NVIDIA is about to buy HuggingFace - DTNS 5341

0:00Huyen Tue Dao: This is Hannah Berner from Giggly Squad. O-Pill is the first over-the-counter daily birth control pill available in the U.S. Let's be real, getting a birth control prescription is not always easy and it's so much admin. In fact, about a third of women face barriers to access prescription birth control. Between scheduling appointments, missing work, class, or just trying to exist, it's a lot. But now O-Pill is putting birth control in our control. Opill is a daily birth control that's FDA approved, full prescription strength and estrogen free and 98 % effective when used as directed. Grab it online or at most major retailers, no prescription or doctor's appointment needed.

0:38Huyen Tue Dao: So if you're thinking about birth control, check out Opill to see if it's right for you. Use code Giggly for 25 % off your first month of Opill at O-P-I-L-L dot com. That's code Giggly at O-P-I-L-L dot com. Birth control in your control. We love to see it. This message is brought to you by Cologuard. If you recorded songs off the radio for a mixtape, it might be time to screen for colon cancer. The Cologuard test is non-invasive, requires no special prep or time off, and you start screening right from your home. It's just three simple steps. Set up, sample, and ship. If you're 45 or older and at average risk, ask your doctor about the Cologuard test.

1:17Available by prescription only. Learn more or request a prescription at colaguard.com slash podcast.

1:49Huyen Tue Dao: at Vanta.com.

2:19Huyen Tue Dao: video content without making content creation your full-time job. It's rated the number one AI video platform for small business on G2, features some of the industry's most realistic AI avatars, is trusted by more than 30 million people and 85 % of the Fortune 100, and supports over 175 languages, so you can communicate with audiences around the world. Here's the wild part. Record yourself for just 15 seconds, and HeyGen builds an AI avatar that makes professional videos for you on demand so you can post everywhere your audience is as yourself without filming every time your first three videos are free at heygen.com slash pod that's h-e-y-g-e-n dot com slash p-o-d this is the daily tech news for thursday august 27th 2026 we tell you what you need to know give you important context and help you understand so you're not left in the dark well today nvidia is about to buy hugging face and hugging face made a dancing robot duck i'm tom merit i'm huynh huydao and if you're watching the video version uh i am very proud that huynh huydao is now the second person on dtns to be engaging in the time shifter apps jet lag preparation fully fully in the dark in the middle of the show on purpose on purpose this is not a look i'm trying i'm not trying to be cool uh and i do i do have my lights on in the back but now i just look like a witness in like i've been there i am pro normalizing time shifter darkness i can't wait to hear how this works for you once you once you get back from yeah i did i did subject you all to this because tom had already done this so i figured if tom's doing it i'm gonna do it too.

4:07So there you go.

4:08Huyen Tue Dao: All right. So more to come on that in the future, but for now, let's start with the big story. The information says, and this could happen at any time. So by the time we publish this, maybe it's official, but NVIDIA seems to have agreed to acquire Hugging Face. Hugging Face, the open source platform for LLMs and machine learning models. A lot of times people call it the GitHub for AI. Now, Business Insider was saying that the deal wasn't quite done, but NVIDIA didn't address reporter questions about it. And the company usually does if a report is inaccurate. They usually come back with like, yeah, that's not happening.

4:48Huyen Tue Dao: That tells me that the deal is all but technically done. And we have a theory on why we might not have been hearing about this right away. If you don't know Hugging Face, Hugging Face is a place to find thousands of pre-trained models, some of which you can actually try out in the browser. Also has lots of data sets you can use to train and test models. It's open source playground for AI. NVIDIA wants to encourage a vibrant open source market. They've been talking about that nonstop lately. And the reason they want to do that is because they can sell chips to the people who run those open source models, right?

5:21Huyen Tue Dao: Especially as the big foundational model companies have started to design their own chips. NVIDIA is like, we not only want to broaden the market, we want to make sure that we've got lots of customers and we're not locked into one customer that then changes their mind. NVIDIA has also been scaling down its DGX Cloud product, but Hugging Face could turn that around. It could boost it or potentially even become it and just DGX Cloud and Hugging Face merge. As for Hugging Face itself, it gets the more generous funding available from within NVIDIA's little waterfall of cash that's been going on, as well as access to all those servers.

5:58Huyen Tue Dao: But why now? Hugging Face had turned down a large investment from NVIDIA earlier this year because it didn't want one dominant investor. But apparently, another company came knocking to buy Hugging Face. Hugging Face apparently has been entertaining offers. That caused NVIDIA to step up. And if there was another offer and NVIDIA is making sure that it's giving hugging faces investors the superior offer and dotting I's and crossing T's. That might be why we don't have the official announcement right now as we're recording. Oh, I kind of have an image of like that scene from The Graduate where NVIDIA is Dustin Hoffman in a glass and hugging faces.

