Kimi K3 & AI’s Price War, What’s Happening To Google?, OpenAI’s Partner Trouble

0:00 The walls are coming down for OpenAI and Anthropic as cheaper models and new open-source competitors from China challenge their ability to make a profit. Google, meanwhile, delays its flagship model and seems to be spinning its wheels. And why can't OpenAI keep its partners from hating it? That's coming up on a Big Technology Podcast Friday edition right after this. Welcome to Big Technology Podcast Friday edition where we break down the news in our traditional coolheaded and nuance format. It's the Kimmy K3 episode. We're going to talk all about this new challenger from China, a 2.8 trillion parameter model that is taking on the frontier and beating basically every model except for Fable. So, we'll get into the implications of what happens after that and whether a price war is really getting underway, especially now that Meta and SpaceX have released cheaper models. We're also going to talk about the state of Google, what's going on there, why is their latest model delayed, and of course, OpenAI. We didn't even get to it last week. By the time we recorded it was it was too early because just a few hours later Apple would announce they've they had sued OpenAI. So of course today we'll talk about the lawsuit and more importantly why OpenAI cannot hang on to its partners at least not for long. Joining us as always on Fridays to do it is Rajan Roy of Margins. Ranjan good to see you.

1:17 >> Happy Kimmy K3 day. It's the Kimmy moment. Are you ready? I sure am because this is a massive week, a potentially earthshaking week in the AI story with the entry of Kimmy K3 into the conversation. So, let me just read the story from Bloomberg and then we can discuss it. Bloomberg says, "China's powerful new AI surprises investors, fueling a tech. A surprise breakthrough from Chinese AI startup Moonshot rippled through global markets Friday, sending AI and semiconductor stocks sharply lower as investors drew parallels with last year's Deep Seek moment. The catalyst was Moonshot's new Kimmy K3 model which the company said rivals the strongest offerings from OpenAI and Anthropic. The launch was quickly dubbed the new Kimmy moment. This is a interesting quote from V Cern Ling who's a managing director at Union Bayare Privy. He said people are worried that if US companies start using Chinese models more and anthropic less anthropic will invest less. That means US firms will lower the capex and and in the end the chip demand will be affected. So basically here's the story. OpenAI and Anthropic own the frontier but there's a set of models that were effectively you know 10 15 months behind them that were basically jumping in and something that you could route like your lower intensity tasks to. All of a sudden, Moonshot, which has had had basically released the very successful Kimmy K2, comes out with this new massive 2.8 trillion parameter models, model Kimmy K3, and it is right there with GPT 5.6 Soul and Opus 4.8 on almost all the benchmarks. In fact, it even beats Kimmy K3 on something called Program Bench. It beat sorry, it beats Fable 5 on Program Bench. It beats Fable 5 on SWE Marathon and it's right there in league with you know all the other top models on on you know all of the benchmarks we look to to assess model quality. So this is a it's it's a big moment because it seems to show that China's open source movement is not the 10 or 15 months behind US AI but maybe four or five months at the very at the most maybe even less than that. And when that happens, you know, your your rationale for going with, you know, the closed more expensive model as opposed to one of these open AI model, sorry, one of these open- source models, gets less and less and you start to wonder, you know, do these is there actually a benefit in building frontier intelligence if you're going to be equaled this quickly. So that is my outlook on it. Ranjan, what do you think about this?

4:07 >> I think this is a massive moment. I do think this is actually on par with the deepseek moment because it's the same principle at work. Again, it's the idea that, you know, like as you said that the frontier model, why invest in it? Why is it so important? Is that truly a moat and a competitive edge versus to actually build things that work? Is there a cheaper and better way? And this reminds us and I mean I've been saying this for a while like is are frontier models required for the majority of tasks that are be going to be done and become agentified and I don't think they are and now having a very powerful model that's just a lot more affordable and is also open source I think is exactly where companies are going to go. I think let's hold off on terms of what it means for the overall US versus China tech I'm not going to call it a tech war but I think before getting there I do think this is going to already in the last few months so many of the conversations I've been in have shifted to model interoperability what is the best model for the best task and now even more so realizing that actually why do I need Fable why do I need 5.6 six when I can actually have all these other options and the model gets a bit more commoditized and it becomes about the harness and the process and the data. I think this is this I don't I don't want to say it's like a transformational moment but I think this is actually going to be a moment where even more so the idea that frontier models are a massive moat really goes away.

5:49 >> Well, here's the thing. So Kimmy is not massively cheaper than any of these other models. and it also doesn't really beat, you know, the latest Frontier models on these benches on these benchmarks. So for instance, it doesn't really beat Fable and in you know most benchmarks here. It beats 5.5 and 5.6 from OpenAI. It beats Opus 4.8, but it doesn't beat the others, right? It doesn't beat sort of the the top-notch Fable models. Okay. So, it's not marketly better. It's also not that much. It's not cheaper than Gro's 4.5 model or Meta's Muspark 1.1. Those are cheaper and those are also competitive in some benchmarks with the with the Opus 4.8 and the GPT 5.5s.

6:40 >> All right. So, >> So, go ahead. Go ahead. I'm just setting the table here, but go ahead. >> No, no, of course. Of course. to to me the benchmark side like the kind of like what is that kind of final delta between you know on the deep SWE benchmark or the front like to me that really is less important than if it is in the general quality range of a 4.8 or a 5.5 you're in business. So, I think on that side, the fact that on some it actually beat Fable, on others, I love that we just actually threw in Grock 5.5. I guess they're still he's still going for it.

