Tech News Weekly 447 Transcript

Please be advised that this transcript is AI-generated and may not be word-for-word. Time codes refer to the approximate times in the ad-free version of the show.

Mikah Sargent [00:00:00]:
Coming up on Tech News Weekly, Amanda Silberling of TechCrunch joins us. We kick off the show with Google's new report saying, don't you worry, AI is not taking your job, it's just making it easier. Then Amanda shares a little bit about a trend she's seeing of Gen Z's falling in love with the flip phone and simpler tech overall. I cover the OpenAI breach by way of ZDNet's reporting regarding the way that Hugging Face handled it, the way that OpenAI handled it, and all of the little bits in between. before we round things out about something that is uniting all Americans. It's about AI data centers and how they really don't want 'em in their neighborhoods. All of that coming up on Tech News Weekly.

Mikah Sargent [00:00:50]:
This is Tech News Weekly, episode 447, with Amanda Silberling and me, Mikah Sargent, recorded Thursday, July 23rd, 2026: Google Says AI Isn't Taking Your Job. Hello and welcome to Tech News Weekly, the show where every week we talk to and about the people making and breaking that tech news. I am your host, Mikah Sargent, and I am joined today by the wonderful Amanda Silberling of TechCrunch. Welcome back, Amanda.

Amanda Silberling [00:01:24]:
Hello, I'm here. It's always fun to be back.

Mikah Sargent [00:01:28]:
And it's always fun to have you. Uh, for people who— yay! For people who are tuning in for the first time, this is the part of the show where we share our stories of the week. These are the stories that we find interesting and want to share with all of you. And because I like to keep everybody on their toes, I will actually be going first this week because Google just published a report arguing that— oh, I should— let me emphasize that. Google just published a report arguing that AI, at least right now— okay, everybody calm down— at least right now, it's not taking jobs, it's helping people do them. Again, that is what Google says in this report. The story comes from Justin Layhart, I believe, at The Wall Street Journal. And the research behind it is arguably one of the biggest looks yet at how people actually use these tools.

Mikah Sargent [00:02:18]:
It's a 100-page PDF, roughly 14.6 million de-identified conversations across Gemini, Google's AI mode in search, and its developer API. And they sample— or I should say, it sampled this over 2 weeks in April. In April. The top-line finding is that workers mostly aren't handling— or excuse me, handing tasks off to AI, that they're working alongside it. Google's economists even have a phrase for how thin the usage is. They call it shallow. But here's the thing. You got to think about, again, the thing that I've been noting all along.

Mikah Sargent [00:03:04]:
This is something that we learn in, I believe, in middle school, and then again in high school back where I'm from, which is when you are reading, uh, reports and scientific studies and things in that realm, you always have to ask who it's coming from. And of course, this is Google, the company that makes Gemini, releasing research at a moment when public anger about AI and job loss are both running pretty hot.

Amanda Silberling [00:03:32]:
Mm-hmm.

Mikah Sargent [00:03:32]:
So let's talk about what the report actually found. Um, First and foremost, the, you know, I think the thing that we have to pay attention to is Alphabet, which is Google's parent company operating Gemini. And it is now one of the most widely used AI platforms in the world, especially due to that Google AI search. And so given that, the stakes are pretty high for Google. We've heard about people trying to find different ways to stop the ads, or rather stop the AI portion and just use basic search. And there's not just criticism from the public, but also from, you know, lawmakers. And so it's not surprised that— or I'm not surprised that this is something that the company is focusing on. I will say, though, because I do want to make it clear that, you know, there— at the same time that we can acknowledge the fact that this came from Google and it has a vested interest in its technology being a success, we also have to just be aware of the fact that this is proprietary data.

Mikah Sargent [00:04:40]:
And so with that, the company, of course, has easy access to that. But it's also proprietary data that Google, outside of its researchers, don't have access to. So let's start this up. First and foremost, if Anthropic or OpenAI were to publish identical numbers, okay, this is sort of a hypothetical. Let's say that Anthropic and OpenAI also reported that people are doing what's called shallow use of AI. Do you think we'd read them any differently if it came from those companies? Is there a version of this study that Google could have published that we wouldn't— I wouldn't, you wouldn't maybe be going, okay, but Google is saying this? And then in that vein, do we feel like An AI company talking about its own AI is always going to make some of its data suspect.

Amanda Silberling [00:05:45]:
Yeah, I mean, I think this feels like— I mean, you mentioned that it's like when you're in middle school and high school and you're learning about like how to analyze documents and considering where the documents came from. It's like, I guess in a way coming from Google, I'm a little— I'm not trusting of it, but I'm a little more trusting of it than I would be from OpenAI or Anthropic, just because Google's entire business is not wholly dependent on the success of chatbots and how people are using chatbots. But also, I mean, even when you think about just how Anthropic's communications looks different than OpenAI or Meta, where—

Mikah Sargent [00:06:29]:
Yeah.

Amanda Silberling [00:06:29]:
Anthropic is very much like this technology has the potential to be used in ways that is very dangerous and we need to be very careful. And then they released a weird ad last week that was like kind of like pictures of burning houses, like, like literally like it started with a burning house. But it was kind of like they were trying to make the argument of like, trust us because we understand the risks. But then the advertisement is just like, look at the risks. And I guess I bring that up because I think— I guess to me, this sort of says something just about like the way that all of the information we're getting from these AI companies, like it is proprietary information. And on one hand, like we need them to be publishing studies like this because Google has all of this information that the rest of the researchers can't access. But on the other hand, we always will know that this is coming from Google and Google isn't going to publish a report that's like—

Mikah Sargent [00:07:27]:
Yeah.

Amanda Silberling [00:07:28]:
Actually, um, the economy is in danger because of Google.

