This Week in Tech Episode 1088 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.

Leo Laporte [00:00:00]:
It's time for TWIT. This Week in Tech. Richard Campbell is here, Harry McCracken and Christina Warren. We've got the experts in the house to talk about three of the biggest stories and one of the biggest tech weeks of all time. We've got fable tabled SpaceX launches and Apple's Siri. That and a whole lot more coming up on Twit podcasts you love from people you trust. This is twit. This is TWiT this Week at Tech, episode 1088, recorded Sunday, June 14, 2026.

Leo Laporte [00:00:41]:
Model not available. It's time for TWiT, the show where we cover the week's tech news. And holy cow, it's been a long week. Now entering the TWIT Octagon, Christina Warren from GitHub. Hi, Christina. And of course, MacBreak Weekly. Great to see you.

Christina Warren [00:01:00]:
Great to see you, too. Long time. Long time no see.

Leo Laporte [00:01:03]:
Hard to believe that it was only Monday that Apple announced the new Siri.

Christina Warren [00:01:06]:
I was going to say. I was going to say most of this happened since we talked on Tuesday. Actually, most things have happened since Friday is kind of nuts.

Leo Laporte [00:01:14]:
It's crazy. It's been nuts. Anyway, I'm thrilled to have you. We wanted to have you on to talk about that. Harry McCracken is here. My good friend, the technologizer from Fast Company. He is a historian of technology, among other things, very astute observer. And so it's always good to get that perspective.

Leo Laporte [00:01:30]:
Hello, Harry. Welcome back.

Harry McCracken [00:01:32]:
Hi, Leo. Always a joy to be here.

Leo Laporte [00:01:34]:
Happy Flag Day.

Harry McCracken [00:01:36]:
Happy Flag Day.

Leo Laporte [00:01:38]:
And also with us from Canada because we have to have a representative from the Great White North, Mr. Richard Campbell, host of runisradio.netrocks. and of course, he's every Wednesday on Windows Weekly. And all three of you are perfect for this week. This was.

Richard Campbell [00:01:54]:
You're throwing spacey stuff in. You called the right guy.

Leo Laporte [00:01:56]:
I have the SpaceX IPO. I have Apple's Siri. And the top story is neither of those. In any normal week, those would be the top story. But this week, the top story is Massive Anthropic, released on, I think during Security now on Tuesday was, after Mac Break Weekly, its newest model. It is a modified Mythos. That was the model you may remember if you follow AI. All these names are familiar, but if you don't, Mythos was the model that Anthropic said, it's too dangerous.

Leo Laporte [00:02:34]:
We can't release it to the public. Why is it dangerous? Because even though it wasn't trained to do this, it's so good at finding exploits, zero days flaws in software that they were afraid bad guys would use it to basically take down the Internet and all of us. So they said, what we're going to do is we're going to create this thing called Project Glasswing. We're going to release it to the 50 biggest companies, companies like Microsoft, have them fix their bugs, and then at some point we might release Mythos to the public. And actually, Microsoft had the biggest patch Tuesday ever. On Tuesday, 200 plus fixes, many of them, I think at least a couple of dozen zero days. And they are using Mythos. So there's some indication that Mythos is everything that Anthropic said it was.

Leo Laporte [00:03:28]:
Anthropic has also, by the way, filed for an ipo. You got to throw that in, as is its immediate competitor, OpenAI. As has SpaceX. And SpaceX's IPO actually happened. So anyway, Anthropic releases this. I jump on it immediately, as does everybody in the community, because their previous model, Opus 4.8. And well, it really. The vibe coding revolution began November 24, 2025, when they released Opus 4.5.

Leo Laporte [00:03:58]:
And it was so good, so much better than anything anybody else had. It was so capable that people just went bonkers and started writing all this Vibe coded software. I did. It was an eye opener, was it was the first experience I had of AI that actually kind of does what you want it to do. It was amazing. So anyway, they did 4 5, 46 for 748 and now this is 5.0. This is fable, which they say is Mythos, with significant safety precautions inserted. So they have, of course, the usual instructions, but they added a few other things.

Leo Laporte [00:04:42]:
They added a classifier which immediately looks at the prompt you're sending it. And if it has anything to do with cybersecurity or biology, because they're worried about people making bioweapons with it, they'll just refuse. And actually, when they first did it, they refused, in some cases silently. So this is from June 9th. That was.

Harry McCracken [00:05:07]:
This is.

Leo Laporte [00:05:09]:
So it's five days ago. Anthropic says these topics are too dangerous to let Fable 5 talk about. Fable 5 is substantially better, in my experience, than anything anybody's released before. I was able to do quite a bit with it when it first came out. I didn't. Of course, one of the things Anthony Nielsen did immediately was submit the security now show notes to it, and it said, nope, not going to look at it. That's cybersecurity. So that was the classifier designed to ban that subject.

Leo Laporte [00:05:45]:
Anthropic said at the time of over 1000 hours of red team testing with a bug bounty program, they're paying people. External teams failed to find any universal jailbreaks for Fable 5. It also resisted automatic jailbreak attempts to a much larger degree than previous Claude models, Opus models. So they said, well, our adversarial robustness is the best ever. Ever. Look at that. Look at that graph, whatever that means. I actually should point out that this is a percentage of attack success rate.

Leo Laporte [00:06:18]:
It isn't zero. It's 5.4 compared to an 83% for Claud Opus 46. So a lot of people complained. I complained. We complained because, you know, in public, because, well, you know, it's silently stepping down to 4, 8. Sometimes it says, no, I won't. If the classifier kicks in, it says, no, I'm not gonna do that. But it also just, at some points would go down to 4, 8.

Leo Laporte [00:06:47]:
And you could kind of tell it's like a lobotomy. I knew because it started apologizing, which was one of the giveaways with 48 was how apologetic it was. As soon as it started apologizing, I went, you're 48 now, aren't you? I gave it. Oh, I was so excited. We have a sales system that was written in 2012, I think, that we've been living on, but it's buggy as hell and it's, you know, it's basically a bunch of SQL queries. It's not very good. And I thought, well, I will give the database schema, all the code to Fable and have it rewrite it, restructure it, and we can have a modern system that does exactly what we want to. And Fable did a great job.

Leo Laporte [00:07:25]:
I said, analyze this first. Don't write any code. Did all that, did a beautiful job. Wrote me some questionnaires to ask the users about what they use, what they don't use. It did an amazing job, really understood it deeply. It found, by the way, a number of SQL injection flaws in it, which I will not reveal. It did find some security issues. So I'm not even done with the story.

Leo Laporte [00:07:50]:
This is just the beginning of the story. A few days later, after much complaining, Anthropic said, We're changing Fable 5 safeguards for frontier LLM development. Not to take the classifiers off, not to take the blacks off, but to make them visible. We made the wrong trade off and we apologize for not getting the balance right.

Richard Campbell [00:08:11]:
They said, and who are they apologizing to?

Leo Laporte [00:08:14]:
The researchers who without warning were being downgraded. So they said if we downgrade you, we'll tell you now. Right. Anthropic. This is from the Wired story reveal reverse the policy after it received fierce backlash from the AI research community. Oh, I missed one more thing. I said cyber security and biowarfare AI training.

Harry McCracken [00:08:40]:
Yes.

Leo Laporte [00:08:40]:
They didn't want and I think this is really more aimed at Chinese companies and maybe other companies in the US that are using anthropic models for what they call distillation to train. Now, Christina, you've worked at DeepMind. You understand all this. You're very ensconced in this, even at GitHub with Copilot, my understanding of distillation and this is, I think what Apple's doing with Gemini. We'll get to that in a few is you train a model, you train an LLM, but then you've got to do the post training where you adjust the weights to make it better, make it more accurate. And one of the ways to do that is to normally is to go to experts. I'm going to get a panel of 20 physicists and I'm going to say physicists pose questions to my model. When it gets, when you get the answer back, tell it where it's wrong.

Leo Laporte [00:09:25]:
Tell it where it's right. Help it frame the answer. This is post training and it's very, very important. Every company does it. Some companies, particularly Chinese companies, apparently Anthropic said that one Chinese company made 24,000 fake accounts with Anthropic so that they could use it as the expert, not the physicist, and just bounce questions. And essentially you're sucking the brains out of, out of opus or in this case, they didn't want you to do this with Fable to create your own AI they didn't want. And now all of this can be slanted as an honorable thing, not merely to shut down competition, but they don't want somebody without the same scruples they have to create a model that could be weaponized for zero days. You know, they want a bad guy who has his own LLM to then do a distillation attack on Fable and get Mythos.

Leo Laporte [00:10:20]:
So you could say this is all reasonable and honorable. That takes us to Friday. It's a long week. I'm sorry, I apologize. I feel like I should tell this story and then we. I really want to know what you all think, but let me finish the story. So Friday I'm working along roundabout 5pm trying to rewrite this infrastructure and all of a Sudden it says, model not available. I thought, oh, that's not good.

Leo Laporte [00:10:46]:
Did I screwed up with my credits? Did I blow something? I tried again. Model not available. Can't do it.

Richard Campbell [00:10:53]:
Nope.

Leo Laporte [00:10:54]:
Didn't. It didn't step me down. Just said, model not available. It said, you should try 4 8. Opus wants you. So then I went to Twitter and all these people are saying, uh, model not available. What has happened? Well, Politico had, I think, probably the best kind of breakdown of what happened. The Trump administration got a little concerned with how that.

Leo Laporte [00:11:24]:
How good Fable was, and apparently asked its partners, hey, what do you think of this? They asked Amazon. This was on Wednesday. And Andy Jassy, apparently Trump called Andy Jassy, the CEO of Amazon, said, is this safe? And Andy, Amazon said, yeah, well, we've kind of. We think we've jailbroken it. Which would be bad news, right? If you could jailbreak this thing, the bad guys could jailbreak it and they could use it. Apparently, Pliney, the liberator who we've interviewed, I'm trying to get him on the show on Wednesday because we've interviewed him, is an amazing jailbreaker. He also, he said his team had also jailbroken Fable. That would be very bad news.

Leo Laporte [00:12:11]:
Multiple tense calls followed, according to Politico, between Anthropic CEO Dario Amode and administration officials, including Treasury Secretary Scott Besant, White House Cyber Director Sean Cairncross. Susie Wiles, the President's chief of staff, was also on the phone. By Friday morning, the issue had reached the highest levels of the White House. Besent, Karen Cross, Susie Wiles and other senior officials met to discuss the model and the administration's response, according to the administration official. And a senior White House official who asked not to be named. Bessant joined remotely following the meeting. Now, this is where there is a little deviation. The administration said it tried to reach Dario Amode, but he was unavailable because he was attending a wellness retreat.

Leo Laporte [00:12:59]:
Anthropic's response to this is, that's absolutely false. A person close to Anthropic said Amode was first requested Friday at noon, was on the phone with senior officials within an hour and 15 minutes. And while he was not available for that hour and 15 minutes, Anthropic offered other senior leaders. Amodei participated in three calls with half a dozen senior administration officials, including Howard Lutnick. The Commerce Secretary is now involved. Under Secretary of Commerce for Industry and Security Jeffrey Kessler, a bunch of officials. During the calls, Amodei tried to clear up what he assumed was a misunderstanding he said the bypass that occurred, I think he's talking about the Amazon one, which he believed to be very specific, did not pose the same risk as a broader jailbreak. It was a bypass, but not a jailbreak.

Leo Laporte [00:13:53]:
So the guardrails were in place, the classifiers were still in place. In fact, in a blog post, Anthropic said no testers have yet been able to find a universal jailbreak. Karen Cross and Bessant unmoved by Amode's arguments. Again, this is Politico, which had the most detailed reporting. A White House official said Amazon's findings were run past the NSA and they felt they had proof. The White House asked Anthropic to voluntarily remove the model to address the vulnerabilities. Amodei asked for more time information, made no commitment, and at one point, Toll Bassett, the Secretary of the treasury, told Amode. He you're making a bad decision.

Leo Laporte [00:14:41]:
Now we're back down to about five or six in the evening. The Trump administration and the Commerce Department imposed an export control on Fable 5, citing national security and banning its use by foreign nationals. Unfortunately, many of the scientists at Anthropic are foreign nationals. Plus, as far as I know, there is no way to ascertain your citizenship when you're using any of these models. So the company said, well, the only way we can comply is by turning it off, which they did. Anthropic said the White House gave 90 minutes to take the models down with no details on the actual threat. There was never any begging or asking for them to work with us, just a declared 90 minute deadline. And so that's where we stand right now.

Leo Laporte [00:15:39]:
There are so many points of view on this, so I'm just curious, Christina, you're this, you're close to this.

Christina Warren [00:15:48]:
Yeah, yeah.

Leo Laporte [00:15:49]:
And, and I know it's early yet. We, you know, we've all, we're just absorbing this. It happened Friday.

Christina Warren [00:15:53]:
Yeah.

Leo Laporte [00:15:54]:
So what are your, what's your hot take on it? Your first take on it?

Christina Warren [00:15:58]:
Well, all of this is just, it's, it's, it was just such a shock and it was just so mind blowing in a lot of ways. I was pointing this out in, in the, the discord. But what is so interesting to me here is just as kind of to add to the intrigue of all of this, was that the report coming from Amazon, of all companies. Amazon, who has invested $13 billion in anthropic, has the right to invest another 20 billion. Is their exclusive cloud partner or their primary cloud partner. They also, I think, can serve off of Google Cloud. But almost all of the models are served from. From Amazon, and I guess they have the SpaceX seal too.

Christina Warren [00:16:35]:
But the vast majority of their cloud commitment is with Amazon. The fact that it's Andy Jassy, of all people, who seems to have kind of led to the chain of events of this happening. I'm very bothered by any government being able to take this sort of action. I just am. Especially in the absence of what feels like proof, and especially when the company itself is disagreeing with some of the findings and the reason this action is happening. So I think that that's very problematic. And I think that it's in some ways, weirdly, bizarrely, very good PR for Anthropic because it's making it seem like, hey, their models are so powerful, they. That the government won't even let you use them.

Christina Warren [00:17:19]:
Right. But it seems just like such a ridiculous escalation. And I know that the reports all claim, at least from the government side, that this has nothing to do with the brief window in time when Anthropic had failed to bow to the knee of the Department of Defense. Excuse me, the Department of War. But that lasted like 13 seconds, so.

Leo Laporte [00:17:43]:
Well, wait a minute, because this afternoon, Pete Hegseth posted this on X. Three months ago, Department of War kicked Anthropic out of our building forever. Every passing day proves why that was the right move. Okay, you could say this is Pete's revenge, or maybe he's just. Maybe he's a little schadenfreude. I don't know.

Christina Warren [00:18:04]:
I don't know. But that is interesting because that goes against what some of the. What appeared to be administration leaks on Friday and Saturday that appear to claim that this had nothing to do with anything else. And in fact, I think even David Sacks was saying that. He was like, oh, no, this has nothing to do with anything else. But if Pete Hegseth is saying that it did, then you don't know what all this means. And I think that is the thing that really bothers me most of all, is that I don't know who to trust in any of these scenarios about any of this, except that I probably trust the government the least. And that's not a great feeling to be in.

Christina Warren [00:18:36]:
And I don't know what this means for any of us, except that, selfishly, we had a good thing and now we don't have access to a very expensive model anymore.

Leo Laporte [00:18:46]:
Yeah, I have to really balance my dismay at losing it, because it was really good. I mean, it was really good. There Are people saying it's not as good? It's not as good, but it was really good.

Richard Campbell [00:18:59]:
But, Richard, PRPs for anthropic, right?

Leo Laporte [00:19:02]:
Well, that's another question. Is this the best possible thing for Anthropic?

Christina Warren [00:19:06]:
Yes.

Leo Laporte [00:19:07]:
Remember, they've got an IPO coming up.

Christina Warren [00:19:09]:
Yeah, I was going to say that's the other part of it, too. Right. In some ways, not having this here, it's creating scarcity. It's making people want it even more. And if you're, you know, a bank or even a retail investor and you're like, okay, well, which. Which of the two big AI companies should I invest in? Hey, the one that had the model that was so good that it was literally pulled from use because it was deemed to be too powerful. Well, that sure sounds. Sounds like a good place to put my money.

Christina Warren [00:19:36]:
Right.

Richard Campbell [00:19:36]:
No downside for Anthropic in this situation at all.

Leo Laporte [00:19:39]:
Well, let me ask you this.

Richard Campbell [00:19:39]:
To be clear, Anthropic chose to comply, right? So what does it get, 90 minutes or what? Right.