7:18 >> No, they are. They are. Grock 5 4.5 actually, you know, was competitive on coding benchmarks. It's apparently more token efficient, and most importantly, it's cheap. So, this was also since the over the past 8 days, we've seen Meta and Grock show up to the game. Not in models that are better than let's say 4.8 and 5.5, sorry, Opus 4.8 and GPT 5.5 and 5.6, but like you said, models that are almost as good that deliver this at 25 or 50% of the price. Well, so but but that's why I actually what I found most fascinating about the way that this has been released is it's not massively cheaper, but it's cheaper. $3 per million token input, $15 output. It's 40% cheaper than GPT 5.6, 70% cheaper than Fable. And again, Meta's already come in hard. Grock's come in hard as well. I think this is I what makes this even more of a significant moment is a Chinese company coming in and saying we're not going in as like significantly cheaper. We're go we're actually battling on quality and a little bit more affordable and you have more control over it which is why I I think it's it's very different. Deepseek's whole thing was it pure price like can we get something that's not even as good but just in the general vicinity and far cheaper. Now this is the first time we're actually seeing no on actual quality it's competitive. again is it better or not as good as Fable 5 or GPT 5.6 I think we'll see over time but I mean this is not dirt cheap stuff. This is not a something you buy off Teeu.

9:03 This is actually good quality and reasonably affordable. And it's just making us realize that like again to me the front and we can get into that idea of like is that going to actually completely distort the investment cycle because Anthropic realizes Frontier is no longer the moat. So they're not going to vest. I think we should get into that. But I think it's a big moment cuz now more and more you're going to hear everyone talking about what's the most efficient and cost-effective model for the task and I'm already hearing it but now it's I think it's going to be far more whereas six months ago it was like >> how could you not do anything on the frontier model it's obviously the best.

9:47 >> Yeah. So there's there's a couple important points here. the first is that you know in the past while you might think okay I need the best version of intelligence so you would h you'd bring in anthropic or open AI is forward deploy to engineers to build something for your company with their best models. this model from Kimmy is going to for moonshot is going to be open weight. So what you could do is basically if you get the right people in you could download the weights you know and sort of build it build your application for yourself with it. although it is massive. So you need a lot of infrastructure to do this, right? So that's going to be for like let's say the governments or the JP Morgans of the world. We've already seen Apple do a version of this with Gemini, right? So the question is does Apple go to Google or OpenAI or Anthropic to do something like it did in building Apple intelligence on a really good model and we've seen the the progress they were able to make there or does it work with an open source model. the other side of it is you're going to get this this model downloaded and put put on all the clouds and that's where the commoditization comes in, right? Because right now the pricing that you listed, that's the pricing that we get from Moonshot, right? Can these clouds find a way to deliver it even more efficiently?

11:01 So giving people frontier intelligence at an even lower price. and that's when that's when you really get into an interesting interesting world where like the the business on the API side, the business of OpenAI and Anthropic was we are going to build a model that is that much better than anything else out there and we're going to charge you a premium to use it. Right? So, we will mark it up in a in a massive way. And when you have these two forces coming at it, you have the Metas and the Gros coming in with comparable models at much cheaper and then you have the the open- source Chinese models coming in and giving you effectively the same performance as some of your better models. you know, it it it it at a price that's competitive, it does get you to a point where you start to ask from the API side, is there profit? Does this all become a commodity? And then let's also recognize the fact that like Anthropic completely shifted its strategy to the enterprise and they did a very good job of that. Open AI, it hasn't really like become as dominant in that way, but they're certainly investing very heavily and they've made significant moves around enterprise. And what you just said right there, the companies like you and I are not going to be cranking out our own version of Kimmy K3 on my MacBook Pro. It's a pretty good MacBook Pro, but it's I'm not 50 of them.

12:31 >> Yeah. Yeah. Exactly. So, but large enterprises will potentially or as you said, the cloud services will be kind of bringing in their own offerings here. So I think OpenAI's story and their pivot gets hit harder than anyone else with this announcement because suddenly why OpenAI becomes a much more salient question than it was just a week ago. >> Yeah. I want to read you an analysis from Gavin Baker who sort of who's an investor managing partner at RCD's management. He's done this the podcast circuit. but his analysis on this was was actually excellent and it kind of shows you where the challenge hits and where the benefit comes. So he says Kimmy K3 may be an important inflection point for AI potentially negative for anthropic and open AI while being net positive for essentially every other company in the world. A world where there's only two to three dominant frontier labs with 90% inference margins is net negative for every other layer while being awesome for those two to three labs. Those labs would become monopsinies for power, data centers, semiconductors, and hyperscalers and would obviously vertically integrate over time into all those layers while also completely subsuming the application software layers. Anything that lowers the margins and increases competition at the model layer is good for every other AI layer. Power, semiconductors, hyperscalers, neoclouds, and yes, even software. I think that captures it.