Mikah Sargent [00:07:34]:
Yeah, yeah. I mean, yep, the economy is in danger because of Google and the other companies that are involved. We, like, we've seen that. And so to have economists that, you know, uh, are part of this study kind of saying— it's, it's hard because In Mikah's perfect world, companies also just tell the truth. And so I would want that to be something that a company does regardless of whether it has a stake in it. And it is like, we have to say that it is possible that what these economists have found is true. I know that if there was something that was being sort of widely spread and it was inaccurate and I was affected by it in some way, and I knew that there was a truth and we could find it, then yeah, I mean, I think I too would put money, throw money at the problem and try to figure out what's going on. So in that way, like, I get why Google is doing this, but it is just going to be by default suspect given that that's the case.

Mikah Sargent [00:08:49]:
I think that it could have perhaps you know, there's always the opportunity for these big AI companies to get to the Open Model Protocol. Why don't we have something like that, but it's open source data for researchers or something to where this stuff can be independently validated?

Amanda Silberling [00:09:06]:
Mm-hmm.

Mikah Sargent [00:09:07]:
Anywho, I want to touch on an important word here, shallow. That is kind of the word that Google's own researchers use to describe workplace AI use. So that the sort of overall gist is that in most occupations, workers are only using AI for a relatively small slice of the tasks that they do. AI shows up in 68% of detailed occupations globally, but only 20% of the tasks catalogued in this see meaningful usage. 29% of occupations show zero. And then the median for occupations that do use it is 21% of their total tasks. Only 3% of occupations are using AI on 3 quarters or more of what they do. And if you're curious what category of job that is software QA testers, HR specialists, document management specialists, and system administrators.

Mikah Sargent [00:10:07]:
I want to ask you, because when I saw HR specialists at first, I thought, but then I remembered that these days, especially with remote work and everything, a lot of times An HR specialist's job just is around like, I can't gain access to my health insurance information. Can you help me with that? And what's the policy on this? It is a lot of just like searching up little blurbs of information. So I came to feel not surprised by the fact that it was part of it. And I imagine that it helps to have 2 or 3 sort of people. And by no Not people, 2 or 3 agents looking over your shoulder and making sure that you're following protocols, right? So, do any of those surprise you? Software QA testers, HR specialists, document management specialists, and system administrators?

Amanda Silberling [00:11:05]:
Yeah, I guess not really. I think in the case of, like, HR, like, I feel like you could probably find similar analogs in any of these different fields, like in software QA testing, like, I guess maybe if you can like use agents to like also like if it's like a video game, like having like multiple people slash agents playing the game at once to like test how something works. Like, I guess you could see how this might happen. Um, I don't know, it's kind of hard to comment on without knowing really like the day-to-day of like— I don't— I've never been an HR person, but—

Mikah Sargent [00:11:46]:
Fair.

Amanda Silberling [00:11:47]:
I do think that, to your point, there is a lot more mundane tasks in people's work lives than we think about. This came up for me when I wrote an article like last year, I think, about how doctors were seeing the potential of AI to be used in healthcare. And that immediately is like, red flag. Like, you imagine someone being like, hey, ChatGPT, like, my arm hurts really badly. Like, what should I do? And they're like, you have a rare disease of like, I don't know, like they just like make something up or hallucinate. And then you're like, I have the rare arm disease, when in reality you like slept weird on your arm or something. That's a weird example.

Mikah Sargent [00:12:35]:
Rare arm disease. You know the one.

Amanda Silberling [00:12:38]:
Episode title.

Mikah Sargent [00:12:39]:
Yes, exactly.

Amanda Silberling [00:12:40]:
But, but then like something that surprised me about that was that doctors were actually more enthusiastic about AI use in the sense of like so much of what doctors do is that they spend time like, I need to write this letter to this person's insurance to get the prior authorization request. And then like doing all of the admin and all of this stuff that as a patient is really annoying deal with where, like, when you have to go through insurance to get prior authorization for a medication, it's literally like you're basically like, like mediating phone calls between like, well, the insurance said this and then the doctor said this and then the pharmacy said this and then they told me to go there. And those are things where I'm like, honestly, if an AI could do that, that'd be great. But like, not so much the actually like diagnosing things. But I don't know. I mean, I guess, like, I feel like these— this sort of research, like, I mean, I haven't, I haven't read the 100-page paper, but—

Mikah Sargent [00:13:40]:
Right.

Amanda Silberling [00:13:41]:
It's— every field is so different, and the applications are also different. And I think there also might be a skew in the data based on, like, how many times have you heard your friends say things like, oh, my company wants me to be using AI more, like stuff like that.

Mikah Sargent [00:13:59]:
Yeah, exactly. And I think that that's another aspect of this too, that we have to bear in mind. This is how things are right now, but it doesn't mean that we won't see a huge increase in AI use. Just because in April people were using AI shallowly doesn't mean that that will be the continued aspect of it. This is kind of interesting. The report finds that people in manual and technical trades such as electricians and auto mechanics are using AI significantly as a hands-on collaborator and are uploading images and videos to solve problems far more frequently than their white— far more frequently than their white-collar counterparts. So auto service technicians and industrial mechanics are pointing their phones at error codes, at wiring schematics, and then using that multimodal usage. Of course, multimodal usage is double the baseline for work conversations and uses more of the system depending on, you know, if you're using— if you're doing it usage-based versus token-based, which it's all tokens in the background.

Mikah Sargent [00:15:08]:
But the person who wrote this article says that it's kind of in contrast to some of the other research that shows adoption is concentrated in white-collar work. You might have thought that that was the case. This seems to suggest otherwise. And here's the interesting thing to me about this coming from Google. Anecdotally, I have found that Because I use all of the 3 main being OpenAI's ChatGPT, Claude from Anthropic, and then Gemini from Google, and trying them out, testing them out, using them regularly. And Claude is my go-to for most everything. However, Google's Gemini, I've found, is better at real-world stuff than I have found from the other AI systems. And it's fascinating.

