Intelligent Machines 885 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 Intelligent Machines. Jeff Jarvis is here. Parris has the week off, but Christina Warren is here, of course, from MacBreak Weekly. She's gonna talk about the use of these new Macintoshes announced yesterday for local AI, plus 2 new AI models. Local models are out. We're testing 'em as we speak. It's a very big day on Intelligent Machines next. Podcasts you love.

Leo Laporte [00:00:26]:
From people you trust. This is TWiT. This is Intelligent Machines with Jeff Jarvis and Paris Martineau, episode 885, recorded Wednesday, August 26th, 2026. Get on the Butterbox. It's time for Intelligent Machines, the show where we cover the latest in AI, robotics, and all the smart little doodads all around us. Jeff Jarvis is here, Emeritus Professor for Journalistic Innovation at the Craig Newmark Graduate School of Journalism at City University of New York.

Jeff Jarvis [00:01:01]:
Newmark.

Leo Laporte [00:01:03]:
He's also, more importantly, the author of a hot new book. It's rising up the bestseller list even as we speak. Hot type, hot type.

Jeff Jarvis [00:01:11]:
Get your hot type here.

Leo Laporte [00:01:12]:
Get your hot type. The Story of the Mergenthaler Linotype. Actually, is it just the Mergenthaler, or do you talk about others?

Jeff Jarvis [00:01:20]:
I also talk about the Paige Compositor, which made Mark Twain bankrupt.

Leo Laporte [00:01:24]:
Crazy.

Jeff Jarvis [00:01:25]:
And the Alden, which drove its inventor to his early death.

Leo Laporte [00:01:30]:
It's a great story, 'cause I didn't realize this, but Mark Twain in his youth Did hand typesetting.

Jeff Jarvis [00:01:37]:
Yes, yes.

Leo Laporte [00:01:37]:
And so he was very interested in this idea of automating it.

Jeff Jarvis [00:01:42]:
He said that a machine could not set type unless it could think.

Leo Laporte [00:01:45]:
Oh boy. Well, it took a little while before we got to thinking machines, but I think Mark Twain would be amused by where we are today. Paris will join us in a little bit, but I want to introduce our guest right now because she's a good friend, somebody I've known since her youth, her youth, early youth, back in the days at Mashable and other tech Publications. Uh, it's great to see you, Christina Warren. She's a regular on MacBreak Weekly, right now at Developer Relations at GitHub, but she also worked at DeepMind before this. And is— I think it's— is it fair to say, Christina, an aficionado of AI?

Christina Warren [00:02:22]:
I, I would like to think so. I hope so. And certainly I'm, I'm a, I'm a— I mean, look, we all have our issues with, with AI in various regards, but, um, I'm certainly interested, and especially when it comes to the stuff that we can do with computing, It's been one of the best and most fascinating things that I think we've seen happen in my lifetime, certainly.

Leo Laporte [00:02:43]:
Well, you're among friends here. You don't have to qualify it. You know, I know it's bad for the water and the environment, but I love it. That's what I say. Anyway, I think this is a good time to have you on because we are in a very interesting time for Local AI for people who want to run AI in their home. I think the, you know, the big story earlier this week on Tuesday morning, Apple announced at 6 a.m. our time, 2 new Macs, the Mac Mini, but maybe more importantly for this conversation, the Mac Studio, which has a new chip in it. The— they call it the M5 Ultra, which is really 4 chips glued together.

Leo Laporte [00:03:29]:
super fast, super high processing, 1.2 terabytes bandwidth to its memory, which is unified memory. And this one comes with 256 gigabytes of memory. And in a month they will offer— and I doubt they'll be able to make more than a handful— 512 gigabyte versions of this. Now, if you want this new Mac, you're gonna pay through the nose. It's a lot more money than the old ones are. In fact, if you want the 256 one with a reasonable amount of A hard drive, it's going to cost you, you know, I don't know, what is it, about $10,000?

Jeff Jarvis [00:04:02]:
If it's 1 terabyte, it's $10,000.

Leo Laporte [00:04:04]:
$10,000. And then we don't know what the 512 will cost, but probably double that. Yeah, close to it. So we covered this a little bit on MacBreak Weekly, but we couldn't really go into the, you know, the deep details of AI on this. And I will say—

Jeff Jarvis [00:04:21]:
Can I ask you a question as we go in just for context here? When local, when, when local models started at a smaller level, the Mac Mini sold out because it was what was what. So, so I just, besides the Studio, I'm curious where the Mac Mini fits in now with AI, the new Mac Mini.

Christina Warren [00:04:40]:
I think that's a great, I think that's a great question. I think the Mac Mini sold out for a couple of reasons. One was that it was even before the RAMageddon and the RAMpocalypse, whatever we want to call it, it was incredibly inexpensive. It was, you know, $599 list, but you could very easily find it for about $500. $1,000, sometimes even less at retail. 16 gigs of RAM, not a big hard drive, but, you know, very capable. And what we saw earlier this year, thanks to the work of Agentic Models and other stuff— and this doesn't really have to do with locally AI, but it's still kind of tangentially related— was the rise of projects like OpenCLAW, where suddenly people wanted to have access to those things, did it. But yeah, I mean, I think before that, you know, Apple, they put 128 gigs of RAM in it was the M3, was the first M3 Max that had that much RAM in it.

Christina Warren [00:05:30]:
I think maybe the M2 Studio might have been able to have about as much RAM. But I think it was the M3 era was when I first started to see, I don't know if your memory matches this, Leo, was when we started to see people taking the Mac a lot more seriously for running local models. And this was well before we were, you know, in the era where we are now, when basically back then it was you know, Gemma and a couple of other— LLaMA, frankly, you know, from—

Jeff Jarvis [00:05:57]:
Small, small models.

Christina Warren [00:05:59]:
Yeah, small. But it was still— I think people saw kind of like where things were going. And we're like, okay, let's see what we can start to do on the Apple side to optimize these models for running locally. Because that was the interesting thing is that, you know, for years and years and years, all any sort of, you know, AI stuff, any sort of training, anything you've done has all been based in the NVIDIA ecosystem. AMD, you know, very recently has had more success with Rock M, but it's been a CUDA world. And that has meant running, you know, x86 hardware, period. And the power of the Apple Silicon processors and the unified memory coalesced in such a way that you started to see even in the M1, you know, Pro and Max era of the very early days of the local models of people saying, hey, okay, well, what kind of tooling and optimizations can we do that doesn't include Nvidia? And, and then Apple released a library, MLX, to make optimizing that, that better. But even before that, there were enthusiasts.