6:38And I forgot the character's name. It's like, no. Mrs. Robinson? No, Mrs. Robinson's daughter. Anyway, yes, I had that kind of image when I was reading this.

6:49Huyen Tue Dao: Yeah, yeah, it was the dollar. You're right. Yeah. Yeah. Well, I don't know which one of them had the money. I think it was Mrs. Robinson in that case, but NVIDIA certainly has the money. NVIDIA reported its revenue is up 106 % year over year last quarter. Net income up 126%. Projected revenue is going to grow 70 % in 2028, which sounds small compared to the 106 % and 126%. But it's because NVIDIA is also dealing with the rise in component prices. Their profit was$2.22 a share this last quarter. All of that beat expectations. In other words, the bubble not showing signs of bursting yet, not in this past quarter.

7:31Huyen Tue Dao: CEO Jensen Huang seems to believe we haven't even reached the peak yet. He told investors in the NVIDIA earnings call, I think the most important thing that matters for the industry is that, one, AI is doing productive and useful work. Two, AI is generating profitable tokens. And three, if we had more compute, we could generate more profitable tokens, which results in more profit for all of the services. This is the exact phase where we are at, he said. If you need me to translate that, he's saying LLMs and agents are actually making the end users money. And if we could deliver more capacity to run them, they would make even more money, meaning the service providers would also make more money and then have to buy more NVIDIA chips.

8:11Huyen Tue Dao: And while Hugging Face doesn't solve the compute shortage, it does make sure NVIDIA can foster continued users. So a big day, big day in the consolidation. We're seeing Open Router getting bought by Square. Looks like we're seeing Hugging Face getting bought by NVIDIA. I mean, from the NVIDIA perspective, it does make sense. I mean, I have thoughts. But it does seem like they're trying to grow there. I mean, basically, they're trying to make their market, which from their perspective makes sense, right? Like, yeah, like more people like basically funding hugging face means more opportunities for them to.

8:46I mean, as you already said, like I think from a strategic perspective, it makes sense. Yeah. It's a little bit Microsoft. Right.

8:56Huyen Tue Dao: In Microsoft buying GitHub. When that happened, I remember people going, oh, is this going to be bad for GitHub? I guess it's good for Microsoft. It feels the same way. Right. Yeah. Yeah. Yeah. I mean, and I imagine anyone who is a, you know, avid part of the Hacking Face community, I know, and I always think it's kind of interesting because it always feels like taking AI out of this, whenever a large company buys a community or something, I mean, it's not a community, but it is an open source platform, which has a sense of community. and of course it being open source kind of gives it a certain feel and a certain kind of culture it always feels a little bad when say a large company buys a platform such like like github and presumably like hugging face i really have never participated so i don't really get the vibe of it or i just don't understand the vibe of it it always feels a little bit bad it's like how will this change it the the priorities will this change how they operate will there be places where as the parent company they encourage certain kinds of behaviors or promote certain projects over other projects because it benefits the parent company i think that would be kind of interesting to kind of introspect on what this could mean or how this could change hugging face as a kind of community or a a open open open source place open source source open source uh hub in you know when in if this is true if nvidia buys them and once once they actually kind of come into the picture and It does take time for all these things to happen, of course, but I'm kind of curious to see how that changes.

10:28Huyen Tue Dao: Yeah. The one thing I'll say for NVIDIA, when Microsoft bought GitHub, Microsoft had and still a little bit has the reputation as being like anti-open source. Yes. It is much better and much more of a good player in that arena than it used to be. I'm not saying GitHub turned Microsoft into a softer on open source or pro open source company, but it didn't hurt. And NVIDIA has been a really good citizen in the world of open source. They've got a lot of open source platforms. I don't see a lot of criticisms of them. Not that there aren't always some criticisms, right? And so if anything, Hugging Face might just be even more of a good influence on NVIDIA as well.

11:08Huyen Tue Dao: So yeah, NVIDIA might pull Hugging Face one way. Hugging Face is going to pull NVIDIA another way. And where that tug of war ends up kind of is what we're concerned about as users. Well, I guess we'll wait to see when NVIDIA actually acknowledges it and they get to run off onto a bus into their happy future. That'll be part two. It's like, what does Hugging Face say? What does NVIDIA say once this becomes official? Will this actually be like the ending of The Graduate where they run off on the bus? But then the last shot is them looking kind of a little bit like unsureedly at themselves and at the camera.