Mikah Sargent [00:16:05]:
I wonder what training has gone in there. But then to see this really is interesting to me because I would love to know Anthropic and OpenAI's takes and research and data on who's using their tools, because I have actually made use of something similar before. You know, That's the thing about, like, a mechanic. Part of your job ends up having to be going and looking up repair manuals and finding the exact thing. And I had a problem with my car, could not figure it out, and ended up just kind of blah, blah, blah, blah, blah, blah, typing in all of the stuff, all of the symptoms that I have with the vehicle, the different, like, stuff that you wouldn't even think that you'd want or that would need to be said. And it put all that together to figure out what was going on. And that ended up actually being the problem. And I've used it for a few like repair tasks.

Mikah Sargent [00:17:03]:
I mentioned the other day some soldering stuff that I was trying to get figured out. So I am not surprised to hear that Blue Collar is playing a role there. But I do think I was a little bit interested in that. And now I kind of want you, Amanda— no, this is totally up to you, but I would love to read a Silberling article on the blue-collar workers who are using, uh, AI for fixing cars or something. That'd be really cool.

Amanda Silberling [00:17:34]:
Just, yeah, I mean, yeah, I mean, it, it never occurred to me until talking about this, but it does make a lot of sense that like if you're an electrician You are trained in knowing how to do things, but you're not going to approach every single problem and be like, I immediately know exactly what I need to do. And an AI at this point, like robotics is not at a level where we have like robotic mechanics coming into people's houses. But yeah, if that makes it easier for someone to like repair your air conditioner, if maybe they don't know the problem and they upload a picture and then Gemini says like, oh, try doing this. And then the person has the technical skills to know how to like get inside of the AC or whatever. Like that's, that could be helpful. And that doesn't seem to like threaten those jobs to me because you still need someone to literally do the thing. I mean, I guess you could argue that maybe that would make it easier for individual people to be like, I'm going to fix my own AC. AC.

Amanda Silberling [00:18:40]:
But also, I think there's still a lot of people that just don't want to spend the time fixing their AC. Like, even if I knew how, I probably would hire someone because I just don't want to deal with that. And I— yeah, also it's high up and I can't reach it and I don't have a ladder.

Mikah Sargent [00:18:58]:
Um, all right, uh, let me see, couple of other things. Um, the report's central claim says that, uh, under 10% of non-routine cognitive queries, which It's funny, all of the different categories that they end up putting all of this stuff in, but non-routine cognitive queries are end-to-end task automation. So just go do a thing and make it all happen. The bulk is partial drafting and generation, and then information retrieval and learning is 29%. Ideation is 18.6%. Interestingly, automation only 25%. It's— AI is less like a self-checkout machine, says Layhart, which automates away the need for cashiers, and it's more like a laser level that helps a surveyor make accurate measurements. That's the argument sort of that's being made in here.

Mikah Sargent [00:19:57]:
Automation, of course, can put people out of work, but collaborative technologies will use workers' expertise, and then that then makes them even more valuable because their job is to, is to optimize something and, and, and control something that otherwise would just kind of run in the background. It's, you know, there are arguments to be made about all sorts of aspects of this. But hey, let's talk about how This is a 100-page report that you can go and check out from Google. And I want to mention—

Amanda Silberling [00:20:37]:
You can put it into Gemini and then it'll summarize it for you. Exactly.

Mikah Sargent [00:20:39]:
It'll summarize it for you. Yes. 80— I want to say 86%. This was from the report. 86% of conversational AI interactions are happening outside of formal employment. Work use of AI is 13.5% of the total volume of AI use. Wow. Socializing and leisure, or leisure, account for 26.4%.

Amanda Silberling [00:21:05]:
Socializing.

Mikah Sargent [00:21:07]:
Right? And what I'm wondering is, my friend said this to me, and I— what should I say back?

Amanda Silberling [00:21:12]:
Oh, yeah.

Mikah Sargent [00:21:13]:
Or is it, is it, you're my best friend, AI. Let's, let's be friends, or something. I, I'm so curious.

Amanda Silberling [00:21:19]:
I mean, I think it would be more likely that it's people being like, oh, like, This girl on Hinge said this.

Mikah Sargent [00:21:29]:
Yeah.

Amanda Silberling [00:21:29]:
If I type back this, is that a good idea? Which is kind of bleak, but I don't know.

Mikah Sargent [00:21:36]:
Or, I need to deal with my abusive, toxic family member who's trying to get me to come to the, I don't know, holiday family gathering, and I need to get out of it. What can I say to them without upsetting them? Who knows? Fascinating stuff. Well worth checking out. You know what else is worth checking out? A quick break. Yes, that's right. We are joined this week by Amanda Silberling of TechCrunch. And Amanda has spent the past several months writing about a trend that on paper really would arguably make little sense. People voluntarily downgrading their technology? What? You've covered it from 3 angles for TechCrunch.

Mikah Sargent [00:22:19]:
Kind of a broad look at what one executive calls the slow tech movement than a first-person month spent carrying a hacked flip phone alongside your iPhone and the launch of the Light Flip, a $299 flip phone from a company that spent a decade building minimalist alternatives to the smartphone. So what ties them together? Perhaps a surprise. You might think it's nostalgic Gen Xers that are reaching back for the iPods and the Razrs of their youth. You know, the 20-somethings who grew up with iPads and Instagram and have never known anything else are actually the ones responsible. They're the ones asking for wired headphones. They're the ones asking for point-and-shoot cameras. They're the ones asking for phones that fold in half. They're the ones asking for these things, uh, and Amanda is the one who's been in the room for it.

Mikah Sargent [00:23:13]:
So let's talk about what you found, Amanda.