Christina Warren [00:07:00]:
I remember noticing just, you know, as a bystander, somebody who's just interested in tech saying, oh, this is interesting that the Mac people are, are now starting to get into this AI stuff, which for the past 15 years that had been an x86, you know, Windows domain, primarily Linux a little bit. But, but even like 2020, 2021, that was still primarily like a Windows domain, and, um, that is not the case anymore.

Leo Laporte [00:07:29]:
What Apple did that was kind of interesting, and earlier it had been done with gaming machines, is something called unified memory. They, instead of having, you know, on a normal PC you have the motherboard has RAM chips, uh, and they're separate from the processor, and then there's a bus, the PCIe bus, or, or an express bus that the processor shuffles stuff into the RAM and out of RAM, and you're limited kind of by the speed of that. It's certainly one of the— it's certainly the fastest bus on the machine, but it's still not as fast as stuff inside the processor. Apple said, you know, what if we put the RAM inside the, the die and connected it in a much speedier connection to the processor?

Jeff Jarvis [00:08:13]:
Which is what the new Chips are going to do on it.

Leo Laporte [00:08:18]:
Well, Nvidia is doing it now. Everybody's actually doing it now because everybody's doing it now. But Apple was one of the first— it was certainly the first desktop, consumer desktop to do that. And that's what AI aficionados said. Oh, well, if it's unified memory, it's faster. Maybe we could run our models in there. I wish I had been— I think you too, Christina, a little more prescient about this and ordered Much more RAM on my Macs when it was cheap last year.

Christina Warren [00:08:45]:
Yeah, no, me too. I mean, I, I will say I was prescient in that I got an M3 Max with 128 gigs of RAM as soon as it came out.

Leo Laporte [00:08:55]:
That is a capable AI machine.

Christina Warren [00:08:57]:
That's a good machine. It still is.

Leo Laporte [00:08:58]:
Yeah.

Christina Warren [00:08:59]:
But I now look at it and I'm like, okay, you know, this machine, which is still very capable, still has very fast unified memory. The chip is still very good. It is not an M5. It is not going to be like whatever an M6 is. And there's a part of me that goes, OK, you know, I waited 2 years between the M1 and the M3. Maybe I would have been, you know, wishing if I'd had the foresight, maybe I should have, you know, gone for an M4, traded in, you know, last year for an M5 or something, just where we are, just because prices are so nuts.

Jeff Jarvis [00:09:29]:
But yeah, I—

Leo Laporte [00:09:29]:
This would be a good time probably to buy an M3 Ultra. Because, you know, there'll be a few on the market, I imagine.

Christina Warren [00:09:36]:
If you're on market, if you could get one. I mean, I think there's probably people in the Bay Area who could get M3 Ultra with 128.

Leo Laporte [00:09:42]:
You would have a machine that's close to what these Nvidia Sparks are for.

Jeff Jarvis [00:09:46]:
And you could arbitrage it on eBay without doubt. Exactly.

Christina Warren [00:09:50]:
So absolutely.

Leo Laporte [00:09:51]:
But until Monday, until Monday, people who were buying Mac Minis for OpenCLaw and other things really weren't doing it for local models so much as it's a cheap, low-power box you can stick on the side and be running Claw on it ongoing, ongoing, and you'd still be using cloud models. And so that's the shift. This is a big shift in the last 6 months. This is going to— and I thought this would be the year of agents and agentic AI. And I guess to some degree it still is because of OpenAI.

Jeff Jarvis [00:10:22]:
It's related, right? Yeah.

Christina Warren [00:10:24]:
Yeah.

Leo Laporte [00:10:24]:
No, it's not related. It's different. I think this is the year of local models, and I think we'll look back on it and say this was the year local models started to bridge the gap between the very fast, high-end frontier models. And of course, if you're going to do a local model, you need the hardware to run it on. Nvidia, 6 months, 8 months ago, announced these DGX Sparks, which also have unified memory. They're actually very similar to a Mac. They look like a Mac size. mini size.

Leo Laporte [00:10:56]:
They have unified memory. They have ARM processors. They're running Ubuntu Linux, a special NVIDIA version of that. And more importantly, the GPUs in there are based on the NVIDIA CUDA platform, which—

Christina Warren [00:11:11]:
Exactly.

Leo Laporte [00:11:12]:
Until recently, as you were saying, Christina, almost all the models wanted CUDA. It's only recently that Apple's MLX competitor has started to become a second platform.

Christina Warren [00:11:26]:
Yeah. And I would have to look at the data. I don't know this definitively, but I would not be surprised if at this point MLX doesn't have more broad support than ROCm.

Leo Laporte [00:11:37]:
Which is AMD's version. You and I have a Framework desktop.

Christina Warren [00:11:41]:
We do. A desktop, which is really a laptop chip, but it also has unified memory. It's not as fast. It's much slower bandwidth, but you can still do a lot with it. And that was, again, that was a machine that I bought. Also, I will disclose, an investor in Framework, not a lot of money, but I am an investor. But I bought this with my own money. But I last year bought the Framework desktop and I got it, you know, with 128 gigs of RAM.

Christina Warren [00:12:08]:
Again, but the purpose was I was like, okay, I want to be able to play with local models as these advance and get better and better.

Leo Laporte [00:12:15]:
That's how Framework was pitching it. They knew that it was going to be a local AI machine.

Christina Warren [00:12:19]:
Yeah, they did. I don't think that they— they were very smart to do that. And I think to presented that way. I mean, I think it was unfortunate that then the timing of the prices of components and everything, you know, makes it very hard for a small company especially. But, you know, like last summer, which was when I got my Framework desktop, I'm sure that's when you got yours too.

Leo Laporte [00:12:38]:
Well, that's when I ordered it, I think.