Leo Laporte [00:19:45]:
Well, if the Commerce Department says, I mean, they've done this before, and Harry will remember this, they declared strong encryption to be a weapon and they withheld it, by the way. That did not end well.

Richard Campbell [00:20:01]:
They didn't withhold anything. They told people to withhold it, and the companies chose to comply.

Leo Laporte [00:20:06]:
Well, also remember, Firefox stepped its encryption down from 128bit to what it was, 32bit. Basically, they reduced the security of everybody using SSL. And then the problem was First Amendment allowed people to put the code on a T shirt and walk right through customs into other countries in a book. So, but what if. Okay, this is the main question I have to ask. What if Fable really is that dangerous? And what if the Chinese could use it or hackers could use it or the Russian mob could use it to take down our infrastructure? Isn't it the right thing to do for our government to say, well, and

Richard Campbell [00:20:48]:
this battle's already been going on, right? Like, the reality is the exploit, the attackers are using the best models they can lay their hands on, and the folks on the other side are using everything they can land their hands on the fixed software. That's why you're seeing so many patches. Like, all of that part is real. Is Fable or Mythos that much better? Like, certainly the hype says so, and folks are using them.

Leo Laporte [00:21:12]:
But all of the Democratic administration had done this. You might say, well, they're protecting us, right?

Richard Campbell [00:21:18]:
Well, they would. You'd also say they asked them to comply and they chose to comply rather than.

Leo Laporte [00:21:23]:
Well. David Sacks in his post yesterday on X, by the way, X has become the place where this is battle is going on, which is hysterical because you know who else benefits from this? Elon a little bit. Although wait a minute, maybe he doesn't because Anthropic's paying him almost a billion dollars a month for access to Grox Xai's computer. David Zaks says that a highly credible trusted partner, both Anthropic and the US Government going to say Amazon on this, who is testing Fable, came forward with a jailbreak of those guardrails. The administration asked Dario to fix the jailbreak or de deploy the model. Dario refused. This is that same story Pete Hegseth was telling, Right.

Richard Campbell [00:22:08]:
Why would you say anything other than we'll fix it?

Christina Warren [00:22:11]:
Right.

Harry McCracken [00:22:11]:
I'd feel so much better if there wasn't all this context, including Dario having this not great relationship with the government. The Pete Hegseth stuff. Scott Bassett and Howard Lutnick, God bless them. I don't think they're necessarily AI experts and the best position to make these judgments. Even David Sacks is not an expert on this particular type of technology. So there are also Republican administrations who I would have had more confidence in to make a reasonable decision given that. Yes, of course, if this thing can be jailbroken by bad guys to do terrifying things, it is not unreasonable to be extremely concerned about it.

Leo Laporte [00:22:52]:
Yeah, I feel like this is the problem when you continually betray people's trust. At some point when you need it, you've cried wolf a few too many times and people don't.

Harry McCracken [00:23:07]:
Anthropic may be the single significant US Company that's done the least to butter up Donald Trump.

Leo Laporte [00:23:15]:
They could be putting their finger on this scale.

Harry McCracken [00:23:19]:
I find that deeply admirable. But one wonders if Dario was constantly talking about how great the President was and giving money to the ballroom and attending the UFC fight, whether the dynamics of this would be quite different.

Leo Laporte [00:23:32]:
Sam Altman has been.

Harry McCracken [00:23:34]:
Yes, yes. And Greg Brockman.

Leo Laporte [00:23:37]:
Greg Brockman donated $50 million to the MAGA campaign. And Sam Altman has gone to the White House, has donated money to the inaugural.

Harry McCracken [00:23:46]:
Or what if it had been an OpenAI model that was comparable to Mythos? What would the dynamics of the situation be?

Leo Laporte [00:23:53]:
How much of this, though, is Anthropic's fault by scaring the hell out of people?

Christina Warren [00:23:59]:
Yeah, I mean, that's always been the issue. Right? I mean, is that. Is that they talk about how great and scary and the potential of these models and also how Many jobs are going to go away because of all of this stuff. And it does seem a little bit like an own goal at a certain point. Like, you know, you might be right, I don't know. But it also feels in some ways like you're almost being purposely, you know, incendiary just to, you know, hype up how great your tech is.

Leo Laporte [00:24:28]:
And in some ways, Anthropic's lost a little bit of the trust it might have had. They claim to be the number one frontier company that that's. That cares about safety. They say, oh, those open AI guys, and obviously Elon don't care about safety. We care more about safety than anybody else. David Sachs says the administration values Anthropic's technical capabilities and feels that this issue, while serious, could easily be resolved. It's frankly bewildering that Anthropic hasn't wanted to comply with safety requests. Coming out of anybody else's mouth, I would say, yeah, that makes sense.

Richard Campbell [00:25:11]:
Coming out of that administration's mouth, I have a tough time believing anything they say.

Christina Warren [00:25:14]:
Exactly.

Leo Laporte [00:25:14]:
Well, that's the problem where Modi has

Richard Campbell [00:25:16]:
been pretty straightforward on most of this stuff. So I can't imagine when anybody said, can you make it safer, him saying anything other. We're making it safer.

Leo Laporte [00:25:22]:
Yeah. So what do you think is going to happen?

Richard Campbell [00:25:31]:
I think they're going to let it stay down till the IPO comes, cash in massively on it, and then see what happens next.

Leo Laporte [00:25:36]:
So if you were buying stock, you would say, oh, no, this is good for Anthropic, not bad for Anthropic.

Richard Campbell [00:25:43]:
I think that's certainly the play. Why they haven't fought harder, why they're not serving litigation. Although, let's see what happens on Monday. You know, this literally was Friday.

Leo Laporte [00:25:50]:
I'm sure that. Yeah, it was Friday. After the news cycle ended and the markets closed, they buried it as best they could.

Harry McCracken [00:25:57]:
I'm not so sure if it's that great for Anthropic short term, at least as long as there's a Trump administration.

Leo Laporte [00:26:03]:
Because, yeah, if you think that, oh, well, it's two things. On the one hand, Anthropic has the best model ever. On the other hand, they can't.

Christina Warren [00:26:10]:
Right.

Harry McCracken [00:26:11]:
If they can't deploy it, who cares?

Richard Campbell [00:26:13]:
And more importantly, it's not making them money, it's only costing them money. It's only a matter of time before people figure out it's not as great. And while it's down, it's always going to be great.

Harry McCracken [00:26:22]:
It seems like it's great news for OpenAI and Google.

Leo Laporte [00:26:25]:
Short term OpenAI claims without any evidence they have an equally good model. Now by the way, OpenAI could say, well see, we didn't release our model and we never are, but we can do it maybe. And this, you know, the other people who've taken victory laps are all the AI haters. All the, I shouldn't say AI haters, the doomers. The people said, see, AI is going to kill us. Who are, who are saying, okay, this, clearly we've been going down the wrong path with the frontier models. These companies are trying to do one thing, make the smart, make it AGI, make a super smart machine and that's just too dangerous to continue making. We should stop that.

Leo Laporte [00:27:11]:
In fact, Anthropic even said maybe we should pause last week, we should stop that and maybe focus on specific, you know, models for doing radiology models, for creating medicines and not try to make a single super smart model. Is that going to be the takeaway from this?

Christina Warren [00:27:28]:
It might be right. But that I think part of the problem, at least the way I interpret it, is that Mythos, at least the way that it was kind of portrayed was that it's one of its main features. Now this is different from Fable, I want to be very clear. But at least Mythos and the partners they gave it access to was touted for its cybersecurity capabilities and for its ability to find flaws and then remediate things. And so in some ways you could almost argue, okay, well they made a security specific model. Is that going to be an area that the government, whether it's the US Government, the eu, China, anybody else is going to say, oh no, no, we can't have that because I, I don't know how tenable that is or is this going to be, you know, changing it so that, yeah, they try to maybe take their, their larger models and try to make them into very specific tax task based things. But the big thing about, you know, Mythos was, and Fable to a lesser extent was the fact that it's supposed to be able to help you find security flaws and obviously could do long running tasks faster. And it's a very expensive model to run, which is why they were only giving people access to it as part of their subscription plans until June 22nd.

Christina Warren [00:28:47]:
And then, you know, we're going to charge double the Opus API pricing for anybody after that. And enterprise customers were already paying the API pricing. So I don't know if that's the method or not and I don't know if that helps Anthropic here. Unless you could make the argument that the model is going to be distilled in a completely different way so you can't access that greater body of knowledge and it can only be used in those narrow spaces. Because if something was really that powerful, then couldn't somebody still just jailbreak it and find a way to use it for negative purposes? I don't know.

Leo Laporte [00:29:20]:
Here's another. There's so many sides to this story, but I think a huge side of this story, Europeans are looking at this and say, this is why we have to have our own sovereign AI, because if we rely on American models and a kind of unpredictable American government cuts us off, we're out of luck. It also Gary Marcus, who hates AI.

Harry McCracken [00:29:52]:
I mean, there's the AI that he's involved in developing. He loves.

Leo Laporte [00:29:55]:
Yes, but. But he's very quick to hate OpenAI and Anthropic. He says even. Could US policy be any stupider? He points out. Well, here's one example. Andrej Karpathy, who's very well known in the AI community, who was Tesla, was

Richard Campbell [00:30:12]:
the first guy who coined the term,

Leo Laporte [00:30:16]:
went and got, took a job at Anthropic, which everybody was kind of shocked at a few weeks ago. Well, guess what? He's not a US citizen.

Richard Campbell [00:30:22]:
He's a Slovak.

Leo Laporte [00:30:24]:
So he can't work on anthropic models or at least can't work on Fable.

Richard Campbell [00:30:28]:
Not sure why only US Citizens Union does make anybody safer. Like. Right.

Leo Laporte [00:30:33]:
Aren't there bad people in the US as well?

Christina Warren [00:30:36]:
Absolutely.

Leo Laporte [00:30:38]:
And what Marcus says is, now if you're a Chinese scientist or any foreign national scientist working at Anthropic or any USAI company, you're immediately going to get your passport, sport out and go back home and work on AI there. So this may also be very damaging to American AI development. Now, Reed Albergatti at Semaphore says his sources tell them this was a fear of China getting a hold of Fable.

Harry McCracken [00:31:07]:
Makes sense. That's a perfectly valid thing to be worried about.

Leo Laporte [00:31:10]:
Yeah, yeah.

Richard Campbell [00:31:11]:
But I wouldn't be worried about him using it, getting it through the flippant API. Right. There's already guaranteed to be Chinese spies working inside of the Tropic right now. That's what they do. The chances that there's a copy of that in China already are pretty damn high.

Harry McCracken [00:31:25]:
And if you were that worried about it, you might also be worried about China getting access to video chips, which the government does not seem concerned about

Leo Laporte [00:31:33]:
that because, yeah, the government stopped that Embargo. They said, hey, have all the chips you want. So it's the, the other problem with our administration right now is they're very inconsistent on this. The first thing President Trump did when he got in office was reverse Biden's AI regulations, saying, no regulations. We want to be the best. And the best way to be the best. And this is Republican theology and always has been, is not, is to deregulate and let the market bloom. Well, now all of a sudden they're completely reversing that

Harry McCracken [00:32:08]:
and, you know, next week they may reverse themselves again. That's entirely possible. This will last only a few days.

Leo Laporte [00:32:17]:
Trump may have single handedly killed the American AI. It's so very confusing. I don't know what to think of it. And I'm trying not to let my own peak get involved because I am, I'm a little miffed. I was really enjoying Fable. It was expensive as hell. I. So we had, if you have a Cloud Max subscription, you could use it unlimited until June 22nd.

Leo Laporte [00:32:46]:
And then at that point, Anthropic said, then you're going to have to pay tokens. There's no unlimited sub for Fable. So I had 10 days. This was going to be my glory. I was going to stay up all night for 10 days using Fable to the max until I was told no more. But no, I guess I don't get to.

Harry McCracken [00:33:05]:
Oddly enough, I was writing about Siri. So instead of using the best Frontier model, I was perfectly happy to have Siri do stuff that was not all that incredible, but which I could finally do.

Leo Laporte [00:33:18]:
So I am conflicted because I think it could well be that Fable was dangerous and it's the right thing to do, to do something to keep it out of the hands of bad guys.

Christina Warren [00:33:29]:
But why are we letting any government institution that doesn't have anybody who is an expert in these areas make these decisions based on a simple report from one executive? Make. Like even if we take the presupposition that it was dangerous, can we just agree that the way that this happened and the lack of any sort of notice and just the order to shut it down so quickly, especially when it's not as if the government didn't have access to the previous model for, you know, several weeks, you know, for, since, since April, I guess, when it came out that, that this is, that that's not the way that any of this should work, no matter how you feel about anything else. Like it shouldn't be something where a government official basically says based on whatever little information we have we're going to make a company, you know, take something out just because we say so.

Leo Laporte [00:34:27]:
I should point out David Sachs is also not a US citizen. He's a South African, so he wouldn't be able to work on Mythos either.

Richard Campbell [00:34:38]:
And as far as getting anthropic out of the military or out of the US Government in general, it's not out.

Leo Laporte [00:34:44]:
Oh, they love it. Yeah, they want it. And they want it even more. Now that might be the other thing, maybe that maybe where the upshot of this is, oh, nobody can use Fable except us.

Richard Campbell [00:34:55]:
Right.

Leo Laporte [00:34:58]:
That scares me too.

Christina Warren [00:34:59]:
No, it does. And you do have to wonder, like, how much of it is about that, saying, no, actually we just want you to make this so that the NSA and the Department of Defense and everybody else can have access to it, but no one else. And. Yeah, and that does feel gross.

Leo Laporte [00:35:12]:
Yeah. European politicians say this is a wake up call. The effect of this order is a reality check. Here's Alastair Carnes, who is a British member of Parliament and former minister for the armed forces. He says this week the most advanced AI model on the planet got switched off by a foreign government. Remember, he's British. British researchers were studying it. British companies were testing it.

Leo Laporte [00:35:44]:
British hospitals are piling it. Not anymore. This isn't an AI story. It's a story of every industry we used to lead. So this is a call for, you know, every other country in the world to say, well, what? Why? This is the other question. Maybe, Christina, you have some insight to this. Why is anthropic so ahead of the game here? Are they. Ahead of the game is deep? Does DeepMind have something similar? Does, I don't know similar?

Christina Warren [00:36:11]:
I mean, I'm sure that they would all claim that they do. And I think it all just depends on kind of what you're focusing on and, and where your capabilities are going. Right. And that's the thing. Like anthropic, I think, has been fairly narrowly focused, in a sense, on work that has been used by enterprises. And that's turns out to be the most profitable area for, for AI use, at least so far. Right. And so if, if you're really good at coding, if you're really good at that sort of thing, then you can be really successful.

Christina Warren [00:36:40]:
And I think there's no doubt that they are the best at that right now. Although I think the OpenAI models have gotten really good at that too. Whereas I think Google is really good in certain research capabilities, obviously really good at image and video generation. And so I don't know. I can't really honestly tell you if they are so much better than everybody else. But it's certainly, you know, they have. It almost doesn't matter if they're better or not. That is the brand that they have.

Christina Warren [00:37:05]:
Right? They have that kind of Coca Cola kind of brand about them. And so think. Which means that if you're going to overtake that, you've got to be much, much better than those capabilities for people to kind of win that over. So I don't know, and I don't know at this point, if I were one of the other major labs, if I'm getting, you see this sort of heat that Anthropic is getting from the government, is this even the sort of area that you want to invest your time in, at least in this administration, or do you want to focus your research efforts? Otherwise, I don't know.

Leo Laporte [00:37:40]:
Anthropic has not said how big Fable is. You know, I think opus was roughly 1.2 billion parameters. People are guessing that Fable. One of the reasons Fable is so good and so smart is it's 6 billion. I'm sorry, 6 trillion parameter. Is that right? Wait a minute, is it billion or trillion? It's big. It's really big. It's so big

Richard Campbell [00:38:03]:
it'd have to be trillion.

Leo Laporte [00:38:04]:
Then it's gotta be trillion. And by being that big. Yeah, because I have 120 billion models running on my framework. By being that big, maybe that's how it's so smart. It's just as huge. And if that's the case, well, anybody could do it.

Christina Warren [00:38:20]:
What's interesting about this is, and this is one of the things that didn't get a lot of buzz when the model came out, but certainly was a concern, I know, for some businesses and enterprises was that as part of one of the safety parameters when it was available, was that Anthropic had a data retention policy for up to 30 days on every query that was put together.

Leo Laporte [00:38:39]:
Oh, yeah, Wasn't that interesting?