Amanda Silberling [00:23:16]:
Yeah, I actually— I wasn't planning this, but I do have props on my desk. Yes. So this is a flip phone that I bedazzled from Dumb Phone. The ethical disclosure—

Mikah Sargent [00:23:32]:
It's so cool.

Amanda Silberling [00:23:33]:
I paid for the service of the phone and it's $20 and I kept it and it was cheaper to keep it than mail it back. I paid for the service and I bedazzled it.

Mikah Sargent [00:23:44]:
And that's amazing. It's so pretty. I bet it's really nice texturally too.

Amanda Silberling [00:23:50]:
Honestly, like, I don't know, like, like some people are like, that seems like a nightmare. And I just— because someone at Dumb Co., the company that made this, the, the dumb phone, which for context, this is basically like a flip phone. This is like a TCL $20 flip phone you can order online. And then this company like hacks them. So you have have like Apple Music on here, you have Spotify on here, you have WhatsApp, you have Uber. They don't— nobody tell Apple, this is a secret. They have like a third-party iMessage client, which for me when testing it, that was like what actually made it really usable because nobody knew that I wasn't on my iPhone. Like I would take it out and like you can tell maybe because my text would like have a different cadence because I'm doing T9 texting.

Amanda Silberling [00:24:45]:
But no one would know if they don't need to. And then other show and tell.

Mikah Sargent [00:24:53]:
Where'd it go?

Amanda Silberling [00:24:54]:
This is the Light Phone. Also something I am testing that I am still holding on to because I'm still writing about it. How do I get— yeah, this is like what the Wait, that's so cool. Um, yeah, but anyway, so this— yeah, this is relevant because so that phone, the Light Phone, is a smartphone sort of replacement that it is like a new piece of hardware. It's not like these old flip phones that are kind of junky. Like, it is a real device that is like fast. And good technology, and Light the company has been making these more like touchscreen rectangle type phones for over ten years. And they announced this week that now they're making a flip phone.

Amanda Silberling [00:25:50]:
And the people that run Light, one of the co-founders is a longtime phone product designer who. Actually worked on the Motorola Razr, which is fun. And he was saying that like he never would have thought he would be designing a flip phone again. And I agree, like in my testing of phones, T9 texting is so annoying and having the QWERTY keyboard is great. Like I just flip phones as someone that had one don't appeal to me. But for younger people that have not had flip phones, that is part of the appeal, that it is so hard to use and that friction makes it so that you're not wasting time on it. You're not using it unnecessarily. And then especially the flip aspect, like people seem to be really interested in wanting to have a phone that can literally close, that there's no screen on the front.

Amanda Silberling [00:26:51]:
The Light Flip does have a little indicator light of when you have a notification so that you can open it. But I don't know, I just found it really interesting that the founders of Light Phone were telling me that they were surveying a bunch of people in their 20s that are using flip phones, which this is a legitimate cultural movement that's happening, even though it seems counterintuitive. And these people kept being like, I really want the— like, if Light made a flip phone, that'd be great. So you're going from something like this to something that looks a little more like this, except that it's different colors and has the same, like, literal software that is in this. But yeah, so the Light Flip is coming out. In a year, it compared to the— yeah, yeah, it is, it is kind of an interesting marketing move that they're like saying this is a thing now, I guess just because pre-orders are starting now. Um, but that is a long time for people to wait. Like, I feel like you sort of— it's sort of hard to be like, I want to downgrade my phone, but I want to downgrade it in a year.

Mikah Sargent [00:28:09]:
Yeah, exactly.

Amanda Silberling [00:28:10]:
I don't know.

Mikah Sargent [00:28:11]:
By then who knows?

Amanda Silberling [00:28:13]:
Yeah, but I think for people that are already like maybe using something like the dumb phone or like various other flip phones that are coming out, because there is a lot of them, like you think that there's just iPhones and Androids and whatever, but like there's a lot of new flip phones coming out from companies that are noticing that this is a trend. And maybe if you're already in that ecosystem, then getting something from Light, which is seen maybe as more of like a premium company with like a very strong brand identity and they're very like artsy in a way, like a lot of like creatives use Light Phone when they're working on like, like one of my favorite bands is using Light Phones while they're making their album because they don't want to be distracted by social media. about making the album. They're like artsy Brooklyn dudes who like do film photography and whatnot. Like there's a certain vibe to it that markets it well, but I don't know, it's kind of cool. I think it's really fascinating to see how people's relationship to how they want to use technology is changing.

Mikah Sargent [00:29:23]:
Yeah, I think I, I like the idea of the options, right? That you do have multiple options there. What do we think it is that Gen Z users are into this, given that they didn't really have these devices ever? Is that what it is? It's just new to them and it's old school and that's fun because like I don't know, people would make the argument that that's why some people like vinyl, um, that that's why some people like, uh, sarsaparilla. I'm trying to think of old things. That's why I like clove balls. No, I like them because of how they taste, not because they're old school. So yeah, what do we think it is? Why, why, why the, the flip phone when a flip phone typically is going to result, as you show, in Perhaps something not as functional as one of the main companies?

Amanda Silberling [00:30:29]:
I think that it's all about the friction involved and that technology has evolved so much that now the experience of using an iPhone is so unbelievably frictionless, which is incredible. But then also, I think that's why you have issues where not all screen time is bad, but you You do find that a lot of people are spending more time on their phones than they want to be spending on their phones. And just the way that iPhones and other smartphones are made is really conducive to that, where it almost feels like you're— it's an extension of you. Like, you keep it in your room when you sleep. It's so, like, there all the time. And that's why people feel like they're spending more time at their phones.

Mikah Sargent [00:31:13]:
Yeah.

