Thinking Machines’ First AI Model, California Loses $3.2B to Texas, TSMC Adds $100B | Diet TBPN

0:01 We have a very special show for you today because we have Tyler Cosgrove guest hosting. >> He's here >> Great to be here. Chair. >> And I know I'm going to make mistakes today because I always throw it over to Jordy. I got to remember that's Tyler. It reminds me of this video from Warren Buffett. We got to play it to show what I'm going through emotionally today without Jordy in the TVP and Ultradoom. Let's pull up this video of Warren Buffett throughout the years at the Berkshire Hathaway shareholder meetings.

0:30 >> Warren Buffett always goes to Charlie after he gives his comment. He he he he he gives his speech and then he kicks it over to Charlie. >> Charlie through the through the years. >> Charlie? >> >> Charlie? >> Charlie? >> Tear jerker. >> Yeah, makes me want to cry. >> It's emotional. >> Charlie, how do you feel about this? Charlie and Charlie and me Charlie. I I Charlie. >> And that's Greg Abel. >> So if that happens today, I apologize.

0:55 but the first big story is of course Thinking Machines new model has released. Mir Murati's AI startup released its first model in bid to loosen AI giants grip. We're going to be talking about open source, closed source, where the frontier is, national, geopolitical model moves. >> And and some people had an idea this was going to happen, some inkling. >> had an inkling? >> That yeah, that they were going to release some. >> they? I didn't have an inkling that they were going to jump into the open source market.

1:21 >> it's actually I think it makes a lot of sense given the Tinker API, right? The whole business is is you know, you're doing fine-tuning on on open source models. It makes a lot of sense that they're going to have their own. Yeah. that you know, you can easily >> It's sort of like it it it they are set up as a business to launch an open source model without it degrading any other piece of their business because the Tinker API, that fine-tuning that they do, that integration with the customers that they have, actually benefits from open source. Yeah. And then they can go to their clients and say, "Look, you know, it's the Red Hat model. At any time you can leave because we are giving you the weights of the model open source. You can do whatever you want with them, but keep working with us because we're helping you a bunch and we're making money in the process. So, Thinking Machines Lab first the first model is an open weights model designed to chip away at the lead of Open AI and Anthropic, says the Wall Street Journal. Former Open AI technology chief Mira Murati is betting on more customizable artificial intelligence models to chip away at the lead of the frontier labs such as their former employer hold over the technology. TML, a company led by Murati, released its first AI model Wednesday and did it with open weights, meaning that others can modify it with their data. Called Inkling, the model has 975 billion total parameters, making it far smaller than estimates of the most advanced closed-source models.

2:37 >> So, only I think the number is 41 billion of those are actually like active >> moment. >> Yeah, so this is definitely on the on the bigger side of open source models. >> Sure. >> but like that number, it's not these aren't like dense models like what you would traditionally think of of the models like, you know, 4 years ago. >> Yeah, yeah. they Mira Murati told the journal, we trained it to be a broad, balanced, foundational foundation model strong across many domains, flexible enough to adapt. Inkling is not the strongest overall model available today, open or closed, which is a different frame of reference for many of these model launches. There's it's been Everyone's been jockeying for the frontier even if they're not world-class at everything. Usually when they launch, they say, "Oh, well, we're best at something or we're best at this." but a different tone, different communication strategy. And I think it's being well received. I think people are are having fun with it.

3:29 The real pitch >> Yeah, I mean, I I think the main pitch here is that this model is like uniquely set up for the tinker API. It's built to be fine-tuned. That's the whole point. >> it. >> DD Das says, "Thinking Machines just dropped the best open weight AI model outside of China." And obviously, that is a big topic of conversation as business leaders in the United States have some policies and some reticence about using Chinese open source models even if they're not worried about the the dystopian you know Manchurian candidate hidden inside the weights, maybe they just want to be aligned with a US-based company for a variety of reasons. Inkling beats Nematron 3 Ultra and benchmarks put it between Kimmy K 2.5 and 2.6. Of course, there's also news today that Kimmy K3 will be launching and is another jump forward, but there's back and forth between some AI researchers around what's going on there, how long that strategy will continue. So, DD says many were contending to this throne, but Thinky has come out on top.

4:31 Really solid release and will pair well with Tinker. So, there's some benchmarks that you can go and dig into if that's your thing. There's another very bullish take from Jack Morris of Engram Labs. He says people are under are underestimating what a big deal this is. This is the only open weight model that that's trained without distilling for Open AI from Open AI or Anthropic. Kimmy distills, GLM distills, Qwen distills, Nematron distills, Kimmy and Deep Seek which count. Basically, a fully different text text stack. The first pure open frontier coding model, very exciting. And there's a community note on this. Can you break down exactly like where are they standing on the shoulders of giants, where are they not?