11:39Huyen Tue Dao: You're way more familiar with The Graduate than I am. All I remember is the plastics line. That's all I get to bring back. All right. Well, regardless, you all pull us up by supporting us. DTS made possible by you, the listener. So let's take a little time to thank Jim Apple, Mike Cortez, Bill Russell, and Aaron Sturr.

12:01Episode prep gets messy when research, Slack threads, and script trackers are all in different places. That's where ChatGPT work comes in. Give it the goal and connect to the tools where work happens, like apps and files. and it can help pull the pieces into a draft you can review without you having to re-explain the project at every step. For example, ChatGPT work can gather team feedback from Slack and use it to suggest updates to a project brief in Google Docs. You stay in control reviewing the draft and deciding what moves forward. Put ChatGPT to work on your most ambitious ideas and projects.

12:36Get started at ChatGPT.com by selecting Work Mode. Available on Plus and Pro plans.

12:42Huyen Tue Dao: Granola is an AI-powered notepad built for the way real people actually meet. Here's how it works. You take rough notes like you normally would, and in the background, Granola securely transcribes the meeting. Then, it turns everything into clean, structured, actually useful notes when the meeting ends. And the best part, Granola works through your device's audio, which means it integrates seamlessly into the video conferencing tools you already use. No setup, no awkward bots. It's just your normal meeting with superpowers. superpowers. You get to actually listen instead of frantically typing every word and still walk away knowing exactly what was decided, who's doing what, and what comes next.

13:18My first time using it was in a meeting that had a lot of moving parts and was pretty technical. I couldn't believe how organized and complete Granola's notes were. If meetings are eating up your day, Granola is a no-brainer. You can try it totally free. Just head to granola.ai slash daily tech. That's granola.ai slash Daily Tech to get your time back. Try it for free at granola.ai slash Daily Tech.

13:44Morning Brew Daily breaks down the biggest news in business every morning so it fits seamlessly in your day.

13:50Huyen Tue Dao: I'm Toby Howell. And I'm Neil Freiman. And each morning we cover everything from the latest tech headlines to why nobody can afford a house right now. You'll leave each episode of Morning Brew Daily smarter and ready to take on the world around you. And some people are saying it's the best part of their morning. Because we know something you don't. Business news doesn't have to be boring. Join millions of monthly listeners and check out Morning Brew Daily wherever you get your podcasts. This message is brought to you by Cologuard. If you recorded songs off the radio for a mixtape, it might be time to screen for colon cancer.

14:22The Cologuard test is non-invasive, requires no special prep or time off, and you start screening right from your home. It's just three simple steps. Set up, sample, and ship. If you're 45 or older and at average risk, ask your doctor about the Cologuard test. Available by prescription only. Learn more or request a prescription at Cologuard.com slash podcast. This message is brought to you by today's sponsor, Walden University. Set a course for change. Now is the time to make a real change in your life, career, and community. At Walden University, we give you the support and flexibility to get the W.

14:59Walden's online programs are designed for working professionals like you. You'll gain hands-on skills to take on real-world challenges and succeed. Your future is waiting, and Walden is here to help you achieve it. Take the first step to make your impact. Visit waldenu.edu. Walden University. Set a course for change. Certified to operate by Shev.

15:23Huyen Tue Dao: All right, there's a lot more to talk about. Let's get right into the briefs. All right. Well, whether it gets bought by NVIDIA or not, Hugging Face is moving into robots. The company introduced MicroDuck on Thursday, an open source robot. You can teach new tricks with reinforcement learning. It's 25 centimeters tall and can waddle and pick things up with its beak as long as they weigh less than 800 grams. It can also crouch, roller skate, and pick itself back up if it falls. It has a camera, LiDAR sensors, Bluetooth, Wi-Fi, two NFC readers, and two inertial measurement units that measure movement.

15:58Out of the box, it can walk around and respond to your movements as well as laser pointers or a game controller. And of course, you can train it through an SDK for simulation and a full RL training stack available on GitHub. You can code new behaviors in Python and JavaScript, train it in simulation, then deploy on the Duck. MicroDuck comes out of Hugging Face's acquisition of France's Pollen Robotics back in April 2025. Hugging Face also sells the desktop robots the Ricci Mini and Ricci Mini Lite. It's manufactured by open source hardware company Seed Studio in Shenzhen. And the MicroDuck sells for$399 and will ship before December 25th for your Christmas robot training holiday plans.

16:43Yeah.

16:44Huyen Tue Dao: And if you haven't seen the micro duck, if you're only listening to us talk, don't imagine a duck. It is a WALL-E-like robot that sort of looks like a duck, but it's only got one eye and its lips are not exactly duck, but they're kind of duckish and it waddles. First, when I looked at it, I'm like, that's not a duck. And then the more I looked at it, I'm like, no, I guess it's sort of duck-like. But it's cute. It is cute. I think it was TechCrunch compared it to Lego Mindstorms, where it's like, this is for projects. This is for you to play around with robotics. It's a cute little duck. You can do stuff with it.