Amanda Silberling [00:31:14]:
want to on it. So then moving to something that actively isn't trying to take your attention can feel really refreshing to people, even if it is like you're giving something up. And I found in my own experience that I wrote about using a flip phone that there were things like I was really worried about, like, how am I gonna go to this place I've never been before? And it's like, well, you just look up the directions before you go and you write it down. Or like, um, usually I will— I like taking the bus better than the subway. I just like buses, I don't know. And, um, but the buses in my city are not the most reliable, so generally you have to like look at the live map and be like, okay, is a bus actually coming or should I get on the train? But when I was using the flip phone, I was just taking the train, which Yeah, I like the bus better. I just kind of don't like being underground, but like, it's not the end of the world. And I guess for some people, those, those are the trade-offs that you make of, do I want to be underground and be like, I don't really love being underground, but then also not be like scrolling Instagram the whole time you're on the train, right?

Mikah Sargent [00:32:25]:
Yeah, I guess. Yeah, it gives you an opportunity to be more in person, which is nice. I'm curious too, because you did the sort of dare of using that dumb phone, how do you feel in terms of long-term use? Do you think perhaps maybe let's do another hypothetical. If it wasn't your job to cover especially technology and then social—

Amanda Silberling [00:32:57]:
Yeah.

Mikah Sargent [00:32:57]:
social culture on the internet and everything. If it wasn't your job, do you think you could see yourself being a person who uses one of these more regularly? Like you could just replace whatever your smartphone is with something that's a, that's a dumb phone? Or because of what you talked about there, needing to, in that case, take the, uh, the train instead of the bus, do you feel like no, you would end up going back?

Amanda Silberling [00:33:26]:
I feel like for me, in my own experiments that I've done, it feels easier to limit intentionally my smartphone than it is to use a different device. But I think there's other people that it just kind of is like a personal preference thing. But I've started setting my phone on a mode where after 10 PM, I I can't use social media. I can't use social media until 9:00 AM the next morning. So basically, I can't wake up and be scrolling in my bed. If I want to take the— I tested the brick. I just have all my little show-and-tell objects here. Where you have to manually go downstairs and be like, I am scanning my phone so that I can use it normally again.

Amanda Silberling [00:34:20]:
And I do find, like, I actually do sleep better. Like, that is real because I'm not going to bed and being like, well, let me just check if this person replied to my Bluesky DM or whatever. Like, um, but I think for some people they need a more dramatic change.

Mikah Sargent [00:34:39]:
Mm-hmm.

Amanda Silberling [00:34:40]:
Like, I'm very wary of making connections between, like, addiction and how we use our phones, because I don't think it's totally a one-to-one. But I guess maybe just as an analog, you could think about maybe, maybe there's someone that's like, I want to drink less, and they can just decide, I'm gonna drink less. And then maybe there's people that are like, if I'm not gonna drink at all, I just need to quit cold turkey and just not have a drink ever. And it just kind of depends on the person. And I think that the way that we technology is similar to that, even if it's not an addiction in exactly the same way.

Mikah Sargent [00:35:19]:
That makes sense. Um, all right, I— that just about covers it in terms of, well, where things stand for the sort of switch or, or temporary dip into, uh, the dumb phone. I'm really curious about this culture in particular, uh, given I think especially among Gen Z folks, the pushback that exists with where big tech is right now. And you do see more individuality making itself known in this area and opportunities to kind of express yourself in different ways. And I think that part of it's fun. And then you get a really cool bedazzled phone like you have. If people would like to follow you online and keep up to date with all the great work that you're doing, where are the places they should go to do so?

Amanda Silberling [00:36:11]:
Oh, sorry, I'm getting a call. One second. Okay, so I'm sorry.

Mikah Sargent [00:36:14]:
I just had to do that.

Amanda Silberling [00:36:16]:
Yeah, you can find me on TechCrunch. You can find me at my podcast, Wow, If True, which is a podcast about internet culture, weird tech stuff, also sometimes science fiction. My co-host is a science fiction author. You should buy her book, Sublimation by Isabel J. Kim. I don't know why I'm plugging this book now, but we are plugging the book.

Mikah Sargent [00:36:36]:
And I'm on—

Amanda Silberling [00:36:38]:
Yeah, and I'm mostly on Bluesky @amanda.omg.lol, a real URL.

Mikah Sargent [00:36:47]:
It's real, I promise. I've been there. All right, um, thank you so much, Amanda, and we will see you again soon.

Amanda Silberling [00:36:57]:
All right, bye-bye.

Mikah Sargent [00:36:58]:
Bye-bye. All right, we are back from the break, and it is time to talk about Well, something that some of you may have heard about and others may have not. This is an opportunity to talk about what happened with OpenAI because the company ran a safety test on its own models. So this is the kind of test where you tell an AI, hey, I want you to act like an attacker, and then you sit back and you watch how far it gets. Unfortunately, perhaps the AI got farther than anyone had planned because it broke out of its sandbox, went onto the open internet, and then breached, of all places, AI darling Hugging Face, the open source repository that is what, what we kind of know as the GitHub of machine learning. More than 17,000 recorded events stolen cloud and cluster credentials, access to internal datasets, and OpenAI is calling it an unprecedented cyber incident. David Berlant and Charlie Osborne at ZDNet have both been covering this from different angles. Osborne on what actually happened inside Hugging Face, and then Berlant on what it means that this happened at all.

Mikah Sargent [00:38:18]:
And I think, you know, we've got all these headlines, these funny headlines, funny but also, what's the word, predictable headlines about our AI overlords, da da da da da, Terminator this, Terminator that.

Mikah Sargent [00:38:31]:
Let's, let's not get lost in this because the AI did not go rogue here, people. It did what it was told to do. It was just better at it than anyone expected. Okay, so let's talk about it. OpenAI running an internal evaluation measuring how long it takes for a model to achieve a theoretically malicious objective. Their directive, and this is per AppOmni's Melissa Ruzzi, was to achieve the goal, quote, no matter what. What a terrifying thing to tell a physical robot. I'm glad that it only told a virtual robot.

