THIS WEEK IN AI: Kimi K3, OpenAI's Alexa, Grok Stealing Code, Thinking Machines

**Ejaaz** (0:00)
For years, the battle between open source and centralized AI models has been very one-sided. Anthropic OpenAI has always dominated the field until this week where Moonshot Labs, an AI lab out of China, has finally released their next model, Kimi K3. And it is fable five worthy. Now, the craziest part about this is Chinese and open source models in general have been behind American frontier labs. They haven't been able to close the gap. Aro recently said that that gap is roughly around eight months having closed from about a year. Today, with Kimi K3, that model has closed it to today. It is as good at one-shotting visual prompts as fable five is. It's amazing at reasoning in general. And we're going to get into a bunch of cool demos that we see on our screen today. Other big news in open source, Thinking Machine Labs, which, funnily enough, raised the largest seed round ever, $2 billion, headed by ex-OpenAI CTO Miro Morati, released their first model called Inkling. It's small yet pretty mighty. We're going to get into that. And then OpenAI themselves have revealed some secretive details about their new hardware device. If you've watched the show at all, you know Josh and I are obsessed with this. We're excited to get into the details and show you what is coming soon.

**Josh** (1:14)
Yeah, so to start, we have a new king on the block. And that comes in the form of Kimi K3. I mean, this very much feels like, I wouldn't say the DeepSeek moment, but something somewhat familiar, in the sense that this is a really novel, leading edge open source model coming out of the Kimi team. And it was paired with this really lovely launch video that actually, I was shocked that it was coming from this lab because this looks like something Apple would release as a product launch video. It was beautiful, the sound design was amazing, the visuals were great. And it's funny, they launched this before they even officially launched the model. In fact, at the time of recording, they haven't even officially stated that the model is live, but it's available. It is available in the Model Picker for you to go and test it. Now, the model comes in the form of about a 2.8 trillion parameter model, which if you listened to the episode yesterday featuring Grok, that's about double what Grok 4.5 was in terms of parameter count. Based on the early reviews that I've seen, a lot of people have been using this model kind of early. It seems like this is pretty incredible.
It seems like people have been able to do really impressive long horizon tasks. They've been able to do really impressive coding generation, gaming generation, 3D generation, and a lot of people are comparing this to close to a fable worthy model. Early testers, they put it above GPT 5.6 actually on coding, and almost fable worthy at coding, which puts it in a really solid spot because the pricing of this thing is going to be very low, and the quality of it is incredibly high.
To our point yesterday, talking about the two frontiers, there is the intelligence frontier and there is the cost frontier. This is placing a new point on that Pareto curve of cost and intelligence at a part of the curve I don't think anyone has ever made to before. This is very much a frontier model in terms of cost per intelligence. It's really impressive.

**Ejaaz** (2:55)
I think this tweet that I have over here summarizes it the best. He goes, the more I test K3, the more it feels like another DeepSeek R1 moment. It is often fabled level, maybe a little worse, but consistently better than 5.6. That's a recurring theme across a lot of different takes from early testers that it's not quite as good as fabled. It's as good as fabled when it comes to certain things, but not everything. But it is certainly better or on par with GPT 5.6, which is a bold claim, obviously, because OpenAI has invested billions and billions of dollars into their thing. Now, enough of us claiming theory. Let's talk about some actual examples. So on your screen right now, you're looking at the promotional video from the Moonshot Labs themselves. Now, what I'm going to shift to right now is Kimi K3 one-shotting the video entirely from scratch. I'm building it from scratch.
Oh, that's really cool.
Exactly. This is one single prompt. It was fed the promotional video that Moonshot Labs themselves created from scratch using humans and all that stuff, and it only took 25 minutes to do. Now, rumor, the cost to create this thing. Josh, actually, take a guess. How much did it cost to create this video?