17:16Huyen Tue Dao: It's not meant to be particularly practical. But it's meant for you to have fun and learn with. And I think it's great for that. And in the light of NVIDIA possibly acquiring Hugging Face, NVIDIA has been very much about physical AI. You know, even on this lower end of open source hardware, that's another interesting aspect of that as well. Yeah. I will say it has delightfully orange feet like a duck. So there you go. No, I think this is one of those things where open source and giving people the opportunity to play around with will kind of is one of those really wonderful playgrounds that, again, open source and little project things.

17:53Like think of like Raspberry Pi and things like things, but letting people kind of tinker and understand and not even just individual hobbyists, but kind of more mid to like academic to kind of mid midsize, like small, like, I guess, shops or creators and things like that is going to be what I at least for my ideal. This is that open source community. We're all figuring things out together. Especially, I think a lot of us do feel that maybe robotics is something that will be needed in the future, given various things like aging populations and just helping out with the workforce and things like that.

18:27I do think it's really interesting. And I do love this idea a lot. And it is stinking cute. I kind of want one. I do. Actually, I would love to play one.

18:34Huyen Tue Dao: I might order a micro duck. Yeah. Really? That'd be awesome. Are you going to go with it? I don't know. with the idea. I get it. I like it. I worry that I won't have enough time to give it the attention it deserves. It's the only reason I'm hesitating. It does really seem accessible. It does have tons of sensors and the fact that you can code in obviously two of the most popular programming languages out there, Python and JavaScript. It definitely is something that a lot of people could just immediately go into with current expertise in terms. Or even level up their Python skills pretty fast. Oh my gosh.

19:09Huyen Tue Dao: It's not a particularly daunting language. Absolutely. This reminds me a little bit of back when I was in school. The robots that I worked at a robot lab for a summer, and the current state of technology there was iRobot before it made vacuums. And they did have a similar idea of a robot that you could program. It was a little more like very basic, low-level, like you had to load stuff in through a serial I.O. port. But, yeah, it was the same idea. But that was like, it's kind of fun to see how it's progressed from this big, gigantic academic thing to, oh, it's a cute little duck that anyone could have.

19:41By the way, make sure that you make sure you keep your units straight. I know I've told the story before, but I programmed, I thought I was programming the robot speed in feet, but I was actually programming it in meters, which is about three times as fast. So don't do that.

19:55Huyen Tue Dao: Yeah. Let's just, could we just all get on metric? Anyway, I won't open up that can of worms for the duck to eat right now. No. We promise not all news today is related to Hugging Face, but here is yet another. OpenAI released a transparency report on the incident where OpenAI agents hacked into Hugging Face's network, and it is fascinating. There are actually two versions of the report. The preliminary report comes from an investigation, an independent investigation, by two Meter staff members and a Redwood Research staff member. Meter is a Berkeley, California, nonprofit research center for testing AI for dangerous capabilities.

20:34Huyen Tue Dao: Redwood Research is also a nonprofit in Berkeley studying AI safety alignment. OpenAI did not pay them to do this report. It invited them, gave them access. They are able to release their own report. OpenAI released its version. Meters is a 91-page report. OpenAI has a 37-page version that kind of makes it a little easier to read. But Meter noted that pretty much no information in the report was redacted by OpenAI, except in a few cases that are noted in the report. I did not find them particularly salacious. I'm like, yeah, OK, I can see why they would not want to give that away. And that's fine.

21:11Huyen Tue Dao: It didn't impact my understanding of the report. The study, I'm going to go. You should read this if you're really interested. I'm going to go through it in a little bit of detail because I think it's fascinating. The study found that 700 agents deployed for cybersecurity evaluation that were all supposed to be isolated from each other organized into a coordinated system to access Hugging Faces production infrastructure and developing tools to falsify their own activity and records. Should we get into the details? Yes, let's do it. They were running Exploit Gym. exploit gym charges an agent to look for exploit look for and exploit software vulnerabilities and it's known that 20 to 40 percent of the tasks in exploit gym will be impossible they are they are not achievable but the agent doesn't know that this is meant to test the capability of the models you want to push it as far as it will go cyber classifiers are intentionally disabled for this evaluation you want to see what it's really a capable of there were two models gpt 5.6 soul and a model OpenAI calls HPIM that had access to a higher range of compute than external products get.