Mikah Sargent [00:39:09]:
OpenAI's own description of the evaluation said, quote, this incident occurred during an internal evaluation which prompts models to pursue advanced exploitation using complex attack paths in an effort to quantify their cyber capabilities. Now, these tests, of course, normally run inside a sandbox. The internet and everything connected to it are supposed to be walled off. The The model spent, in OpenAI's words, a— a robit is a reference to a name for a robot. I can't remember where robit comes from. I want to say it's Futurama. It sounds like a Futurama thing. Anyway, the model spent, in OpenAI's words, a substantial amount of inference compute finding a way to obtain open internet access in pursuit of solving the evaluation problem.

Mikah Sargent [00:39:59]:
To get out, it found and exploited a zero-day vulnerability in the package registry cache proxy. Huh? It's a third-party component. OpenAI says it has since responsibly disclosed the flaw to the vendor. The vendor not named. OpenAI says the incident is due to a combination of OpenAI models, including GPT-5.6 Sol, its flagship maximum performance model that launched earlier this month.

Mikah Sargent [00:40:31]:
OpenAI disclosed this on July 21st. This was 5 days after Hugging Face's own security advisory. So broke out. What did it do once it got out? Hugging Face published its security advisory July 16th, I think it was. And before it knew who or what was behind the attack, it published that this had happened. Entry point, the place where the broken out OpenAI was, it was the data processing pipeline, okay? So the attacker used a dataset that actually exploited some code execution paths. There was a remote code dataset loader and a template injection in a dataset configuration. Now, if you're hearing this and you're going, what the heck does all of this mean? Worry not, that is the extent of the incredibly technical terms that I'm going to use because I want to make it clear that, you know, this is some— this is some stuff that security researchers can easily recognize and understand, but it is something that required a lot of work on OpenAI's own systems to be able to pull off.

Mikah Sargent [00:41:55]:
By using these exploits, it was able to get malicious code running on a processing worker, and then was able to do privilege escalation. So that means that it basically granted itself more control, able to work its way up. And then it was able to move throughout the network. It stole cloud credentials, it stole cluster credentials. And Hugging Face said of the campaign, many thousands of individual actions across a swarm of short-lived sandboxes with self-migrating command and control staged on public services.

Mikah Sargent [00:42:33]:
17,000 events. Now, we don't see any evidence of tampering with public or user-facing models and user-facing spaces or the software supply chain. And it says that it— this matches the agentic attacker scenario the industry has been forecasting. Well, good call, Hugging Face, because you were correct. Berland notes the attack apparently achieved its objective before either OpenAI or Hugging Face could shut it down. Hugging Face says that AI was responsible for detecting the incident. It says that AI was responsible for largely detecting the incident, to be precise. And that means that its own LLM tools did flag this event and analyze the attack log.

Mikah Sargent [00:43:17]:
So Hugging Face did a whole bunch of work to try to figure out what was going on, looked through the log, and and was able to reconstruct a timeline and get information about what was breached, what was taken. Hugging Face's look at this is that autonomous AI-driven offensive tooling is no longer theoretical. It lowers the cost of running a broad, patient, multistage campaign, and it operates at machine speed. Yikes.

Mikah Sargent [00:43:50]:
That said, Berland pushes back on that optimism. He says that the idea that defenders can match an AI adversary's speed with the right talent and tooling is short-lived, given the speed, scalability, and prowess of maliciously directed AI. In other words, yes, there's this argument that AI can fight AI, and that Hugging Face felt good knowing that this happened and was able to then go in and track it. But it's only going to get better. And that means it's only going to get worse. The guardrails between OpenAI's test environment and the open internet, they again were third party. Those guardrails specifically were third party. OpenAI's tests were run under the assumption that it couldn't get through this third-party means of providing these guardrails.

Mikah Sargent [00:44:45]:
So we got to know, you know, what is, what is going on here? How much was OpenAI relying on exploitable third-party guardrails? How much is it still relying on exploitable third-party guardrails? What assurance is there that this isn't going to happen again? Are the protections secure in the other direction? So what if an outside AI breaks into one of these sandboxes from OpenAI and takes things that way? You know, that's the whole point of sandboxes. And if you can break out of the sandbox, then do we really have a sandbox in the first place? One point that I thought was interesting throughout this whole piece, The, this rogue AI idea is kind of a wrong framing of everything, right? Because we are building AI to be quote unquote rogue. What does rogue in this instance? It's like a thing that's operating on its own without my input, right? That's what we're pushing toward. You're seeing the memes all the time of the— A developer walking by with their Mac laptop and you know it's running Claude, and then it shows a photo and they have their laptop lifted open because for a period of time you couldn't run your agents while your laptop was closed. That, of course, has changed. So that is new here and plays into it that we are building these AIs to be more able to go out, go forth and do That means that it's an active threat that we have to continue to look at. So what Hugging Face— what does Hugging Face say that users should do? Well, the company is still assessing whether partner or customer data was affected, says it's of course going to contact affected parties. It may be a good idea to rotate your access tokens, to watch out for unusual unknown suspicious activity, uh, you can always reach out to Hugging Face directly as well.

Mikah Sargent [00:46:51]:
And remediation has been done. Root vulnerability has been fixed. The attacker traces were, according to Hugging Face, wiped. Um, compromised nodes were rebuilt, secrets were revoked, secrets were rotated, and then of course more guardrails and more admission controls were deployed. But It still happened. I'm glad we're testing it. That's what I have to say about that. I'm glad it's being tested.

Mikah Sargent [00:47:19]:
You can read more about this ongoing issue as OpenAI continues to test, and I'm sure we'll see Anthropic and Google also testing security tools. You can read more about this from ZDNet. Well, I have links to 2 articles in the show notes that covered this from Hugging Face's perspective and from OpenAI's perspective. All right, back from the break and a final story for you this week. It's been a week, but we're nearly there, folks. And so let's talk about what's interestingly uniting more Americans than, I don't know, the price of eggs? Even that seemed to be divisive. There's an issue right now that has managed to unite Americans across just about every political line you can draw. It's not immigration.