13:58 >> Yeah. And also monopsis. Monopsin is one of my favorite words. I still remember it from undergrad grad econ is where there's one dominant buyer rather than dominant seller. And I think like it is the I'm so curious man. Anthropics S1 I just want to see it so badly. I'm sure they'll be able to tell a reasonable story. But like again is it the story was 90% inference margins. do the training, invest the money, your model becomes dominant, and then you make a ton of money. This cuts so directly into that. And I think this is why this makes the entire AI story completely new and brings in just a lot more opportunity and a lot more players. and and like I I mean the vibe shift on Twitter has been wild in terms of like again 6 to 8 months ago everyone who's just like Claude and anthropic there's unstoppable greatest thing in the world everyone is now obsessed with what is the best model for the best task Kimmy 3 harnesses everything like which is good competition is good this is exciting >> yeah more Baker he says an open source model requires the exact same amount of compute to run as a closed frontier model of a similar size and architecture. Kim K3 is roughly the same size as GPT 5.6 terra on a per token basis which actually suggests that it's less computation computationally efficient. that being said lower margin at the model layer. So going back to this question of like will the OpenAI anthropic not be able to you know command those 90% margins? What happens?

15:38 Baker says it's more margin at every part of the infrastructure layer and it's a godsend for software. This can happen either through open source models like K3 at the frontier or having vertically integrated models like Meta, SpaceX or Google at the frontier, which is exactly what we're talking about. We're seeing that Google to come. He says both outcomes result in a lower margin at the model layer and as vertically integrated model companies don't really care where the margin comes from. This is why it was so painful for OpenAI and Anthropic was Google when Google was right there with them from a model competitiveness perspective and why Grock 4 4.5 and Muse 1 1.1 were just as important as Kimmy K3. Right. So this is all happening in conjunction at this week.

16:25 >> So what do you think OpenAI and Anthropic need to do? I mean I guess open AI you make a speaker and physical devices or clouds or all these other business lines but anthropic I mean this is their business right now this is >> inference margins that's the entire game so they start to feel even more under threat if you're either of those companies what do you do >> okay so I'm going to read some more Baker because he talks a little bit about it and then I'm going to give my own perspective he goes the reason Kimmy K3 is only potentially negative for anthropic and open AI is one the claude and chatpt products and harnesses may be more important than their models today and to the hypothesis that they have much more advanced model checkpoints internally that are already being used for recursive self-improvement and the latter scenario reaching recursive self-improvement even a few months ahead of the other labs might be enough to cement a permanent lead. Okay, I'm going to tackle the second one first and then the first part. Okay. so basically there's a theory that they have like you know self-improving AI in internally already and then therefore that will help them open up a gap you know far ahead of any other competitor. I don't believe that and I don't believe that that's actually you know defensible given how far we've seen these models the open source models and the competitor models start to catch up or how quickly we've seen them catch up. This is from Ryan Greenblad who's a researcher. He says, "I now expect an openweight AI which is straightforwardly mythos level at cyber in like 5 months supposing Kimmy and the others don't change their openweight model policy." All right. So that is how close the open source world is to the frontier right now.

18:10 and so any like you know sizable tech advantage or model int advantage or intelligence advantage I don't believe in I don't believe in anymore. And it was it was always treading this way. I mean, but this is where the whole AGI and I like that we made it this far without actually saying AGI yet, but I mean, when I've spoken with people at these companies and spoken with others, like there still is this belief, and we debated this last week, like all the things that you can do about model efficiency and the right model for the right task, if you just get smarter and smarter, you can just subsume the need to even think about all that. And the model is just so good, it does everything. But I don't know, like to me it feels more and more like no one's talking about that now. Even hearing recursive self-improvement at these labs that's going to give them some significant edge. I don't know. Is that significantly different or real? If Kim is able to or Moonshot is able to do this, I don't think it is.

19:10 >> No, there's been no evidence that you could hoard that. That's the whole point, right? So Nick Kle, who is, you know, executive at Medic, came on this show a couple years ago and basically said, "I don't understand why any of these comp like where are the profits going to come from pursuing super intelligence since I don't think you'll be the only company that's going to have super intelligence when you get there?" And I've never gotten a good answer in terms of what the response is on that.

19:34 >> Well, yeah. Or sorry, go. >> So I was just gonna say, so let's assume that intelligence is commoditized, right? That's sort of what this is all building building to. If you look at what's happening with Meta, with Grock, with Kimmy K3, as Baker put it, if you have two companies that have this, it works. If you have five companies that have this level of intelligence, it's a price war. It commoditizes. So, I think we should assume, and we've talked about this on the show, that intelligence is going to commoditize.

20:08 Well, it it I mean going back to Baker's point, I actually think it is interesting that chat GBT I still believe the product and UI was as important as the underlying model and intelligence. again like I remember this is back in 2023 see feeling the difference between typing something into chat GBT and it actually looking like it's thinking and kind of like streaming the text out versus just getting like a chunked API response as a block of text felt more intelligent and AI and like and we said this for a while like OpenAI and Chad GBT was is a great product but it feels like both anthrop ropic in OpenAI have kind of been moving away from the product in the UI side of things again like actually kind of you know bringing it all back down to just a command line experience only moved away from that chat GPT the new Mac app they're removing more of the actual chat function like the chat >> they're adding some of it back they're adding some back this week yeah they >> there's there's an outcry but like they have kind of like forgone the entire UI battle versus and just focused on the model is going to be so smart. So they've given up and seated some of that ground and if I think that's a the right point. It's the product it's the I mean okay sorry of course I think that's the right point because I've always said it's the model. Yeah I know I like as I was saying that out loud I was like oh yeah now it feels even more real. Are you team product now over model? I mean my perspective was always that and I guess my perspective was more like kind of a getting to AGI is important right because once you get there the product experience is much better and because they were able to improve the intelligence they've been able to build better products right but if you're going to ask me today are they going to compete based off of building the most intelligent model or are they going to compete based off of the best product I would have to say who Who's they? Because because this is where things get very interesting. If you assume that intelligence commoditizes, then does open AI or anthropic have that big of an advantage over anybody off the street who would take these type of models to build their own product that competes with them? So basically, I think they're going to differentiate on product. But instead of it just being open AI and anthropic competing to sort of corner this market on intelligence and everybody depending on them now if intelligence is abundant and available to be accessed through multiple providers it will come down to who builds the best product.

