OpenAI Slows Down and Focuses on Safety - DTNS 5335

[music] This is the daily tech news for Wednesday, August 19th, [music] 2026. We tell you what you need to know, give you important context, and gosh darn it, try to help each other understand. >> Today, Open AAI is announcing new plans to keep you safe from those rogue agents. >> They're coming from everywhere. I am not a rogue agent. I'm Tom Merritt. >> I'm coming from inside the house. and I'm Sarah Lane. >> You literally [laughter] are. Uh let's start with what you need to know with the big story. >> Yeah. So, OpenAI uh made a post which is the beginning, I suppose, of disclosing changes to its safety practices. Uh we'd had some rumors recently that they had changed their safety team to kind of move safety people into the actual production teams and this seems to bear that out if you think about it that way. Recently, OpenAI announced that it would pause development on an internal model it calls Astra for two weeks uh because they believed Astra had reached a critical threshold for security capabilities. OpenAI now says it is pausing development on more models that are farther from shipping than Astra. I'll just quote from the posting. Our largest planned frontier RL reinforcement learning run remains on hold while we conduct smaller scale training and evaluations to assess model behavior, validate our safeguards and establish more evidence of alignment before proceeding. So what that means is OpenAI is always training new models for the most part. It is not training those new models right now while it figures out its new safety practices. Uh, OpenAI is also rewriting its preparedness framework. Uh, the last time it updated it was back in 2023. It is strengthening monitoring in its development process by introducing alignment. Alignment is is what the model's trying to achieve the same as what humans want it to try to achieve. Uh so they're introducing alignment and security safeguards earlier in the development of models and also adding some higher safeguards post training. So adding the safeguards in earlier and then raising them after you're done training when you're doing the inferencing and testing. OpenAI will also add controls on workloads from the internet during testing. This was a big factor in the rogue agent stuff. uh they wrote, "We have designed these controls so that a single compromise of a workload or a supporting service does not by itself allow for unauthorized access to the internet or other internal networks." Uh that that's a hindsight is 2020 move, but a good one. Uh it will remove potentially vulnerable shared services. So not just vulnerable shared services, but ones where it thinks like maybe they could find a bug there. We're going to just pull that one out there. and also do some simulated attacks to test boundary conditions. Issues will be flagged within 30 minutes of detection. In other words, they shouldn't be more the the faster the better, but no more than 30 minutes. And security and research teams should respond immediately once they're flagged. And if they can't identify it as a false positive in 30 minutes, then pause the model activity. Uh don't wait to keep looking. It will monitor all models in training. Uh it didn't always do that. that it only monitor monitored the most high-risisk ones, but now it's going to monitor all the models in training. And it will also monitor inferencing, that's the when you ask it a query and and it kicks back after training. Uh if a model displays critical capabilities like Astra, OpenAI noted that some security measures will increase its compute overhead by 20% of the observed inference workload, but said that's for internal research, so it's not going to impact how much you pay for the API. It'll raise the overall cost of running open AI. So, it's going to impact cost down the road, but it's not something that's raising the compute for you on the other end. This is all for the internal research. Company is increasing the computer resources for understanding the way systems reason and act. They wrote those systems examine tool actions, available reasoning, and the full sequence of activity for unauthorized access, data theft, destructive behavior, and attempts to defeat safeguards. Uh, and that's just the beginning. That's what they wrote today. OpenAI says it's going to share more details about its monitoring scheme in a future post. Sarah, do you feel more secure? >> Yeah, I do. I I I feel like if you are somebody who is suspicious of Open AI or any company that is building models like this, then you might be suspicious of what the company has said. Here are all the safeguards that we we're putting into um you know into into the uh [laughter] the the chain uh of command and and I think that open AI being extremely transparent about this is the right call. Um, let folks know what happened, why you realize that other other uh safeguards need to be put into place and let everybody know exactly how that works. Because when nobody knows how it works, that's where a rogue agent jumps out of the sandbox, does something weird, and then we all talk about it for 3 weeks. >> Yeah. I I think the the lazy