Google Won’t Have to Sell its Ad Business - DTNS 5345

[music] This is the daily tech news for Wednesday, September 2nd, 2026. We tell you what you need to know, give you some [music] important context, and help each other understand. >> Today, Google will not have to sell off any of its ad business. >> Ah, good for them. I mean, for them. Just for them. I'm Tom Merritt. >> Poor Google. I'm Sarah Lane. We all feel so bad for them. Let's start with what you need to know with that big story. Yeah, this one's going on for a couple of years now. US District Judge Leani Brinkma issued her decision for remedies in the antitrust case against Google's advertising exchange. This is one of the two big US antitrust cases. There was the one about search. This is the one about Google's practice of selling ads. Trial began in September 2024. Oh, so long ago. The trial concluded in April 2025. The remedy hearing started this past May. And here we are just two years later with that remedy decision. How are they going to remedy the fact that Judge Brinkma found that Google abused its market position in ad markets for ad exchanges and ad servers? These are the technologies that help websites sell ad space and then determine where those ads are going to be placed all over the web. Google tightly integrated its doubleclick for publishers ad server with ad exchange. It's its ad exchange adx which the judge had said enabled the company and I'm quoting now to establish and protect its monopoly power in these two markets. Google was not found to have abused its position in tools to purchase ad displays. So what do they have to do as a result? We don't know. We do know that Google will not have to devest any of its ad businesses. However, and I'm going to quote again, most of the parties proposed behavioral remedies as modified by this court be and are accepted. In other words, most of the stuff that the Department of Justice was asking for, they're going to get, just not the devestment of any businesses. A full decision with reductions will be released later this month, and we'll get some more details. Uh but apparently Google convinced the court that making some of these changes public would reveal trade secrets. So that'll be interesting. Uh it is being reported as a win for Google, but either Google or the US Department of Justice may appeal the case. So all those headlines you saw years ago saying that Google would be broken up. They might have to sell Chrome or Android or their ad business have not yet proven to be true. >> [snorts] >> Yeah, we talked about this uh quite a bit over the last few years, sort of wondering, you know, what is the worst case scenario for Google, you know, spinning off Android, spinning off its ad business, spinning off Chrome. None of that uh has come to fruition. I sort of wonder what uh the redactions were in the full decision. Google saying like, hey, it's tra it's trade secrets, right? So that that kind of makes sense. Um, this is definitely a good win for Google, but as you mentioned, Tom, what changes is not clear at this point. It's sort of like Google was found to be uh negligent, I guess, you know, as far as uh >> abusing their their monopoly position. So, antitrust, violation of antitrust. >> Yeah. So, so there's that. Um, so it wasn't as if uh a judge said, "Nah, everything's cool, Google. you just go on [laughter] with your bad self. But uh but it but it is unclear what what we're going to see change going. >> Yeah, I'm I'm interested uh in what got redacted. I mean, I use the term trade secrets. Google didn't use it, so I'm not I'm not sure what the actual justification for the redactions are, which I guess we won't know because they're redacted. Uh but yeah, it'll be it'll the the proof will be in the pudding that we get to see, but we will not get to see all the pudding. >> No. Although I like pudding. >> Pudding is delicious. >> Pudding is good. Pudding is good. Uh, you know who else is good, Tom? >> Who? >> DTNS is made possible by all the listeners who may or may not like pudding. Uh, we want to thank everybody who supports us and especially today Tim Deputy Brandon Brooks, Simon Daniel Jerk, Daniel Jerk rather, and Luke Perier. We also have some new patrons and we love those. Hey, we got AK MZ, Meline, Shane, Matuses, David, William, Joseph, Zack, Hank, and Helzinki. And by the way, if I mispronounce any of your names, write us. >> There is more we need to know today. Let's get to the briefs. >> Let's do it. Open AAI released details on its Astra model, which it previously said was the first of its LLMs to meet a critical cyber security threshold, which was upsetting to a lot of people. Open AAI claimed a claims that Astro now Astra achieved a perfect score on