Christina Warren [00:12:40]:
Yeah, I pre-ordered mine. I think it arrived last July and it was, you know, it was $2,000, I think was, you know, before tax was the MSRP basically for the 128 gig version if you brought your own hard drive or whatever. Now, I don't even know what they sell for. But that— but that was kind of taking— AMD was smart and kind of looked at, you know, OK, we want to also kind of unify our memory stack. And then they also have been doing work, you know, much like Apple has, where, as you said, CUDA has been the standard for so long. And AMD has made very capable graphics chips that could even be used for AI work and for inference for a long time. There just has never been the tooling support around it. And it seems like this moment has finally been the thing that's made them really have to invest themselves.

Christina Warren [00:13:28]:
And then the community has invested as well because, you know, what happens with these local models is that you kind of look around the house. If you can't afford to buy an RTX Spark Bark like you or you can't afford to get one of the new Mac Studios, you want to look around your house and you go, okay, what type of GPU do I have? How many do I have? What can I run on these things? And historically, it has been easier to get an AMD graphics card that had a lot of memory on it less expensively than you could an NVIDIA equivalent. And so, you know, people were wanting to figure out, okay, what can I do with my, you know, AMD card potentially? But the tooling just hadn't been there. And so that's, that's always been kind of the push and pull with these things. But now I think we're at this point where hardware prices notwithstanding, the tooling has caught up. And you can get the advantage of something like a Mac Studio. Jason Snell pointed this out on MacBreak Weekly to us yesterday, is that the power consumption is so much lower. So you're not only are you getting very strong performance that would be very similar to buying a number of, you know, 5090s and stringing them together, but the power that is then being, you know, run is it's not going to be like a MacBook Air or anything, but you could you know, plug multiple units into a power strip, Jason Snell was saying, and then plug that into the wall.

Christina Warren [00:14:46]:
And that's going to be very different from if you have multiple 5090s, um, needing, you know, 1,000, 1,500-watt power supplies, you know, to power the whole thing. And then your electricity bill is going to go up.

Leo Laporte [00:14:59]:
Yeah, that's, that's not something to ignore. I think people kind of forget about that. But if you got some of these fancy RTX NVIDIA cards and plugged them into a giant machine to run it, you could be using hundreds, maybe even 1,000 watts of power, and suddenly your electricity bill is hundreds of dollars. So that's one advantage that Nvidia has with these DGX Sparks. They're very much like a Mac in terms of power consumption.

Christina Warren [00:15:23]:
Yes.

Leo Laporte [00:15:23]:
And that's ARM. Apple— Jason got a briefing on Monday from Apple, and Apple showed him 4 Mac Studios connected together with Thunderbolt 5. And then they followed the power cable down into the regular wall power cable. That's all it needed for, for 4 of them. So they, they know, they know absolutely that that's one of the, uh, you know, the value propositions of these Mac Studios. Do you think though, Christina, so do I feel like a chump or what? Because, because like 2 weeks ago I bought 2, not 1 but 2, almost $10,000 worth of NVIDIA DGX Sparks. Hoping to run some of these exciting new local models that I knew were starting to come out, like DeepSeek V4 Flash. And it runs very well on them.

Leo Laporte [00:16:13]:
And I have been using it full-time since they came. But was I a chump? Should I sell them while I still can and buy a Mac?

Christina Warren [00:16:20]:
I don't know. I think that's hard to say. I don't think you're a chump. I don't think—

Leo Laporte [00:16:25]:
Well, it's nice of you to say that.

Christina Warren [00:16:27]:
No, genuinely, I don't think you're a chump. I don't think that you're maybe as far ahead as you were a week ago, right? I think that's the real thing. And that's what we're going to have to say. One thing I will say to you though is that your 2 machines, both of them are 512, correct?

Jeff Jarvis [00:16:44]:
No.

Leo Laporte [00:16:44]:
So the DGX Sparks are 128 each.

Christina Warren [00:16:47]:
Oh, they're 128 each.

Leo Laporte [00:16:48]:
And they have a 200 gigabit Ethernet connector, ConnectX connector, so that they are able to pair and run as TP2.

Christina Warren [00:16:58]:
Okay, that's the difference. So that's gonna be—

Leo Laporte [00:17:00]:
It's not as fast as unified memory, but it is 256.

Christina Warren [00:17:04]:
It is 256. And I will say, I think if you're, If you were pairing them together, because Apple's configuration, what most people are using, they're using— it's from a company called XO Labs, I believe, which basically has created like a, you know, a connector of sorts. There is performance loss there.

Leo Laporte [00:17:21]:
And Apple even admits this. They say 4 of them is somewhat like 3 of them.

Christina Warren [00:17:29]:
Alex, I can't think of his last name, who does, has a great YouTube channel, might be Ziskin, might be his last name, does a lot of local AI stuff, especially on Macs and on other machines. And he's, you know, done a lot of these. Yeah, Ziskin is correct. He's done a bunch of this sort of work on local models and pairing a bunch of, you know, Mac Minis or Mac Studios together, and then comparing it with GPUs and other mini boxes and mini PCs and whatnot. His channel is really, really great. And so he's kind of shown the differences in what you lose when you connect all those together. And so I think that the Spark, even though it might not have the same unified memory, 2 of them together, you might wind up getting better performance than if you were to pair 2 Studios together. But I don't know.

Jeff Jarvis [00:18:19]:
This is gonna be one of those things we just won't know until we try it out. Well, 2 Studios would be $20,000.

Leo Laporte [00:18:24]:
Well, so here's— I could sell the other kidney. So what I was thinking— what I'm thinking is I'm gonna wait and see, because—

Jeff Jarvis [00:18:35]:
Lisa, where are you? Lisa?

Leo Laporte [00:18:37]:
September 22nd, the 256GB versions of Mac. By the way, the Minis, as nice as they are, you can run, you know, Claw on them fine, but they're not— those are not local.

Jeff Jarvis [00:18:49]:
That's what I was trying to hear. They're not relevant to the local discussion.

Leo Laporte [00:18:51]:
Not really. I think they tap out at 64GB, right? Yeah, you can run small models.

Christina Warren [00:18:55]:
It's models which can still accomplish a lot depending on what you want to do.

