will you tell it? What would you ask?

32:47

Speaker A

I would ask it so many questions. So, if it could be talking it I don't know off the bat, I would ask it for all the secrets.

32:56

Speaker A

Yeah. What? Like what secrets? like uh I I would honestly ask him like sorry you're you're a bit far from the mic. Can you come closer so I can hear you?

33:06

Speaker A

I would be like what is consciousness? Is it something that's programmed into us? Is it something that can be expanded on? Is it something that's interdimensional?

33:17

Speaker A

All right. So, as you can hear, this sounds super realistic. What's fascinating is that it learned to speak entirely by listening to human conversations without any text or pre-trained models. So basically, it learned to talk just by listening to

33:31

Speaker A

people. Kind of similar to how babies learn language. And because it learned from only speech instead of text, like large language models, this can even pick up subtle things in speech like accents pauses breathing sarcasm humor, etc. And the crazy thing is, at

33:46

Speaker A

least according to them, they only used one Nvidia GP200 GPU to train this. Now, the unfortunate thing is they're currently offering access to a small group of AI safety researchers. If you are a safety researcher, you can get in

34:00

Speaker A

touch with them and request access. But for the rest of us plebs, we don't have access to this yet, and I'm not sure if this will be open source. But if you're interested in listening to some more demos, I'll link to this page in the

34:11

Speaker A

description below. Also, this week, Anthropic introduces their new AI assistant called Claude Duet. This doesn't just wait for you to ask questions, but instead this is a proactive agent that can check your work, identify things that need your

34:25

Speaker A

attention, and complete tasks on your behalf, even if you turn off your computer. So, it's exactly like OpenAI's Dots or MetaMuse or Grockbot or OpenClaw. Think of it as like your personal assistant that can work throughout the day. So, for example,

34:39

Speaker A

every morning you can get Claude Duet to prepare a summary of your important emails, messages, calendar events, outstanding tasks. It can also monitor your connected apps and alert you when something important happens. You can ask it to manage emails, provide financial

34:53

Speaker A

forecasts, manage invoices, or even run an online store. And you can connect it with over 6,000 tools and integrations, including Gmail, Slack, Stripe, QuickBooks, and more. So, this uses Claude Opus 5.5, which is currently their best model, and it has its own

35:09

Speaker A

virtual computer and browser. So, it can do tasks independently even after you close the app or turn off your computer.

35:16

Speaker A

Now, unfortunately, this is only available by weight list and you do need to be in the US with a pro or max plan.

35:23

Speaker A

But if you're interested, I will link to this page in the description below where you can apply to join the wait list. In humanoid robot news, Unree releases another crazy demo. This time of their latest humanoid robot, which is called

35:37

Speaker A

H2X. Here you can see it doing some crazy things like ramming into this car with a knee and denting the car door and then it proceeds to destroy the car further. Now this is positioned to be more of a heavyweight serve and guard

35:50

Speaker A

humanoid rather than just a general household assistant. So it's able to do some heavy duty things like lifting a car or destroying a car or even intervening against an intruder. And apparently it has a standing endurance of around 8 hours. Currently this is all

36:05

Speaker A

we know for now. I will update you in my future videos once we hear more specs about this robot. Also, this week we have a ton of new open model releases.

36:14

Speaker A

So, let me rapid fire through them. First, we have this open model called Sage A1.5 by Alpha Innovator. This is a 35 billion parameter mixture of experts model based on Quen 3.6 and the earlier agents one model, but this one is

36:28

Speaker A

optimized for scientific research. So, you can ask it to research a complex scientific topic. It can split the work between multiple agents, gather evidence from different sources, even create figures and diagrams, and then put everything together into a nice detailed

36:43

Speaker A

and cited report. They also demonstrated how you can use this to design molecules or calculate how strongly they bind to other molecules. Or they also got it to analyze brain research data or even optimize chemical reactions. It can

36:56

Speaker A

actually review its own progress, identify mistakes, and change its approach if something isn't working. And if you look at these benchmarks, Sage A 1.5 beats many similarsized models across scientific research, coding, and complex agentic tasks. In some categories, it even outperforms much

37:12

Speaker A

larger frontier models. The awesome thing is they've released the model and code to this. So, if you click on this hugging face link, note that they released several versions of this. The full one is around 72 GB in size.

37:24

Speaker A

There's also a smallest Q4 GGF version, which is only 22 GB. And then on this GitHub, it contains all the instructions on how to download and run this locally on your computer. If you're interested in reading further, I'll link to this

37:37

Speaker A

main page in the description below. Also, this week, Miniax drops their latest model M3.1 Flash. And this is actually an upgraded version of the stealth model called Space Bunny, which appeared on Open Router a few weeks ago.