23:01 and so it be it goes from a twoerson race, right, or a two company or a three company race, OpenAI, Anthropic, maybe Google, to like now you in order to expect OpenAI and Anthropic to win, you basically have to expect them to be the best AI product builders in the world. And that is a much tougher bet than expecting them to be the best intelligence builders in the world. >> But then, do they even make sense as a business? Like the way these companies have structured their entire business is they have to win on intelligence. They can they make some good products and you know like they're usable and they got some good features and they but like that's not the story. That's not the I was just I saw some like bank analyst note that was saying anthropic should come out at $6 trillion. Like I mean come on. like the absurdity of the story is all built around super intelligence or AGI at the at a minimum versus we make some pretty good products. We're going to build a good vertically integrated company. We're going to be the next Google. That's not their story right now. That's not the way they're coming gonna supposed to be coming out to market. That's why to me the most interesting part of this week is I don't want to say it's the nail in the coffin, but like that story I think we're both agreeing doesn't work as well as it did certainly three months ago and even last week. So I I I'm not sure unless OpenAI gets a really good Johnny IV pin and suddenly they become and launches a little bit of a Neocloud business. I'm not sure what what do you do you think there is going to be some kind of story other than first stage EI?

24:51 >> Yeah, it's a much tougher story without being able to hoard intelligence, but it doesn't mean it's an impossible story for them. And I'll point you to two interviews I've done with folks at Anthropic over the past year that sort of shows the line for these companies to be able to make it work. Right? So last year when I was with Daario he agre he he confirmed that more than 50% of anthropic revenue was coming from the API. This year when I was with Bor's churnney who runs claude code he would not confirm that and in fact he said that the claude products have contributed meaningfully to the company's revenue and the company's revenue has 10xed pretty much since that time that I was with Dario last year. So there is an advantage of being that close to the intelligence that like you don't need to sort of guess on how it works or you can be you can sync it into your products better than anything else, right? So you can sort of have that integration in a way that it's going to be harder for other people to do and you know what's coming next. So which Anthropic has built off of. so so that to me is the path here is that there is still a way where you can be the developer of AI and then have a product sense that enables you to still be a massive company which Anthropic is effectively doing even though the API is still important to it. it has really made the company's really made a lot of headway with the products that it's selling and and we're gonna have Paul Kadski on next week and and basically the anticipation from him is he's an investor and analyst is that these companies will continue to go up market and try to like remember there was cursor before there was cloud code cloud code is built to Anthropic has shown that they can build a tremendous business by doing it themselves and how many other areas are there for a company to build AI native products themsel for a company like Anthropic or OpenAI to build AI native products themselves and then you know start to profit tremendously from that direction. So I think that's still open for them.

26:57 >> Okay. I I I will say claude code even though it's just in the command line is was and is an incredible product. So like and the product was effectively the harness and like along with the model itself but like the experience. So even in the command line, they did create something that was dramatically different or better than everything that was out there before. One thing that that actually brings to mind though is like I mean I've been seeing a lot more around I'm sure most of our listeners saw the you know like Figma with a plugin to claude and then claude design comes in cursor running a lot on anthropic models and then claude co coming out like at what point do other companies actually avoid anthropic where it becomes clear that especially if they are completely dependent on going not up market creating this suite of AI native products which they are able to do because they're being fed all of the data of these companies that are plugging into these systems. At what point do people actually just say no like sorry we see what's happening?

28:07 >> Well, it becomes a lot easier when you have a model like Kimmy K3 that works just as well that you can customize. But again, it cost a lot of money. >> That's what I mean. That that that's exactly what I'm mean that like now with that option, if you have any fear that they're going to take everything that you're giving them because you're using their models and then recreate your business, you now have an option. And I think that's going to that can slow things again. The speed at which they just completely replicated Figma and did it better. The speed at which cursor was replicated and they did a very good job probably better. I think people will start questioning that a bit more.

28:52 >> Yep. we should actually we should actually this is a good time to just bring in quickly although maybe not quickly this idea of the reverse information paradox that Satya Nadella wrote about this week. he said he said you know because they Microsoft also wants to come in and offer this to its customers basically like we'll protect we will let you develop AI and we won't take your stuff. He wrote he writes in the age of AI the buyer risks giving away knowledge just in order to use what they bought. You essentially pay for intelligence twice. Once with money and again with something even more valuable. The proprietary knowledge you must reveal to make that intelligence useful. The better you want the model to perform, the more knowledge you have to feed it. Over time, the information is symmetry becomes increasingly skewed.