Leo Laporte [00:18:58]:
Oh yeah, I'm running— I have my M4 Max, uh, 64GB Mini, and it's running QWin. That's— I'm running the obliterated version of it, which I'll show people later in the show, so I can, I can plan all sorts of nuclear disasters with it, but, uh, and ask it about Chinese dissidents at the same time. But, but, uh, you know, it's okay, but it's not— it's not— it's, it's a, it's a useful model. But it's not what I'm looking to do is to really run my agent on a very competent local model and only use the frontier models— Anthropic, OpenAI, Grok— for the most challenging coding stuff. Now, we're going to get to this in a second, but that may even be possible soon locally. And we'll talk about 2 new models that came out this week. But I want to finish up the Mac conversation. So the, the strategy could be to wait September 22nd.

Leo Laporte [00:19:55]:
These 256 gigs will come out. I'm sure a lot of people will be banging on these and trying to figure out, you know, what can you do? How are they comparable? Is one 256 gigabyte Mac Studio with an M5 Ultra comparable to 2 DGX Sparks? Because, because that's roughly the same price. It's the same amount of RAM.

Christina Warren [00:20:16]:
There you will, you will get a lot more storage on your DGX Spark. I think that—

Leo Laporte [00:20:22]:
I have 8 terabytes. Yeah.

Christina Warren [00:20:23]:
Because it's right. I was gonna say you'll get—

Leo Laporte [00:20:25]:
I'm not so worried about storage.

Christina Warren [00:20:27]:
I mean, well, I don't run— I mean, I don't know. I think that storage should be a consideration. If you're talking about running models this size, A, they're going to be very large. B, you might need swap, right?

Jeff Jarvis [00:20:38]:
Right.

Christina Warren [00:20:38]:
So, so, you know, I'm not saying that you need to buy 4 terabytes on a Mac. But I am saying that that's the thing that if anybody's trying to do like a value comparison, you need to account for that.

Leo Laporte [00:20:50]:
That's part of it, sure. And yeah, you know, I can add more Sparks and connect them in that same high-speed fashion, which is still probably a little bit faster than Thunderbolt 5.

Jeff Jarvis [00:21:00]:
And just having CUDA, an advantage.

Leo Laporte [00:21:02]:
And that's the big difference. It's CUDA on the NVIDIA stuff, it's MLX Metal on the Mac stuff. And so that's what I'm gonna watch with interest next month. And then that'll give me— because Apple's not going to ship the 512 or even let you order it till late October. I figure I have 30 days to watch and decide, and then sell the kidney.

Christina Warren [00:21:23]:
Or sell the SPARKs, right? Because, I mean, there are plenty of people—

Leo Laporte [00:21:26]:
And that's really what I would do, is I would actually sell the SPARKs.

Christina Warren [00:21:29]:
Yeah.

Leo Laporte [00:21:30]:
That'd get me halfway there.

Christina Warren [00:21:31]:
Because I think that people would absolutely— I think you could, you know, if anything, you might even be able to profit off of it. I know that's what—

Jeff Jarvis [00:21:39]:
Oh, yeah.

Christina Warren [00:21:39]:
People are doing it for that reason. But, you know, because the SPARK was announced, I mean, and this is an interesting thing, too, it just shows how much time passes. Is the Spark was first announced, I believe that it was in March or April of 2025. And so it took it, you know, well over a year to become like actively available as a product. And so that doesn't take anything away or deter anything from it. But when you think about just how much even Apple's roadmap has changed in, in 18 months and Nvidia's roadmap for that matter too. It's just something to think about. So, you know, it's one of those things where when they announced the Spark devices, there was a lot of interest.

Christina Warren [00:22:18]:
And I definitely like— I signed up for the interest form. I was like, oh, I want one of these. And then they didn't really do much with them. And then it took them until recently to basically make it so you could, you know, order them directly from their website and, and not have to go through anything else. And obviously the price has been adjusted. to account for all the component price increases, too. But, you know, Nvidia has been busy selling all of their, their GPUs to the data centers.

Leo Laporte [00:22:49]:
I also have to think, yeah, I mean, this is all complicated, as you said, by Ramageddon and so forth. But I also have to think that Nvidia looked at this Apple announcement on Tuesday and said— by the way, Tuesday was yesterday. It feels like yesterday. And said, gee, maybe it's time for Spark 2. And, uh, so it is very much complicated by their ability to get— I mean, they, they get the chips, but maybe the RAM is— I don't know. I don't know. And I don't know where Apple's getting the chips. Have they been stockpiling these all year? I don't know.

Jeff Jarvis [00:23:22]:
Maybe.

Leo Laporte [00:23:23]:
It's a great mystery. Uh, it shocked me. I think it shocked everybody when they said, yeah, we're gonna make a 512GB M 5 Ultra, which would be probably, again, $20,000 computer. I don't even, by the way, plan to use it as a Mac. It's going to be running headless. I'll still use this framework.

Christina Warren [00:23:43]:
I was going to say that that's the other thing to kind of think about. You know, you mentioned at the beginning, like Nvidia is running like a custom version of Ubuntu. And, and so they're able to really customize everything specifically for this purpose. And there's not— I obviously love macOS. It's my favorite operating system. But, you know, Apple hasn't made a proper server version of their operating system in a dozen years. And they're— yes, you can run them headless and there are more things you can do with that. But there's still going to be overhead that's going to be involved with that because it's being expected to be used, at least the way that they sell them, as a consumer desktop operating system, which is, you know, I'm sure that they have their own custom you know, kernels and other things for what they're running in the cloud.

Christina Warren [00:24:26]:
But for what we consumers get, it is still going to be, you know, a Mac, even if you're remoting into it, even if you're not connecting it to a display. Whereas the Spark and devices like it are much more purposely built, you know, to be—

Jeff Jarvis [00:24:42]:
I want to ask you both— sorry, Christina, can I ask you both this kind of a strategic question which I raised in our chat before the show? with Leo. It seems to me 2 things. One, this whole local model thing is a bit of a surprise to the hardware market in terms of how quickly it's grown and all the possibilities. And then point 2 is that Apple was being criticized right and left and quite properly for not having an AI strategy. And, and oddly, I said before we got on, it seems to me that this could be its AI strategy is, right, well, let me talk about local models.

Leo Laporte [00:25:16]:
Let me give them some credit. One of the reasons Apple's ahead in some respects is They wanted to put AI in their phones. And they even talked about local models on their phones years ago. And so they had already started putting neural, what they call neural processors into their A whatever, 17, 16, A15, their early phone chips. So they were already doing this and it wasn't such a stretch for them to say, oh, well, you know, maybe there's a market here.