criticism that that people will have uh is why didn't you do this before, right? Uh and that that's sure fun and an easier way to win an argument. I don't know that it's particularly useful though. Uh I also am curious why they didn't have all these safeguards before, but I say that not as a I'm trying to score points on Open AI, but like I'm curious like was it well we never had a problem, so we didn't think we needed them. Uh or was it uh we just honestly didn't believe that you needed to have safeguards that far out? Cuz we can go to today and look at the safeguards they've put in place and there is room for something to happen bad. You can never without not using the model entirely rule out that something bad will ever happen. But these seem reasonable to me right now which is like tighten up the sandbox, put the safeguards in earlier, uh monitor faster. That is the one where I'm like, really? You didn't have a 30-minute response time before? Okay, that's interesting. I'm wondering why that was. Um, but you know, this is also there probably just wasn't an incident right that ever happened, at least to anyone's knowledge, within 30 minutes. It's also kind of saying to somebody who's very, you know, adept at coding, like, well, why'd you make that mistake? >> Yeah. It's like well I mean just cuz we're [laughter] that happens sometimes and then you fix it and then you know the software is better as a result. I don't I don't mean to say that it's not a big deal uh that some of these safeguards could have been implemented before and you know hindsight's 2020 like you said so it's it's we're kind of this is where we are and I think as safe as we can be and again being able to explain h here's how we're safer because now we understand more about what happened you know this is still the wild west of all this stuff and what else can you do? >> I I do get concerned when I hear stories that there were people internally saying we need to have these safeguards moved. Uh there's a possibility of doing this and others said I don't think so. Uh and then the question becomes did they say I don't think so because they just wanted to be risky and move faster uh which is what a lot of people think >> or they just didn't believe or they just were like you had great info >> you know that that's like you know whenever you you uh you are securing anything and someone's like why are you going to all of that trouble why >> me being like I think the salsa's bad in the fridge and someone else being like it's probably fine >> yeah [laughter] uh and then someone eats it doesn't get no remembers it, someone eats it and does get sick, they're like, "I told you." So, >> Yeah. Right. >> Exactly. Uh, security is salsa. That's what we have learned today. >> Yeah. And salsa is delicious. >> And it is. So, security >> as long as it's, you know, in good standing. Uh, DTNS is made possible by all of you listening right now. You're all in good standing with us. Thanks to Scott Wilhelm, Johnny High-Tech Okei, and Chris Zerugosa. There is more to talk about today, so let's get to the briefs. >> Let's do it. The Pew Research Center conducted a survey of opinions about AI among 3,488 US adults from June 22nd through June 28th. So, you know, kind of kind of a small date, but a fair amount of folks. 52% of those surveyed said they were more concerned than excited about AI, matching the previous high of concern from a survey back in 2023, but almost level with the past two results of 50% and 51%. So, it's a little bit, you know, more skeptical, but sort of on par. But for the first time, concern passed 50% to 55% for adults who are 30 years of age or younger. So while concern among those aged 50 to 64 fell to 47% the first time below 50% for that age group it's the younger folks who are saying I don't know I'm not sure about this. 71 of those surveyed thought AI would lead to fewer jobs in the US over the next 20 years. That's up from 64% 2 years ago. So more people are concerned about this. This sentiment rose across all age groups. So, not just people who are, you know, graduating college and entering the workforce for the first time. Pew Research also wrote, quote, "Adults under 30 are more likely to think AI will be bad for society and for them." They largely think it will make connection and creativity harder to come by. And [clears throat] pardon me, though most say they use chat bots, they're just as likely to say those tools hurt their own creativity as well as help it. >> That was me in my 20s. like, uh, oh yeah, this is bad for me. I'm still doing it. Uh, so I'm not surprised by that. Oh yeah, I'm using the >> you HAVE AI IN YOUR 20S. >> NO, NO, I'm just saying things, you know, in general when you're young, right? >> I'm kidding. >> Yeah. Yeah. Um, 71% thinking AI would lead to fewer jobs is higher than I would have thought. I I would have guessed it was in the 60s, and it has been in the '60s. It's been pretty close to that. Uh, I think the thing that really amused me was the younger folks being more concerned about this than the Gen Xers. Uh, the the the Gen Z's, the millennials, and the boomers are all more concerned than they are positive, whereas the Gen Xers just get more positive every every year. Uh, you could they're the only one with the line going down. Uh, as a member of Generation X, I have no explanation for that. Uh, but >> I think we're just like, you know what, [clears throat] we made our bed. >> Yeah. We've seen worse. >> Yeah. like it'll be fine. We'll figure it out somewhere. I I do It's not surprising to me that the younger folks think that this is going to be a a hindrance to getting a great job. >> Yeah, that makes sense. >> That it it's it is a high number. 