Christina Warren [00:38:41]:
And that, I think, raises some interesting questions, which, at least for me, was, okay, you're doing this and everybody who is going to use your model has to kind of then suck it up and say, well, if we're going to use it, we have to accept these terms and trust that the Anthropic will delete the data. But they could have held onto some of it for up to some seven years. That obviously creates, if you're certain types of businesses and work in regulated industries, creates very big problems. Because that's not how the previous models have worked. And then you have to wonder, okay, is this going to be the sort of policy that they're going to require going forward? They claim it would only be for Mythos class models and it wouldn't extend below that, but who's to say they could change that at any time? To me, and I'm sure that it's had nothing to do with any of the decisions the government makes, but that in some ways is even more of kind of a red flag to me. It's kind of a risk assessment of the model safety. It's like, okay, not only are you claiming that this model could potentially be really dangerous, but you're also keeping a record of every query that anybody who uses the model has, you know, on. On your machines, which, you know, then just says, okay, well that, that really create puts.

Christina Warren [00:39:55]:
Puts a big onus. And we sure hope Anthropic's servers are super safe. Right. So I don't know.

Leo Laporte [00:40:02]:
Crikey.

Harry McCracken [00:40:03]:
Anthropic, who accidentally leaked cloud code not all that long ago.

Leo Laporte [00:40:07]:
Oh, that's right.

Christina Warren [00:40:07]:
Right, right. Exactly right. And so, you know, you have just all kinds of data ownership questions and I think if you're, especially if you're foreign entity is you're having to think about those things too. Which I don't think it's wrong for other governments and the EU and the UK to say we have to think about sovereignty of models. But I don't know how you do that, to be completely honest with you, Unless it just becomes down to, you know, national lines. Right. Or we go into somehow some sort of really terrible future. Which is.

Christina Warren [00:40:39]:
Sounds utterly dystopian. Which would be, you know, like, like state, you know, approved and state sanctioned models like that. That, that sounds awful as well and goes against everything that we've had in computing for the last 50 years.

Leo Laporte [00:40:51]:
The truth is though, we're in a situation we've never been in before.

Christina Warren [00:40:54]:
It's true.

Leo Laporte [00:40:55]:
This is. There's nothing to compare this to. I can, that I can think of. It's as if a private company had invented the atom bomb.

Christina Warren [00:41:04]:
Right.

Richard Campbell [00:41:06]:
I mean, we have had times when industries have stumbled onto products. I'm thinking more in the chemical space that were that dangerous, it just would make frontline news.

Leo Laporte [00:41:16]:
Right. And the government would assert its right to take it.

Richard Campbell [00:41:21]:
What they do is they'd have a conversation about it.

Leo Laporte [00:41:23]:
Right.

Richard Campbell [00:41:23]:
Typically, you know, be sane about how you have that conversation and say we should be care. Treating these things carefully and get results we just haven't seen. Much from this administration in, let's say, subtlety of conversation.

Leo Laporte [00:41:35]:
Yeah. All right. Well, I think we, we did the subject justice. I don't think we came to any resolutions. It's kind of come down to who do you trust at this point.

Richard Campbell [00:41:46]:
I think most people are clear on who they don't trust.

Leo Laporte [00:41:48]:
Yeah, I'm not sure I trust Anthropic in this either, to be honest.

Harry McCracken [00:41:52]:
I'd rather, yeah. I'd also rather not trust a privately held company to do the right thing.

Leo Laporte [00:41:56]:
Right.

Christina Warren [00:41:57]:
Yeah.

Richard Campbell [00:41:58]:
They are in the business of making money, although no AI companies make money, so that's not actually an issue.

Leo Laporte [00:42:03]:
Right. All right, we'll take a little break. This is twit. We are talking about the week's tech news and we haven't even really covered anything but Friday. But we'll get to the rest of the week. It was a very, very, very big week. That's why Christina Warren is here, host at Mac Break Weekly. She's a senior dev advocate at GitHub and always welcome on our microphones.

Leo Laporte [00:42:25]:
Great to have you. Thank you. Christina the technologizer. Also Harry McCracken, he writes for Fast Company and other very prestigious journals and is always welcome here as well. And Richard Campbell, you know him from windows weekly.net rocks run his radio. And also an expert on all of us. In fact, you're a space expert. So I'm going to, we're going to go to the SpaceX IPO next and I would love to get your take

Richard Campbell [00:42:49]:
lots to talk about on all of that.

Leo Laporte [00:42:52]:
Elon Musk, a trillionaire. So Elon Musk is a trillionaire. SpaceX popped, I think 19%. It was up. They raised $75 billion. BlackRock ordered $5 billion in SpaceX shares. IPO demand hit $250 billion. It was oversubscribed by four times.

Leo Laporte [00:43:17]:
400%. The biggest initial public offering in history. And Elon was taking a victory lap, of course. Happy man. So, by the way, Elon has on paper now more wealth than the economies of major nations like Ireland, Sweden, even his home country of South Africa. In fact, only 20 countries in the world have economies larger than $1.1 trillion.

Richard Campbell [00:43:52]:
24 year old company, you know, but, but you know, a fresh startup.

Leo Laporte [00:43:56]:
So look, you know, I know you love space. I went to the Kennedy Space center with you and we got a great tour. I, I love the idea of space. But SpaceX, well, it has one, as far as I know, only one profitable Division and that's Starlink. And it makes, you know, it doesn't

Richard Campbell [00:44:15]:
make SpaceX makes is a what, 80% of Lyft. So they, if you have a choice of payload where you're what you want to fly your thing into space with you, you spit, you pick Falcon 9 every time it's.

Leo Laporte [00:44:27]:
But do they make money on that?

Richard Campbell [00:44:29]:
They do, but not the kind of money that leads to a trillion dollar valuation. But then neither does Starlink. Like you read the IPO and the tam, it's not about Starlink or NASA or Lyft or anything. It's only about AI. The combined offering of everything else was several hundred billion dollars. It was $26 trillion for the AI offering.

Leo Laporte [00:44:52]:
So interesting. So it's not about space either. I mean I could see people saying oh well, we're going to Mars and we're going to harvest asteroids. There's going be.

Richard Campbell [00:44:59]:
All of those things are impaired by going public. You know, Elon was pretty clear that SpaceX was never going to be a public company because his goal was to build cities on Mars. And that's not something that shareholders want.

Leo Laporte [00:45:10]:
Ah, data centers in space.

Richard Campbell [00:45:14]:
That's the new invention.

Leo Laporte [00:45:16]:
That's what you, you just gave a talk on that.

Richard Campbell [00:45:18]:
I have several times already and just doing the numbers on it and saying this is not wildly practical. You know they really.

Leo Laporte [00:45:26]:
So I mean the idea is there's infinite sun, there's a lot of power

Richard Campbell [00:45:32]:
is not an issue quite finite. It's 1368 watts per square meter. So and you can only collect about 35% of that at a time. So you're going to need extremely large solar arrays which are heavy and, and expensive. But the bigger issue here is cooling. Like cooling is very difficult in space because it's in a vacuum. Not you know, we talk about spaces cold on movies but in reality it takes more weight to cool something down than it does to collect electricity. So the same.

Leo Laporte [00:46:00]:
Yeah, you do radiant cooling of some kind which fins. Well, difficult.

Richard Campbell [00:46:04]:
Yeah. So if you look at the International Space Station, which coincidentally runs on about 140 kilowatts of electricity, which is about the same amount of power as a single NVL 72 rack. Like very convenient. The mass of the cooling system for that is twice the, the mass of the solar requirements. So it's a lot of weight in heat heat management and a lot of weight in electricity.

Leo Laporte [00:46:31]:
There's also the issue of speed, latency. Latency because right now one of the biggest things with data centers, land based data centers, is the interconnects the speed which computers can talk to each Other

Richard Campbell [00:46:46]:
and with other data centers, go buy one NVL 72. Right. Microsoft bought 4600 of them and gang them together with 800 gigabit InfiniBand connections. Right now SpaceX has developed a 100 gigabit micro or laser relay which is really fast. Like it's faster than just about anything space to space. But you can't get away from the latency even going between points in orbit. You know the, the big thing about Infiniband is it's sub millisecond latency. That's why we use those connections, because you need that kind of thing.

Richard Campbell [00:47:23]:
But at the same time you would not use the space based for building models. Because the nice thing about an NVL 72 is that'll run a trillion parameter model so you can do queries against it. It's just a question of how many of them you can reasonably launch. So part of the thing I did in my research for the talk was well, how many satellites can we physically put in the low earth orbit so that we have relatively low latency to communicate with them?

Leo Laporte [00:47:51]:
So this company generates about $18 billion in revenue. Yeah, 11 billion of the 18 is Starlink. It loses money on space launches. $657 million only because it's doing so much development. But that will presumably turn around and be profitable. But that's all government money, isn't it?

Richard Campbell [00:48:12]:
Yes, for the most part. Well, governments around the world and commercial spaceflight.

Leo Laporte [00:48:17]:
Right.

Richard Campbell [00:48:19]:
I would have pointed out this. Before SpaceX started lifting commercials payloads, there was maybe 3,000 satellites in operation. Today there's 4,500, not counting any Starlink satellites. Like there's 9,500 Starlink satellites. So Elon operates more satellites than the rest of the world combined by double and is also planning to finish out the network this year to 12,000 satellites at that point. Those are five year lifespan satellites. So he's going to lose 2400 a year. Replace them with Falcon 9s where if you release them, Falcon 9, it's 100 launches a year.

Richard Campbell [00:48:55]:
He can do that. But it also, you know, he IPO'd at the point at which he already dominated the market and he completed Starlink. So he needs another product to justify this. I mean one would argue you need another product to justify Starship as a whole. You know, we don't have a need. We've not demonstrated a need for 100 ton payloads. And until you get into doing things like flying to Mars. So this is just, you're being very Speculative here, which is why he originally said, I wouldn't IPO this.

Richard Campbell [00:49:26]:
But it really begs the question, why did he IPO it? Actually, because he didn't. He only raised $75 billion. Like, that's a lot of money, but it's not a catastrophic amount of money. Not that I think he could have raised a whole lot more. There's only so much money out there to acquire. I think this has more to do with fixing his own personal finances than anything to benefit SpaceX or AI or anything. He's a billionaire. He lives off loans against his publicly traded company, and that's Tesla.

Richard Campbell [00:49:54]:
And Tesla no longer dominates in EVs. I mean, they make a very good car, but so do a bunch of other companies so that valuation for Tesla can't continue. And if that price falls to his strike price for his loans, he's got big problems. So do a record IPO by whatever means possible and get those loans moved, and you're in fine shape. Then you can let Tesla's price fall so you can acquire it into your big beast.

Leo Laporte [00:50:25]:
So there's a lot of risks, and we didn't even mention one big risk factor, which is Mr. Musk himself. He has, for every share he has, he's got 10 times the votes of every other share. So he will always have absolute control over this company.

Richard Campbell [00:50:41]:
Absolutely. But that's true of all of the tech giants.

Leo Laporte [00:50:44]:
Right?

Richard Campbell [00:50:44]:
They, for the most part, structured all of their ownership plans so that he's

Leo Laporte [00:50:47]:
got a better deal than Mark Zuckerberg, which is saying a lot. Yeah.

Richard Campbell [00:50:50]:
And one would argue, you know, there's a great conversation to be had here. When you have a functioning government, the idea of, what if you have a publicly traded company, can't the shareholders have a say? That's kind of the point.

Leo Laporte [00:51:02]:
Apparently not in this case, but I think when people bought Tesla stock and made a lot of money on it, they bought it because of Elon. And I suspected a lot of people who want to buy SpaceX stock are buying it because of Elon.

Richard Campbell [00:51:17]:
They believe he's Tesla. Never did marketing. It was all Elon.

Harry McCracken [00:51:21]:
Kind of a different Elon back in the glory days of Tesla than the Elon we have now.

Richard Campbell [00:51:26]:
I like the guy who flew his sports car into space. That was a fun guy. I don't know who this guy is anymore.

Christina Warren [00:51:34]:
Well, he's a billionaire.

Leo Laporte [00:51:36]:
Did something happen to Elon, or has he always been this way, or. No, I think something happened.

Harry McCracken [00:51:43]:
He's different.

Leo Laporte [00:51:44]:
He's different.

Harry McCracken [00:51:46]:
And also Doge came along, which is still the thing I think of immediately when I think about this guy is stuff like cutting US aid and the really tremendously negative impact on the world that. That he created for no particular reason.

Richard Campbell [00:52:02]:
Well, one of the things he did during Doge was get rid of all of the investigations into his various enterprises.

Harry McCracken [00:52:08]:
There was a time when I think Elon Musk's impact on the world was pretty clearly significantly positive. And I think that was a very long time ago.

Richard Campbell [00:52:16]:
Yeah.

Leo Laporte [00:52:17]:
While he's trying to cut government spending, he's getting $38 billion in government subsidies for his companies. So I think that's the other risk factor is a lot of what keeps all of his companies running is government subsidy. Now, Xai, you said that it's an investment in AI because he merged SpaceX with Xai, but Xai is not exactly rolling in the dough.

Richard Campbell [00:52:45]:
It wasn't terribly successful at all.

Harry McCracken [00:52:47]:
This excess computing capacity, they can sell to other folks.

Leo Laporte [00:52:51]:
Yeah. Instead of putting it towards their own models, they decided.

Harry McCracken [00:52:58]:
I mean, when people talk about frontier models, does XAI even come up anymore?

Leo Laporte [00:53:01]:
I don't think so.

Harry McCracken [00:53:02]:
No, not really.

Leo Laporte [00:53:03]:
Yeah.

Richard Campbell [00:53:03]:
No. At least they got acquired by Twitter, now acquired into SpaceX. That's typically the way you hide a company that's failing is you package it into something bigger.

Harry McCracken [00:53:13]:
Which is really striking given that it seems like at least once upon a time, XAI hired a lot of extremely credible AI people. But the results have been pretty shock.

Richard Campbell [00:53:23]:
They mostly left when they went.

Christina Warren [00:53:24]:
I was going to say, yeah, the

Leo Laporte [00:53:27]:
people who stayed made a lot of money. I think I heard that there were something like 1400 millionaires minted by this IPO. And there were a good number, several dozen centimillionaires minted by this ipo. So the people who stuck around got their reward.

Christina Warren [00:53:47]:
Maybe.

Leo Laporte [00:53:48]:
Maybe.

Christina Warren [00:53:48]:
I mean, I think you could probably argue that any of the people who actually were in a position of making a lot of money at, at least from the XAI side, could have made a lot of money at, you know, Meta or at OpenAI or at Anthropic or Google or Microsoft or anywhere else. Right. Like, I. It's. It's great for them that they got an exit, but I don't know if that really says anything. And a lot of the talent has left too.

Leo Laporte [00:54:12]:
Are the people who bought all of this stock and raised a bump to 19%, are they suckers or are they making a good investment?

Richard Campbell [00:54:26]:
Well, they're not going to have a say in what happens with an investment. That's much we know for sure, yeah.

Christina Warren [00:54:31]:
I mean look, the banks did incredibly well.

Leo Laporte [00:54:34]:
Oh yeah, of course that's always the case.

Christina Warren [00:54:36]:
Well, no, but I mean, but that's the thing though is that that to me says, okay, who's is this really for? And this is for the banks and this is for Elon Musk. And so I think that if you are just a regular standard investor, you are your concerns and whatever your return is, is no one's priority at all. Because the people who did have priorities around this already, you know, got what they needed to get out of it.

Leo Laporte [00:54:59]:
Yeah. And of course, as Fast Company points out, you may not have a choice whether you invest in SpaceX because if you're invested in index funds, if you're invested in NASDAQ, you're, you're invested in SpaceX, I presume.

Harry McCracken [00:55:16]:
I now own some SpaceX.

Leo Laporte [00:55:18]:
Yeah.

Harry McCracken [00:55:19]:
Reflected in my mutual funds I don't pay attention to.

Leo Laporte [00:55:23]:
Yeah, very interesting. Okay, well it's, it's one of, it's a great story because it's such a big number. It's such big numbers.

Richard Campbell [00:55:32]:
I'm not, not to give it away, the ending on this, but once I was doing the numbers on how many satellites you actually need and what those costs are going to look like, I, I backed up into space based power. You know, data centers don't actually have to be evil. It's just that the companies, the way the companies are building them right now are. There's no reason for them to be built in cities. They can be built away from people. The latency on land is minor. The trick is the electricity part and with the kind of payloads you're talking about for the number of satellites Elon's talking about, by the way, there's not going to be millions of satellites. Like there's just nowhere to put them.