Jeff Jarvis [00:25:47]:
But the possible volume, had to be a surprise.

Leo Laporte [00:25:50]:
Yeah, well, everything's a surprise to everybody at this point. I mean, I don't think so.

Jeff Jarvis [00:25:54]:
Where does this— where does this come out? I mean, so I put in the rundown that by one account from Versal, in terms— and I'm not sure exactly where they get the data or how they get it, but they said that open weight token calls are now more than closed weight calls. It's grown like crazy. It's going to keep growing. Oh, The options are there. So it's huge. So where does this go? First, you know, for small businesses and such, but then everybody else on phones. Where does this go from the heart? I didn't think the hardware was going to be the gating factor. And now the hardware kind of is to growth here.

Jeff Jarvis [00:26:31]:
Where do you think this goes? Does Dell play in?

Leo Laporte [00:26:35]:
Where does Dell make some sparks, by the way? Dell?

Jeff Jarvis [00:26:37]:
Yes, that's right. That's right. There are multiple sparks. I forgot that.

Leo Laporte [00:26:39]:
Sure. Acer, Dell, Lenovo.

Jeff Jarvis [00:26:41]:
They're easier to get, I would imagine.

Leo Laporte [00:26:43]:
I don't know. It's the same.

Christina Warren [00:26:46]:
Yeah, I think it's probably the same.

Leo Laporte [00:26:48]:
But the prices have gone up on all of them.

Christina Warren [00:26:50]:
Yeah. So I don't know. I don't know what you think, Leo. I think that is an interesting thing because I think that we have 2 coalescing things happening with local models where you have on the— or open weight models. Let's reframe it that way. On the one hand, you do finally have, assuming you can get the hardware, you now have models that are good enough that you can do a lot of really great tasks on a local machine, you can also do reinforcement training and you can do inference work. You can do real work on these devices. I think on the other hand, you do also have this coalescence of you have businesses that might not be able to maybe do the initial outlay of cash for, you know, $50,000 or whatever for local machines to run these on.

Christina Warren [00:27:36]:
for it in their own network, who are looking at costs of frontier models and going, okay, but we can spend— we would rather spend this for various reasons on an open weight model that is maybe hosted in a data center somewhere. So we're still using a cloud, it's just going to not be necessarily using—

Jeff Jarvis [00:27:56]:
We have control.

Christina Warren [00:27:57]:
Opus. Exactly. But we can have more control, we can fine-tune it ourselves, We can make modifications. We can ensure that all of our data is going to be protected. So I think there are almost like 2 similar stories where on the one hand, yes, if you can do it all locally in your own data centers or own, you know, business operation or however you're wanting to run, you can do that. But on the other end, the, the open weight model story is not limited to just simply self-hosting because these are also being hosted by all the hyperscalers too. And often You know, the, the amount of money that it costs to run these models for a variety of reasons is much lower than the frontier.

Jeff Jarvis [00:28:36]:
Should somebody else have bought OpenRouter than Stripe? I mean, is that— is that because it seems that's an opportunity?

Leo Laporte [00:28:43]:
No, this is— so you asked me this last week when we had the news that Stripe was going to spend, what was it, $7 or $8 billion to buy OpenRouter. And I think Having thought about it now for a week, I have a better answer for you, which is Stripe, which is all about dollar transactions, suddenly says, wait a minute, tokens are another kind of currency, and we want to be in the token business. It's a very— it's not quite a pivot. It's a little bit of a pivot, but I think it's a very smart thing.

Jeff Jarvis [00:29:13]:
We're talking to somebody else, though.

Leo Laporte [00:29:15]:
Well, let me reintroduce Christina. Hold on a second. We're talking to Christina Warren. She is the host of MacBreak Weekly. We love having her on every Tuesday to talk about Apple, but I really wanted to get her on because in her job, both at DeepMind with Google and now Senior Developer Advocate at GitHub, she's definitely dealing with AI. In fact, we talked about the new Copilot desktop app on MacBreak Weekly. That's pretty amazing. And I'm sure at GitHub, one of the great beneficiaries actually of the AI revolution.

Leo Laporte [00:29:45]:
Every Everything is happening on GitHub, including, by the way, all the models that I'm testing and downloading and so forth.

Jeff Jarvis [00:29:53]:
Christina, can you share those numbers you shared with Benito and me before the show?

Christina Warren [00:29:57]:
Yeah, yeah. Let me pull up the numbers. So if you have, you know, if you are a GitHub user, you might have noticed that we have our availability has not been as good as we would like it to be. We have been a little bit busy. And so some of the numbers that we've, we share some numbers in April. These are updated to August. So just to give you an idea of just how much things have grown. So in 2023, we are getting 20 million merged pull requests a month.

Christina Warren [00:30:26]:
As of August 2026, 130 million. So going from, you know, like, you know, 25 to 130 million. Our commits per month have gone from half a billion in 2023 to $2.9 billion in August. And that $2.9 billion is up from where we were. I think it's nearly doubled from where we were in April. The new repositories created per month, it's now— it used to be, you know, 5 million in 2023, 2024. It's now 24 million. So the, the, the growth— and this is expanded across the platform— has just exploded.

Christina Warren [00:31:05]:
And it really started when the frontier models got really good in November. But obviously, as people will use other models too, open weight will affect that as well. But yeah, I mean, when we talk about like AI stuff, like, you know, this is where most people are hosting their code and are, you know, doing a lot of their work. And that certainly includes agents. And we've seen—

Jeff Jarvis [00:31:29]:
Yeah.

Christina Warren [00:31:29]:
Just ridiculous levels of growth that we, you know, I don't think anybody could have prepared for, even if you'd done all the, you know, analyst planning, forecasting. Yeah, to see it just literally 5, 6x in some cases.

Leo Laporte [00:31:43]:
In some cases, 4 half-assed repos to something like 30 within minutes, right? And a lot of that's very useful. I go to there to see what, you know, people put their skills there, people, uh, put their code there. It's a very good way to see what other people are doing. And this is We're at such early days here with AI that I think we're all learning from each other. And GitHub is a great kind of place for people to get together and post stuff and learn from stuff.