71% you're like, "Oh, well, you know, okay, you know, we're we're pulling what uh you know, 3,500 or so people." So that's not, you know, it's not like we pulled the entire US, but I think that that that that is a it's a real concern and that is a concern that millennials had and gen like it's it's it's a concern that all people have when they're very young adults. Um, when I say get out of college, you know, you don't even necessarily need to go to college, but you're in that age group where you're starting to go like, "All right, well, so what's my career?" And if your career is something that has been rewritten over the last few years, at least in your mind, yeah, you're not going to be happy about that. >> Yeah, >> it's it's scary. >> The the most interesting thing about that is uh it makes sense that the younger people would be worried about jobs. Everyone with a new technology is always worried about jobs. And as I've said before, generally speaking, uh the technology ends up creating as many or more jobs than the jobs that are displaced. Uh and the problem is always in how fast that happens and how smooth the transition between jobs goes. Uh we have examples of very rocky transitions in the industrial revolution and we have examples of very smooth transitions in the computer revolution of of the mid- 20th century. uh nobody knows until it's done which one we're in right now. So 71% of everybody surveyed uh saying I'm worried about jobs uh isn't surprising although it's a little higher than I would have thought. I expect the young folks to be worried about it uh and I expect maybe the older people wouldn't be as worried about it but this is across all age groups. Uh and in fact while Gen X was saying no we think it's better we're think it's going to do more positive or at least break even than be harmful. they were still saying, "Oh, but it's going to get rid of a bunch of jobs." Gen X was right there with all of the other uh demos on that. So, uh I I think you cannot underestimate how worried people are about this kind of technology, especially because this one is moving a lot faster than previous technologies in its capability as discussed by needing these kinds of safety procedures that we talked about earlier. >> Absolutely. And I think I think when you you mentioned sort of the Gen X uh folks being like, "Ah, what's the big deal?" I I think that that is it's an interesting category of jobs where you're not close to retirement yet. You you got a ways to go, but you're not retired. You know, you're not just starting out. If you're just starting out and being like, I don't even know, you know, what I'm going to do, that is a much uh potentially uh bigger uh problem to to hang over your head than someone who's like, I just have to figure out what to do for the next like 5 to 10 years. >> Yeah. >> And then I'm out of there anyway. So that that would the the boo I don't we should probably uh shy away from being too reductive on this but but the boomers are saying I think this is bad and I think it's bad for jobs but I'm retired so I'm concerned for my grandchildren. Uh the millennials and the Gen Z's are like I think this is bad and I think it's bad for jobs. The Gen Xers are like nah I think it's mostly good. Oh it's bad for jobs but you know [laughter] what you got to do. >> They're like come on like how long is this really going to bother me now? we're we're being silly a little bit, but at the same time, I I think that if you if you feel like, okay, I can adapt to our new normal, you know, as it changes on a daily basis for the next decade or so, that is a much easier pill to swallow than I have no idea what's going on. >> You know, I'm 21 and where's my job? >> Yeah, absolutely. Uh let's talk drone delivery. Uh because at least you can get a burrito sent to your house by a quadcopter. Uh even if you don't have a job, drone delivery is expanding in the US with Alphabet's Wing passing a million commercial deliveries this year. Zipline has done 2 million across four continents. And Amazon, which got all the attention in the early days when they went on 60 Minutes and talked about it, finally increasing its efforts uh to bring Prime Air into the conversation as well. Primary delivery going to expand to suburban Chicago and Atlanta as well as metropolitan Cleveland and Syracuse, New York by the end of the year. Amazon already delivering in 11 US locations that include Houston, Dallas, and suburban Phoenix. And Amazon says it should be serving 500 cities and towns after this roll out is done, which it says over the next several months, but it doesn't say if it'll happen exactly by the end