NVIDIA CEO AFFAIR, Fable 5.2, Gemini Carbon, Meta Muse Ember, GPT Image 2.6, Minimax H3: AI NEWS

0:00 Last week, rumors started swirling around that Jensen Hong, the CEO of Nvidia, one of the most important people on Earth right now, potentially was having an extrammarital affair. And on

0:13 Sunday, he admitted it. He basically said, "I had an affair. I was unfaithful. I apologize to my family. I apologize to Invidians and shareholders, but I'm staying." And here's the thing,

0:28 he is an incredible CEO, an incredible technologist, again, one of the most important people on Earth right now, but this is a really bad situation. An affair is one thing. An affair with

0:40 potentially a Chinese spy, and that's what the intelligence agency in the United States says is a major, major deal. Now, the board is hiring an independent investigator to basically

0:52 figure out what happened. Jensen says, "No, I did not leak any information to this person, but we will have to see." And he did say, "Yes, I'm staying at the helm of Nvidia. I hope this ends up just

1:05 being an affair because if he truly did leak sensitive information to a Chinese spy, this is going to send absolute shock waves through the AI industry. Nvidia is as strong as ever. They are as

1:20 important as ever. And I hope that this entire situation doesn't distract the industry more so than it already has. And next, just yesterday, the best model on the planet just dropped. It is Fable

1:33 5.2. I made an entire video reviewing it and showing you a bunch of the most insane demos I've ever made. So, I'm not going to do that again. But, we have some new independent benchmarks from

1:44 artificial analysis that I want to go over. And the reason I want to talk about Fable 5.2 too is because right after this and just today, Gemini 4 Carbon dropped. And Gemini 4 Carbon is

1:57 the absolute frontier of what Google has to offer. And I want to show you the differences between these models because the benchmarks do not tell the full story. Let me get to that in a moment.

2:08 So, this is the artificial analysis intelligence index. And what we have here's CloudFable 5.1 Max down here at previously 49. a massive jump all the way to a substantial lead at number one.

2:22 Claude Fable 5.2 Max. Interestingly, the first, second, and third place models are all Claude Fable 5.2. Then right after it, Gemini for Carbon. Now, there is basically a twopoint difference

2:36 between Carbon and Fable 5.2 High and Fable 5.2 extra high. But if you're actually using these models, and I'm going to show you some demos in a second, there is no way that that is

2:48 justifiable by artificial analysis. Either there is some benchmaxing happening or something is going on because when I'm using the model, there is such a distinct difference between

2:59 the quality of Fable 5.2, to its ability to do basically anything and make great decisions on my behalf and make assumptions that make sense to me and how I'm using carbon. Unfortunately, I

3:13 have had early access to carbon. I've been testing it and I just keep going back to Fable. So, we also see of course Fable is very expensive. So, this is intelligence index versus cost per

3:25 intelligence task. that is how much does it cost to complete a task and that is a combination of how many tokens it takes plus the cost of those tokens. So really intelligence density plus cost and what

3:37 we see here is cloud fable 5.2 is the absolute frontier as I said but it is also the most expensive. Now here's Gemini 4 carbon right here sitting a number of points below Fable 5.2 to, but

3:52 also a fraction of the price. And this is really important. Not all tasks take the absolute frontier. Now, I have to admit, I have become addicted to Fable 5.2. Even going back to Opus 5.5 seems

4:06 like a pretty massive downgrade to me, but Fable is very expensive. If it turns through your quota very quickly and even with the three free resets that Enthropic just gave out, which you know,

4:19 that's fantastic. It just feels like you're just going to burn through your tokens, especially if you're trying to get, you know, extra high or high thinking effort to really ek out that

4:30 performance out of the model. It's not easy to keep up with all of these model drops, but of course, I'm excited to do it. And one way that can help me do it is by using the sponsor of today's

4:39 video, Victor. Every team has the work around the real work and that's what today's sponsor Victor is all about. So here's an example for a team like mine. The obvious job is the overnight AI news

4:53 pass. We have to know about each release and I have to know was it actually shipped? Is it vaporware? Is it just announced? What's actually happening? And so Victor as a full AI employee can

5:06 do that for me. He lives in Slack and Microsoft Teams, connects to over 3,200 tools, and actually does the work rather than just answering questions about it. And my whole team shares him, and he

5:20 stays within guard rails. He proposes something the team approves. Nothing goes out until the team says so. Go hire Victor right now and get $100 in free credits with code Matthew Berman 100.

5:32 I'm going to drop the link down below. Let's get back to the video. And it's basically the number one model against every single benchmark. HLE Diamond, AA Omniscience, Terminal Bench, Frontier

5:44 Code. Now, there are a couple exceptions. AA Briefcase, Sonnet is still number one. That is a knowledge work benchmark. And Sonnet 5.5 was really built to be a knowledge work

5:56 expert. So that's why it's number one. Here we have automation bench carbon number one. Here we have GDP. PDF, which is professional document reasoning, and Muse Wonder coming in at number one.