5:13 >> So, I think this tweet is is not exactly true. In the blog post, they say to bootstrap post training, we ran initial supervised fine-tuning on synthetic data generated by open weight models including Kimmy K 2.5. >> Okay. >> So, I think that's like generally how people think of like distillation that they mean something related to this. >> Sure. >> So, I think that's actually is not that different than what the people like you know, at Nvidia with Nematron did. So, this is not like very new, I think.

5:40 >> sort of like the lightest touch of distillation that could happen cuz it's It's one piece of the pipeline, one small amount of data. It's not one of these scenarios where we're like, "Why is it identifying as Claude?" Or why is it Why Why is it just saying that it's ChatGPT? This >> But it is funny cuz you can kind of say like, "Oh, well, if this is like kind of distilled on on Kimi and Kimi's kind of distilled on closed source."

6:02 >> Yep. This sort of round trip loop. But at the same time, there's probably something to be said for the more layers of abstraction, the the the more, you know, ingredients you pour in, like the distillation becomes weaker and weaker. >> Yeah, and I I think there's also an important question of like, well, okay, that they're doing some level of distillation, like why? Because you can either be like, well, they're just doing it to save time, whatever. Like obviously they have these capabilities, but there's no point in in in, you know, doing everything over again. You might as well just just use what's out there already. Or is it because these capabilities that they get from this this, you know, distillation light, whatever it is, are those actually super imperative to the model like being good?

6:42 >> Mhm. And Grok says our founder, Jack Morris, recently issued some unfounded claims that got community noted. We deeply apologize for the confusion caused by his original post, the follow-up post, and the follow-up to the follow-up post. Nevertheless, we stand by his conviction in his own takes and in strong open-source models like Inkling. And there is a question of like, distillation is a vague term where it's not a binary thing. And if it's not in the pre-training data, does it count?

7:07 I don't >> I think it's also very much this this meme people love to talk about on X. They like to kind of, you know, scapegoat it. Oh, you know, it's all distillation. That's the only reason Chinese models are good. >> Yep. I mean, Anthropic's head of national security policy, Taryn Chabbra, accused Zhipu, Z.ai, of distilling both Claude and OpenAI models for GLM-5.2 at the Aspen Security Forum earlier this week. This is from Vincent Chao, senior AI reporter at SCMP.

7:34 He said it's the first time that they've named Zhipu specifically after previously calling out DeepSeek, Alibaba, Moonshot, and MiniMax join the join the club at this point. Also accused they also accused Deep Seek of continuing its adversarial campaign of distillation. Anthropic is now shutting down distillation accounts on the order of millions accounts of per week. That is crazy scale. You have to I I mean, you always think about it as like, "Oh, there's like shut down that one company or shut down that one block of IP addresses." But when there's a really really distributed attack. We've even heard about whole companies that just like resell Claude tokens or or GPT's 5.6 tokens. And that looks like a reasonable business because it's just a wrapper company. Of course, you want to work with them. But then you don't realize that on the other side, who are their customers? Why do they Why do they get to a 100 million run rate so quickly? Well, maybe it's a lab that's trying to distill through this pass-through entity. And of course, it's hard to like watermark the tokens once they go out the API. And they get passed through some other system. And they can go through other countries, all sorts of things. So, millions per week, that is crazy. That's got to be really difficult to It's a game of whack-a-mole.

8:50 They say, "GLM is {quote} probably the most advanced Chinese model on the market now, which poses significant cybersecurity challenges." They hinted that Anthropic will expand access to Mythos to ensure fair fight for cyber defenders. And they said that distillation challenge is real in shrinking USA a lead in AI, suggesting that the US government more to clamp down on Chinese model adoption globally by working with allies similar to trusted telecom efforts like Huawei and ZTE. So, obviously a hot topic, and people will be debating how how how exactly how heavy of a hand the government should be putting.

9:26 >> Yeah, I think this this release is also makes a lot of sense in I think it was a week ago there was that article about like Beijing is like at curbing overseas access to Chinese top AI models. Yeah. Right, so you're not going to be able to access the Chinese open source, right? It makes a lot of sense to to start doing American open source, Western open source. >> Yeah, it really does feel like there's a >> Like it feels very well timed.