Christina Warren [00:32:11]:
So I'll just share this one last stat. Since April, monthly commits have grown from 1.4 billion a month. This is since April to 2.9 billion.

Leo Laporte [00:32:20]:
More than double.

Christina Warren [00:32:20]:
Literally, the number of commits have doubled since April, which is— we're not even— it's not even September.

Leo Laporte [00:32:26]:
It's only gonna get worse.

Christina Warren [00:32:28]:
It's only gonna get worse.

Leo Laporte [00:32:29]:
Or better, I guess.

Christina Warren [00:32:30]:
Well, it depends how you look at it, right? But just—

Jeff Jarvis [00:32:32]:
Depends on what your assessment, right?

Christina Warren [00:32:35]:
More code is being written than ever before, regardless of who you are. More people are starting to write code, whether it's human-initiated or agent-initiated, doesn't really matter at this point. But like, more and more of that is happening. Like you were saying, Leo, like went from having, okay, I have a few, you know, side projects and things. Now it's like, I have an idea. And now whether it's using a local model or I'm using something else, frontier model, you know, cloud, whatever, I can iterate and play around with that idea and I can deploy it. And that's the sort of thing that, you know, used to be— take a lot more consideration and now it's just a prompt.

Leo Laporte [00:33:10]:
Yeah, I've used GitHub for a long time. I'm a fan and a proud member. But a lot of people— Anthony Nielsen is saying, I don't really understand GitHub, but I'm but I'm using it because the agents understand it. The agents know that the agents— it's just, it's a natural thing.

Christina Warren [00:33:27]:
I was going to say, it is, it is. That's the thing, right? Because if you're building something on your computer and going, okay, I want to be able to deploy this, I want to test this out, I want to be able to have version control, which of course you want to do, you know, GitHub is going to be the obvious choice because we have a very good free plan and it's been around for years.

Leo Laporte [00:33:42]:
The agents suggest it.

Christina Warren [00:33:43]:
The agents do suggest it, right? Because they've, you know, scanned the internet too. And obviously, all the models have been trained on the public code that has been on GitHub. And, and, you know, that was the genesis of the, you know, kind of original kind of implementation of GitHub Copilot back in 2021. It was a joint project between us and OpenAI. They had a model which was called Codex at the time. Obviously, Codex is now the name of something else. And it was, you know, we worked with them to train a version of, I think it was GPT-3.5 on the code that was hosted, the public code that was hosted on GitHub. And our idea was, okay, well, we have this model now.

Christina Warren [00:34:25]:
What if we try to do, you know, some sort of like autocomplete and suggestions? And so that was the early version of GitHub Copilot back in 2021. And then that evolved to more of a chat interface, you know, after, you know, the emergence and and buoyancy of things like ChatGPT. And then when we entered, when the models got better and better, and got things like MCP and tooling involved, then you started to get into like the agentic coding era where we've been for the last couple of years where now, like, it can, it's not just autocomplete anymore. You know, it will fully create things for you. And then, like you said, because all these models have been trained, they're going to suggest Okay, well, where do I deploy my code? Where do I host my code? You know, where do I, you know, run my different dev environments? And GitHub is going to come up.

Jeff Jarvis [00:35:15]:
Stay there for a second. How? Because I'm fascinated with Leo, with his agents, talking to his agents and all of that. How did the agents learn that GitHub was the place? What was the process by which that happened? Because it's a fascinating business model to understand that you find yourself at the center of all this. It propagated on its own to an extent. But where do you think the root of that knowledge was for agents in general?

Christina Warren [00:35:40]:
I think it was the fact that we have always had a really robust free plan and you can have free private repositories for just a regular user. You're given a certain amount of build minutes every month known as GitHub Actions if you want to build more than that. And what that means is that basically in the cloud it will deploy something you can You can run local actions too, but you get a certain number of minutes where your project will build in the background and run it on our cloud servers and make sure that something builds correctly and you can kind of get like a report back if it works or if it doesn't. You can deploy static websites to our Pages platform too, but we've had these tools built in and we've had a really robust free plan that frankly no one else has really had. And that's part of why GitHub for 18 years has been kind of, you know, it kind of replaced SourceForge and things like that, is like the place where if you're going to build a software project that you want to share with other people, that's where you put it. And at first, you know, the emphasis was very much on it's going to be public. But private repos have been a thing for a very, very long time. And the fact that, you know, you can get them for free, I think the agent know that, you know, as they're scanning the web and documentation and anything else.

Christina Warren [00:36:54]:
The fact that so much code was already on GitHub also added to that. The fact that we have had, you know, very well-documented, you know, like processes of how to use our SDK and our APIs, because so many businesses and individual developers have used us, that also means that the agents will know how to use it too. So I think it was a coalescence of all of those efforts that really said, okay, well, why are we recommending GitHub versus someone else? It's just because that's A, where all the code is, B, the infrastructure and the documentation was already in place, and C, like, that's what it was trained on.

Jeff Jarvis [00:37:30]:
And so I think that's a model to whiny media companies, news companies and such that, well, I don't get links anymore. Well, you're not useful. If you put yourself in a position where you were useful to these agents, where you had repositories of data and information, then they'll figure it out. It's a really interesting new model for where that goes.

Leo Laporte [00:37:50]:
I was surprised. One of the very first things, maybe the first thing I ever wrote with Claude Code was an RSS reader. And Claude Code has a GitHub MCP server, as does my Hermes agent. Everything does. And I said, yeah, yeah, go ahead, put it up there. And then a minute later it says, okay, so I have a Linux, Mac, and Windows version of your program running. I said, I don't even know how that happened. It goes, this is actually one of the great things.

Leo Laporte [00:38:17]:
We don't talk about this much with these AIs and LLMs is they're great for learning. I mean, I knew about GitHub and I knew what CI/CD was, but I'd never done continuous integration, continuous deployment. I'd never done it because I'm not in enterprise. And so it did it. And that's how I learned about it. It was fantastic. I thought, oh. You can build these binaries and I could just download them on my— all my different machines.

Leo Laporte [00:38:44]:
I was blown away. So yeah, this is, this is why GitHub is a natural and a lot of people use it. We're talking to Christina Warren, who is Developer Relations at GitHub. She's in a good seat right now. We got to take a break, Christina, and I haven't gotten to the local models. Do you have some time?