29:36 The seller learns more and more about you. as you use what you purchased, while you learn very little about what the seller is learning in return, I think of this as the reverse information paradox. Wink wink. don't buy directly from OpenAI and Anthropic. >> Yeah, I think actually how do you feel about these Satcha X and LinkedIn posts? I'm curious. This is the second kind of >> Do do you do you think >> I'm always? >> Yeah, probably with help. But yes, I mean I'm always for I'm always for more executive communication because like as a reporter it like helps me at least understand the mindset here and we could definitely understand the mindset of Microsoft here. They seem furious. I mean, they really don't like the fact and it's interesting like, didn't we talk about it last week or a couple weeks ago about how OpenAI and and Anthropic were like the single point of failure because only they were getting the products right and they are the ones getting the products right and they're the ones profiting and what was going to happen everybody was going to come in and try to knock them down a peg and we're watching that now. We're watching that with Meta. not just knock down a peg. I think this is like kind of fundamentally pushing against the entire I mean I I I can tell you so at writer where I work like we have our own foundation models they're trained on synthetic data >> and we have seen when you're not and we will like in be ingesting all the data of the users it does change the entire trajectory of how models the advances and the training and I mean I think anthropic and opening eye are very open that all non-enterprise data at least in theory is not is used for training and that is one of their biggest advantages and it's kind of that flywheel and I think SATA is certainly making it clear again I I can say this like a few months maybe six months ago everyone stopped caring about these companies training on your data and now everyone is talking about it again and it's cra again like the vibe shifts are so wild right now Like it literally just how much the conversation changes. But I mean Sachi is completely right. Like it's completely right on this.

31:50 >> Yeah. I mean this is sort of this by the way this is like this is normal business cycle right. It's like companies get ahead to certain this stuff compressed to like >> but but compressed but compressed. I agree compressed like things that would typically take years are now taking weeks. so it's it's very very interesting. Okay. couple more interesting things. First of all, it is interesting that even with the restrictions, I don't think we should gloss over this point. China has been able to catch up the way they have.

32:17 this is from a user named Matt on Twitter, who I'm sure is not a Chinese bot. >> This was my favorite. This was my favorite >> since his his handle is Matt503A5S95. Doesn't sound body anymore, but but body bot E. But anyway, >> he wrote, "How is Kimmy running on a bunch of 14 nanome nanometer Huawei toasters beating SpaceX AI with multiple data centers of Blackwell chips?" >> That was my favorite thing I saw I think on Twitter this week. But I think again like what do you think this means for the not just US versus China but also within even the US like actually going back to you had made you had kind of read this at the very beginning around like this worry that anthropic invests because they believe frontier models are going to be the battle and that powers so much of the current trade or the US and I don't want to say the entire US economy but the whole AI ecosystem. So it's technically bad if they don't believe that which I I find that ridiculous. I find that >> I don't Yeah, I don't agree with that.

33:33 Well, here's my perspective is all right, let's say, okay, let's say anthropic and open AAI go to zero but like the industry whatever achieves AGI or something close to it, right? So what happens is Amazon, Google, Microsoft, they buy all the data centers and they buy all the compute that these companies have. In fact, a lot of the compute that these companies have are already effectively, you know, kind of paid for in conjunction with these companies, >> right? Like Amazon >> just paid they are paid paid to these companies counted as revenue that >> invested in by Come on. It's not circular financing.

34:10 >> No, you're right. Or circular financing laws. All right. I'm Andy Jasse. and I have I'm running Amazon and all of a sudden Anthropic goes to zero and now like as a big shareholder I get to collect and basically own at 100% these data centers or I don't have to meet like commitments to compute that I would previously for anthropic but my customers can use Kimmy K17 and get AGI like performance out of them. and I just get paid by delivering the infrastructure. So that would encourage me to invest even more, you know, in AI infrastructure, even though Anthropic and OpenAI aren't hoarding it anymore.

34:51 >> It's rare that I will ever look at Amazon as the company that tells a good story about a competitive economy, but you just did and I agree with it. I think that was a good Andy Jasse impression. I think that's like that that is like a very logical way that and I don't want to say likely but logical way this can play out and would play out in this case. >> Yeah, it's one possibility. All right, let's talk about it quickly before we move on from the US government perspective. David Saxs allin podcast host and Urswile AISAR from the US government who was sort of instrumental in making sure that this fable ban happened when it happened. he writes this is concerning for the first time a Chinese model Kimmy K3 has taken number one on the frontend code arena and is scoring at at or near frontier on other benchmarks. Meanwhile, America is tying itself in knots. Politicians and bureaucrats are banning new data centers, piling on state regulations and pushing for new federal agencies to preapprove frontier models. This is how you lose the AI race. The rest rest of the world won't play by our rules if we bog ourselves down. Permissionless innovation is how America won the internet and became the technical technological envy of the world. We can do it again with AI. I I appreciate a lot of what David Saxs has to say and I've invited him on this show. but this is like hilarious and so rich and ironic that he he being part of the Trump administration, which did like the most interventionalist policy to ban fable is now saying we need permissionless innovation to beat China.