9:45 >> Yeah, there's it seems like there's a pretty wide gap with between at least what's reported preferences from Beijing from the actual government and the companies. The companies are like, "Send us all the Nvidia chips. Let's distill everything. Let's and then let's open source these models and compete internationally." And Beijing's like, "Hey, maybe we need like, you know, an indigenous supply chain here. Maybe we need to, you know, lock down these models, keep our lead over here, go work internally."

10:12 I don't know. This was an interesting post from Grace Lee. She's She asked the question, "How did OpenAI's Soul finally learn design taste?" She projected 1,000 websites by GPT-5.6 Soul into a design manifold and discovered big holes. These holes were where GPT-5.5 previously generated outputs outputs with quote bad AI smell. So, there were you know, there's these tells in any AI model that it's not this, it's that, the M dash.

10:42 Once people start identifying those as, "We don't like that. It's too AI. It's too generic." One way it appears to actually sort of beat that out of the model is to actively avoid those specific things. And then she calls out three particular areas that have been avoided as anti-patterns. One, the Bento box layout in dashboards. Two, large typefaces in hero images. I did realize that that sometimes you would ask for a website and you would just get a massive block of huge text. And that's just not the way you when you land on a beautiful website, it's usually there's more wordsmithing, there's more >> Yeah, yeah.

11:19 >> first language. >> you know, you you make your first website with with 5.6 Soul or whatever and it looks really good. And then you make 10 and you're like, "Oh, wait. Okay, there's there's actually a lot of patterns I'm I'm seeing." >> Totally. Totally. >> And then you can start clocking them like everywhere you see you see a lot lot like cloudisms whatever on general design. You see them everywhere. >> if you don't come with any >> it's like the the the the, you know, high border radius on the edges. There's a little color on the on the side.

11:41 >> Yeah. especially if you don't come with like an opinion. If you come like we made a whole vibe-coded website in Codex for just the latest episode of Nick Bostrom on Joe Rogan. And I wanted to look like a UFC fight card and a fight promotional website. And it doesn't look like any like normal AI slop. I mean, there's still like AI-generated images. It looks like AI, but it doesn't look like, oh, yes, that's the Bento box layout, or that's the offset layout, or that's the purple, or it's stealing from linear.

12:12 >> Yeah. It's a completely different style. So, if you at least inject like one reference point, you'll usually land somewhere. >> Yeah. I mean, it's just interesting though this makes it seem like the new model is not necessarily doesn't have like higher variance with outputs it gives, but it we basically just found like, oh, there's certain examples that people really don't like. Let's just remove those. >> Mhm. >> But you're not necessarily like making the model more creative by you by removing these like patterns that always comes to you.

12:36 >> Yeah. Well, you're giving like the the flavor of creativity. And maybe that's >> Yeah, but you can imagine if we kind of keep the same model for 6 months, we'll just notice new patterns. >> Totally. >> And and you'll have this kind of same problem. >> the same time like Midjourney had like a very distinct look, and people like that look, at least some people. And so, if you can if you can quickly personalize and customize and land in a place where someone whose job is designing dashboards is happy every time with the layout. Like there is somewhat of a platonic ideal for some of these design patterns. And at the same time if you're yeah, working on certain like there's certain designs that are just like solved. Like, you know, make the call to action green or blue, not red, right? and so some of those like do you need to be consistent?

13:29 And then also I imagine that many folks who are using these tools like in enterprises are doing even if it's not a fine tune they're uploading a reference for everything that they're designing. So it's consistent with the brand that they've designed, right? Anyway, California forever lost a $3.2 billion shipyard project from defense startup Saronic after the company chose the Port of Brownsville, Texas over Solano County. I was like, "No." You got to give us a clap for that.

13:56 >> >> We got a Texan in the studio who's happy about that. This is bad news for California. We want California have a whole bunch of amazing stuff. Brandon Grahl who wrote the newsletter tbpn.com today was very disappointed about this. The automated ship shipyard known as Point Alpha, Port Alpha is expected to create roughly 10,000 permanent jobs along with thousands of union construction jobs. Supporters say California's lengthy approval process ultimately cost the state one of the first marquee tenants that California forever had pointed to as evidence its planned city could anchor a new era of American shipbuilding.