37:51

Speaker A

Now, this is a flash model, so it's designed to be fast and affordable. But what's impressive is that it can compete with some of the leading open models out there across various benchmarks on agentic coding, computer use, knowledge

38:04

Speaker A

work, and automation. Here, if you look at its cost per task versus its performance on this automation bench, you can see that it is by far the most efficient model out there. And if you look at Deep Suite, it's almost as good

38:16

Speaker A

as the latest Deepseek model, which I featured a few days ago. Now, as with the previous Miniax models, they say that the model weights will be released later. For now, you can use it through their API or their Miniax plans. If

38:30

Speaker A

you're interested in trying this out, I'll link to this main page in the description below. We also have this new open model called Mem 1.0 by MEM Tensor.

38:39

Speaker A

Now, what's special about this model is its memory. It's designed to remember information across conversations without having to reread your entire chat history every time. Here's a simple example. You can tell the model that you have a kitten named Mei who's 3 months

38:55

Speaker A

old and loves shrimp. Then you can remove that conversation from its context. But when you ask about the kitten again, it can still remember its name, age, and favorite food. That's because it saves important information inside a compact internal memory state.

39:10

Speaker A

So it doesn't have to keep the original conversation around. And if you look at these benchmarks, both MEMT models perform extremely well on tests of persistent AI memory, especially compared with other models that store memories internally. The awesome thing

39:24

Speaker A

is they've released the model weights and instructions on hugging face. So you can try this out locally. There are actually two versions of this. There's the smaller nano model with 3 billion active parameters and a much larger flash model with 288 billion total

39:39

Speaker A

parameters and 17 billion active parameters. The nano model takes about 70.3 GB to download, while the full flash model takes about 576 GB. And then further down, this hugging face page contains all the instructions on how to download and run this locally. If you're

39:57

Speaker A

interested in trying this out, I'll link to this main page in the description below. Also this week, Ant Groupoup drops their latest open medical model, LING 3.0 Flash Sante. It's built to understand medical information, reason through complex clinical questions, and

40:13

Speaker A

help researchers analyze scientific literature. Now, this is quite a large mixture of experts model with 124 billion total parameters, but only 5.1 billion are active at a time. And it's actually a specialized version of the existing LING 3 flash model. And if you

40:30

Speaker A

look at these benchmarks, it's very competitive with other much larger AI models across clinical knowledge, medical reasoning, and healthcare safety. The nice thing is they've released the model weights on hugging face. Now, the full model takes about 255 GB to download. So, this one's

40:46

Speaker A

definitely meant for powerful servers rather than an ordinary laptop. If you scroll down this hugging face page, it contains all the instructions on how to download and run this on your computer.

40:56

Speaker A

If you're interested in trying this out, I'll link to this main page in the description below. Also, this week, we have a new acceleration method for the best open video model out there, Miniax H3. So, it's called Fast H3 Omni. And

41:09

Speaker A

the nice thing about this is it's specifically designed for reference to video, not just text or image to video.

41:15

Speaker A

So you can just provide reference images of characters or scenes and it can generate videos while keeping these characters consistent the whole time.

41:23

Speaker A

What's especially impressive is how fast this is. So they managed to reduce this down from 50 steps of the base model to just eight steps. And they achieved up to 12 times lower latency when they ran it on 8 GP200 GPUs, which I'm sure you

41:38

Speaker A

don't have. So you probably won't get the same amount of speed up, but it's still going to be a lot faster than the base model. Here's a table where they tested out different reference inputs.

41:48

Speaker A

And the right column shows how fast this is relative to the base model. So if we input like three images, then it's like 5.8 times faster. And if you input a reference video, then it's like 6.98 times faster. Anyway, at the top, they

42:02

Speaker A

released the models to this already. So if you click on this hugging face link in this folder, it contains a ton of different variants of this fast H3 Omni.

42:10

Speaker A

They've made some models for a comfy UI. There's also an FP8 version, an FP4 version, and even a version for Apple Silicon. If you're interested, I'll link to this main release page in the description below. And that sums up all

42:23

Speaker A

the highlights in AI this week. Let me know in the comments what you think of all of this. Which piece of news was your favorite? And which tool are you most looking forward to trying out? As always, I will be on the lookout for the

42:37

Speaker A

top AI news and tools to share with you. So, if you enjoyed this video, remember to like, share, subscribe, and stay tuned for more content. Also, there's just so much happening in the world of AI every week, I can't possibly cover

42:50

Speaker A

everything on my YouTube channel. So, to really stay up to date with all that's going on in AI, be sure to subscribe to my free weekly newsletter. The link to that will be in the description below.

43:02

Speaker A

Thanks for watching, and I'll see you in the next one.

Topics: AI news open-source AI vision language model speech transcription image editing DNA language model Kimi Tencent cybersecurity Hugging Face

Study this video

- [Flashcards from this video](https://sozai.app/tools/ai-flashcard-generator/?from=transcript&id=86278)
- [Quiz on this video](https://sozai.app/tools/quiz-generator/?from=transcript&id=86278)

Free, made by AI from this transcript. No signup.


---
This is the markdown twin of https://sozai.app/transcript/ai-news-cybersecurity-dna-kimi-claude-minimax/ — the same content, without the markup.
Published by SozAI (https://sozai.app). Reuse and quotation are allowed with attribution and a link back.
Machine-readable index: https://sozai.app/llms.txt · data API: https://sozai.app/api/