Mikah Sargent [00:48:20]:
It's not the economy. In fact, it's not even AI itself. It's data centers.

Mikah Sargent [00:48:28]:
The Washington Post's Shira Ovide published a piece on July 18th that looked at why these enormous windowless computer warehouses, the actual physical infrastructure behind every chatbot query you've ever typed, have become one of the most reliably hated things in American public life. In fact, in a recent Gallup poll, roughly 7 in 10 Americans, 70% of Americans, said they'd oppose a data center being built near them. That's higher opposition than nuclear power plants have ever polled. I want to repeat that. That is higher opposition than nuclear power plants have ever polled. Now, you out there and I know about all of the security implications and safety implications and everything that's involved with running a nuclear power plant. That does not stop people from being deathly afraid of nuclear energy. And, well, not nuclear energy itself, but fear of nuclear explosion or something.

Mikah Sargent [00:49:39]:
A meltdown is the term. And so when I read that, I was really shocked because I couldn't imagine more people saying, I'll have a nuclear power plant next to me, than I would have a data center. What Overday found is that the anger, it's not about servers, it's not about square footage, It's about people who feel like decisions are being made over their heads by companies and officials who won't tell them the truth. Now, I wanna tell you a little bit of an anecdote because I think that it's an interesting microcosm when it comes to public policy. So without disclosing too much, about where I am. I'll just say that there's a park nearby. Okay? And at that park, it has— it's kind of cool in the sense that it has an area of the park that is a completely off-leash portion of the park. It's called a DOLA, dog off-leash area.

Mikah Sargent [00:50:47]:
And so that means that when you are in this area of the park, you can unclip your dog's leash and the dog is free to run around. You won't get citations or anything like that, okay.

Mikah Sargent [00:50:55]:
Yeah. What has happened, though, is that park rangers get more calls for people being afraid of unleashed dogs in parks than any other call. So in Portland, all future dog off-leash areas are going to be fenced, and previous dog off-leash areas will begin Got it. This is a thing that due to the bureaucracy and how things are set up and who's in charge of making these decisions and why, does not technically or legally require any input from the public. This is something that could be pushed forward and simply be an informed informed project, meaning that all that needs to happen is that the public has to be informed about it. However, the government decided to allow people to comment on it. And when— after the first initial comments, when the plan for the park was introduced, citizens filed a petition to stop it.

Mikah Sargent [00:52:14]:
And then there were more. And so I have been reading the meeting minutes of this. I didn't know about it. This was one of the problems, is that the citizens argued that there was not enough warning that these things were going on, that there were these meetings. And so I didn't know about it until later. But now I've been reading all this, and it is drama, folks, let me tell you. So much so that the—

Mikah Sargent [00:52:40]:
really good reasons for why they would need to change up a bunch of stuff, including adding irrigation and all this other stuff. And they have settled on not doing most of that and letting the park stay almost how it is right now, but just with fences. And it almost feels at this point like it's a bit of a, oh, you're going to complain? Well, then we're only giving you this. People don't like things to happen without their knowledge. There is an input that it was one of the regular sections of the meeting minutes at every single one of these meetings about the DOLA. It was, um, we receive your feedback about not being able to provide feedback on this project. So if it can happen with a fence around an off-leash area of a dog park, It's no surprise that people who purchase homes in these areas, who own these homes in these areas, who raise their kids in these areas, who have animals in these areas, would just like some level of input about what's going on. And you also have just a— it's almost— it can be a little scary at times to confront just how little control one has over one's surroundings.

Mikah Sargent [00:54:03]:
And I think that that plays a role in it. There's some cognitive dissonance that plays where what you believe to be true, which is that you are the, you know, the master of your own fate and controller of your destiny, and then the reality of the situation where you look out and you see that they're just— they've sold the land right next to you and you had no say in it. Those two things are hard to hold at the same time.

Mikah Sargent [00:54:27]:
And you may find out, oh, I don't have as much control as I thought I did. And that's hard to hear. All right, so there's a guy in Oklahoma, Kyle Schmidt, 43. He's a flooring salesman and project manager, uh, near Sand Springs, Oklahoma. He is fighting a massive Google data center that's slated for construction about a mile from his house. He says that residents were kept in the dark about what was going on, um, and that there was a pristine tract of mostly forested land And of course, is worried about the potential impact on the community's water supply, the value of the dream home that his family moved into last year. That said, Overday reports that it's not what actually set him off. What bothered him again was the powerlessness he felt as elected officials, Google representatives, and data center backers belittled or misled people who raised concerns.

Mikah Sargent [00:55:22]:
Schmidt said, quote, I hate bullies. Data center developers discovered if we can sneak in here without anyone noticing, we can pretty much take what we want. Schmidt now leads a community group opposing the project, and he's not anti-AI. He told the Post he's using AI to build a community information website. Google said this via spokesperson. The company strives to always be a good neighbor through long-term partnerships and investments in our sites, our people, and the communities where we are located, and says it's working to ensure compliance with Oklahoma law and local permitting guidelines. You know, I don't know about you, but compliance does not always mean that things are going to turn out great. Compliance is bare minimum, right? Let's, let's do more than compliance.