exploit bench and can find unknown security vulnerabilities and exploit them without limited human in input. It was able to compromise web browser sandboxes and also execute commands on the browser's host. OpenAI said also says it developed a test informed by the hugging face incident to create honeypot tests of whether a model would try to compromise surrounding security infrastructure instead of solving its assigned task. Astra made no such attempts in these tests. These features describe behavior under the test conditions without cyber safeguards, not normal production use. So this is something outside of what you or I would do on a daily basis. The information sources say Astra uses a recurrent depth technique that improves efficiency. It lowers costs andre and increases performance, but in a way that may make it harder to monitor going forward. Reoccurrent depth can save compute because it doesn't need to access the model weights as it loops information through multiple layers to improve accuracy and deepen reasoning. But this hides some of the reasoning from chain of thought logs that indicate how a model reached its conclusions. So open AAI is planning to make Astra available soon. doesn't say exactly when, but soon. But with limited security capabilities, additional chain of thought monitoring to stop and spot bad behavior, and restricting the model's responses to prompts from accounts accessed as higher risk. Limited number of testers are going to get its full security capabilities. Won't be me, but uh yeah, this is a this is a this is going to be an interesting one. people. You got a lot of people kind of across the board saying like, "All right, they're doing the right thing." And other people saying, "This is the beginning of the end." >> Yeah. People say all kinds of things, don't they? Uh, OpenAI definitely trying to be more transparent in why it's doing what it's doing. Uh, and you can, like you say, judge whether you think they're doing it well enough or not. I think they are also trying to walk a thin line between government regulation, voluntary regulation, what the government would like to restrict from being accessible not just to the general populace but also to foreign adversaries. Uh and so they want to say, look, Astra is real capable. Here's what we're doing uh to prevent it from falling in the wrong hands. Here's what we're doing to prevent it from going rogue again. Uh specifically, we tried to make it go rogue and it didn't do it. Uh, and also on the same day as Enthropic announced a new model, which we'll get to later in the show, uh, saying, "Well, we've got a new model coming." Because I I don't know if it was lost on anybody, but OpenAI said, "This is what we're going to do to safeguard Astra when we release it." They did not yet release Astra. That's not the news. >> Yeah. Yeah. Well, and you know, the hugging face incident has not only was that sort of like, oh crap, here's where we are now. You know, that that has spawned so many sort of think pieces about uh from from security experts about here's how we need to not let that happen again. and OpenAI happened to be the company that that was associated with. But that could have happened in a lot of different situations. And you know at at this point it's it's I I think my real question is who is the trusted company or I don't know just you know whoever organization that does get access to something that is potentially very dangerous in the wrong hands and that's something that open AI can can decide for themselves you know >> well they're not deciding it for themselves they're deciding it with the cooperation of the United States government because the US government has to vet that list. >> Well, I meant deciding by by you know in the sense of just making it in in the first place. But yeah, it's uh yeah, it's >> it's a whole new world. >> I think it's also interesting to watch whether this gets these openweight models from China closer to the capability uh of what open AAI and anthropic have been putting out. uh it hasn't yet but it is slowing things down. I mean in the past we would have already had Astra. Uh we would have had you know it would have been called something else probably not a code name but it would have been out there but they are taking great care and I'm not saying they shouldn't. Uh but it's definitely slowing down development, slowing down releases. Uh and the other thing that that you probably shouldn't get lost on is if it's that good at doing vulnerability exploitation, then it's that good at finding and patching those vulnerabilities and it's probably good at other things that aren't dangerous. uh things like scientific research uh and things like like health improvement and and things