36:32 I mean, come on. And let's not forget like regulatory capture was kind of the entire business or that like that was this project Stargate and whatever else that he helped coordinate. Maybe he was kind of resistant to a lot of this stuff early on. But you know Oracle, Stargate, all of these things which I mean the Trump administration I think it is rich. It's ironic. It's I don't I agree that or actually do you think we are overregulating at the current state and that that is a danger and competitively against China?

37:15 >> Yeah, I think so. I mean >> why >> I don't think well we could we Stamos explained it to us at the summit. It wasn't like Fable had cyber capabilities beyond that weren't available, you know, from like open source. >> But hold on, but Fable wasn't that again rich and ironic. The Fable ban felt as much political as it did. >> Yes, >> that's my point. >> Yeah, it was. You asked me if we're overregulating. That is like a pure case of poorly thoughtout overregulation.

37:46 >> Okay. I guess I differentiate I'm still thinking of like overregulation around actual safety concerns and like are we too concerned with like wellthoughtout but potentially overly aggressive things around we should not release models purely from the safety side versus Daario's beefing with someone at the secretary or the defense department. Like to me let's ignore that. Let's ignore no >> the pure political no because I mean that's just that's not good in any situation. I'm talking about >> like right >> should the US be versus China more of a leader in terms of we are going to have safe equitable regulated AI like a like Europe.

38:31 >> I mean I don't know what safe equitable regulated AI. >> I don't know I don't even know what that means. I was trying to think of like just smartly regulated AI, but it's not free for all anything goes >> purely permissionless innovation. Here's my perspective. If you have a model that is going to cause cyber security problems for companies if it's released to the public right away and and not just companies, companies, academic institutions, governmental agencies, if you know, if you can see in your testing that it's going to cause these issues, if it's released to everybody right away, I would try to release it in a somewhat controlled way in the early going, and then release it to everyone. Like I think the Fable initial launch of Fable made sense.

39:21 Same with GPT 5.6. But I don't I don't want the I don't think the way the government has been involved recently has been smart because it has been largely political. >> Wait, but so you're a glasswing guy. >> You're >> I'm a glasswing guy. I've always glass. >> All right. All right. >> I'm a glass. >> Hold on. But you just said that the roll out was good. So what what don't you agree with in terms of the the roll out >> the government banning the model for no reason?

39:50 >> Oh yeah. I guess after that yeah yeah yeah. Oh that whole I don't even remember last week man. >> Yeah I know right. It's crazy. All right let's take a break. I want to come back talk a little bit about Google and then we can end with OpenAI beefing with Apple and all of its partners. We'll be back right after this. And we're back here on Big Technology Podcast Friday edition. We come to you amid a rapid a rapid deployment of AI models. Of course, we have the latest models coming out from Anthropic with Fable and OpenAI with 5.6 six and Meta with Muspark 1.1 and SpaceX with Grock 4.5 and China with Kimmy K3 and Google with dot dot dot. This is from Bloomberg.

40:42 Google Gemini launch delayed as tech falls short of internal goals. What's going on at this company? Apple links Google is months behind schedule on delivering Gemini 3.5 Pro, its most powerful flagship AI model, because the company has been taking time to try to improve its capabilities, particularly in coding. The delay has been a source of frustration for Google engineers, AI researchers, and managers, many of whom are concerned the company risks losing an edge in the market as rivals Anthropic and OpenAI produce models that exceed Google's capabilities. Google has multiple layers of stakeholders involved in preparing models for for release working to weave AI across a vast product portfolio including search maps and YouTube which can cause delays. Both OpenAI and Meta Platforms recently released new models that further outpace Google's current offerings in AI for writing code. Late last month, Google updated the data being used to train Gemini in an attempt to improve these skills, but the results were disappointing. shares slipped as much as 3.2% on Thursday.

41:44 Oh, this is this is I mean we've known that Google has been behind for a while, but this is like getting to the point of embarrassing. What do you think about this? >> See, I actually I I was just looking up Gemini 3 was launched November 18th, 2025. Let's not forget there was a few months where everyone's like, "They're back." We were like, "They're back." that Gemini back and they were back. Yes. Okay. They they were back and like >> it was on par with everything else and it just stopped. And I I I genuinely wonder what's going on cuz they caught up. They felt maybe they could even be ahead and they have distribution like no other. And suddenly like Gemini itself, even the standard consumer version like has just kind of gotten a little bit maybe not worse, but everything else is getting so much better. You can feel the difference. Nothing in the last seven or eight months feels like it's significantly improved. I though except for AI overviews, I have found myself again and I always feel very basic using them more and more and actually doing a Google search and following on with the and maybe they're just going to go all in on that and make a ton of money, but in terms of like the actual frontier battles, they feel like they're they're out.

43:08 >> Yeah. So, here's a theory. I mean, you know, the world's perspective on AI seems to seem to shift last last year, January or February, when Deep See came out, right? And there was a perspective that if you can deliver intelligence that was on par with the intelligence that existed then, which was pretty smart at a cheaper price, more people would use your models and you would be like the beneficiary of the Jevans paradox. And for Google which had you know not just a model but cloud services to sell it figured maybe if I bundle the smaller flash models with the cloud services I will enable people to do more. Then the world shifted and the bigger models began to do this coding autonomously starting in December, January this year, right? So a year later and a big company moves slowly, right? so Google just took a long time to catch up and this idea perhaps that you know a big model is all you need and we see the problem of developing these big models maybe didn't really you know catch on within Google and you know instead it might have just like instead of training these these like unifying to train these big models it might have just decided smaller flash models is going to be the way to sort of make our make make the most out of this and it's also helpful for our products which could use these sort of smaller purpose-built AIs to enhance what they're doing.