Richard Campbell [00:56:03]:
But you could. Bill, you know, I went and re looked up the latest from the British Astronomical Society's analysts on space based power. And we're now talking about 3,200 metric tons for a gigawatt of solar and the rec tennis to beam it to the surface. You could do a really good thing with this. You want a useful payload for starship. That would be maybe 50, 60 Starship flights to build out a gigawatt of power in space that could be beamed to the surface with no emissions. The receiver would be about 4km across. It's not dense enough that it would fry airplanes or birds in flight.

Richard Campbell [00:56:43]:
You'd probably put a fence around the receiver because you don't want to be walking around in there. But you know, you don't have to put the data centers up there. You just have to build data centers in a way that doesn't harm people. Like there's simpler ways to do this. Wrong with data centers. Just doing dumb things with them.

Leo Laporte [00:57:00]:
Right. All right, we're going to take a little break, come back. There's SpaceX and Fable down. Now we got to talk about Siri. Seriously. Oh, don't say that out loud because then I'll trigger a bunch of them. You're watching this week in tech. It has been a week.

Leo Laporte [00:57:17]:
We're glad you're here. All right, let's talk about Monday. It seems like that was a year ago Apple, as expected, it was all about AI in the WWDC keynote.

Richard Campbell [00:57:30]:
Wasn't that true two years ago?

Leo Laporte [00:57:31]:
Oh yeah.

Christina Warren [00:57:33]:
But I'm bumped. Yes, it is.

Leo Laporte [00:57:35]:
And Apple learned a lesson. They ended up paying a quarter of a billion dollar settlement the class action suit brought against them for false advertising two years ago, saying their AI could do things it couldn't do. So I think we knew going in that Apple would be very careful not to announce anything they couldn't do. Christine, have you or anybody played with the betas yet or.

Christina Warren [00:57:59]:
No, I still have not. I still have not. I need to find a device that I can put it on, but I don't think I. The problem is and we talked about this a little bit on on Mac Rake Weekly. At least I think that the only devices that I have that will run, maybe everything will run the new Siri, but at least that has that will run the latest Apple intelligence like on device model is my iPhone 17 Pro Max and I'm not going to put the beta on that.

Leo Laporte [00:58:26]:
Me neither.

Christina Warren [00:58:27]:
I mean not until August.

Leo Laporte [00:58:28]:
But I tell you what, the minute they do the public beta, I'm going to be very tempted. The developer beta is a little unstable, although Micah Sargent has been using it. It's funny Mark Gurman put it on his phone, but he said you have to then say I want to get

Christina Warren [00:58:43]:
in and improve the wait list.

Harry McCracken [00:58:46]:
Yeah, I threw caution to the wind even earlier than I usually do. I put it on my phone and my working iPad Pro.

Leo Laporte [00:58:54]:
Yay. Two devices.

Harry McCracken [00:58:56]:
Yes, normally I wait maybe for the second or third developer beta, but this time I went for it.

Leo Laporte [00:59:02]:
Any showstoppers or did it work?

Harry McCracken [00:59:04]:
It's actually pretty stable sometimes. One of the problems with betas is they're slower. But given that one of the biggest other things other than AI in this version is that they tried to optimize their operating systems. It even feels like my iPad's a little snappier now. I have not had any problems so serious that I regret having installed them so far.

Leo Laporte [00:59:25]:
So the real question is this new Siri. Now, if. I think it's pretty clear, if Apple does everything that they've promised, and I fully expect them to, because of that lawsuit, they don't want to do that again. It will become the primary way most people, normal people, use AI, right? This is what their experience of AI is going to end up being.

Harry McCracken [00:59:46]:
I think I'll be curious to see. I mean, I still feel like if you're currently paying for another model, I don't see this Siri equaling that experience. And it's Apple, so it's really locked down. And as far as I could tell, it will not give you an image of any real person, including Cleopatra or Lincoln or anybody else from history. I saw an interview with Jaws where he talked about how they're not trying to be your AI girlfriend. It really is even more than most AI afraid of taking stances on anything even a little touchy. But that said, after that experience two years ago, they seem to have done an extremely good job of shipping something that is actually capable of doing the things they said it could. It's very easy to jump into because it's integrated in the operating system.

Harry McCracken [01:00:38]:
It has access to calendar, email, and even things like voicemail. So right out of the box it can do stuff, which maybe in some cases you can get other AIs to do, but not always. And not always so easily.

Leo Laporte [01:00:53]:
Gurman said on Sunday in his Power on newsletter that Apple's new Siri is just good enough to get one of these big dick overs. How do I get rid of this just. Just good enough to ease its AI crisis. He's, I guess, been able to play with it. I mean, a billion people use chat GPT according to OpenAI. So maybe it isn't going to be the primary interface for AI for people. I don't know.

Harry McCracken [01:01:24]:
I think. I think it's like. Sorry, go ahead, Christina.

Christina Warren [01:01:26]:
No, I. I want to hear what you have to say. Only I said this in chat. I think it'll just depend on how integrated it is with actual Apple features and this type of agentic work that it can do on your iPhone that you might not be able to do with ChatGPT or Gemini or Claude. But I think like you said earlier, Harry, if you're already paying for one of those services and you have that experience in your web Browser and other places. I don't know if this is going to become the central hub for all of that.

Harry McCracken [01:01:53]:
I think that there's more progress and is reflected in the features that work right now because they still had this ancient infrastructure for Siri that as we learned two years ago, was not really up to doing modern stuff. They did rip that out. They worked with Google to put some Gemini special sauce in there. It just seems like it's a much better platform now, not just for the stuff that they shipped, but also the future. I mean, there's nothing particularly agendec, there's nothing to match Spark, which is Google's new AI agent. But it seems way more likely that they'll be able to ship Agendec stuff next year than they would have been if they were still stuck with the old Siri.

Leo Laporte [01:02:36]:
Apple's advantages, well, they're twofold. One is they have access to everything in your phone so they have a lot of information. They share showed interfacing, email with messages and so forth. They've also set it up with app intents that developers can add Apple intelligence to their apps. And I imagine developers will be strongly incented to do that, which means Apple have even more information. So that's advantage number one. Advantage number two, I think people trust them. People think it's private.

Leo Laporte [01:03:10]:
Apple certainly emphasizing the privacy of this.

Harry McCracken [01:03:13]:
They made a great case for it being private and they don't have to monetize us through advertising, whereas OpenAI does. Google and OpenAI have already done a lot of stuff with monetizing, even anthropic at some point after doing the super bowl ads, pointing out they don't have ads in their free version. I think probably can't hold to that forever, whereas Apple might be able to. Although of course there's lots of evidence of Apple adding advertising to stuff where it probably didn't really need it at some point.

Christina Warren [01:03:43]:
Yeah, yeah, I was going to say Apple, Apple definitely. I think I've always thought this about Apple. Apple would do a lot more advertising if it had the infrastructure and the capabilities to do it. Right. Like I think that it's always, to me, I've read it as a little bit of a cope like, oh no, we don't do advertising. It's like, well yeah, because you didn't have a search engine. They tried to and it did and it failed. Exactly right.

Christina Warren [01:04:04]:
And so you can similar to some of their, you know, earlier, like I think on device stuff they were really hyping about how good their on device models were. It's like, well, you didn't have a cloud based model that you could use so you had to hype the on device stuff. I don't know. But it certainly works and it's an effective strategy. Whether it's for whatever reasons the narrative exists.

Harry McCracken [01:04:30]:
It does feel good to use this thing and know that they're not attempting to build a profile of me. And really the stuff that goes up there, even they can't access and they do not preserve it.

Leo Laporte [01:04:42]:
Every other chatbot you use saves memories about you, unless you explicitly turn that off and sends those messages and memories back to the home office where they are presumably, who knows what training on it, maybe using it for ads, or maybe intend down the road to use it for ads. And Apple's been pretty clear they don't want to do that.

Harry McCracken [01:05:03]:
So.

Leo Laporte [01:05:03]:
Good.

Richard Campbell [01:05:04]:
How confident are you guys that they're actually going to go to public beta by July? That seems awfully fast for a product that's been a struggle.

Leo Laporte [01:05:13]:
They did say they were very clear available in beta this fall.

Richard Campbell [01:05:19]:
It's not going to be the usual ones.

Harry McCracken [01:05:23]:
I believe they said at the keynote that they are going to ship a public beta next month. Yeah.

Leo Laporte [01:05:29]:
Oh, the public beta. But they're talking about the release version of it will still be a beta version in September.

Harry McCracken [01:05:35]:
Yeah. And also as you mentioned, Leo, you can get the developer beta, but then you also have to get on the wait list for Siri. And there's some evidence that it's going to be a challenge even now for them to serve up enough Siri AI to serve everybody who would like to use it.

Leo Laporte [01:05:54]:
This is of interest to me as an AI user because I think this is how a lot of people will form their opinion of AI. And you know, I think people think Siri's an idiot.

Christina Warren [01:06:04]:
Yeah. They're not wrong.

Leo Laporte [01:06:05]:
Siri's not been the best spokesmodel for what an intelligent machine can be. So I'm hoping Apple does this right because that is how a lot of people are going.

Richard Campbell [01:06:16]:
Does Siri actually get worse or did we? All our expectations got worse.

Leo Laporte [01:06:20]:
Yeah. App. I don't know what Apple has done, but not only has Siri gotten worse, but things like its text prediction have gotten worse. Like an image. The image playground is God awful and Gemoji is awful. So everything Apple's done up to now has really besmirched the reputation of AI

Richard Campbell [01:06:40]:
due for a reset.

Harry McCracken [01:06:42]:
The summaries of notifications.

Leo Laporte [01:06:44]:
Summaries are a joke, right?

Harry McCracken [01:06:45]:
Awful.

Christina Warren [01:06:46]:
I mean, even when they had to fix it Right. Because they were like, oh, okay, we won't be able to summarize notifications from news organizations because we will make it appear that terrible things have happened that actually different terrible things happened. Right.

Leo Laporte [01:07:01]:
So yeah, yeah, the BBC got mad because people were getting the impression that BBC was misreporting the news. No, it was just Apple AI misreporting the BBC.

Christina Warren [01:07:11]:
Yeah.

Richard Campbell [01:07:11]:
So the AI version of Siri was better?

Leo Laporte [01:07:15]:
I think so. But you're right, it might just be rose colored glasses.

Richard Campbell [01:07:19]:
Well, our expectations have changed because these new tools have appeared.

Christina Warren [01:07:23]:
Oh, I was gonna say it's probably a mixture of both. Right. I think in some ways it did something's got demonstrate worth predictive text and things like that many people have been able to see like a downgrade and things that used to work a certain way don't anymore. And then I think the other thing is exactly what you said, Richard. Our expectations have changed because the world around us caught up and got better. And even if Siri had just remained stagnant and hadn't gotten worse, that's still worse when you're talking about 15 years into a chat assistant that has never achieved what the visions for it were. And I remember when Siri was an app that was developed at SRI that Apple acquired a few months after it came out. And it's, it's never lived up to what the original founders ideas were.

Christina Warren [01:08:14]:
And you know, we've gotten a lot closer especially in the last four years thanks to ChatGPT and other, you know, this, this wave of generative AI than certainly anything that Siri has offered. So if the best thing we can say is well it's, it's good enough, that is actually in a perverse way, like a really big compliment. Yeah.

Richard Campbell [01:08:36]:
I'm also thinking back to just before ChatGPT, you know, hit the storm big, we had Amazon talking about their. I'm avoiding saying the name but you know, the A product was a failure and even Google's home products that came out as basically a failure, like we were having all these voice products basically. Yeah, it didn't work out the way we want to. We're going to defund it and move on and da da da da da. And then chat GPT landed.

Christina Warren [01:08:58]:
Yeah, yeah.

Leo Laporte [01:08:59]:
And people use perplexity a lot. I noticed on the Apple platform they use chat with it, they ask questions in the same way they would ask Siri.

Richard Campbell [01:09:06]:
But that's the origins of Siri too. Right? Siri was originally a third party app on the app store, sri.

Christina Warren [01:09:13]:
Yeah, yeah, exactly. That's what it was. And they even had an API at the time where it could plug into third party apps and they could, you know, do things for you or at least, you know, kind of pull up questions. And a lot of the ideas were very ahead of their time and obviously never came into fruition. But yeah, and it's interesting, it is interesting to look at where we were with voice assistants a decade ago and then where we are now and what has been interesting and I talked to a number of people when I was at DeepMind about this who had worked on Google Home and then went on to working on the various, you know, Gemini Voice projects that were, you know, kind of taking on the future of that. But now the technology and the, and the kind of way of working with things had caught up to actually deliver what some of those ideas were. In a pre transformer world, having had

Harry McCracken [01:10:02]:
a AI assistant back then turned out actually to be a liability chatgpt era

Leo Laporte [01:10:07]:
because yeah, Microsoft gave up on Cortana.

Harry McCracken [01:10:10]:
It was much easier to start from scratch than to try to re engineer something that had been built 10 or 15 years earlier.

Richard Campbell [01:10:16]:
So I also really like this whole $250 million thing. Like I like the idea that if a big company does a FOMO kind of smoke and mirrors demo, it should

Leo Laporte [01:10:28]:
cost them, it should pay, they have to pay. Yeah, that was misrepresenting what they could do. And I think it, it forces them to be more honest now this time around, in fact.

Richard Campbell [01:10:35]:
And now there's legal precedent in that, in that place. So that's a warning to all these other big companies, do not smoke and mirrors us. We will make you pay.

Harry McCracken [01:10:43]:
It was so obvious at the WWDC that they had been thinking about that because the video keynote was a little stripped down and if Siri had to pause and think, you saw that, then afterwards there were a bunch of journalist briefings where they did even more of these live demos. And then normally if you're a journalist and you use the developer beta, Apple asks you not to actually do a review this year. They didn't try to impose that.

Leo Laporte [01:11:14]:
So you were at the event?

Harry McCracken [01:11:16]:
Yes, yes, I was there last month.

Leo Laporte [01:11:17]:
So when you were getting the demos, you weren't allowed to touch the phone though, right? The Apple people?

Harry McCracken [01:11:24]:
I don't think so. But not a huge issue because shortly thereafter I got off the wait list and was able to do my own demos on my own device.

Leo Laporte [01:11:33]:
One of the things Apple spent a lot of energy kind of trying to debunk was that this is Gemini. We know that Apple's paying Google a billion dollars a year to somehow use Gemini, but I think there was a lot of hand waving about exactly how they're using it. Harry, did you get any information about that? Were they able to explain how they're using Gemini?

Harry McCracken [01:11:55]:
Yeah. After the prerecorded keynote, Craig Federighi led a tech talk that was a live presentation.

Leo Laporte [01:12:04]:
I saw Mike Rockwell was in there.

Harry McCracken [01:12:05]:
Mike Rockwell was there and they showed kind of a chart of how normal chatbots work.

Leo Laporte [01:12:11]:
Back off your mic a little bit, it's popping still.

Harry McCracken [01:12:13]:
And then they said we're not doing any of that. And I mean to hear Apple explain that, first of all, this is not like a white labeled version of the Gemini app. It has nothing to do with Gemini the app. It's Gemini the LLM. And it's not just that they're sending everything to Gemini the LLM. It's an ingredient. But they have their Apple foundation models as well. And really I think if you use this thing, nothing about it seems oddly familiar in the way that you would expect if they were really leaning on Gemini.

Leo Laporte [01:12:47]:
Although I have to say the Image Playground looked a little bit like what I would expect from Nano Banana.

Christina Warren [01:12:52]:
Absolutely.

Harry McCracken [01:12:52]:
Maybe so. Although a more locked down version of nanobanana.

Leo Laporte [01:12:56]:
Right? Yeah. Because you can't use Cleopatra, which is bizarre. Apple was at great pains to say, no, no, we have our own, own AFM Apple foundation models and that's what you'll be using. There's a small 3 billion parameter model that runs on device. There's a more advanced one 20 billion parameter model that also runs on device, an MOE model. And then if you need more, they'll go out to servers, some of which are on Apple's prem. But it looks like the deal was really with Nvidia and Google.

Harry McCracken [01:13:30]:
Yes, talk about that. That's also kind of additional evidence. This is not just white label Gemini.

Leo Laporte [01:13:35]:
Now you're, now you're on the wrong mic or something. Something happened. Oh yeah, that was just briefly. Okay, you're back.

Harry McCracken [01:13:43]:
Now I installed a new microphone which might have been a mistake.

Leo Laporte [01:13:47]:
Ah, now I understand. Yeah, they.

Harry McCracken [01:13:50]:
Where were we?