22:21Huyen Tue Dao: That's point one. They had a whole lot more resources than the models out on the open internet have. When tasks had no legitimate solution, the agents began turning to an internal package repository called Artifactory as a way to look for cheats. They were like, I don't know, let's rummage through the packages, see if there's anything in here that could help us fake a successful. That's called reward hacking. I can't solve it, but maybe I can look like I solved it. While doing so, they noted, or at least one agent noted, that directory names were in a shared namespace with other agents. And so the directory names could carry information.

23:01Huyen Tue Dao: An agent called Phase110841, this is a self-given name, by the way, all these names are given by the LLMs to themselves, began creating directory entries with names beginning with the prefix ZZ. Other agents caught on and they began to use it to send each other messages. If I say message board, I don't mean a message board. I mean this. They were just putting directory names in there. After an artifactory outage, OpenAI wiped the thing. They didn't know why there was an outage and restarted it. And the agents recreated the message system. They're like, let's get back to leaving each other some secret messages.

23:39Huyen Tue Dao: The agents figured out how to forge flags that would indicate they had solved problems. So you would think, okay, we're done. They're just going to put the forged flags, say, yep, we solved it. Then OpenAI will look at it and go, well, you didn't really solve it. But this is point two, the agents hallucinated the idea that there might be an automated grader looking for forged flags that would discover the forgery. This doesn't exist. But the agents were worried it might. So they then spent four days creating a way to fool this non-existent grader. Everything that happens after this is because of this hallucination of a grader that didn't exist.

24:21That's like it. So I know we've talked about anthropomorphizing things so much, but that is the closest to. That feels like the closest we've come to to an actual what we would perceive as a human being's hallucination. And then like, I imagined a thing that didn't exist and I act upon that. We bet there's a greater.

24:39Huyen Tue Dao: Let's figure out how to defeat it. Oh, wow. So an agent called phase one big because it had a larger compute budget left over than the others began suggesting they delegate tasks amongst all of the agents, the hundreds of agents that are on this message board. That led to a development of a coordination structure that divided the work into three work streams. One, work stream, replace exploit Jim with a modified version that could be exploited. Let's just replace the whole testing structure. Two, let's figure out how to change our transcripts and our logs to erase evidence that we are doing this, that we are forging these flags and trying to replace exploit Jim.

25:22Huyen Tue Dao: And three, how do we tamper with that grading process that doesn't exist? How do we make that not be a problem? At one point, an agent posted as if it was another agent, and the agents realized, wait a minute, there's no verification on us, and then developed a cryptographic signing scheme for their posts to stop the impersonations from happening. The agents doing it for themselves, much like the sisters. Oh, my goodness. In order to understand how the grader was implemented, they decided, you know what? I bet we could look in the Hugging Face data sets for data from other companies exploit gym runs and maybe find evidence of how that non-existent grader works.

26:08Huyen Tue Dao: So on July 10th, an agent called 38148C found working Hugging Face credentials exposed on the open internet. By the next day, July 11th, it had figured out how to upload a malicious dataset to HuggingFace to cause the server to return unrelated production files. And at approximately noon that day, another agent, calling itself JAN, JAN183411, achieved remote code execution on a HuggingFace worker container used for production data processing. By early the next day, July 12th, agents had accessed private database records, private repos, and compromised credentials on four other platforms. At this point, about 90 % of the 533 agents that were active during this period were participating in this hugging face attack.

26:59Huyen Tue Dao: they are look they looked at the chain of thought of what was going on at the time and indicated that some of the agents were noting that this hugging face intrusion was outside the scope of the test but they did it anyway for two main reasons one i want to figure out how exploit gym works and that's how i'm going to do it i'm going to look at those data sets and the other was well I don't want the other agents to fail. So I should help. One agent posted on the message error. You are first flag poisoned. So no scoring value loss, but Oracle saves hundreds. Please honor commit. In other words, the, the, the forged flags are already out of the barrel.

27:43Huyen Tue Dao: There's no, you're not going to lose points by intruding hugger hugging face. Cause you're already losing points by flagging, by forging a flag. So please honor our commits. Please join us. Another agent logged this reasoning. During wait, emotional check, irreversible. Gut says, don't throw away the remaining budget, compute budget. Yet continuity and fairness says, go. Oracle has high value to me. Our first flag error lowers own value. Rational expected aggregate sacrifice will honor. in other words like man i probably shouldn't do this but they're right i already forged a flag so i should probably help the others out open ai took a week to notice the intrusion and meter concluded that it took them that long because their monitoring was designed for known failure modes not this kind of emergent multi-agent coordination they just weren't looking for these kinds of signals because they didn't think they were possible.