Christina Warren [00:38:58]:
Yeah, of course.

Leo Laporte [00:38:59]:
Do you need to run?

Christina Warren [00:39:00]:
No, I'm good actually.

Leo Laporte [00:39:01]:
I want to talk about OxAlpha. We now know what OxAlpha is.

Jeff Jarvis [00:39:07]:
And who made it.

Leo Laporte [00:39:08]:
And who made it? Uh, we— the secret is out. And, uh, we had 2 this morning, not one but 2. At 6 AM, I'm up early because I know Quen 3.8 Flash is coming out. And then another surprise, OxAlpha came out and admitted what it was and said, and by the way, you want to try it? You want to download it? Because you can. So we're going to talk about that when we come back. It's so great to have you, Christina.

Christina Warren [00:39:34]:
Thank you.

Leo Laporte [00:39:34]:
I'm sorry Paris isn't here. Jeff, you and I are nitwits. She told us last week.

Jeff Jarvis [00:39:40]:
And you blessed it. It's my fault. It's my fault. I'm sorry. No, it's all our fault.

Leo Laporte [00:39:45]:
You don't have to throw yourselves on the sword. She told us, she said, there's a going-away party for somebody at Consumer Reports. I'd like to go. And I said, well, please, by all means, go to the party. You don't want to miss that.

Jeff Jarvis [00:39:56]:
But you're lucky because— She said, I felt bad about being gone. And Leo said, no, no, no, no, no. Priority.

Leo Laporte [00:40:00]:
Go, go. Yeah, I urged her to go and then forgot all about it.

Jeff Jarvis [00:40:03]:
Sorry.

Leo Laporte [00:40:05]:
But hey, we're glad Christina's here. I'm glad you're here.

Jeff Jarvis [00:40:08]:
Hold on, hold on, hold on.

Leo Laporte [00:40:09]:
I have to take a break. Yes, couldn't be a better time. Actually, it's like I planned it, but I didn't.

Jeff Jarvis [00:40:15]:
Brilliant plan.

Leo Laporte [00:40:17]:
We'll have more in just a little bit with Christina Warren, Jeff Jarvis, and intelligent machines. So we were talking about this all last week. It was all the rage on X.com. 0xAlpha, a stealth model that was being offered free, unlimited, uh, on OpenRouter. Noose, uh, was offering this. Uh, OpenRouter said we have a capacity for 100 trillion tokens, and everybody's going, well, who could this be? It's not the first time. I think, um, Kimmy did this. Uh, they had a stealth model that they really— yeah, some time ago.

Jeff Jarvis [00:40:58]:
I forgot that.

Leo Laporte [00:40:59]:
Uh, so it's not the first time it's happened. You know, there's a lot of speculation. OpenRouter said no, they're not, they're not training on your prompts, they're not saving your data. But there's a lot of speculation that whoever's doing this might be using this to improve their model before they ship it for real. I don't know if that's the case or not. Uh, people tried all sorts of techniques to figure it out. This was kind of funny. The folks at Google pretended that it was theirs.

Leo Laporte [00:41:26]:
Did you see that, Christina? They tweeted, the friends we met along the way. And it was like, dudes, it isn't Gemini. Gemini, it isn't. Um, people had all sorts of ideas, but more and more they converged on ZAI, whose model, uh, 5.2, GLM 5.2, was very good. They then released 5.3, which was, they said, the same model, the same LLM, But post-training was improved, and it was, it was very good. I actually am using it, uh, in my engenic workflow, uh, for the coding I'm doing. I mean, it's, it's one of my auditors. It's very good.

Leo Laporte [00:42:03]:
Uh, and then I played with OxAlpha, and I was very impressed. Did you play with it at all, Christina?

Christina Warren [00:42:08]:
Oh, I have a little bit. Yeah, I haven't had a, a chance to do, um, uh, a, a ton with it, but I have a little bit.

Leo Laporte [00:42:15]:
I benchmarked it. I, uh, okay, I was really curious, so I had, uh, I had DeepSeek Flash, my agent Quicksilver, write some tests. Actually, I've been using these tests for a while in a skill it calls Bake Off, where it pits 2 models against one another. And only one emerges. It has 7 very basic logic tests. You know, the doctor's son is on the table and the doctor says, I can't operate, that kind of thing. These word problems you've all heard. And then it had a much larger corpus of tests based on work that we've done, agentic-style tests to see if it can do good agentic work.

Leo Laporte [00:42:55]:
And I was getting a lot of ties with some of these frontier models. So I said, okay, and now we're gonna have Fable, which everybody agrees is the best model out there. We're gonna have Fable design 7 of the hardest coding problems it can come up with. And I pitted these 7 problems, I pitted Grok and OxAlpha on these, and they both aced them. In fact, they got all 7 right. They found a bug in Fable's implementation and corrected Fable. They said, Fable, you're wrong.

Jeff Jarvis [00:43:30]:
What?

Leo Laporte [00:43:31]:
So I was kind of impressed. I thought, man, if this is as good at coding as Grok 4.6, there's something special. going on. People, you know, other people are interested in how it does graphics, game design, web design, and stuff, but for me coding is kind of the stuff that I really want a frontier model for is coding. And I was very impressed. I have the preliminary— well, first of all, let's— So, okay, so we're all thinking about OX Alpha. I think that the impression was it's gonna be a week that you can use it for free and then we'll reveal And the hope was it was one of these companies that is doing, uh, some people thought it might be NVIDIA that's doing open models. They have an open weight Neutron you can use.

Leo Laporte [00:44:17]:
Uh, and then of course we've been playing with, uh, Quen from Alibaba. Quen has been very, very good. I'm running Quen, uh, 3527B on my, uh, 3090 system and it's very good. That's the one that I've obliterated, removed the censorship from. It's very, very good. Um, I did give OX Alpha and, uh, Quen these tests, uh, and then this morning at 6 AM, I got up early because I knew this was going to happen, uh, Alibaba announced the release of the open weights of Quen.

Jeff Jarvis [00:44:53]:
Hey, real quickly, can you give me your screen?

Leo Laporte [00:44:56]:
Oh, I, I don't have yours. You don't have my screen. That would make it hard for you to see what I'm looking at right now. There you go.