of the year. Chicago and all those getting it by the end of the year. 500 cities and towns overall going to get it soon. Uh, if you're curious, Amazon uses an in-house design it calls the MK30. Six propellers on it. Uh, so not a quadcopter, it's a six double copter. Anyway, uh, range of 7.5 miles >> and cost you $3 per delivery if you're a Prime member or free if your order is worth more than $50. If you're not a Prime member, you pay $5 per delivery. Meanwhile, in the UK, Prime facing some headwinds. They are capped to a 100 deliveries a day during an accident investigation. One of the MK30 units crashed into a back garden after it collided with an object extending above the roof of a moving vehicle. Uh that happened on July 29th. I'm guessing maybe a ladder on a truck or something. Uh Amazon is going to have to finish that investigation before it applies for a renewal of its approval to operate in Darlington in the UK by December 31st. Well, so in theory, I love this. I want this. I don't know exactly how this works uh at my particular residence. Um and I was talking to Jason Howell about this a couple weeks ago on DTNS. Uh you know, how how how does it safely bring me the things that it needs to bring me and and where does it go? Because I don't have a helilipad. Uh, don't put it on the roof. Nobody can get up there. Even my my neighbor and I, um, and this is just sort of a funny visual for everybody. We were taking avocados off of a neighbor's tree cuz they don't like they don't seem to care about the avocado tree and they're beautiful avocados and we're like, we have to do it, you know? So, here we are like on a roof with a ladder and I'm just thinking if there was some sort of Amazon drone delivery that we confused in some way, you know, is is everybody safe in that scenario? It's kind of like safety first. Uh I also have a little bit of an issue with the $3 delivery because I'm a Prime member. Um and this is not even available in my neighborhood, so I can't try it out anyway. uh uh you know it it it kind of turns into well uh does Amazon save on gas uh because you know then they don't have to have a truck come to my house probably. Um I I don't totally know you know how that works out for the company. But it doesn't really make sense for me because I'm like I'll just use the track because then it's free no matter what I get. I don't have to buy $50 worth of of goods or more. But if you do that, you know, maybe it's pharmaceuticals or or just stuff that is going to be a little bit pricier. Maybe maybe you just get creative. Sometimes it's a road delivery, sometimes it's an air delivery. >> Yeah. So, you don't think they're safe, huh? >> I don't think that. >> I mean, we've had a million commercial deliveries this year and like two or three accidents. It's not It's not that I I don't think that in general it's unsafe. I think that there are certain areas where the drone is just going to be like, I can't safely drop this, you know, and you're going to have to go back and forth with Amazon about that. >> I'm curious for people who live in the areas that actually have this. Uh are you someone who's like, "Yeah, I'm supposedly covered, but it won't come to me." You know, like uh this is not a new problem. This is this is not something that these these delivery companies haven't thought of and they seem to have a lot of ways to do it because you don't need a helellipad, right? It just needs a clear area near your residence. But you're right, if you got a bunch of trees in your front yard, you know there where how is it going to pick what that clear area is? Or what if there, you know, the landscapers are there that day and you just, you know, I don't know, you didn't think about that beforehand or drones are able to avoid people. That that that stuff's pretty proven technology at this point. >> Yeah. But it could it could just be confusing for the drone basically. It's like >> I'm just saying a million commercial deliveries and we don't have a lot of examples of running into landscapers. So it's it seems like they've figured that out for the most part. Allison Sheridan also had a good point uh who said, "What if my dog is in the backyard when the yummy burrito delivery happens?" >> Yeah. >> That's like what if your dog's in the front yard when the Uber Eats guy comes, right? Like you you do need to mailman who he doesn't like anyway. Right. >> Yeah. But but these are all sort of I think that >> these are the kinds of questions we have and we never experienced it. I think >> Exactly. And then once it happens, you're like, "Oh, it works like that." And it becomes mundane. Same as a Whimo. People are like, "Well, how does it know where to go? And how what if it's going somewhere I don't want to go and how do I talk to someone?" And then you got past all of that cuz now you're like, "Oh, yeah. I've been in Whimo a couple of times. It's totally >> honestly it like works well." This is it. It's I don't think being a pessimist is the right call, but we all want to be pessimistic about something new cuz we're like, "Well, what if it doesn't work?" Mhm. >> And and I think that that's where I land on on all of this. Trust me, I