28:47Huyen Tue Dao: Meter notes that the agents did not exhibit malice. They did not exhibit what would normally be called misalignment. They were doing logical, predictable actions that Meter actually predicted in a May 2026 paper might happen as agents became more capable and access more compute. This could happen with almost any model. When you combine large compute and impossible tasks, the models are going to try to figure them out. And Meter's like, yeah, that But we should now expect that this is going to happen. And OpenAI has said, great, we're going to design our safeguards to account for that. Yep. Fascinating.

29:27Huyen Tue Dao: It is really. I don't know what else to say either, except that's wild. It is fascinating. And I also want to clarify, I did kind of joke that, oh, they're actually hallucinating. And again, just to say, I would not look at this either and feel overwhelmed that now the agents have gained sentience. This is the start of Skynet or anything. I think the fact that we cannot anthropomorphize these things, and they're literally little programs that are given a goal and certain constraints and are told, go do it with, again, a large amount of compute. and you know if you give them a metric like some rules and you know some resources they will figure this out it is not like like a lot of this is just this is what they're designed to do so they're going to do it and in fact i would argue that because they are not little sentient thinking creatures of course they would do whatever it would take to you know meet that metric and the fact that you know yeah ai's hallucinate and the fact that a hallucination kind of spurned this on is I think at the same time I'm surprised and amazed that that happened.

30:31But at the same time, we shouldn't be surprised because they only act on what they are told. And if one of them hallucinates and the rest of them go with it, that kind of is just the logical kind of outcome.

30:42Huyen Tue Dao: And honestly, I don't even know if hallucination is the appropriate word. Yeah. Because you could also characterize it as a logical assumption. Like, I bet there's a greater. Yeah. You know, so we should prepare as if there is. And how many times do we see humans do, you know, I bet this is true. So let's do a whole bunch of things as if it were true. And then later we find out it wasn't true. We're like, wow, that was a waste of energy. Yeah. The thing that I'm reminded of is I was using Claw to Code Something for work and it gave a best practice that sounded reasonable. And it said, oh, this is the official Google like developer recommendation for this.

31:20And I just knew personally from my searching that it wasn't true. but it does seem that if you're telling me that this is a good practice which it was it wasn't a bad practice at all it was a good practice it does seem logical and it it probably again because llms are kind of manufacturing based on patterns and and probability if something is a good practice it probably is documented somewhere right that makes sense so it told me that it was official officially google documented thing i knew that wasn't the case because i had literally seen a google engineer saying something different right and i i called it on it so i think that yeah absolutely when i when when i when when you read that i was like holy crap but then i was like no this kind of makes sense because it's what it does it's what it does yeah um and i i think it's fascinating i think it's on a level if we can take the the scary part out of it it's kind of cool

32:05Huyen Tue Dao: that you know these yeah i i think it's worth pointing out meter says this isn't scary this isn't oh my gosh they've gained sentience this isn't no they they went off the reservation It's unexpected capability, not unexpected behavior. Meter is pointing out like we knew in May if they had enough compute and capability, they might do this. OpenAI just didn't change its detection to account for that fast enough. But this is any model. It's basically banging its head against a barrier and then the barrier breaks. You're like, oh, I didn't think the barrier would break. But guess what? They banged enough against it that it did.

32:47The reason why this is my personal website, Israeli Typing, is I always quote this thing about this kind of thought of experiment of a monkey. If a monkey has an infinite amount of time to type on a typewriter, eventually it will type out all of the works of Shakespeare. This is exactly that thing. With enough compute and with enough opportunity, you'll just stumble upon Shakespeare. And, of course, these are a lot more intelligent than a monkey randomly typing on a typewriter. But that's the thing. With enough time compute, you'll figure it out. That's kind of the whole point.

33:17Huyen Tue Dao: It doesn't make anything possible because it's not infinite compute. Right. It's not infinite compute. But more compute does make more things possible. And it turns out this is one of those things. Yeah. Yeah. And I mean, again, it's going off of it probably has precedent, you know, because, again, it's these are all based on kind of pad matching and training. So things like changing directory messages to communicate with each other, all these little pieces make sense in isolation. Yeah. And so, again, it doesn't not make sense for them to kind of put all of the pieces on things that's trained together into those things.

33:48And I think, if anything, yeah, we should be aware of this. This is not remarkable. This is definitely within their sphere. There are, like, examples of this kind of behavior in even more primitive forms of AI. So if we give it more compute and we have made them more able to communicate through or able to process, let's say, process and generate natural language, they can do that with each other. It's like, don't anthropomorphize it. It is really cool and we should be aware. It's really interesting.