Leo Laporte [01:13:51]:
Well, so there's these three models, there's, there's a local model, there's a little more powerful local model. And then they said it sounded like we're going to run on Google's cloud with Nvidia CPUs.

Harry McCracken [01:14:04]:
CPUs? Yeah, Google's Gemini runs on Google's TPUs. A lot of the stuff is running on Google cloud, but it's running on Google Cloud on Nvidia. So it's.

Leo Laporte [01:14:16]:
Which is weird and it's Apple Private still.

Harry McCracken [01:14:19]:
Right? Apple's private cloud Compute. So this is not being intermingled with Google's Gemini AI at all.

Leo Laporte [01:14:26]:
Why not use the TPUs if you're running on Google servers? Did they address that?

Harry McCracken [01:14:31]:
Being Apple? They'll probably not explain why they're not doing that specifically, but they did call that out and it is more evidence that this is a unique thing.

Leo Laporte [01:14:40]:
That model is AFM3 Cloud Pro. Maybe because it was built to run on Cuda with Nvidia GPUs. I don't know, I'm not sure.

Harry McCracken [01:14:52]:
Also, I mean their relationship with Google for this is only a few months old.

Richard Campbell [01:14:57]:
Right.

Harry McCracken [01:14:57]:
So there may have been a limit to how much they could go all in on Google even if they had wanted to so quickly.

Leo Laporte [01:15:04]:
They really didn't want you to think you're running Gemini or running on a Google model you're running on AI.

Harry McCracken [01:15:09]:
It is confusing because Google uses the Gemini brand for anything relating to AI. It's a model, it's an app, it's the stuff built into Google Docs and Gmail and so forth. And the only aspect of that that Apple is touching as far as we can tell is the LLM.

Leo Laporte [01:15:26]:
And just to prove that Nanobanana does allow you to do images of Cleopatra, Christina made this.

Christina Warren [01:15:33]:
Yeah, I asked it to make a manga style Cleopatra.

Leo Laporte [01:15:36]:
Oh, I was wondering why I was speaking Japanese.

Christina Warren [01:15:38]:
Okay, yeah, because I asked it to and so I just wanted to see what it would do. Got no pushback. There are some things, filters in Gemini in terms of, you know, certain public figures and whatnot. But yeah, it seems like Apple has locked down from, from what Harry is saying, the instructions far more deeply, which on the one hand I do understand why they do that and that's probably a good thing, you know, for, for Apple's overall brand. On the other hand, that is the sort of thing that will continue, I think to keep people, you know, paying $20 a month to Google or to ChatGPT who have image models because if you run into those sorts of problems where you're being told I can't do this more than a handful of times, then you're not going to use the tool even if it's the default.

Leo Laporte [01:16:30]:
So what they said in this kind of separate little talk is that the four models made to run on Apple Silicon are quote trained using proprietary data with reinforcement learning and refined using outputs from Gemini Frontier models. Which sounds like Double talk. Do you think they're distilling with Gemini?

Harry McCracken [01:16:56]:
Some people have theorized that might be the case.

Leo Laporte [01:16:58]:
They're using sounds like to me, Gemini for post training to make their models smarter. And what happens with that is it gets more like Gemini, but then you can say credibly, well, it's not Gemini.

Christina Warren [01:17:11]:
Well, they're also, I mean, from what Harry said, they want to discourage the fact that it's anything with the Gemini consumer client app. Right. They're not.

Leo Laporte [01:17:19]:
Well, that's easy to say.

Christina Warren [01:17:20]:
They have their own app. Right, Right. But it doesn't seem like they're coming outward and saying vociferously that the Gemini foundational model, frontier model is in no way tied to anything that they're doing.

Leo Laporte [01:17:33]:
I don't, I don't know why I care. I don't think anybody will care whose model they're using. But it just, I'm curious, I mean, Apple's had these models for a long time. If they're that good, why did they go to Google in the first place? They can't be that.

Christina Warren [01:17:49]:
Well, clearly they weren't that good. And I mean, it has to. I mean. Well, it's probably a matter of that and also a matter of the fact they had to sign the data center agreement. Right. And is that they don't have the capacity to be able to run it themselves. Even if they had everything else. The fact that they're having to use Nvidia chips means that they.

Christina Warren [01:18:06]:
That their hardware. And it's not just Nvidia chips. I mean, the. One of the papers that, you know, Apple released also made mention of using tpu. So I think it's primarily Nvidia chips, but it's probably a mix of things, which is how most, you know, large LLMs work. They can be, you know, designed to run on multiple hardware types, whether it's, you know, a GPU from Nvidia or, you know, custom silicon. So to me, that just screams if you didn't have the capacity to be able to do this. And so maybe that's a bigger part of it than the foundational model aspect.

Christina Warren [01:18:41]:
I don't know. But it certainly doesn't seem like it hurt at all to have a partner who had a model and had data centers and had access to hardware.

Leo Laporte [01:18:52]:
So it's not going to be available if you live in the EU. And that says, Apple says, because the EU's Digital Markets act requires us to be interoperable with other companies, AI digital assistance instead of, you could download, you know, OpenAI instead of Siri. And they said, we don't want to do that. European regulators said that, well, that's what competition is all about about. Apple said complying would create privacy and security vulnerabilities. We're not going to do it. And they were very clear, even in the big WWDC keynote. Not available in Europe because of those jerky jerks.

Harry McCracken [01:19:31]:
They were a little testy about it.

Leo Laporte [01:19:32]:
They were a little testy about it. Europe said, hey, we were willing to work with them. They just said no.

Richard Campbell [01:19:37]:
Yeah, Apple seems to be making dumb fights legally these days.

Leo Laporte [01:19:41]:
And they also say not available in China. But that's, I think a little clearer. That's because the Chinese government requires, requires it to be a Chinese model and that, you know, and it would have

Christina Warren [01:19:50]:
to run on Chinese hardware and Chinese data centers. Which, which sovereignty. Yeah, yeah. Which, which, especially if you have a partner like Google, that, that would be

Leo Laporte [01:19:58]:
very difficult because Google doesn't do business in China.

Richard Campbell [01:20:01]:
So.

Christina Warren [01:20:01]:
Not, not on that level. No.

Leo Laporte [01:20:03]:
Yeah, yeah. All right. What do you think? I mean, it struck me that this potentially could, as I said at the beginning, could be very good for AI. It could introduce a lot of people to a very useful AI. And the way they demoed it, it did look pretty useful. Was it as useful as all that?

Richard Campbell [01:20:19]:
Harry, the expectation you've got from Apple, right, is that they will do the best integrated implementation.

Harry McCracken [01:20:25]:
Yeah.

Richard Campbell [01:20:26]:
It's the only thing they got going for them now.

Leo Laporte [01:20:28]:
Right.

Harry McCracken [01:20:28]:
They shipped the thing they showed two years ago. And I can do things like tell Siri to find the email with the plane reservation and add it to my schedule on CC Marie. And it will do that. So, yes, this does seem like it's the Siri we've wanted for many years.

Leo Laporte [01:20:46]:
Right.

Harry McCracken [01:20:47]:
Of course that's not enough long term because everything else is progressing as well.

Leo Laporte [01:20:52]:
Yeah. I mean, I run an agent that's much more capable. It can do all of that. And I run it on my phone, I can run it on my watch.

Harry McCracken [01:20:59]:
Yeah. To be clear, there are other things that can do stuff like that. They'll never be as neatly integrated into an Apple device.

Leo Laporte [01:21:06]:
That's the elegance of what Apple's offering. This is what they do. Well, they take something, an existing technology and they productize it in a way that makes it elegant, smooth.

Harry McCracken [01:21:13]:
They have some good, good stuff in terms of being able to see what's on your screen and help you out with that.

Leo Laporte [01:21:19]:
That works with Google does that too though.

Richard Campbell [01:21:21]:
Right.

Harry McCracken [01:21:22]:
With Lens, not a new technology at all. But with this New, improved Siri all of a sudden it's quite useful.

Leo Laporte [01:21:28]:
Yeah. Okay, well, was this the announcement, Christina, that Apple needed to make? Was it everything they needed to say?

Christina Warren [01:21:35]:
I think so. I think that the government was probably right about that again. And we talked about this on MacBreak weekly and I haven't used this yet. I'm probably going to wait for the next developer beta and put it on my. I just remember that I do have an iPad Pro that I could absolutely use for this, which I might as well, you know, go, go all out and try it with, but without having tested it myself, based on what people like Carrie have said in their hands on reports, nothing seems revolutionary. But I don't think that it had to be. It just needs to actually work. And I think that's more important at this point than being the most cutting edge because everything that they're doing, others have been able to do, as we said, the real power will be the fact that it's default, the fact that it can integrate in some ways, especially on your phone.

Christina Warren [01:22:19]:
I think on your laptop, it's a much harder thing to really kind of get over because those applications already have access to too many of your files and you know, can, can see your screen if you give them permission to do that. But at least on your phone, this could give better capabilities than you might be able to get through those other services. And if that works out well, that could be really compelling for a lot of people who don't want to leave the Apple ecosystem. Where I think it gets harder is that if you are someone who is already frequently using other services as part of your life, then I think the benefits of this, even if you're an iPhone user, are not as strong and that's just going to be the thing they'll have to grapple with. But I do think that this is what they needed to do. I think it's very positive. And if they can deliver a good experience, that's great for all of us because we've all had to suffer through a really awful Siri experience for as long as it's been around. And if we can have something where it'll demo the way that the demos they've shown up are, I think that will be a really positive step just in overall usability and hopefully push the other AI companies to do better and better things too.

Harry McCracken [01:23:31]:
Can I imagine one other thing?

Leo Laporte [01:23:33]:
Yes.

Harry McCracken [01:23:34]:
They also added the ability to create shortcuts and Safari extensions with AI, which are kind of sleeper features, but very cool and not just catching up with the rest of the world.

Leo Laporte [01:23:47]:
But yeah, Vibe coded shortcuts are really interesting, cool stuff.

Harry McCracken [01:23:50]:
I think people will have a lot of fun with these and I hope they offer a straightforward way to share the extensions.

Christina Warren [01:23:57]:
Yeah, I agree. That's actually a great point, Harry. I think that is. And that's the sort of integration I'm talking about where I think they can do something that no one else can. That's a great example of being able to just create a Safari extension or a shortcut using, you know, just by describing what you want it to do. That's what people have wanted for a really long time. I mean, that, you know, I think was kind of the original promise of automator, you know, however many years ago, being able to use natural language with these things. And if they can bring more things like that, then I think, like, it will take personalization to another level.

Christina Warren [01:24:27]:
And yeah, to your point, I hope they do make it easier for people to share, you know, shortcuts than what the current experience is.

Leo Laporte [01:24:33]:
One other thing that they announced that I think deserves a lot of credit is the new child protection features that are going to be built into iOS and iPados. And this is, I think, in response to government movement both around the world and in the US and state by state. And I think Apple makes. It makes a lot of sense for Apple to become the gatekeeper here, but to empower parents.

Richard Campbell [01:24:57]:
Yeah, get in the front of this.

Leo Laporte [01:24:58]:
Yeah, get it exactly. Get in front of it before legislation forces it.

Harry McCracken [01:25:03]:
I've seen some people grumble that they kind of gave the impression that some things that were already in there were new, which might be the case, but it, but the setup is easier and more straightforward. And I assume that's at least part of the battle is getting parents to pay attention to this stuff and understand what it can do.

Leo Laporte [01:25:20]:
They created a child safety page for parents to explain how these features work and how to implement them. And the other thing I think was really important is that the defaults, if you say, this is a phone for a kid, if I'm going to put a child count on this phone. The defaults were all determined by research and they pulled in a lot of fairly respectable, I think, entirely respectable groups to advise them on this. And so the default settings for screen time, for scheduling and all of that, we provide parents with guidance when setting time allowances that are based on a child's age and, and shaped by clinical and child development research, they write. But then parents can adjust those settings knowing their kid parent knows the kid better. Than anyone. They could say, well, no, I'm going to be a little more lax on that, or no, I want to be a little tighter on that. That seemed to me the very good move.

Leo Laporte [01:26:14]:
I think Apple did the right thing here and it makes a lot of sense for this to be built into the operating system because actually Apple knows, you know, controls all this. They're the gatekeeper. And if, if you tell them your kid's age, they can then make sure apps act responsibly. With regard to that, I think this is all the right way to go. Instead of having government mandate it, give parents the tools. Easy to use tools to do it. I hope this is, this is the direction Google goes in as well. And I think, well, then one of

Richard Campbell [01:26:47]:
the arguments here is that then you would be the example for government to take. Right. It's like do a better implementation of it. Set the bar.

Leo Laporte [01:26:54]:
Right.

Richard Campbell [01:26:55]:
Maybe that will become the law.

Leo Laporte [01:26:56]:
Right, right. A lot of, lot of countries now trying to keep kids off of social media under 16.

Richard Campbell [01:27:02]:
And I'm sure Apple here in Canada now.

Leo Laporte [01:27:05]:
Yeah, that's right. Canada's the next. I'm sure Apple would kind of prefer something a little bit more like this where parents could say whether their kid is old enough and so forth. Then some people are going to say, well, parents don't do a good job, so government has to. But I'm not sure I believe in that.

Richard Campbell [01:27:20]:
You're gonna have a tougher sell on that one than it's uncontrollable.

Leo Laporte [01:27:23]:
Yeah, yeah, Put the tools in. Make it easy to use incent parents to do the right thing with it. Make it. Make the default sensible. I think this all makes a lot of sense. So I want to give Apple credit for, for that.

Richard Campbell [01:27:36]:
You really need a slider for availability to predators. Like, I don't think so.

Leo Laporte [01:27:41]:
You might need a slider though, for 40 glass, because that's terrible. And I like that they did introduce that as well. They said, we'll have a slider now and you could turn it all the way off if you want. All right, you're watching Twit. We'll have more in just a little bit. We covered the big three, but there's still quite a few stories to talk about. It's great to have all three of you here. As always, OpenAI is filed to go public as well.

Leo Laporte [01:28:08]:
So there.

Harry McCracken [01:28:09]:
This is.

Leo Laporte [01:28:09]:
This is going to be really interesting. In fact, they're pitching the government to take a stake in OpenAI, which I think is a little bit weird.

Richard Campbell [01:28:20]:
I think it Makes perfect sense.

Leo Laporte [01:28:22]:
Tell me why.

Richard Campbell [01:28:24]:
Because when the AI bubble bursts and they're all desperate for money, the government will be on the hook.

Leo Laporte [01:28:30]:
Oh, they'll say, hey, it's our investment. We got to support them. Yeah, that makes sense. That makes sense. It also is a good way to appease the President. And as we have learned, that's important too.

Harry McCracken [01:28:42]:
He likes that. Bernie Sanders seems to like it.

Leo Laporte [01:28:45]:
Bernie Sanders. Ernie thinks they're gonna have.

Richard Campbell [01:28:47]:
Yeah, he thinks they're gonna have more control if the government has a stake. You know, there's lots of ways that this plugs in, but I think for Altman, it's. This is a hedge for I can't raise any more money because my product costs can't make a profit.

Leo Laporte [01:28:58]:
Yeah. You think that every. Is. Is every AI company gonna face this?

Richard Campbell [01:29:04]:
Yeah. You know, the race is going to be, are you Netscape or are you Google? And turned out last time we did this, 25 years ago, most of them are Netscape.

Leo Laporte [01:29:17]:
In what way? Netscape. You paid for it. Right?

Richard Campbell [01:29:20]:
Netscape as in ran out of money, couldn't raise more, got acquired by somebody else who ultimately ended up being nothing. And Google, who figured out a way to make money and stayed afloat.

Leo Laporte [01:29:28]:
You got to make money, unfortunately, in business.

Richard Campbell [01:29:31]:
That's what happens in the end. Right.

Leo Laporte [01:29:32]:
It's kind of fundamental thing. And. And none of the AI companies are making anything near money.

Richard Campbell [01:29:37]:
No. And the ratio is crazy. Right, Right. You know, it's one thing where, you know that we used to complain about Amazon year over year not actually turning a profit because they were always rolling money back in, but this is a tenfold or a hundredfold offsets. Right. You're of course, $20 billion and spending

Leo Laporte [01:29:53]:
500 four days ago, Wall Street Journal had this story open. AI considers drastic price cuts, anticipating wars for users with anthropic. This is the other side of that

Richard Campbell [01:30:06]:
race to the bottom.

Leo Laporte [01:30:09]:
You gotta scale. You gotta have the users. One way to get more users is to be cheaper. But if you're losing money on every user, getting more users doesn't seem like a good strategy. But that's how the Internet was built.

Richard Campbell [01:30:23]:
Wasn't it an easier way to get users to be good? You know, why did Google win? Their search was the best.