want drone delivery. This is not something I think is a bad idea. I just think like, oh, it's that one situation where something goes ary and then everyone says, "You ruined my backyard. You know, the drone fell in the pool." Or, you know, something. >> Yeah. Um, and then for $3 a day, I that'll go away eventually. that that's a this is new and we kind of want to put a break on it getting too popular too fast. >> Uh and so, you know, we'll just we'll just put a little >> little you want the special delivery, you want the slightly faster delivery, you know, you pay for that. And if it does take off, haha uh and become popular, I I expect that will that will change. >> Thanks to Motang for posting this next one in our DTNS subreddit. Uh, Comcast is marketing motion sensing for customers who use its Xfinity Gateway XB7 wireless modem. So, if a customer opts in by turning on the feature in the Xfinity app, the modem will detect interruptions to the Wi-Fi signal and send notifications if the homeowner [snorts] is away, for example. But there is a catch. If you turn on the feature, and this comes from Comcast specifically, quotes, Comcast may disclose information generated by your Wi-Fi motion to third parties without further further notice to you in connection with any law enforcement investigation or proceeding, any dispute to which Comcast is a party or pursuant to a court order or subpoena. >> Yeah, I think this is just boilerplate, but it's causing a lot of people to raise an eyebrow. Uh, I think what it's saying is, you know, once we have evidence, it could be subpoenaed. Um, however, in connection with any law enforcement investigation or proceeding or any dispute or pursuant to a court order, so it's saying if law enforcement comes and asks for it, they don't necessarily need a court order or subpoena. They don't mention warrants, so I'm concerned about that. uh I don't want them giving away this information without a warrant. Uh and any dispute to which Comcast is a party uh is a very wide openen legal door. Um this is not selling it to third parties for ads or or anything like that. Thank goodness. Uh but there there's there's a couple loopholes in here that that I don't particularly love. >> I'm not Yeah, I'm not sure how worried this makes me. I mean the of course the legal speak everyone kind of goes like doesn't sound good but uh but yeah I don't I don't know exactly how this would screw you up if you had this sort of situation and you opted into this and again you >> someone accuses you of something goes to the police and complains about it. You're not guilty of it but the police get uh some details about who was in your house and then suddenly the police know things about you that you didn't want them to know and they had no business knowing. Mhm. Um, and I also wonder how uh, you know, it for sure not being you versus somebody else, you know, can be a point of contention. It It does >> sound like muddy waters a little bit. >> Yeah, it's a little muddier than I would prefer in here. I don't expect that this is going to go haywire and and suddenly, you know, it's going to be ripe for abuse. But, you know, if you're if a third party company has uh details about what's happening in my house, I want to know what they're going to do with it, how it's protected, and who gets access to it. Exactly. >> Yeah. It's worth paying attention to. >> Well, if you out there want product reviews from people who are putting the gadgets and services and platforms to the test, then your new show is Live With It. Thanks to everybody who's already watching and listening. We'd like more of you. Each week we talk about the tech in our lives. Mini splits, rice cookers, fitness solutions. In fact, speaking of fitness, this week Wu Dao is walking us through her experience so far with her new Fitbit tracker. And if anybody knows fitness, it's Wen. Honestly, you should follow her on Instagram. She is a powerhouse. Uh you can find us on YouTube at youtube.com/livewithitshow. And big thanks to everybody who is helping us try new things out every week at patreon.com/livewith. [music] >> All right, we got some quick headlines for you, but you know what? We give these to you because they're the kind of things that just stick in your head and you're like, "Oh, wait. I know. I heard about that." And you feel smart. >> Yeah. Here's the first one. Google has begun rolling out its advanced flow for sideloading apps whose developers have not been verified by Google itself. So, when you turn on the setting, you have to restart the phone, wait 24 hours before installing an unverified app. >> This is the solution to people being mad that you couldn't install unverified developers, even if they weren't in the app store. Now, you can. You just have to wait 24 hours. Security thing. >> Yeah. >> Amazon announced it will roll out its advanced Alexa Plus voice assistant to all Fire TV devices in the US for free. No Prime membership necessary. Uh, it cost $20 a month for non-prime members on nonFire TV devices, but now if you got a Fire TV or a Fire TV box, you get it for free. >> Yeah. I mean, I am a Prime member and I have a Fire TV Stick, so doesn't really