14:33 Joshua, executive director for the California Alliance for Jobs said California failed to move with the urgency the product required. Quote, "While Texas moved quickly and aggressively, thank you Jackson. California could not provide clear expedited approval process needed." He said calling the decision an enormous loss for Solano County, California workers, and our state's manufacturing economy. Earlier this year, California forever signed a 40-year construction labor agreement covering 70,000 acres and labor groups later backed legislation to fast-track environmental review and permitting for the proposed shipyard. The legislation has yet to advance. Instead, Texas approved a $211 million tax abatement package in June to secure Saronic's investment at Brownsville, roughly 20 miles from Starbase. Labor leaders said they warned that without expedited expedited approvals the project would leave the state, and that is exactly what happened. A project insider told the San Francisco Chronicle that California Forever itself remains on track, but acknowledged that losing a major defense contractor sends a powerful signal about the state's ability to compete for large industrial investments. Very disappointing, but I like Jan, I like the California Forever project, and I'm excited for where he takes it next. I'm sure he's on the hunt for the next major tenant.

15:47 >> Got to talk about TSMC. >> Yes, TSMC. TSMC both beat earnings and raised their capex guide. They're spending a lot more money, and >> pledged to invest an additional 100 billion in the US. >> Yes. >> Plans to spend a record amount, cementing its position atop the global semiconductor supply chain. >> Yes. but and and yeah, and they're they're they're investing another 100 billion in in Arizona fabs, but people are worried about overspending. The news is that the Nasdaq dropped 1% on TSMC's spending plans offset by strong results. Very very odd story that in in a time when even TSMC, which was not a particularly AGI pilled company for a long time since they've been through the smartphone boom, the so many booms and busts, so many cyclical buildout cycles, that when they are finally like, "Yes, now is the time."

16:37 People are like, " I don't know." >> Yeah, it's too late. >> They're they're skeptical. >> >> in creator world, there is some news from Colin and Samir. Lexus is now the official car of Colin and Samir. What does that actually mean? They made four ads for them that roll across the roll out across YouTube. They're sponsoring four videos on their channel. It's the first of its kind deal that represents a broader shift taking place in media. The aperture of what it means for a brand to work with the creators is changing quickly. It's very cool to see because obviously they've been on YouTube for a long time. They've done a lot of like host-read ads, mid-roll ads, but this is a much deeper integration and something that I think will be hopefully replicated all over YouTube and be a new source of revenue for for creators of all kinds. So, I was excited to see this. in other entertainment news, Jake from Econompic says this is almost hard to believe. Disney spent $129 billion acquiring Marvel, Star Wars, Pixar, ESPN, and Fox, which is $182 billion in today's dollars. Throw in all their legacy assets and the entire company's market cap today is $169 billion. Wow. Do you know what this picture is missing?

17:55 >> which What do you mean? Picture? >> So, so they're saying they acquired all these assets and the company's only worth $169 billion. What's missing from this analysis? The cash that's been returned to shareholders. Disney across dividends and buybacks has returned like 70 billion, maybe more, to shareholders, which is I don't know. I just thought this and and it is it is it is an interesting angle because they have spent a lot acquiring and the company is not worth more than what they acquired. So, there's this question of like where those where where those acquisitions are creative or destructive or dilutive.

18:31 But there is a whole separate picture, which is that a lot of cash has been returned to shareholders throughout this journey. >> Yeah, also I mean the the mechanism with with with which those acquisitions were funded also >> Yeah. >> should >> Yeah, matters a lot. I don't know. It was sort of interesting. Sean Frank has a pitch. He says you should move to New York City. Bro, you got to move to NYC. The weather, horrible, 100° easy.

18:53 >> >> AC, F that. Taxes, so high. Rent, highest in the country. Air quality, some of the worst in America. Tech, bro, we banned we Did they really ban Waymo in New York? >> I believe so, yeah. They didn't allow Waymos. >> Wow. That's very wild. yeah, if you can make it here, you can make it anywhere. so, I don't know. Do you ever have aspirations to move to New York City? >> at some point it seems >> to New York City?

19:17 >> Yeah. What do you think? >> It's a nice city. That's the thing is that all of this is true and it's still a great city to hang out in. It's so fun, so dense. You can see so many people walk around. It's beautiful. It's just like I don't know, it's unlike anything else. Still great, but yeah. You never lived in New York City. City, but I've spent like a lot of time there. So, I've had a good time. Thank you to everyone who tuned in in the chat. Thank you for positive reviews of Tyler.

19:42 Let us know what you think of Tyler. Leave us a review on Apple Podcast and Spotify. Write us an email. tell us how he did. I think he did fantastic. Have the best Thursday of your life. Yes. Have There you go. That's a good impression. That's a good impression. But but thank you. Sign up for our newsletter at tbpm.com and we will see you on Monday. See you. Goodbye. >> Incoming flashbang. >> Oh, we got the flashbang. There we go.

20:07 >>