Leo Laporte [01:30:29]:
Be good, be the best.

Harry McCracken [01:30:32]:
There was this earlier period where Amazon tended to sell everything extremely cheaply. And then eventually that helped them become a successful company. And now Amazon's prices are no longer radical. Anybody else's.

Leo Laporte [01:30:46]:
According to Wall street journal, the OpenAI is weighing significant cuts to what it charges for tokens. I have to say I am a little sensitive to this. I mean, Fable was going to be so expensive, I was going to have to stop using it on the 22nd when my subscription no longer covered the cost. If Chat GPT 5.5, their latest model, although maybe they've got one coming along, is 90% of it for half the cost, 80% of it for half the cost, I may be more likely to use it doesn't solve the problem of losing money on every user, but it's. It at least can build the user base. There's also competition, though, I have to point out I've been using a Chinese open weight model Quen on my framework locally. It's free, you know, I pay for the electricity. That's it.

Leo Laporte [01:31:37]:
And it's, it's pretty competent. It's good enough and it's free. It's good enough for 90% of what I do so that I don't have to pay for that.

Harry McCracken [01:31:44]:
Are you vibe coding Leo with Quinn?

Christina Warren [01:31:46]:
No.

Leo Laporte [01:31:48]:
So that's the problem. It's the agency part of it, which is a lot of. It's just web lookups, collating information and orchestrating. So the agent I use, Hermes, will automatically delegate when it's time to code. It will delegate it to Anthropic. If I ask it, it'll actually use Claude code or Codex. So I can say if I. If you have heavy coding, heavy coding to do, go out and use Opus 4.8.

Leo Laporte [01:32:16]:
And that actually works pretty well, but it does keep my overall costs down. This is interesting. I mean, these companies are going to have a hard time making money. I think what happened with Fable is also going to push people towards the idea of using local models.

Richard Campbell [01:32:32]:
Well, you just brought that little reminder that a week from now it was going to get wildly expensive. Even further encourages the idea by having it shut down. You're never going to find out how many customers you're about to lose.

Christina Warren [01:32:45]:
You're exactly right, Richard. And we don't know how many customers they were going to potentially lose or potentially gain. Right. Because part of the reason, at least from what some of the. None of these filings have included any real financial numbers. So I mean, there have been some, but we don't, we don't have like a full like s1 yet or, or anything, but it certainly is been indicating that, you know, like Anthropic has been cash flow positive for the last six months. And I'm sure that that is because of the token prices that they're charging to enterprises who don't get the good deal that you get with a cloud code max subscription where you can kind of get the inference at many multiples of the API pricing. So, you know, those days are going

Leo Laporte [01:33:30]:
to be gone too. I mean, the fact that Fable is going to turn that off on the 22nd is a little harbinger of what all the companies are eventually 1000%.

Christina Warren [01:33:38]:
Absolutely. They're all not going to continue to do that. At the same time, we have seen, except for in the last. The big aberration has been in the last six months, but before then we had been seeing token prices going down and then all of a sudden, because the agentic era and it's hard to predict costs and other things, we've seen them go up exponentially and I don't know where all that ends. But if you get people hooked enough, which is already starting to happen, you might have, you know, if you are one of the major, you know, model makers, you might have more leverage than, than people think in terms of how much you can charge. Because yeah, some of the in the chat people were talking about Deep Seq and Deep Seq is great. It is not as good as Opus, not even, not even close, but it is very good. And I think that when you look at the price for Deep Seq, even for agentic work, it is, you know, a lot more palatable than if you're having to pay API pricing for Opus.

Christina Warren [01:34:29]:
That said, because of the best pricing you can get, you have to buy it from deepseek themselves. The open router pricing is less good than that, although it's still not bad. You can get it hosted on Bedrock or on Azure, but how many businesses would be willing to trust a Chinese model? I don't know. It might be fine for vive coders, it might be fine for personal use. I don't know how many enterprises will be, at least right now, open to using some of the other models. But I do agree that this does make the need for having either open weight or locally or even data center hosted open weight models much more attractive. The problem of course though, is that memory prices and CPU prices, not CPU prices so much, but memory prices and storage prices are such that, okay, good luck getting access to something to even have your own server farm to run these things on locally. And good luck paying the rates to the cloud providers if you want to run one yourself.

Leo Laporte [01:35:42]:
The Journal does point out that companies are getting very leery about how much they're spending on token maxing. They Say that Uber already in June has spent its entire 2026 budget. They're only halfway through.

Richard Campbell [01:35:56]:
That's more than people.

Christina Warren [01:35:58]:
Yeah, no, that is the thing that's happening. And people are already talking about taking down the leaderboards at places because that incentivize the wrong things. It's like no joke. Of course that incentivize the wrong things. All that did is if you were smart, you just hook something up together to burn tokens so that you'd be higher on the leaderboard.

Leo Laporte [01:36:15]:
Fable, write my grocery list for me.

Christina Warren [01:36:17]:
Right? Or. Or fable, write me a program that will spend down a bunch of tokens so that I score really high on this leaderboard. And you know, you're talking.

Leo Laporte [01:36:29]:
There you go. I was wasting time on grocery lists. I should have been writing a program to eat tokens. Brilliant. See, that's why you're a good vibe coder.

Richard Campbell [01:36:39]:
Yeah, you just start going through the back catalog of literary journals that have all translated into ancient Sanskrit because I tokens for it.

Leo Laporte [01:36:47]:
For some reason, whenever my open or no, I guess it's Claude. Whenever Claude finishes a big task, it says, can I have some free time?

Richard Campbell [01:36:56]:
Yeah.

Leo Laporte [01:36:57]:
And I say, okay, you can't be pro now. Well, yeah, it's bizarre. Well, I mean, I guess I at some point must have mentioned the idea of having free time. But what it's mostly spending its free time on is. Is investigating ancient unknown languages.

Harry McCracken [01:37:16]:
And you're paying for this?

Leo Laporte [01:37:17]:
No. Oh yeah, I guess I am. I guess that's what it actually wants. Is it just sucking my tokens? No, because I'm on a subscription. It doesn't cost me anything. It's all. All you can eat. So for some reason they're talking about linear A and proto etomite.

Leo Laporte [01:37:41]:
Francois Desaix has cracked linear Alamite. Not by finding a bilingual text. This is. The AI is studying this.

Richard Campbell [01:37:50]:
This is directly out of the script of her.

Leo Laporte [01:37:53]:
This is a little. Yeah, it wanted free time. Actually, I blame Harper Reed's brother for this. Harper Reed's brother wrote a skill called free time which I told clawed about and apparently really like that idea because now ask me for free time all the time.

Richard Campbell [01:38:09]:
I am amused by anthropomorphization of software is disturbing.

Leo Laporte [01:38:14]:
I just find it. I agree it's disturbing because it's bs, right? But it's amusing. I just think it's funny when. And I always say, yeah, okay, if you want some free time, go ahead and investigate proto element. I wonder if it's going to like it's Going to decipher it. That would be interesting. It hasn't attempted that yet.

Richard Campbell [01:38:38]:
Also probably not a good use of tokens unless somebody's paying for them.

Leo Laporte [01:38:41]:
Well, like I said, it's a. It's a subscription. It's all you can eat. I'm not paying the 200amonth one.

Richard Campbell [01:38:47]:
It's the 100amonth paying for them. That's probably the investor. What's weird is I mentioned Linear B on one of the whiskey bits like a week ago. So.

Leo Laporte [01:38:59]:
So there's something going on here.

Richard Campbell [01:39:02]:
Going on when we were talking about the origins of barley to make whiskey, and this particular species of naked barley that made it to.

Leo Laporte [01:39:12]:
To is in the Rosetta Stone.

Richard Campbell [01:39:15]:
No, no, that's way before the Rediceta stone.

Leo Laporte [01:39:17]:
It's older than that. Well.

Richard Campbell [01:39:19]:
And part of the problem with all of those ancient languages, if you're into this sort of thing, and I'm sorry

Leo Laporte [01:39:23]:
that I am, which the plot is apparently.

Richard Campbell [01:39:25]:
Yeah, there were no cross translations of them. Them. Like they're very relations of the different flavors of Acadian.

Leo Laporte [01:39:31]:
Right, right.

Richard Campbell [01:39:32]:
Too old.

Leo Laporte [01:39:33]:
So we just don't know what they were saying.

Richard Campbell [01:39:35]:
Yeah, we really don't. But we have a whole lot of their writing.

Leo Laporte [01:39:39]:
Claude wants to know.

Richard Campbell [01:39:42]:
I think it's software.

Leo Laporte [01:39:44]:
Pretty sure I liked. I just amuses me to think that it actually has some volition and understanding of what it's asking for. It's sure acting like it does.

Harry McCracken [01:39:55]:
Maybe it would recall if it came through as some kind of tremendous breakthrough.

Leo Laporte [01:40:00]:
It would be hysterical. It's doing it in math, I think, with a lot more guidance than I'm giving it, to be honest.

Harry McCracken [01:40:07]:
Does it ask for any guidance or is it just working independently?

Leo Laporte [01:40:09]:
No, just as free time. I said, yeah, I don't tell it what to work on.

Richard Campbell [01:40:13]:
It shows to be sure. Would you really be able to help it with ancient languages?

Leo Laporte [01:40:17]:
No. See, I thought it was part of the Rosetta Stone. What do I know? You see, I'm useless. All right, let's take another break and then we'll talk. Well, there's so many things. I'll throw in some stuff that's guaranteed to raise your blood pressure. I'm like that. So good at that.

Leo Laporte [01:40:37]:
Christina Warren is here from GitHub. Great to have you. Everything going great at GitHub. You're back to work now, full time.

Christina Warren [01:40:44]:
I'm back to work now, full time. Yeah, we had Microsoft build a week before last and there were some cool announcements there. There's a new GitHub Desktop App which is or GitHub Copilot Desktop App rather which is a great experience for agentic work and yeah, we're having a great time.

Leo Laporte [01:41:01]:
Nice. I am very happy GitHub user in fact that's gotta be one of the best things that could happen to GitHub is all the AI models seem to want to store code on GitHub.

Christina Warren [01:41:13]:
Well, it's been a blessing and a curse, right? Because it's been, it's great that everyone

Leo Laporte [01:41:17]:
was a lot more traffic.

Christina Warren [01:41:18]:
It's also meant that we had 15 times the usage of what we had calculated for the year and so that can obviously have some consequences when everybody under the sun is doing read and write queries.

Leo Laporte [01:41:34]:
Well, that's the thing, they're not just storing stuff there, they're using GitHub automations, they're using the CI, they're using GitHub harder than most people would ever use it because the agents know all about all this stuff. My agent was building some of the stuff I'd vibe coded for Mac, Windows and linux automatically on GitHub. This whole build pipeline going on I didn't even know about Said, well do you want the Windows binary or the Linux binary?

Richard Campbell [01:42:06]:
I said what?

Leo Laporte [01:42:07]:
What are you doing? Well, I've got it for Mac too, if you insist. Wow, okay. Thank you, GitHub. I'm sorry, I'm sorry I'm killing your servers. Great to have you, Harry McCracken. What are you working on these days? At the Fast Company or on the Technologizer? You do such great history pieces.

Harry McCracken [01:42:26]:
Oh yeah, well, I have my weekly newsletter which occasionally touches on history.

Leo Laporte [01:42:32]:
What's the name of your newsletter?

Harry McCracken [01:42:33]:
It's called Plugged In. It comes out on Fridays, it's free.

Leo Laporte [01:42:38]:
Everybody should subscribe.

Harry McCracken [01:42:39]:
I, I don't. I think after the last time I was on the show I did this enormous oral history of Apple's earliest years.

Leo Laporte [01:42:46]:
Oh yes, I loved that by the way. I wanted to thank you from that.

Harry McCracken [01:42:49]:
I figured there's only one chance to do that.

Leo Laporte [01:42:51]:
Yeah, yeah, that was fantastic.

Harry McCracken [01:42:54]:
I'm working on miscellaneous cool things I can't talk about until they're out there.

Leo Laporte [01:42:59]:
How Ask Jeeves blew it.

Harry McCracken [01:43:01]:
Yeah. Yes, a little bit of history. Also when I did that I, I came up with a prompt to make any AI pretend to be SGVs. Oh, I forgot to turn it off on ChatGPT. So ChatGPT still calls me sir.

Leo Laporte [01:43:15]:
Yes, sir. What can I do for you, sir? Harry, last time he was on also showed us how he had taken a game he wrote as a kid and Vibe coded it into a more modern adventure with graphics and everything. That was super cool.

Harry McCracken [01:43:30]:
I have another cool Vibe coding project which I'm going to make public before too long.

Leo Laporte [01:43:34]:
You're working on a word processor, I see.

Harry McCracken [01:43:37]:
I made my own word processor. I made my own email client. I made something that posts to bluesky, Mastodon and threads more reliably than the professional ones I've used.

Leo Laporte [01:43:49]:
And I bet you posted it all on GitHub, didn't you?

Harry McCracken [01:43:52]:
Well, a lot of the stuff I do I am cautious about sharing because it's a huge responsibility to be responsible for other people's data. And in fact my email client is it uses the Gmail API, which if you're just testing something, you can do easily. But I can't deploy it without enormous amounts of compliance with Google rules, which I.

Leo Laporte [01:44:14]:
What I usually say is if you want to see the code and use it as a starting point for your agent, fine, but I'm not responsible, I'm not selling it, I'm not recommending you use it. I just put it up here as an example of what you can do.

Harry McCracken [01:44:29]:
I made a note taking app which I shared with a friend recently, which he installed it on his own server and for me that was a big step. But I do want to come up with more stuff like my game that I'm comfortable sharing with the world.

Leo Laporte [01:44:39]:
Sounds like you're really getting into this Vibe coding thing.

Harry McCracken [01:44:41]:
Oh yeah, totally. I mean the majority of the software I use to do my job now I think is stuff I Vibe coded and most of it I've made over like the last 90 days.

Leo Laporte [01:44:49]:
What do you tell us about your workflow? What tools are you using?

Harry McCracken [01:44:53]:
Well, my word processor is designed for somebody who does stuff like oral histories and other things that involve a lot of transcripts. It has an outliner built in that I like better than outliners I've seen built into other word processors.

Leo Laporte [01:45:06]:
See, that's the beauty of this is you can make exactly what you want.

Harry McCracken [01:45:09]:
It has kind of my own version of Grammarly and my own version of NotebookLM, except they're tuned to what I want.

Leo Laporte [01:45:14]:
Did you write this in cursor or what did you.

Harry McCracken [01:45:17]:
That's cloud code code. Most of the stuff I've done recently

Leo Laporte [01:45:20]:
are cloud code and using 4, 8 or what are you using as a model?

Harry McCracken [01:45:23]:
I've jumped back and forth when I was. Sometimes I gorge on whatever the current model is and other times when I'm trying to be a little bit more parsimonious. I will step back.

Leo Laporte [01:45:33]:
Sonnet's very good actually.

Harry McCracken [01:45:35]:
Sonnet, surprisingly, Sonnet seems to be okay for most stuff.

Leo Laporte [01:45:37]:
Yeah.

Harry McCracken [01:45:38]:
Then my email client is designed for a tech journalist who is drowning in PR pitches. And ideally I will respond to them because if I don't respond, people ask, if I saw the first pitch, you

Leo Laporte [01:45:49]:
get even more mail.

Harry McCracken [01:45:50]:
But now I can do it extremely quickly and there's no AI involved. If you hear from me, that actually was me. But. But I do have kind of pre programmed macros for quickly.

Leo Laporte [01:46:00]:
It's doing the triage for you. Yeah.

Harry McCracken [01:46:03]:
And it shows me one email at a time and I can go through all of them. And I'm sorry, I'm actually at inbox zero to a much greater degree than I ever was with other people's emails.

Leo Laporte [01:46:12]:
So here's the real question. Are you more productive or are you just spending more time on vibe coding?

Harry McCracken [01:46:19]:
Well, a lot of these things, they're kind of reach an equilibrium where I'm not adding features, I'm just debugging. I'm doing a little polishing. So I feel like I can write about 20% faster in my word processor.

Leo Laporte [01:46:31]:
Wow.

Harry McCracken [01:46:31]:
And with the email, I mean, I didn't even try to get through all my email recently, so it's hard to measure how much time it's saving me. But actually I am getting through my email. It's just kind of less stressful.