like education and research and and experiment design and all this kind of thing. So, you know, I think I think I'm not trying to justify uh being unsafe and I think it's important for the conversation to focus on safety. Uh but these models are capable of doing great good as well and we shouldn't lose sight of that. >> Yep. Acer announced lots of hardware ahead of EPHA in Berlin. The Predator Atlas 7 is a 1 in smaller version of the Predator Atlas 8 cuz seven is less than eight. Uh so it's got all the same power and performance, but it's going to be easier to hold and not just because it's smaller. They also put an anti-slip texture uh along with better grips, a larger palm rest and recessed keys. comes with a 7 inch full HD 120 Hz display and 24 gigabytes of LP DDR5X RAM. No price yet, but coming in November. Acer also showed the Vero 16, a heavyduty laptop with milstandard 810 certified chassis and somewhat modular design for improved repairability of batteries and SSDs for up to two generations. I mean, it's no framework, but you know, a little bit of modularity is always welcome. It is a uh there's also a lighter 799 gram Swift blade 14. That one is 0.51 inches thick running Intel's Wildcat Lake chips. There's an Acer Swift Air16 that has an OLED display and quad speaker audio and the SFFF RTX Spark, a desktop for running foundational models locally on Nvidia's RTX Spark chip. That's their desktop LLM machine. Uh, but the star of the show seems to be the return of a concept piece called the Project Dual Play Mini, which Acer has shown versions of before. We saw it at CES 2025. We saw it at Epha 2024. Uh, so it's it's been a minute since uh we we saw it, and this is its 2026 appearance. The 8-in handheld gaming device can function as a mini laptop. That's the concept. You basically dock the handheld in a backlit physical keyboard and then you can use it as a Windows 11 laptop. But, uh, still a concept, Sarah. No date, no price, no release. They're just showing it off again. >> Well, so, okay. So, it's like EA 2024, CES 2025, and here we are in September 2026. Still a concept, >> but maybe cool >> if it ever becomes something that [laughter] Yeah. Like, do you like it enough for us to actually make it? >> Yeah. like offer it for sale at whatever price. We don't know yet. >> Yeah. And the price is not getting cheaper, right? With component prices being what they are. >> Certainly not. And and I wonder if that is what uh makes a lot of these altogether pretty cool models, you know, with the specs that people would want. Um you know, where >> [snorts] >> uh Acer might just be sort of like, well, uh you know, coming soon. Coming soon. Not totally sure when yet. Uh we'll wait for Apple's announcement and then we'll figure out, you know, what things are costing. >> Yeah, it's all part of the EHA announcement. We we'll have another uh laptop coming from Dell that we'll mention later on in the show, but uh it does seem like everybody's in the Apple MacBook Neo space now. Uh everybody wants to make that slightly powerful but maybe affordable and certainly thin and light laptop. >> Yeah, indeed. Well, Anthropic released an update for Fable 5 called Fable 5.1, as well as an upgrade to the more capable Mythos called Mythos 5.1. Anthropic says Fable 5.1 is as good or better than Fable 5, but will also cost you about 25% less to run on most workloads. Anthropic also believes it may be 45% more efficient for highly agentic work. So part of this is efficiency, but it's also because Anthropic reduced the price of Fable 5.1's cache reads. That's where the model reuses context. It's already processed where every usage is build by token. That's actually that's a that's an interesting sticking point that I know myself and a lot of other people have had because, you know, it's going through context that it doesn't need anymore. So anyway, Anthropic says Fable uh and and Mythos will be better at this going forward. Specifically, Anthropic said Fable 5.1 will be better at long programming tasks and also code reviews. It should work better at scientific tasks like experiment design, simulations, reading dense diagrams, reading tables. Anthropic also says sometime later this year, mentioned the fall, it will let business partners retain data in their own cloud when using these new top models rather than requiring them to be on anthropic servers. The enterprise frontier safeguard system is new, still monitoring for misuse by agents or humans, but enterprises will control how that monitoring takes place. Finally, Mythos 5.1 is the same model with modified safeguards to allow for work in security and also life science. Safeguards have been improved over Mythos 5. Anthropic wrote, quote, it cooperates with human