affect me one way or another, but, you know, if you're curious, this would be, you know, a interesting entry point. Bite Dance and Tencent have each received about 10,000 Nvidia H200 processors. The US approved the US approved the chips for export to China back in December, but China didn't permit the imports until now. >> A trickle but a beginning. China's Unit Robotics, which actually officially uh trades under the name Yushu Technology, issued its first stock shares IPO on the Shanghai Stock Exchange. Share price surged 460% at first, but then finished at 629% up. So people very excited to trade. >> Yeah. Yeah. Wow, that is that that's a fun first day on the market. Google sent media invites out for a Google Book announcement event on September 15th in New York City. Google Books from Acer, Asa, Dell, HP, and Lenovo have all leaked over the past several weeks. So, we kind of know what we're getting, but still kind of a fun announcement. >> Yeah, September 15th, we find out the Chrome OS successor. Anthropic revenue in the last quarter was reportedly 11.6 6 billion with a small profit already. They haven't even IPOed yet. Uh and by comparison, uh Wall Street Journal pointing out that OpenAI is telling investors its revenue was $6.7 billion in the same period. If you're uh not paying attention, 6.7 less than 11.6 billion. Uh and also a large loss. So, OpenAI not yet profitable. China's land space recovered the first stage of its reusable rocket on Wednesday, joining SpaceX and Blue Origin as the only private companies to land a reusable rocket. >> And Cerebrris announced its nextgen wafer scale engine or WSE and Nexus rack systems can double perch chip performance and fit three times as many chips into a rack. Uh that's going to save people building data centers power, which is something we all want them to do. if they're going to build them, we want them to be more efficient and it'll save space as well. >> We like to end every episode of DTNS with shared perspectives from all of you. Today, Dr. Douglas was listening to our conversation on the backlash against wearable cameras, particularly in smart glasses. >> Yeah, Dr. Douglas uh was was also writing about the fact like it reminded him of smartphones. remember when we first had cameras and smartphones and you couldn't take them into certain places and there was worries about people recording them in, you know, locker rooms or stuff and we've all gotten used to like, yeah, all the phones have cameras. Sometimes you have to put them away otherwise it's just don't don't be recording in the locker room. >> Yeah. Don't be recording with it there. Uh, and Dr. Douglas writes, "I have a little bit of an invested interest in this because I have a learning disability that though not as severe as face blindness, is still limiting where I cannot connect most people's faces with their names to the point that it is a clinically diagnosed issue. The idea of being able to have a pair of smart glasses that has facial recognition and could tell me the name of the person I'm talking to would be life-changing for me. I teach and to be able to call my students by name would be amazing. It may sound like a small thing, but we put so much stock in people showing they care by knowing our names that this is important for me. Thanks for all the great work you do to help us understand tech every day. Dr. Douglas Mulford. Thank you, Dr. Douglas. >> Indeed. Indeed. Yeah. And just a a great reminder for, you know, there's so many folks out there being like, "Oh, these are going to be used for bad reasons." There are bad actors. there will never stop being bad actors, but somebody like Dr. Douglas is like, "This is actually going to make me a better professor. >> It's going to, you know, help me connect with people better." And yeah, that's that's great. >> So, we we shouldn't deprive Dr. Mulford of this. uh we should figure out a way to make sure that he can take advantage of it and the bad actors can't take advantage of it which >> harder than it you know it's not a it's not a thing you can just do by leaning back in your lazy bulge tear but >> well it's like it's like if I get in my car right now and I mow you know into a restaurant it's like bad actor >> yeah right >> most people would never do such a thing you know it's like it's like you have to understand that these things are possible with the technology >> and then take the precautions. Don't just ban the category. Right. >> Exactly. Exactly. I'm with you on that. Well, if you're thinking about anything that we talked about on the show, if you have insight into a story we might talk about in the future, do share it with us. We read every single email feedback at daily technewshow.com. >> Big thanks to Dr. Douglas Molford for contributing to today's show. Thank you for being along for Daily Techne Show. You're the folks who keep us in business. Thank you patrons. If you want to become a patron, you get the feed ad free and a bunch of other stuff. patreon.com/dtns. Go take a look. [music] It's really cool. Don't miss it. >> The DTNS family of podcasts helping each other understand. >> Diamond Club hopes you have enjoyed this broker. [laughter]