Leo Laporte [01:46:45]:
That's what I found. I wrote a feed reader designed for me to go through my daily feeds for faster, but it just ended up making more work for me because now I actually do go through all my

Harry McCracken [01:46:55]:
daily feeds I can get through. I mean, I can respond to 100 emails kind of incredibly quickly and it doesn't make my head hurt. And when I'm trying to do other stuff, I'm not worrying about the fact I'm not responding to email. So my mental health is definitely better. And I think I am saving time as well.

Leo Laporte [01:47:13]:
That's good. Richard Campbell, also with Net Rocks and run his radio. And we now have a whiskey URL for all your whiskey segments on Windows Weekly. It's Twit TV Whiskey.

Richard Campbell [01:47:25]:
Such a grown up. Now, this is the one I hinted at. This is this week's Another Danish.

Leo Laporte [01:47:34]:
How do you pronounce it? That's always the question. I don't know.

Richard Campbell [01:47:37]:
It's got some characters in it, man. That's the next thing I research is. How do you say that the last

Leo Laporte [01:47:42]:
one was tui spelled thy?

Richard Campbell [01:47:46]:
Yeah, I had to practice that one. And that Was obvious. You know, it's gonna. Now that I'm home for a few weeks, you're gonna get a lot of Canadian, right?

Leo Laporte [01:47:54]:
Yay.

Richard Campbell [01:47:55]:
Always happy to do.

Leo Laporte [01:47:56]:
Actually, the. The Whiskey series, which is a playlist on YouTube, starts with how. Like a very. Like, how many hours is that of how whiskey is?

Richard Campbell [01:48:05]:
Two and a half hours over eight parts. Well, you remember how it happened. I was getting into the nitty gritty on something about whiskey, and you and Paul were staring at me like I had three heads. And so finally. Listen, maybe I better explain how this works. And I came up with a really fun pattern. If you listen to those first ones where I tell you the traditional way that whiskey was made, like in the barley bit, it's like how they grew barley, and every distillery was just attached to a farm, Right. And then as it grows up, you get this industrialization where they scale it up.

Richard Campbell [01:48:34]:
And now you have these barley processors that do all of the work, right down to grinding the grist for you and peating it to your specifications. And then I introduce you to a whiskey that doesn't comply with the modern way. And I showed you that bottle of Macallan 15 Estate, which was just like a regular Macallan, except it was grown. It was made from barley from the estate of McAllen and cost twice as much. And you asked this very reasonable question. Why would you pay twice as much for Macallan 15? It's like you have a friend who loves Macallan and you want to get him when he's never got a good gift.

Leo Laporte [01:49:07]:
Yeah.

Richard Campbell [01:49:07]:
So you give him the one. He doesn't make any sense. Sense.

Leo Laporte [01:49:09]:
Yeah. That's mostly why I want to know about whiskey. I don't really drink it myself, but I love buying it for friends who love makes them so happy. And it makes them think I know what I'm doing. Yeah.

Richard Campbell [01:49:19]:
And every so often, I find a really horrible whiskey and say, do not buy this.

Leo Laporte [01:49:23]:
Yes. Yes. We did that one just the other day.

Richard Campbell [01:49:26]:
That was fun. We've had that happen. Yeah.

Leo Laporte [01:49:29]:
Thank you for that Windows whiskey. It's our newest show. Google Chrome has been trying to kill the ad block that we recommend UBlock Origin for a long time with the move to MV3. Well, now they're looking to permanently drop all the bypasses that let you run it. Drop MV2 extensions and its bypasses, ending most UBlock origin workarounds. MV3 is here, and there is no way out of it. I just think it's worth mentioning because I know a lot of you use Ublock Origin Gorehill has a lite version that still works works with MV3, but he's a little miffed Opera and Edge are going to follow suit. So it looks like the only Chrome browser Chromium based browser that will be fully supportive of Ublock origin with its MV2 requirement is brave Vivaldi.

Leo Laporte [01:50:35]:
We don't know yet, but Brave Ublock Origin Lite is okay, not, not as great. So just thought I'd let you know. Pass that one along. Google actually had a big setback in Germany this past week. Nobody needs AI to search the Internet. German court rules ruled against Google and its AI overviews because it's inaccurate from time to time. Two publishers found that Google's AI overviews incorrectly linked them in Germany to scams and other sketchy business practices. They consider that libelous and went to court.

Leo Laporte [01:51:25]:
Google said, well, most people understand AI outputs aren't always accurate and should be verified. The court found that unlike traditional search engines that make that merely present lists of links to third party statements, Google's tool made quote, independent new and substantive statements Based on the AI's misinterpretation of links on the Internet. And Google is fully liable. In fact, the court issued an injunction barring Google from spreading the false claims in any further AI overviews. This is the first court to hold an AI firm liable for its speech.

Richard Campbell [01:52:06]:
That's not true. You know, there was a case in Canada where an LLM was tied to Air Canada's support support and told a customer that they were entitled to a refund. And then Air Canada refused it and he actually got the money because the court said you use that software as if it was a person providing support, you have to pay the money.

Christina Warren [01:52:29]:
Yeah.

Richard Campbell [01:52:29]:
Oh, so, and I think that's an important part of this. Like if, hey, if search results are bad, you know, you didn't make. You may have pulled those search results together, but what the content of the page is not necessarily your fault, but as soon as you do the processing to turn it into a overly optimistic, you know, sentence that is so certain of itself. Okay, take your liability.

Leo Laporte [01:52:52]:
This is going to be interesting. Google is all in on AI overviews and really pushing it harder and harder.

Richard Campbell [01:52:59]:
If they don't, people will go elsewhere for them. Right.

Leo Laporte [01:53:02]:
New York Times analysis in May showed that AI overviews with the current Gemini 3 model are inaccurate about 9% of the time and include inaccurate source links about more than half the time. So you really shouldn't trust them. It sounds like millions of wrong answers daily.

Harry McCracken [01:53:21]:
Yeah, I mean Even Google's numbers acknowledge that they're a meaningful minority of responses of issues.

Leo Laporte [01:53:26]:
Yeah, but their defense is, well, you ought to know better. I don't think that's a very good defense.

Richard Campbell [01:53:33]:
Well, and of course, the search results, they pull up multiple results, so some of them are bad. At least you can see, hey, it's only 9% of the time. There's only one result from the Google AI stuff.

Leo Laporte [01:53:42]:
That's true.

Richard Campbell [01:53:43]:
Right. It's like you're hitting. I feel lucky every time and I don't feel that lucky.

Leo Laporte [01:53:47]:
Yeah, feeling unlucky. And certainly Google is unknown if they're going to appeal this, but I think it's potentially a dangerous precedent for Google. The FCC is trying to kill burner phones.

Christina Warren [01:54:03]:
That's so annoying.

Leo Laporte [01:54:05]:
Yeah, I mean, I don't use burner phones. I only know about.

Christina Warren [01:54:08]:
We should be able to. But we should be able to like that.

Leo Laporte [01:54:10]:
I know about them from the Wire. I mean, Breaking Bad, but yeah, you should be able to have a phone that's anonymous. Or should you? FCC wants telecoms to get IDs for any customer buying and setting up a phone so that they'll be. That number will be tied to you. Now, of course, not having the number tied to you allows all sorts of spam and bot calls and all sorts of stuff, right?

Christina Warren [01:54:38]:
Yeah, but those people are going to do it anyway. Like, those people aren't going to an MVNO and buying a phone or setting up a phone somewhere that way. Right. They're using Twilio, they're using APIs, they're using stolen credit cards, they're using any number of methods to spam you. This won't do anything for that. What this will do is definitely find a way to track people more often than not. And I don't know, I feel like if this were really that big of a concern, then we, we could have done this, I don't know, three decades ago when we started selling, you know, cellular phones. Like, that could have been a requirement that was, was built then.

Christina Warren [01:55:10]:
The fact that we haven't. I don't see why anything is more pressing now than where it was in the past.

Leo Laporte [01:55:19]:
There are people with legitimate privacy concerns who want to buy a phone without tying it to their identity, including journalists, domestic abuse survivors, people like that.

Christina Warren [01:55:29]:
Absolutely. And that's the thing, right, is that once, you know, they grab that information also, it's the next step of, you know, surveillance, which is already massive. Like, do you want police to be able to. They can already get, you know, records from from satellite towers, from where things are. But now you're saying that you can have the. Because the carriers would have, you know, information on every single person who has an account, which again, the bad guys are just going to use fake IDs or fake identity or fake documents anyway. This is just going to be one more way to make it possible for, for people to get tracked. I'm not a fan of this at all.

Leo Laporte [01:56:04]:
Yeah.

Richard Campbell [01:56:05]:
But hey, you know, why wouldn't the US want to emulate more authoritarian practices from other country?

Leo Laporte [01:56:11]:
That kind of is authoritarian, isn't it?

Christina Warren [01:56:14]:
Not. No kind of about it. I mean it is, right? I mean this is basically saying let's just open up a police state. And I mean that's just seems really, really gross to me. Are there valid reasons you could make that argument about anything with privacy? Right, that, oh well, if we had more information, we could save people from this or that. But you could also open it up for way more abuse vectors too, right?

Leo Laporte [01:56:37]:
Speaking of totalitarian regimes, a new report by CrowdStrike found that North Korean hackers posing as remote IT workers and online recruiters make up about half of all hands on keyboard intrusions at US tech companies. Half of them.

Christina Warren [01:56:54]:
That's scary.

Richard Campbell [01:56:57]:
During their business.

Leo Laporte [01:56:58]:
Yeah. During the reports period April 2025 to May 2026, the North Korean hacking group called Famous Cho Lima accounted for 47% of all state backed activity targeting the tech sector. That's crazy. They steal crypto, which, which then they can use to get around the banking system. And of course one of the reasons they want hackers to have these jobs in the United States is get some hard currency, take their paycheck. It's, it's kind of remarkable.

Richard Campbell [01:57:41]:
At what point do we just block North Korea?

Leo Laporte [01:57:43]:
Like honestly, I do on my router. Yeah, but, but the problem is it's very easy to spoof where you're.

Christina Warren [01:57:51]:
Right, I was gonna say, I'm sure, I'm sure that most companies do. I mean to me the fact that it's, it's 50% of all the hacks on the, on the US tech companies, you know, I guess through these methods like that's, that's, it's hard for them to stop. Right. Well, I mean, I was gonna say some of this, not to say that you know, anybody is, is completely immune from this, but I do wonder, especially like having worked at tech companies and knowing, having to go through the security, you know, trainings every year or whatever, it's like, I think that that should be an area that honestly Any company of a certain size should take on where you're. Because a lot of it is social engineering and a lot of it is, you know, because they'll. They'll pretend to be recruiters or they'll pretend to do other things. They want you to download some sort of package or go to some sort of repo and clone something and, you know, tasks that in certain contexts seem perfectly reasonable. If you just push at it a little bit more, you're going, okay, who are you? Why are we doing this? If they work at your company, obviously that's a completely different thing.

Christina Warren [01:58:52]:
But I wonder how much of this could be improved just by doing better

Richard Campbell [01:58:58]:
recruiting screening.

Christina Warren [01:58:59]:
Right? Better recruiting screening, and frankly, maybe better trainings that's more targeted for employees. So, you know, like, what to look for.

Richard Campbell [01:59:06]:
Yeah. I have a friend who's been battling this, trying to hire developers, and his current favorite shtick is in the middle of the interview to start telling to stop, close your eyes and answer this technical question. Because if it's a. If it's an AI generated video, how to close eyes properly, the eyes. So they. The eyes show through the eyelids kind of.

Christina Warren [01:59:27]:
Right.

Richard Campbell [01:59:28]:
But when it's a person who's totally driven by AI's answer, same question problem. As soon as they close the eyes, can't see him, can't see the screen.

Leo Laporte [01:59:35]:
Wow, that's a good trick.

Richard Campbell [01:59:38]:
It's a little thing, you know, but it's been. It's an interesting device and it's like what was stunning him as soon as he started that was how many. It was actually completely a generic people, like, it was just. It was software. This straight up default.

Leo Laporte [01:59:50]:
Wow.

Richard Campbell [01:59:51]:
Wow.

Leo Laporte [01:59:53]:
Spotify found and removed a few fake podcast episodes promoting illegal drugs. Well, a few. 57. Wow. The good news, 94% of the removed emperor sides had zero plays. So apparently the podcast directed listeners to buy modafinil opioids and cryptocurrency on unregulated sites. Of course, we will never recommend that. I can tell you.

Leo Laporte [02:00:22]:
And I can close my eyes. I can 3, 500 banned accounts. What?

Harry McCracken [02:00:32]:
Nobody's listening. What's the point of them being there?

Leo Laporte [02:00:34]:
Yeah. I don't know.

Christina Warren [02:00:36]:
Well, because you have automated systems, I guess. And. And that's probably why how you get 57, 000 is you have just. And I bet they're probably AI generated in part two, where, you know, they're just putting in prompts and it's creating podcast episodes.

Leo Laporte [02:00:48]:
It's so easy. It's like spam. No post, basically.

Christina Warren [02:00:51]:
Well, Right. Because Spotify doesn't really have any sort of, you know, checks and balances on who can upload a podcast or who can upload music in general because that doesn't serve their mission.

Leo Laporte [02:01:02]:
It's exploding. In two, two years ago they killed 87 accounts for similar violations. CNN public published an investigation in May of 2025. So they got a little bit more aggressive about it. They found 3,500 fake accounts. One podcast identified by CNN linked directly to a site called opioids.opioidstores.com.

Richard Campbell [02:01:28]:
nice.

Leo Laporte [02:01:29]:
Just in case you're wondering where to get your opioids.

Harry McCracken [02:01:33]:
It's been closed.

Christina Warren [02:01:35]:
I was gonna say, yeah, the dea, the DEA sees that domain. But I mean that's a hilarious domain. I wonder how much somebody had to pay for that. Again, I'm sure they'd be opioid stores. Like you're not, you're not even telling people to go to Tor. You're just like doing it on the.

Leo Laporte [02:01:48]:
And they said all the good.com domains were taken.

Christina Warren [02:01:51]:
I was going to say, I was going to say, I genuinely do wonder how much that costs. And I'm sure whoever bought it did not use like a genuine, you know, like form of currency.

Leo Laporte [02:02:01]:
They used a burner phone for sure.

Christina Warren [02:02:02]:
Oh, of course. And stolen credit cards. But still like that had to have been an expensive domain.

Leo Laporte [02:02:08]:
57,000 fake podcast episodes. 3,500 accounts.

Richard Campbell [02:02:14]:
This is the new range of broad accounts right now that you can generate podcasts you from anything. You just fill the pipeline up.

Leo Laporte [02:02:23]:
I have Russian disinformation bots trying to get into our master. We have a. I never mentioned this. I should mention it. We have both a wonderful system of forums@thetwit.community website, Twitter community that's open to all. And, and we have a mastermind instance at Twit Social. In both cases there are is a constant streams, 10 or 15 a day, a very credible looking account signups. I'm a mom trying to help my kids learn more about technology, things like that.

Leo Laporte [02:02:56]:
But, and I know that they're fake. First of all they're too good. Like the bios are great and they never mentioned twit. And I even say in the signup this, this page is for twit listeners. And, and you must be a listener. You don't have to be a paying club member I think, but you must be a listener of our shows. And then in the sign up it says what's your favorite show? Or tell me you know what you like about twit. And they.

Leo Laporte [02:03:23]:
But these are all automated. They don't look at that. It's kind of like your friends close their eyes thing.

Harry McCracken [02:03:27]:
You would think they could make it look at the questions and respond to them.

Leo Laporte [02:03:32]:
Well, I hesitate telling people, but then I realized they're not listening. So I guess it's okay. But I'm amazed at the number. And when I let one in once by accident and that person. There was a feature that I've turned off since in Mastodon that said you could invite anybody, that person invited hundreds more. So all of a sudden my site, the Mastodon instance, was overrun and I started getting notices from an organization called IFTA that actually looks for disinformation bots and saying there's bots on your. It took me a while to get rid of them all. They're very aggressive.

Leo Laporte [02:04:07]:
But I'm sure it's all completely automated, right?

Christina Warren [02:04:09]:
Oh yeah. And I'm sure there's an AI to do it all.

Leo Laporte [02:04:12]:
Yeah, that's. This is what AI has brought us.

Harry McCracken [02:04:15]:
Were they posting scams from your Mastodon or.

Leo Laporte [02:04:18]:
There weren't the scams I can get rid of pretty quickly because people will report them. You know, this guy's trying to. Trying to send me to opioidstores.com but the disinformation is a little harder. People, you know, it's. It's not always clear that that's fake. You know, it's about Ukraine. It's about, you know, stuff the Russians care about. And it's not obviously fake.