6:08 Terminal Bench 4, surprisingly, Kimmy K 3.1 Pro, but honestly, these benchmarks I start to have less and less faith in given how much I'm using all of these models right now. And it is so clear

6:20 Fable is far and away the best. And that leads me to this next important point, something that I've talked about pretty extensively over the last 6 to9 months, which is how is anybody going to catch

6:30 up with Anthropic at this point? I made a video a while back where I talked about their incredible flywheel that they have. They hyperfocused on building the best coding model on the planet.

6:41 Then they sold that coding model to enterprise and that was the first best use case. So they started generating insane amount of revenue. They also got to collect an incredible amount of data

6:54 because everybody was using it. They took all that revenue. They took all that data. They bought more compute and then they also trained their next model. And just as I covered recently and a

7:04 while ago, recursive self-improvement is in play right now. So everything they're doing, all of their additional research, building of new models is all powered by the previous generation of model and

7:16 anthropic tends to hold on to models for longer at least seemingly I don't know this for sure than any other company. meaning they probably have Fable 6 by now and they're probably training Fable

7:29 7 by now. And so we only get the models once they are far enough ahead where nobody can catch up by using their models. And so I really think that's the world we're in right now. And Carbon is

7:39 a good model, don't get me wrong, but it is nowhere near fable. It is like another level difference. All right, so let's talk about Carbon now. Carbon is now available. We did get the benchmarks

7:51 a few weeks ago and then they did some early access testing. I was lucky to be a part of that. They also gave it to security researchers to do some testing. But now it's officially out. You can use

8:00 it immediately and it's a good model. There's nothing wrong with it. It's not great. It's not near fable. It is not even near Opus to be perfectly honest, but it's a good reliable model. Now,

8:14 this first demo I'm going to show you is actually quite impressive. This is why I say Carbon is good but not Fable level. So we have a full Microsoft Excel clone. And if you're just looking at this

8:25 screen right now, it is identical. And if you look, most of the features that you would expect in Excel are here. This is fully browserbased. It is snappy. It works quickly. So I actually asked

8:36 Gemini how many tokens and how much time did it take to build this? And it told me in total it looks like it took about 3 hours across multiple agents but end to end about an hour and 40 minutes. Now

8:48 total tokens 26 million which is like nothing. That's a couple dollars to build a clone of Excel. So it it is quite good quite cheap. All right. Next I had to build a

9:01 fall guys clone. This is a common test that I've had all the recent models do. So, I'm going to throw up on the screen a few other examples of this so you can see the comparison, especially Fable

9:10 5.2, which was by far the best version of this I've seen made so far. All right, let's play and let's see how it does. So, you can already see some glitching.

9:20 Like, the level changes. Uh, the little beans look pretty good. There's only minor collision detection amongst the beans. There's a lot of visual clipping and weird stuff. The physics don't feel

9:30 great, but, you know, it's certainly better than a lot of other models I've tested it on. Here's a Rube Goldberg machine that I had at Build. This looks pretty good. Uh, I'll just show it to

9:39 you really quickly. And you can see the ball hitting all the dominoes and kind of moving each part of the Rube Goldberg machine along. This is pretty impressive. Definitely not a bad

9:50 demo in the slightest, but again, like once you get Fable 5.2, is there really anything else? So, go try Carbon. It's not that expensive. All right. And next, we have a bunch of new open- source

10:02 models. And I'm really trying to include more open- source models in my videos. So, this is from a company called Lancat, and they released LongAT 2.5. It is a 1.6 trillion parameter model, which

10:16 is a really big model, like for a company I had not heard of before this, it is quite a big model. 48 billion active parameters, multimodal, stronger visual understanding, and it's good at

10:29 creative design and 3D tasks. You can download this model right now. You can try it out. At 1.6 trillion parameters, most people will not be able to run this locally. Next, another company I hadn't

10:40 heard of, Recursive Intelligence, and this is Ripple 65B preview, a 65 billion parameter dense LLM. It is a US-based company, and it is an open-source openweights model. And so, shout out to

10:54 Recursive Intelligence. I am always appreciative when we have a new open weights model. some of the benchmarks. Here it is on terminal bench and they are only comparing to other open- source

11:04 open weights models and it is extremely strong. It is a very strong open weights model compared to other open weights models getting an 80.5% on terminal bench 2.1. Really good. This model

11:16 supports up to a million token context length which is excellent. It's a small model so you can fit it locally including they say on a MacBook or DJX Spark. So very very cool. It is

11:29 extremely inexpensive. 20 cents per million input tokens, 60 cents per million output tokens. But one thing I'm noticing is we're getting multiple US open-source openw weights companies

11:40 building these kind of small to mediumsiz models that are really good. And one of them is Mirror Morati's company thinking machines and they just released a great open weights model. So

11:50 introducing Inkling too small scoring a 31 on the artificial analysis intelligence index. It's actually a really good small model. Look at this. This is kind of like a Sim City style

12:03 game. Here is a from scratch PC emulator written in Python. Here is a little video game, kind of an origami style video game. This looks fantastic. And let's look at some of the benchmarks.