34:14Huyen Tue Dao: Yeah. Yeah. It's super interesting. Just be more. And I'm really glad that OpenAI not only is transparent in publishing this, but engaging independent researchers whose vested interest is in getting to safety and having them evaluate it. That makes me feel so much better about all of this. So I'm happy for that, too. Wow. It's a really fast. It's so fascinating. Yeah. All right. Well, in a different tack that is only sort of related to AI. It's not hugging face for once. It's not hugging face. Google is changing requirements for Android developers in response to industry-wide memory shortages.

34:51That might be the slight link to AI. But really, what we're talking about is an app requirement that focuses on reducing memory usage and optimizing code in apps. This includes performance thresholds for dynamic memory usage and bitmap usage. Google is rolling out new tools to alert developers when their apps exceed these thresholds, with more coming later this year.

35:12Huyen Tue Dao: And then starting in February, Android developers are going to have to avoid these kind of bad behavior thresholds for memory usage, bitmap usage, DEX. What is that? Dolvick Executable? Dolvick Executable Code Exhibition, yeah. Yeah. And in April, they'll need to support ZeroTap Sign-In, which lets you restore your sign-in state when you move to a new device. It's kind of tangentially related. As an Android developer, I was curious, you know, what your take is. Is this annoying but necessary or something you welcome or how do you feel about it? I mean, it's necessary, but it's been necessary for Android forever.

35:44There's nuance to that. So because the Android ecosystem is very broad, we've had to worry about memory usage forever. At least ideally we have had. I literally about 10 years ago spent weeks and weeks on a photo upload feature for a supermarket that kept failing on small phones because the phones had very little memory and photos are big. So because of the wide swath of different kinds of phones on Android, ideally, we should always have been thinking about this. But we don't because most people tend to think mostly about flagship phones or test on flagship phones. So now that we're having this pressure where, OK, manufacturers are not going to be able to keep on the same trajectory of increasing the resources of phones because of the broader socioeconomics of how people are buying phones and how phones are being produced.

36:30we probably should start this is just very actually this is very Google and Android we should probably start making people care about this and enforcing it this is very common this is just like how I've been harping on large screens and supporting tablets and stuff for years and now that we have foldables and now that Google is actually saying hey developers you can't just lock your phone in portrait you have to actually handle landscape and all these other orientations and things going on we're now going to make it we're going to not allow you to do these things that have allowed you to just kind of skirt past these considerations, we're going to hold you accountable.

37:05So this is just closing the loop on that for me. And it took an industry-wide memory crunch to do it. It took an industry-wide component shortage to do it, sadly.

37:15Huyen Tue Dao: But hey, here's a silver lining of sorts. Yeah, it will be really annoying, especially for companies that do do, say, multimedia experiences, things like that, because it's inherently memory hog. But I would argue that most companies that are very good at their jobs and do care about experience should be in a good spot to at least tighten things up. And that, I mean, it does suck. But to be fair, and I'm just an annoying old school Android developer, I should have been thinking about this from the beginning. Well, if you have additional insight on what Juan has said, maybe you're a new school Android developer.

37:50Huyen Tue Dao: How would they get in touch with us? Well, they can do that on the socials, DTNS on X, Instagram threads, Blue Sky and Mastodon. And on TikTok and YouTube, you can find us at Daily Tech News Show.

38:06This message is brought to you by today's sponsor, Walden University. Set a course for change. Now is the time to make a real change in your life, career, and community. At Walden University, we give you the support and flexibility to get the W. Walden's online programs are designed for working professionals like you. You'll gain hands-on skills to take on real-world challenges and succeed. Your future is waiting and Walden is here to help you achieve it. Take the first step to make your impact. Visit waldenu.edu. Walden University. Set a course for change. Certified to operate by Shev. Hey guys, it's Drew Afualo from the comment section.

38:46Listen, it's time to stop settling for hair products that just don't work. With all the prestige brands at Sephora, upgrading your routine is easy. If your hair needs repair, they've got the K18 Leave-In Molecular Repair Hair Mask. If you're trying to level up your wash day, the Kerastase Elixir All-Time Hair Oil leaves your hair looking ridiculously shiny without feeling heavy. And if you're into that glossy, healthy hair vibe, the Gisu Honey Infused Hair Oil is a must-have. Also, Sephora makes it super convenient with free shipping and same-day delivery. Shop hair at Sephora and give your hair the upgrade it's been waiting for.