misuse and accepts unverifi unverifiable claims of authorization somewhat more readily than Opus 5, but is less likely to ignore explicit constraints, hallucinate inputs, or falsely claim to have completed tasks than previous models. So, Methus 5.1 only available to trusted partners. Yeah, I think this two big significant things in this announcement. Uh an announcement of a model that is shipping. Uh one is the the enterprise thing. Uh for safety, they were telling enterprise they couldn't store their Fable 5 data uh locally. Enterprises hate that. They do not like that because it's unsafe for the enterprise to have its data anywhere but on its own infrastructure in a lot of cases. So fixing that I think was an important move for anthropic especially on the enterprise side. Uh, but more widely, I think efficiency is becoming the new selling point for these LLMs, saying, "Look, we're going to let you get more done with fewer tokens and with things like what you were describing with the cash reads, uh, not even have to use tokens in in order to get the performance out of it because it's just reusing context. It's not going back to the well. Uh, and that saves you compute. But that that is one of the biggest advantages to these foundational models over the openweight Chinese models is the fact that they are more efficient. Uh yes those openweight models are open and they're very close in capability but they usually don't surpass that capability and then when people try to run them themselves they often find out it's incredibly costly uh both both in in API token cost but also compute if you're trying to run those open weight models locally >> costly or you just get shut down. You know for example I I don't I I'm not using anthropic models uh currently but I am using codecs on open AI you know on a daily basis for variety of things and uh sometimes I get shut down and it's like yeah you just right tomorrow >> on a consumer level that that that's true I was I was thinking more about a company that takes one of those openweight models and like colllocates it and runs it locally you know on on their infrastructure but it's the same kind of problem right yeah >> well Yeah. And I think that you know the the models the companies behind the models uh you know saying hey we can do this you know however many% faster uh for less money that's like yay it's a it's a big win. Um at what point does you know like you you can't get cheap enough to be free. So, it's it's I feel like we're in an era right now where it's a way to get more companies to buy into the system. Um, and then, you know, we'll see what everybody's paying later on. >> Yeah, I think we're we're in the competition phase. Uh, we've gone from the customer acquisition phase where things were free and some stuff is still free obviously. Uh, but we are now in the phase of everybody's paying too much. uh companies have stopped token maxing and they're starting to look at like how do we actually use these things responsibly and and because of that you're seeing companies battle over we are the more efficient way to spend your money. You're still going to spend money but you're not going to spend as much money and what you do spend your money on is going to be productive for you instead of wasted. >> Yep. Well, um we hope that uh if you're wanting to waste time then you won't feel like your time is wasted >> when you have all that extra time. Right. >> Yeah. Right. I know. Look at us just sitting around all day. Nobody nobody has a lot of extra time. But um if you want a new podcast, please do check out our product review show called Live with It. Each week we talk about the long-term tech in our lives. Home internet, vacuums, energy efficient solutions. We run the gamut. This week we're going all in on Tom. Can you guess? Can you guess what I'm about to say? You're never going to guess. >> Is it a phone? It's not a phone. It's babies. >> Of what? >> Human babies. >> Human babies, even. >> That's right. What is the best way to keep breast milk fresh and also portable for new families? Mhm. Amos is actually living with it in his household and he has all the deets. It is fascinating stuff. Uh, you know, the baby tech. Baby tech. >> This is good. This is a good one. Yeah, >> it's a big one. Yeah. Livewith is on YouTube at youtube.comlivewithitshow. And big thanks to all of our patrons who keep us going every week. If you'd like to become a patron and directly support us, go to patreon.com/livewithit. >> All right, let's get to the quick headlines that are going to make you smart. >> Let's do it. Wired has an article about the rising popularity of a company called Pangram with a product that promises to detect how much LLMs were used in the generation