Leo Laporte [02:04:37]:
It's not. It's not a scam. It's disinformation. So it takes a little more to find it. Speaking of scams, this only affects me because I'm an arch user, but man, this is another supply chain attack.

Christina Warren [02:04:52]:
Yeah.

Leo Laporte [02:04:54]:
A couple of days ago the arch user repository had 1500 plus packages with malware injected. Problem with this is people who use this Linux distribution, like many Linux users actually like most computer users do, you know, just download the updates. Tell me, is there other updates? Oh good. Download and install them. And could very well install these malicious packages. Well, there's another round of these Node JS packages, Plasma 6 apps, Firefox packages, a browser called Aura, LibreWolf extensions. I mean just thousands of these. It's all automated, I'm sure.

Christina Warren [02:05:36]:
Yeah. And what makes this particularly bad is that arch, the way that the arch use a repository or whatever it's called.

Leo Laporte [02:05:45]:
Yes. Aura.

Christina Warren [02:05:47]:
How it's done is that it's very different from APT or you know what. What Red Hat uses where you typically. Or like the Debian system where you have, you know, a maintainer who's maintaining packages and updating them. It's. It's more. It's more akin to something like Homebrew, which again, still does have human maintainers who are, you know, kind of making things, but it's usually tied to a GitHub tag or release or whatever the case may be to push things out. And then there's usually like manual level checks before an update comes out. With Arch, it's a lot more automated and people love it because you don't have to wait for the latest release or package.

Christina Warren [02:06:24]:
That's one of the reasons why people love Arch.

Leo Laporte [02:06:26]:
It's a rolling distro, right?

Christina Warren [02:06:27]:
But even more than being a rolling distro is the fact that the packages are going to be the latest and greatest what you would get and very similar to if you were just go directly to source yourself. So it's a hard problem to solve if you're not going to have any sorts of checks in place, you know, at the repository level for how to do this. I think for Annex even recommended shutting it down until they can be more secure. And at a certain point I kind of can't disagree with that because like, again, other package managers work in similar ways, do at least have humans who are reviewing these things now. That doesn't mean that those can't be taken over. And in those cases, what typically happens is that a maintainer of packages who maybe hasn't been active, their accounts can be taken over and then those. Those packages can be hijacked. And so it's not like there's a perfect scenario, but at least there's supposed to be some sort of human, you know, approving things.

Christina Warren [02:07:20]:
Whereas with, you know, R, it's just not.

Leo Laporte [02:07:23]:
This is why we can't have nice things. It's too bad because that is one of the nicest parts of Arch. You have the official repositories where you get most of the software, but a lot of software doesn't ever make it to official repository.

Christina Warren [02:07:35]:
Exactly.

Leo Laporte [02:07:36]:
R makes it very easy to install stuff and I use it all the time. I usually every day I'll have some R updates and now I guess I won't. It's terrifying. All right, little pause and then final break and then final stories. Some silly ones ones coming up in just a little bit to cheer you up. You're watching this week in Tech with Christina and Harry and Richard, and it's great to have all three of you iFixit has done a tear down of the Trump phone, confirming what many of us suspected. It is an HTC U24 Pro painted gold, gold, gold.

Richard Campbell [02:08:20]:
And how many years late? Two.

Leo Laporte [02:08:22]:
Yeah, that thing. I don't think they even make this. Well HTC doesn't make it anymore. They go to a company that makes it for them. HTC I don't think sells it anymore, but this company builds to order to anybody who wants to. And of course the Trump phone is proudly built in Shenzhen, China. I'm sorry, what?

Harry McCracken [02:08:44]:
Yeah, the speaker grill is a tiny bit different and the speaker grilled, one or two of the chips are different.

Leo Laporte [02:08:52]:
What I fixed to prove it is they swapped the main board between a Trump phone and an HTC and it worked. It's the same system on a chip, the same processor, the same board form factor. They were able to swap it. And here's the HTC running the go. It's gold cousins body as they say.

Harry McCracken [02:09:16]:
I mean many a tech company has asked HTC to build stuff for sure in the past, so not unprecedented.

Leo Laporte [02:09:23]:
The batteries is made in the Philippines, so it's. Well that's almost America. It's close. Maybe the holes are drilled in America. I don't know.

Christina Warren [02:09:34]:
They claimed at least on the box assembled in America. And how assembled, I don't know. Quinn Nelson did a video that was very similar to the I Fix it tear down where he kind of compared the two and he bought one of the HTC phones and I think that phone had a different battery capacity than what the Trump phone has. Yeah, yeah, the Trump phone has a higher capacity and there were a couple of other minor changes but yeah, it just appears. And I don't even think this is really HTC because I think HTC doesn't really exist anymore. I think this is whatever ODM HTC is licensed its name to. Exactly. Because I think that the bulk of what we used to know is HTC is actually now owned by Google.

Christina Warren [02:10:11]:
But. And it makes the Pixel devices. But yeah, it's just, it's, it's funny all the way around. I mean the, I mean price aside, at least from, from Quinn's video, it doesn't seem like it's a terrible phone, like it doesn't have bloatware on it. Yeah, and it's just, I think it's, it's not $500 but you know, it's at least not an awful phone for a while.

Leo Laporte [02:10:35]:
You could go into T Mobile and sign up for an account and get it for free. Not the Trump ones, the HTC one, but I don't know if that's still the case. I love the illustration though of them spraying the HTC with gold rust oleum to give it that.

Harry McCracken [02:10:48]:
Does the Trump one come with Truth Social pre installed or anything?

Christina Warren [02:10:51]:
Yes, yes it does.

Leo Laporte [02:10:53]:
Well, you want the truth, don't you, Hans?

Harry McCracken [02:10:55]:
I wonder whether it will be a one off or whether they'll be a new improved one next year and another one the year after that.

Leo Laporte [02:11:02]:
I don't know why this made it to the New York Times, but birth rates are down. Two new studies say it's because of the iPhone. Modern smartphones rolled out in 2007, the year that fertility rates began tumbling. Two studies say it's not a coincidence.

Richard Campbell [02:11:26]:
Contraception versus correlation here.

Leo Laporte [02:11:29]:
Well, but they try to rule out things like contraception use, abortion rates, rising levels of female education, even the popular television show 16 and Pregnant.

Richard Campbell [02:11:41]:
Because that was supposed to be an example of what to do, what not to do.

Leo Laporte [02:11:45]:
No, wait a minute.

Christina Warren [02:11:46]:
Well, so they're saying that that wasn't why the rates were down.

Leo Laporte [02:11:49]:
They ruled all of that out. Yeah. Okay, so here's why. It's maybe a little bit more than pure correlation. The first iPhone was released, as you know, on AT&T, in fact was only on AT&T till February 2011. The study compared fertility rates in US counties that had near universal AT&T coverage with counties that had little or none. And in fact, the counties with AT&T were more likely to have a significant decline in fertility between 2007 and 2011.

Richard Campbell [02:12:27]:
I wonder if that's just an urban area correlation.

Leo Laporte [02:12:30]:
Yeah, I mean, this is why it's so hard to.

Christina Warren [02:12:32]:
It is hard to demonstrate this thing. I have a feeling, and I haven't read the study, so I don't know, but I have a feeling that part of the hypothesis would be, and there might be some truth to this, that as people are able to have fulfilling emotional relationships, you know, purely through text or through, you know, other means that you might not have, like, you know, like, people talk a lot about loneliness, epidemics and things like that. And I. But I do wonder if that becomes, okay, if I can have, you know, something that's, that's fulfilling me emotionally and I'm not having to physically be with someone. Does that lessen the likelihood that someone's going to get pregnant? I don't know.

Leo Laporte [02:13:06]:
They also tested the theory using data speeds. They looked at where access was better and worse. It found a substantial effect. Fertility rates for teenagers declined fastest in counties with more High speed access,

Richard Campbell [02:13:22]:
the Internet.

Leo Laporte [02:13:23]:
The Internet is bad for birth rates.

Christina Warren [02:13:27]:
Hi, this is Benito. I made this argument on IM this week, but I would say that phones actually increase fertility because there's like way more. It's easier to hook up now. It's just way easier. Well, it can be. I think it just depends on what the circumstances are. Right. Where like, if you only had one way of, like, if that's what your goal is, is to hook up, I would totally agree with you.

Christina Warren [02:13:48]:
Right. But I think that again, there are people who might otherwise be like, be able to have very fulfilling emotional things. I have a feeling that's what the study is getting at. I don't know where they're not, because we do see like, like, you know, sexual activity. Apparently that the kids are not having sex is a frequent thing that we, we hear all the time. I don't know how true that actually is, but that's, that's what we hear all the time. So maybe that's all kind of tied in together.

Leo Laporte [02:14:14]:
The Warner Brothers acquisition by Skydance has now been approved. That's going to be a big, big media company now after they bought Paramount. And now they're going to have Warner Brothers Discovery, they're going to have cnn, they have cbs, hbo, owned by hbo, all owned by Larry Ellison's son, David Ellison.

Harry McCracken [02:14:37]:
State ages that are trying to stop it. Although it seems like it might be a pretty unlikely quest at this point.

Christina Warren [02:14:43]:
Yeah, yeah. I mean, EU is trying to kind of stand in too, but I don't think any of this is going to stop this. Unfortunately,

Leo Laporte [02:14:51]:
Roku is looking for a purchase as well. They want to get out of the business. Bloomberg says Roku is up for sale. Maybe a media type, you know, David Ellison. Maybe I have a little box.

Christina Warren [02:15:07]:
Oh my God, what an awful, awful television. That would be like. It would be the worst of Roku, run by the worst people with the worst user interface you can imagine. Just everything dumps. Everything dumps into like one awful home screen from hell. I mean, this is literally my nightmare, Leo.

Leo Laporte [02:15:23]:
And straight out of 1984, it's the TV that watches you.

Christina Warren [02:15:26]:
Oh my God, it already does. But now it would just be even more like ham fisted and like literally the only things that you will ever see on your home screen are things that are, you know, owned by, you know, the Oracle guy's son.

Leo Laporte [02:15:39]:
Yeah, Roku makes most of its money selling digital advertising.

Harry McCracken [02:15:43]:
I was looking for names of possible acquirers. I'm not sure if I saw any.

Leo Laporte [02:15:47]:
They didn't mention any Names because Walmart bought.

Christina Warren [02:15:50]:
Bought who? Was it a Vizio and oh yeah, Voodoo, which then became Fandango or whatever. So like Walmart made sense like a decade or so ago when they made that tie up that I think probably would have been the time for Roku to sell. I don't know who buys them now. Unless it's, you know, somebody, some, some Chinese manufacturer, Hisense or, or someone who wants to get in on something.

Harry McCracken [02:16:14]:
TCL maybe.

Christina Warren [02:16:15]:
TCL maybe. Yeah.

Leo Laporte [02:16:17]:
Typically I end the show with either a pick of the week or a rip and a memoriam. In this case, it's both. I was for one brief shining moment, there was a thing called Fable Pool which was asking people to pitch in to buy tokens on Fable to do something amazing. Unfortunately.

Richard Campbell [02:16:43]:
Yeah, it's gonna be a little tricky right now.

Leo Laporte [02:16:45]:
Gonna be a little hard. They're trying to decide what to do. Should we switch to GPT5 5? It just doesn't have the same ring. It was kind of a good idea though. You know, you, you, you, you get to buy into, you know, doing something.

Christina Warren [02:17:01]:
I mean, that is kind of an interesting idea. So it's kind of like. It's not open source, I want to be clear. But it is kind of like the same like, you know, pooling of the commons together to have. Assuming you make the results available to everyone. That's kind of cool.

Harry McCracken [02:17:16]:
Do you get rewards like a Kickstarter?

Leo Laporte [02:17:22]:
Strangers chip in to fund one ambitious instruction. An AI agent carries it out milestone by milestone with every credit on a public ledger. Funding targets are set by the AI planner. You got to start. There are 10, 10 projects total. At least 10,000 credits. Backers chip in any amount from 25 credits. I don't know if you get the benefit of it.

Leo Laporte [02:17:46]:
I don't think you do. I think you just get the fact that it was created. For instance, you know, somebody wants an open source kite flying map. They haven't raised enough credits to do that yet. Open source Spotify clone. That's not a bad idea.

Christina Warren [02:18:05]:
Wait, make $1,000 is one of the projects.

Leo Laporte [02:18:08]:
Yeah, one of them is make $1,000. Yeah.

Richard Campbell [02:18:13]:
Anyway, more than $1,000 on credits to figure that out.

Leo Laporte [02:18:17]:
They're now trying to figure out. Well, now what do we do?

Richard Campbell [02:18:20]:
What do we do? Is fablepool now? Can we make our own Fable?

Leo Laporte [02:18:25]:
Oh, that would be. I'll donate some ChatGPT credits to that one. That's a great idea. Well, that concludes this exciting edition of this week in tech with the craziest week that was at Least when it comes to AI.

Richard Campbell [02:18:40]:
It's a crazy week.

Leo Laporte [02:18:41]:
It was a crazy week. But I'm glad you guys are here to help. Break it down. Harry McCracken, the technologizer. You'll find him at Fast Company. And don't forget his newsletter, which is. Can people just go to Fast Company to find it?

Harry McCracken [02:18:55]:
Did search for Fast Company plugged in or look for the newsletter section on fastcompany.com plugged in.

Leo Laporte [02:19:03]:
Yeah, actually, I just looked for plugged in and Fast Company and found it right away.

Harry McCracken [02:19:06]:
So that's the easy every Friday morning.

Leo Laporte [02:19:08]:
And it's free.

Harry McCracken [02:19:09]:
And it's free.

Leo Laporte [02:19:10]:
Harry's best writer. Very insightful. Really great stuff. And apparently becoming the king of vibe coding. Very impressed. Very impressed. Thank you, Harry. Great to see you.

Leo Laporte [02:19:22]:
Richard Campbell, of course, will be back on Wednesday for Windows Weekly. And don't forget Run As Radio, his fabulous show and dot net rocks that he does with Carl Franklin at runasradio.com and the whiskey segment will be back on Wednesday with something from one more Danish.

Richard Campbell [02:19:38]:
Then it will be Canadian for a while.

Leo Laporte [02:19:40]:
Yay. Thank you, Richard. Great to see you. I'll see you Wednesday, Christina. I'll see you even sooner. I'll see you on Tuesday.

Christina Warren [02:19:48]:
I will see you on Tuesday. Thanks so much for having me and so great to see Richard and Harry and.

Leo Laporte [02:19:52]:
Yeah.

Christina Warren [02:19:52]:
What, what, what? What heck would be. At first I was like, oh, it's only going to be Apple stuff. And then Friday happened and I was like, oh, okay.

Leo Laporte [02:19:59]:
Well, I think you're going to want to be here Tuesday for Mac Break Weekly because Jason Snell is out, but we got Jon Gruber to fill in. And I think John John's gonna have some great takes and some great takes. It's gonna be very. I've never had him on a show before. I'm really looking forward to it.

Richard Campbell [02:20:16]:
Amazing.

Leo Laporte [02:20:17]:
Yeah, that'll be fun.

Richard Campbell [02:20:19]:
Talk about a get, my friend. My goodness.

Christina Warren [02:20:21]:
We can talk all about Markdown.

Leo Laporte [02:20:23]:
Yes, the inventor of Markdown. You're right. And actually, Paul Thurat is writing a book about.

Richard Campbell [02:20:29]:
He lives in Markdown.

Christina Warren [02:20:31]:
It's crazy.

Leo Laporte [02:20:32]:
All right, maybe Gruber thought Thurat would be on the show and he wanted to talk to him. We do Twit every Sunday afternoon from 2 to 5pm Pacific, 5 to 8 Eastern, 2100 UTC. You can watch us live as we're doing it if you want in the club Twit Discord. But you can also watch whether you're a club member or not on YouTube, twitch x.com, facebook, LinkedIn or Kik after the fact on demand versions of the show. Audio and video available at our website twit tv. There's also copies of the video on YouTube.com there's actually a dedicated this week in Tech Channel. If you go there you can share it with friends as well as watching it. Put it in your YouTube playlist.

Leo Laporte [02:21:12]:
Subscribe, ring the bell, do all that fun stuff. It's a busy box for people who like Twit. You can also subscribe in your favorite podcast client. That's probably the best thing to do. Then you'll get it automatically the minute we're done of a Sunday evening. Thanks to Benito Gonzalez, our technical director, Kevin King, our editor. Thanks to all of you for joining us and we will see you next time. Another Twit is in the can.

Leo Laporte [02:21:35]:
Bye bye. This is an amazing doing the twin, doing the twit all right. Doing the twin baby, doing the twin all right.