44:40 >> Actually, you you can even feel that big company under like in the same article it talked about Google counter co-founder Sergey Brin and others were advocating for Google to move faster to seize opportunities in AI coding but slowed by competing factions. two former employees said both cloud computing unit Google cloud research lab Google deepmind and the team between behind the Android operating system were all building AI coding tools so you can picture I guess I can I mean that's a mess and like again you have developers who have the opportunity to build their developer tools they're going to do it but to try to do that in a unified fashion that just feel that feels like old school Google, not not the lean fighting machine that Sundar just somehow reorged into. This is back to like Google Chat, Hangout, whatever the product names whenever they all just were ridiculous and kind of like stacked on top of each other. This feels like that Google.

45:46 >> Yeah, there there's some crazy stuff in here in this Bloomberg story. efforts to win at coding have also been up against some engineers at Google with the more purist hints who believe what all important code should be human written to adhere to Google standards. I mean obviously you don't want AI to write like the core Google software but to have these purists who are like it must be handwritten where like nobody's writing handwritten code anymore is sort of where you get into trouble.

46:14 >> Yeah. This that actually I mean they're a giant organization. It is funny though because remember Sundar I think said like 95% of code is written by AI. >> 75% >> 75%. Okay. I guess the the AI holdouts are still handcrafting their code. But >> yeah, it's Do you do do you think they're going to come back? Do you think we're back? >> Yeah, they'll come back four is going to blow us all away. And >> Google don't we know like Google will inevitably come back. so they I mean they have they have the talent and they have the compute, but the the one thing that I can't understand here is for the life of me, if you're Google, you should never run out of compute, right? The fact that they've run out of compute, by the way, they're licensing a lot of their compute in their cloud business.

47:02 where maybe I mean the cloud business is doing great, but maybe that should be going to your AI development if you think this is the most important technology in history or one of them and Sundar certainly does comparing it to fire. So I don't get how you >> Yeah, but that's like a perfect >> seems like poor management. >> No, no, but that's a perfect example of like Google Cloud can't I mean it was one of the fastest growing businesses of the last 15 years.

47:27 It's gigantic. It's run very separately from the rest of Google. So like that idea that you can reallocate resources without any massive complexity in internal politics, I can only imagine how difficult that would be. And that's a perfect example of like trying to move stuff around to where it's most effectively allocated is got to be difficult. >> Yeah. But that's your job, right? When you're running a company like this, that is your job.

47:58 >> Come on, Sundar. >> You got to make decisions. >> Mckenzie this once again. Let's reorg. >> They'll see it. I mean, they've they've shown that when they get the whole company focused on a goal, they can accomplish it. But I mean it it is crazy watching them to go watching them go from unfocused to focused and effective to whatever this is now. It's not good. >> Yeah. I'll say I'll give them credit that still on like multimodality image and video. They still kind of own it right now and everyone kind of puts them far in a way. Like I do I really never hear that much about chat GBT image 2 or any of these others. like out of the big players, they still kind of own multimodality. So, they're still doing good there, but the rest of it, they're definitely something is up, >> right? Okay, we can't leave today without talking about what's going on with Apple and OpenAI. So, if you listen to the show, you've already heard that Apple sued OpenAI for stealing trade secrets from it. And if you've done any of the reading or if you've watched any of the coverage, you know that this is the most boneheaded corporate espionage attempt maybe in history where the Apple employees on Apple issued laptops were discussing plans to exfiltrate Apple data to bring Apple parts into interviews to you know access through a bug but like Apple's road map and future plans and then Apple caught them red-handed. So, I don't know. Like, I'm just going to turn to you quickly on that, Ron John. Like, this is Do you agree with me that this is like one of the the like legitimately dumbest moments in corporate espionage history?

49:38 >> I mean, if you are going to an employer's whose goal is super intelligence, don't do something this dumb. I mean, I come on. Like I still cannot believe I I Yes, I would firmly agree. This is one of the dumbest corporate espionage things I've ever seen. Even though even the the Uber Whimo stuff back in the day had a little bit more like cloak and dagger elements to it like this. It's literally like you're on Slack or whatever other chat just hey how do we steal information from our employer? How do we >> Right. But you're not just on Slack.

50:16 You're on Slack on Apple's, you know, we're being figurative on Apple's computers. Stupid. >> Yeah. Yeah. >> Anyway, Apple, go ahead. >> Do you think like these are career technologists? Like, how do you end up here or thinking this way? Seriously, >> some people are really smart in some ways and really not smart in other ways. >> Okay. I you know what the one bright side I'll say is >> it's probably clear they had not stolen information or participated in corporate espionage prior to this. So that's the that's what I'll give them. That's the good side.

50:52 >> When you make a career and potentially company ruining mo moment like blunder and you start this blunder with lol I think you're >> you're just like completely out of your depth, right? Like one of the people at OpenAI apparently on the in the chat said like lol I found all the network files. It's like so stupid. So stupid. >> Come on guys, up your espionage. >> So the latest is that Apple has now sent dozens of OpenAI employees legal letters asking them to to preserve their documents. And Apple has also said that this they only found the tip of the iceberg. And the AI lab said that while it had OpenAI had said while it had taken the allegation seriously, it was not aware of any evidence that the complaint has merit. so that's where it goes next is Apple's going to take OpenAI into discovery. You would imagine not settle and just get about as much information as it can. And this is probably going to look much worse when all is said and done.