of any text. Substack uses Pangram to help readers determine if text was generated [laughter] by human or something else. Pangram uses LLM and a technique called synthetic mirroring to create a data set used to train a model to detect generated writing. >> Yeah. And uh it is it has caused some controversies out there but I am waiting for the controversy where they get it wrong where they get it provably wrong. Uh that will be the test of whether pangram can weather that because nothing is 100% right. So >> sure >> it's a it's a good buzzword to have in your mind if you hear people talking about pangram that that's what it is. We have more MacBook Neo competitors to talk about. Dell announced a new budget line of laptops the Dell 14S line. more color options and running on Intel Wildcat Lake processors. One of the themes at EHA is Intel showing off all the things that are running on those more efficient Wildcat Lake processors. Uh we don't have prices until later this fall. If you're interested in the Dell 14S line, just wait for the leaves to turn the >> Nobody wants to announce prices just yet. >> Not in this economy. [laughter] >> Huawei announced the successor to its first trifold, the Mate XT2. So, instead of the zigzag folding of the original, Huawei's new trifold follows the form factor of Samsung's trifold with both screens folding inward. >> Oh, Samsung won a format war, even though it gave up the format, at least for the time being. Uh, ahead of the Google Book event that's coming on September 15th, GBook Hub noticed that a Lenovo Pen Pro for Google Book has passed through Bluetooth SIG certification, so stylus lovers, take note. The largest and most popular video game modding platform called Nexus Mods has acquired database and player account service Steam DB. Steam DB's owner XP Paw says it quote will continue to work with Nexus Mode Mox mods rather over the coming months as well as to make sure it's a smooth transition. Yeah, apparently Xpa just uh burned out, tired of of the internet the way it is today and wanted to find someone that Exp. So they that's why they sold it to Next month. >> Good for them. >> The largest school system in the United States will ban large language models for use in classrooms by students through the eighth grade. New York City schools will say that use of such tools in high school will be restricted to literacy, career and certification programs and five pilot programs in some schools. So the the biggest school system to say no LLMs. >> And I love in a high school it's like they didn't explicitly say it this way, but it's like no math class for you. You better learn it. >> LLMs aren't great at math anyway. It would be more the English class that I'd worried about. >> Yeah. Law firm Eden PC added 30 additional complaints to its seven lawsuits against Open AI on behalf of victims and families connected to the Tumblr Ridge mass shooting in Canada this past February. >> And Norway's digital minister Kerianne Tung told the AFP that the government is preparing to regulate camera enabled wearables like Meta's smart glasses. I don't think that will be the last government we hear talking about that. We like to end every episode, sometimes we wear glasses, but not today, uh, but of DTNS with some shared perspectives. Today, Mohan shares a lucky experience he had with the Nancy Grace Roman Space Telescope. >> Yes, my glasses are dumb. All they do is bend light, nothing else. >> Yeah, my glasses [snorts] are just in my eyeballs. >> Mohan wrote, "I just wanted to share that I was at the launch of the Roman Space Telescope bright and early on Sunday morning. A family member of mine worked on the telescope at Gddard Space Center and invited us and other family members to see the launch. It was one experience that was well worth the wait. >> Awesome. >> I'm jealous. >> Well, maybe you get on up there. >> Yeah, Moan, next time you know if you got an extra invite, let us know. >> Yeah. Yeah. Where was our invite? [laughter] Uh yeah. Uh no, that that's awesome, Mohan. Thanks for writing in. Um, if you are thinking about anything, something we talked about uh sparked your interest, you have insight into a story, share all of that with us. We read every email feedback.com. >> Yeah. And no, all all seriousness, uh, very cool, Mohan, that that you got to experience that. Thanks for sharing that with us. Uh, thank you for contributing to today's show. Thank you all for being along for Daily Tech News Show. You're the folks who keep us in business. So, become a patron if you can right now. Super easy and you get cool [music] stuff. patreon.com/dtns. >> The DTNS family of podcasts, helping each other understand. >> Diamond Club hopes you have enjoyed this broker. [laughter]