12:15 So, here's Deepseek V4.1 Flash, which currently is the best open source openweights model on the planet right now. And it actually competes very well. And it is a much smaller model than

12:27 that. So inkling too small, Frontier Code 25 versus 28, Deep Sui 72, which is kind of nuts for a small model versus 77. We have SUBench Pro 86, Post Train Bench 31. So yeah, go try this model. Go

12:43 download it. You can use it right now. You can probably even fit it on your local computer. So go try it out. And next, Meta has released Muse Ember, their smallest, best model. And this is

12:54 intelligence versus cost per task. And this model is competing against the smallest of small models. So Claude Haiku GPT6 Luna and this is the Paro Frontier right here. It scores highest

13:07 than Clad Haiku but it is much less expensive about a third of the price which is great. Here's Automation Bench absolutely dominates Haiku and Luna. Agents last exam scores a few points

13:20 higher. Terminal Bench 4.0 and know barely eesks out a win but still a win. OS World which tests the model's ability to control a computer and it definitely scores the win but just barely. And on

13:33 artificial analysis intelligence index ties haiku 5.5. So these are the workhorse models. This is the type of model where you plan with a fable you plan with maybe even muse wonderful full

13:45 and then you delegate out to this small model and then you just get so many more tokens. Use that strategy. I highly recommend it. OpenAI released GPT image 2.6, their new image model, and it seems

13:57 like maybe OpenAI is the only company with a good image model or the best image model of all the Frontier Labs right now. One of the coolest new features is that you can generate GIFs,

14:09 which I am very excited about. It also has faster image generation. So, you can see right here it is about 50% faster. Here are two GIFs. And so now you can generate gifts. And this new model is

14:21 going to be available in Chad GPT codecs and the API. Now there are two versions of this model. Sunburst and flare. Sunburst is if you want higher quality. Flare is if you want speed. And look at

14:33 this pricing. It's interesting that they have input and output pricing. $8 per million input. I don't really understand how that works because image models don't have tokens as far as I know, but

14:45 that's the pricing right there. Google released Gemini Omni 1.2 Flash. This is their video generation model. And Google is one of the only companies, one of the only Frontier Labs that has a video

14:57 generation model at this point. Remember, OpenAI shuttered Sora. Faster, 30% lower cost, you get longer clips, you can edit directly in the conversation, more natural speech, more

15:08 controllable shots. And so if you're looking to do video generation, this is a fantastic option. And next we have Bite Dance Seed Dream 5.5. Probably the best image generation model out there

15:20 right now. I mean, look at the quality. It is hyperrealistic. The images look flawless. Like if you're just looking at this, you I would probably think this is absolutely real.

15:32 This is so good. You could do multi-turn editing, so it keeps the image consistent and you just tell it exactly what you want changed, which is fantastic. I mean, look at the detail on

15:42 the hair. Look at the skin textures, the coloring, the fabrics on this woman's sweater, even on the fleece on the outside. Look at You can see the waiter walking by gets some blur where the two

15:55 people at the dinner restaurant are completely in focus. We have more blur on the background even though they're not moving. I mean, the quality of these images is unlike anything I've seen

16:05 before. And remember how I was just talking about Inkling Small from thinking machines? Well, Cognition took that model and post-trained it and made it even better. And this is their own

16:16 sui 2 flash, our new fast and efficient coding model. And here's inkling small. Here's Beam 1.0, Musepark open, GPT6 Luna, Haiku 2. This is a really good, really fast model meant to be delegated

16:33 to. It is meant to be used by sub agents in coding tasks. So again, if you are planning with a big model, an expensive model, then delegating to one of these smaller models, you're just going to be

16:45 able to maintain your quota for much longer. And next, we have an open weights image generation model based on miniax, which is interesting because miniax has been fine-tuned, post-trained

16:57 multiple times to have image generation models, video generation models, and I mean just last week we had Pika, which came out, which is based on miniax. So, it seems like some of these models just

17:09 come out and every other company gobbles them up and tries to make their own version of it. The most well-known version of this is Quen 3.8. Every single open-source open weights company

17:20 took that model and just made it their own, which is amazing. And now seems like everybody's doing that on the visual side with Miniax. We have Miniax H3 image. This is a fully open- source

17:32 open weights model you can download right now to generate phenomenal images. Now, it is 68 GB, so not everybody's going to be able to run this natively, but you can get a nice quantized version

17:44 at 18 GB, which most computers can run nowadays. So, I'm going to drop the link to this down below. It has all of the instructions of how to set it up, but of course, you can just point your Asian at

17:55 this hugging face page and say, "Download it and get it running for me." And so a bunch of awesome open- source models, some great closed source models to choose from, but really thing that

18:06 I'm thinking about all the time is Fable 5.2. And you should definitely check out my full review with all of the most insane demos I've ever been able to create. Go watch that video right