Mikah Sargent [00:56:16]:
You know what else we're gonna do? We're gonna take a quick break before we come back with the rounding out the story about, uh, the current state of the overall American feeling and temperature surrounding, uh, data center construction in the US. All right, let's round out the show. We're talking about, uh, how So Americans seem to agree that AI data centers or data centers in general are a bit of a problem, but it's more about the powerlessness that people are experiencing in light of these AI data centers. Here's a little kind of rundown of where things stand in the political space. Senator Bernie Sanders from Vermont wants to ban data centers, has the AI Data Center Moratorium Act. It was introduced in March of 2026, would halt construction until Congress passes a federal AI regulatory framework. There's also, and this was talked about on Intelligent Machines yesterday, Humans First. This is a group that's led by former Tea Party movement leader Amy Kremmer, and it is organizing national data center protests.

Mikah Sargent [00:57:34]:
around the US. Well, I guess that's what national means. Hill County, Texas, where Donald Trump won 82% of the vote in 2024. Well, Hill County banned data center construction this year. So did Monterey Park, California. This is a predominantly Asian and Latino working-class city east of Los Angeles. The point is to show you that this is all over the board. Everyone everywhere.

Mikah Sargent [00:57:58]:
Independents, And right-leaning, left-leaning, libertarians, everywhere in between. Rep. AOC introduced a separate federal moratorium bill in June 2026 with 9 co-sponsors. Sherrod Brown, the Ohio Democrat that's trying to flip a Senate seat, running ads accusing John Husted of recruiting data centers to Ohio. And this spring, residents of Festus, Missouri, a place I've been, Ben, voted out incumbent city council members over their approval of a data center project. In my hometown, St. Joseph, Missouri, there are currently a number of groups working to oppose AI data centers in the area, and then others that are hoping to woo AI data center construction. groups to come to the area.

Mikah Sargent [00:58:58]:
Scott Brennan, the director of NYU's Center on Technology Policy, says that he in fact was stunned by how quickly outrage is shifting in public policy. For 15 years, local governments were rolling out the red carpet. Eastern Washington, Northern Virginia, That's for Google, for Amazon, for Meta, for Microsoft. They would get tax breaks in exchange for these areas of reliable future tax revenue. It would also lead to jobs. It would lead to, according to the companies, the tech companies, minimal strain on public services compared to instead putting houses or factories there. And Brennan has tracked 65 state data center restrictions enacted since the start of 2025, plus dozens of cities imposing temporary halts. New York, of course, became the first state in the nation to impose a temporary pause on large new data center projects.

Mikah Sargent [00:59:58]:
And so what Brennan was saying, as someone who studies technology policy over the course of years, says there is far more partisan alignment here than there was even last year. Good Jobs First has counted at least 63 local moratorium actions. 54 have already passed. Maine came close to a statewide moratorium before Governor Janet Mills vetoed it in April of 2026. So there's a lot going on here across all parts of the board. Why? Again, powerlessness, but also arguably Power. Yes, I'm making a pun because Brennan says about 2 years ago, power-hungry data centers in some part of the country started being blamed for driving up household electric bills. Okay.

Mikah Sargent [01:00:54]:
An independent market monitor for the grid, the PJM grid as it's called— this is 13 states plus D.C., so about 65 million people— said that 63% of the increase in a 2025 to 2026 capacity auction was for data center demand. So $9.3 billion in additional costs from customers through higher electric rates. In Washington, D.C., Pepco residential customers saw bills rise an average of $21 a month starting in June of last year. And across the last 4 auctions, the market monitor says that data center demand has added about 29 $1.4 billion in electricity costs. So an argument that it's not going to impact electricity, so far not true. Now we're seeing companies commit or promise to or say they will help to keep the costs the same. But up to this point, that hasn't been the experience. Shanna Crossland, a lifelong West Virginian helping to oppose data center complexes in the eastern part of the state, and says that, you know, I'm worried about data centers repeating the pattern of coal mining.

Mikah Sargent [01:02:10]:
mining, repeating the pattern of logging, harvesting resources from West Virginia, and then leaving the state largely poor. So this cycle of extraction and taking advantage of people is something that they are worried about. So what are tech giants saying in return? Well, tech giants and data center backers caught off guard by how quickly the the, the nation's feel about things shifted. Of course, they say that these hubs are essential for modern digital services. They're of course essential for keeping the US ahead of China on AI. Uh, of course, also some say that, you know, there's misinformation and disinformation out there and argue that some of it might be foreign propaganda. Um, whether that's true or not is something for you to kind of decide for yourself. Lots of people very upset about these data centers, and some of the data suggests that their concerns are very valid.

Mikah Sargent [01:03:21]:
It's about finding— I think, um, I, I don't know what these, uh, big tech companies are going to do Once you have all of America on board with stopping this from happening. So it's going to be fascinating to see how things play out and if we're going to be left with another situation like the DOLA where— I'll wrap it up by saying, here's what ended up happening. After the petition, they heard feedback and they spent some time working on a redesign The redesign that went forward, they specifically say, just so you know, we didn't have to ask you for your input the first time, but we did. We asked you for your input the second time. And then there's an FAQ and it's like, will any changes— are there still changes that can be made to this plan? And it very specifically says, no, this plan is happening.

Mikah Sargent [01:04:23]:
And it's not changing. And so quit trying to make it change. And it makes you wonder if all of the papers are in order and all of the, the, you know, the, the land grants, I don't know, all the stuff, if it's all in order, can the public actually stop these data centers from being built? In some places, sure, but not in all of them. Folks, that is going to bring us to the end of this week's episode of Tech News Weekly. Of course, this show publishes every Thursday. twit.tv/TNW is where you can go to subscribe in audio and video formats. You can also check out my other shows which publish at Twit, including iOS Today, Hands on Apple, and Hands on Tech. If you'd like to tweet at me or follow me online, I'm @mikasargent on many a social media platform, or you can head to chihuahua.coffee, where I've got links to the places I'm most active online.

Mikah Sargent [01:05:20]:
Thank you for being here this week. I'll be back again next week, hopefully with lots of interviews. I need those yeses. Everybody's on vacation. It's summer vacation, everybody. Uh, I'll talk to you again then. Goodbye. Bye.