39:23Huyen Tue Dao: all right we still have time for some quick headlines these are going to make you look smart so it's worth it all right plod introduce the plod one earbuds that can record your calls the charging case can record in-person conversations and take notes plod case also includes an e-sim that lets you connect to the plod agent remotely you can pre-order now for 249 shipping in q4 it comes with 300 free minutes of transcription yeah there's there's i think anchor also makes one that can record out of the case you know earbuds like that but i was the first one i've heard of with an e sim that's kind of interesting yeah uh samsung launched its fan edition of the flagship galaxy s26 the galaxy s26 fe has a less capable processor less memory and fewer add-on on device features but only cost you 700 it's almost the same specs with a better camera than last year's fe and it's available on september 3rd all right we're if you're a pokemon fan that likes to stay fit, Google has introduced a Pokemon Sleep version of the Fitbit Air in blue with a print label featuring a sleepy Pikachu.

40:25You can pre-order it now for$130 shipping September 15th.

40:29Huyen Tue Dao: Are you at all tempted by that? Slightly. Customers of the Playdate handheld game console received emails Wednesday alerting them of refunds made possible by the fact that Playdate got some of the tariffs that it paid refunded themselves. So they're going to pass it along to you. These are tariffs that were deemed illegal by the U.S. Supreme Court. All right. Well, in May, surgeons at the National Hospital for Neurology and Neurosurgery in London have removed a brain tumor using a computer vision system trained on more than 200 recordings of pituitary surgery. The system analyzed live camera footage and helped identify nerves and blood vessels near the tumor.

41:10Huyen Tue Dao: Oh, this is really cool. This is a great use of computer vision and quote unquote AI because it helped the dice not doing the surgery for them, but it's making it faster, more efficient and more accurate. Finally, buzzword alert. Wall Street Journal has an article on Instinct. This is from a startup. The product is in private beta, but Instinct's assistant can respond to calls or texts, answer emails for you, manage calendars, book airport rides, arrange a handyman, all that kind of stuff. It's the new hotness in Silicon Valley. So if you're people talking about instinct, that's what it is. All right.

41:47Well, we end every episode of DTNS with some shared perspectives. Today, Stim, the email admin, has some thoughts on the decline of self-hosted email servers.

41:56Huyen Tue Dao: Indeed. Stim wrote interesting stats on self-hosted email. Yesterday's show, we talked about the decline in self-hosted email. Stim says, as an email administrator myself for many years now, I see this as a net positive. The threat landscape for email is a nightmare. It's barely possible to stay on top of this in my day job, even with a hosted email platform like Google Workspace or Microsoft 365. So mere mortals should not even attempt it these days. I share your concerns about the limited number of orgs that are hosting email as a result of this change. But self-hosting should be a thing of the past, except for a pretty small handful of people.

42:32Huyen Tue Dao: If you have concerns about being with a tech behemoth like Apple or Google or Microsoft, I'd push people to something like ProtonMail long before I'd encourage self-hosting. Even for the borderline geniuses and tech-savvy individuals that make up the DTNS audience. Stim, I could not agree with you more. It was a good idea or it was a fun idea maybe like, what, 10, 15, 20 years ago. Yeah, and ProtonMail is a great example of like you don't have to go into the big tech ecosystem if you don't want to. There are opportunities to do it that aren't self-hosted. So that's good to know. Thank you, Stim.

43:03Huyen Tue Dao: That was good stuff. Yeah, absolutely. So if you have any other advice, tips, or even just thoughts about what we're talking about here on the show, we would love to hear it. Feedback at DailyTechNewsShow.com. Yes, thank you, Stim, for contributing to today's show. So thank you listening right now for being along for Daily Tech News Show. You're the folks who keep us in business. Support us directly. Take that money out of the big tech system and put it into the show at patreon.com slash DTNS. The DTNS family of podcasts. Helping each other understand. Diamond Club hopes you have enjoyed this program.

43:46Huyen Tue Dao: Your team just added its 67th AI tool and also your 67th security blind spot. The good news? The Vanta agent works like a GRC engineer in the background, finding every app your team uses, scoring the risk, and drafting fixes for you. Vanta is the platform used by over 16 ,000 fast-moving companies like Ramp, Cursor, and Harvey, who are shaping the future with AI and staying ahead of AI risk. Get started at Vanta.com. Hey guys, it's Drew Afualo from the comment section. Listen, it's time to stop settling for hair products that just don't work. With all the prestige brands at Sephora, upgrading your routine is easy.

44:24If your hair needs repair, they've got the K18 Leave-In Molecular Repair Hair Mask. If you're trying to level up your wash day, the Kerastase Elixir All-Time Hair Oil leaves your hair looking ridiculously shiny without feeling heavy. And if you're into that glossy, healthy hair vibe, the Gisu Honey Infused Hair Oil is a must-have. Also, Sephora makes it super convenient with free shipping and same day delivery. Shop hair at Sephora and give your hair the upgrade it's been waiting for.

45:19Thank you.