51:54 >> Well, I mean 40 employees is a lot and I think OpenAI's probably got like seven or 8,000. I mean, >> oh, they have 400 from Apple. So, >> yeah, but if 10% of those employees were engaged in something like this and that's the tip of the iceberg, that's actually that's wild. That's wild. >> Like if if it's 400 from Apple and 40 have already been, you know, like directly receiving some kind of communication around this. This is this is going to get fun. This is going to get very fun.

52:26 >> Not for Open AI. Not for Open AI. And if you're open AI, you have to think like what what am I doing that makes my partners my enemies? Elon Musk, you know, founder of OpenAI now and about doing >> Dario Dario an early OpenAI employee now an OpenAI enemy. Microsoft biggest funer of OpenAI for a long time at least and the champion of this company. Now we just read what SA said an open AI enemy.

52:58 Apple, a partner with OpenAI to build Chat GPT into Apple intelligence, now an OpenAI enemy. I'll just say one thing. I'm going to turn to you in tech. You need friends to win. You can't do it yourself. You need friends. Look at Apple and Google. They should be enemies. Google helped Apple save its business by putting Gemini into Siri to a degree. I don't want to overstate things. OpenAI's loss of these friends over time is going to add up to something really bad. who knows what that exactly is, but there's going to come a time where OpenAI is going to need Satya or it will need John Turnis or it will need Elon >> and they won't be there for them.

53:40 >> So, I I I keep thinking right now, who do you think is it Tim Cook or is it Turnis who's going to be the one? Could this be like Tim Cook's final act? just the head of Sam Altman in his head hands like or is this Turnis coming in like killer Turtis? Who who's gonna be spearheading this effort? >> Great question. So, it's got to be Turnis, right? Cuz Cook is going to step down. This case will like last for a long time after he leaves. And from what what I would guess is that Cook brought this to Turnis and said, "John, we got a problem." John, we got a number. A number of your former employees, including people that reported to you, stole our >> Oh, wait. Did they report to Turnis even?

54:27 >> A one reported to Turnis, but he did. He ran hardware engineering. These people are stealing from hardware engineering. >> Oh, man. Oh, he's going to come strong. This is >> Oh, he's coming strong on this one. >> And Cook probably said, "We want to do this." And the thing is this may take up a lot of energy as you get your as you get going. but ultimately it's sort of up to you in terms of whether we should sue Open AAI. And Turnis probably looked at him and said Timmy boy sick the lawyers on these >> The betrayal. The betrayal.

55:07 H. All right. This is okay. This is this is now becoming one of my favorite stories of that to see what happen. Forget forget Kimmy K3 and the entire future of the AI ecosystem and economy. >> I just want to see what Turnis is going to do to open AI right now. >> I just want to be clear. I'm not the one saying that OpenAI folks are I'm just saying that that's probably what John Turnis would have said.

55:33 >> I think John Turnis doesn't swear. He's >> John Turnis definitely. He's a very polite, upstanding citizen. >> No. Do you do you think Okay, tell me. We'll we'll end on this. >> Do you think that there's any of the folks at the top ranks of these companies who doesn't curse during the day? >> no. But this is a good >> There's so much stress involved in these jobs, you almost need swear words as a as a way to let off. I would like to say as the parent of a seven-year-old child, I do feel and as someone who has sworn many, many times in my life, proudly, I still feel weirdly weird how normalized it's become with adults swearing in like very public communications and forums.

56:21 Obviously, the president and others and like it's just become very normalized and it's weird to me like there's still there's still bad words. We should all know and Turnis should be dropping fbombs left and right, but >> not in front of the kids. That's all I'm asking. >> Let's take a moment to reprimmend John Turnis. >> John prefers swearing in front of kids. Apparently you say >> the fact that a imagined fanfiction version of you swore on this show is deeply upsetting to us and our listeners and you should you should really think twice before using that sort of language.

56:57 >> I think the good thing the good thing if I am to swear on this show I think the population of seven-year-olds listening to the big technology podcast is one of our smaller if non-existent demographics. Well, >> it's definitely not non-existent. That's why I try to keep it as clean as I can because I know the parents play it in the car and and honestly I commend those parents. You know, we're here, you know, all right in the part as an educational endeavor to make sure that the youth of the world knows what the AI industry is going to look like when they grow up. We're here >> not to mess with John Turnis. Do not mess with John Dis.

57:36 >> One of those lessons is Yeah. Stay out of the stay out of the bad side of John Turnis because you never know what he'll do. Is this going to be a running joke that will just have bad >> John? Well, I think I I know so little about his personality that I can only create extended fanfiction around it. So, sorry, John. Get ready. >> Okay. Well, this is this is a new thread for us and it's a new meme. So, we're gonna we'll run with it.

58:10 >> And I guess that's it for this week. >> That's it. >> Kimmy K3 episode ending with a meditation on language as we typically do here on Big Technology Podcast. We leave you with that to think about. Thank you, Ron John. Great to see you as always >> and thanks to all of you listeners and viewers. We'll see you next time on Big Technology Podcast.