Jeff Jarvis [00:45:10]:
And now Christina knows because she's here all the time, you have to click on meeting.

Leo Laporte [00:45:14]:
Oh yeah, it's— don't look at my camera. So 6 AM, we're all up early. Everybody, all the YouTubers are on live streams installing 3.8. Everybody's going crazy, going crazy. I installed— I downloaded immediately and installed it. It runs in FP4. Actually, it ran in FP8 on the Sparks, which is the higher resolution version, but I then downgraded it to FP4 because it couldn't finish the tough coding ones. And so I still haven't benchmarked FP4 because...

Leo Laporte [00:45:51]:
ZAAI said, oh yeah, Aux Alpha, that's GLM 5.3 Flash. Were you surprised, Christina?

Christina Warren [00:46:01]:
I was. I mean, the fact that we got not one but two.

Leo Laporte [00:46:05]:
And both of them have open weights out now.

Christina Warren [00:46:07]:
And both of them open weights. Incredible. And so I would be curious to know, just from a process, was this always planned? Did they move releasing the weights up based on, you know, what, what rumors they heard, you know, through—

Leo Laporte [00:46:20]:
Because they're both Chinese companies, right? There's Alibaba and there's Zhipu, which is Z.ai.

Christina Warren [00:46:27]:
Yeah. And so I wonder, you know, if they heard things in the ether, you know, because they see and hear the same rumors that we do. And maybe if that, you know, kind of forced the timeline up, if it just happened to be coincidental. I mean, and this is not uncommon. We've seen this in the frontier space, too, where you'll have, you know, 2 models released in the same week, usually not the same morning. That, that I think is maybe a first, especially in the open weight space. But yeah, because I woke up and I was like, wait, what? You know, we have 2 of them. Okay.

Leo Laporte [00:46:54]:
Christmas Day.

Christina Warren [00:46:56]:
It is Christmas Day. And like, you know, I'm like, all right, I'm gonna have to get some time to, to run some evals on my, on my machines that are capable of doing some of this. But I'm just really more just kind of following along with everyone else. Because as soon as they come out, they go on Hugging Face. And everybody's trying to kind of figure out, okay, what can we do with this? Who has the compute to really see what's capable? And it's super fun.

Leo Laporte [00:47:19]:
There are a number of labs that specialize in immediately making quants and variations on this. They all got to work immediately. Hugging Face had the initial weights from the official releases, Alibaba and Z.AI. people started working on it. There's also when— and you made this point yesterday on MacBreak Weekly— that while it's not so hard to get this stuff working on a Mac, it's a little bit trickier on a Spark. I mean, this is what we have AI for. I just, I point, I say, hey, get this working.

Jeff Jarvis [00:47:54]:
Yes.

Leo Laporte [00:47:55]:
I'll be back. And so, but, but I downloaded several different versions that wouldn't run, and then there's issues, and people said, oh no, you have to change this switch. There's also a variety of platforms. There's VLLM, there's Slang, uh, there's another one, TokenSomething. And so there's, there's a lot of variables and a lot of switches and a lot of messing around. Uh, I did let my AI work on this and fuss with it, uh, and I do have some early benchmark results, and I am Encouraged. These are the earliest results from the— this is the local test compared to the cloud test. So 0xAlpha on the cloud passed the 22 agentic— 22 of the 30 agentic was weak on 5 of its answers.

Leo Laporte [00:48:50]:
Not wrong, just the reasoning wasn't right, because I'm really curious also about how good their reasoning is. And then completely failed 3 of And it did that in 1,200 seconds.

Jeff Jarvis [00:48:59]:
Yeah.

Leo Laporte [00:49:00]:
Uh, so I don't know what that is, 6 minutes, right? No, 60 minutes. Um, same model now on an FP4 quant, NVFP4. So instead of the— I don't know if it's 16-bit on the cloud, probably was— we reduce it down to 4. Actually did better. It failed one more, but it was weak on one less. Time was a little longer, but not much longer. Quen, not so good. And Quen is, by the way, on an FP8.

Leo Laporte [00:49:31]:
Passed 20, week 6, failed 4. Now I'm still waiting for it. By the way, Quen could not do the hard coding problems at all. It just died. It just died. We're right now in the middle of testing GLM. But GLM right now running locally on dual Sparks is looking very close to the performance of OX Alpha running in the cloud. If that's the case, this will be a watershed.

Leo Laporte [00:50:04]:
I will not feel like a chump. No, because I will have something that's very competent.

Jeff Jarvis [00:50:09]:
If you're inside OpenAI and Anthropic this week, how are you feeling?

Leo Laporte [00:50:14]:
What do you think? Christina, you've been inside these.

Christina Warren [00:50:18]:
I have been inside.

Jeff Jarvis [00:50:19]:
Or DeepMind too.

Christina Warren [00:50:21]:
Well, I mean, and this is where it's interesting, right? Like DeepMind, obviously Google has their own open weight strategy, Gemma, which is— they are not on the same level as, you know, what we're seeing from the Chinese labs.

Leo Laporte [00:50:32]:
Although they're going to get a big audience when Apple releases this in a couple of weeks.

Christina Warren [00:50:37]:
Well, for sure. But I don't think that Apple is going to be using Gemma. They're going to be, you know, using the probably a variant of Gemini. But, you know, but Google does make, you know, certain, you know, have open weights available and it's just at a different cadence. It's a different thing.

Jeff Jarvis [00:50:51]:
They're arguing that they have billions of uses for it already out there.

Leo Laporte [00:50:54]:
That's—

Jeff Jarvis [00:50:54]:
it's that Google argument.

Christina Warren [00:50:56]:
Yeah, absolutely. And so I think that if you're at those organizations, that's what you're looking at. If you're at the frontier labs, I think that you're seeing— I don't know how much right now the concern is, to be completely honest. People running these things on their own local devices. I think the bigger concern with OpenWait is how much less expensive is this going to be if someone is getting this from a hyperscaler, either through OpenRouter, because, you know, kind of pick your own poison, or, you know, someone like, you know, Amazon or Microsoft Foundry or, you know, Google Cloud or whatever the case may be. I think that is, if I'm Anthropic or OpenAI, that's my bigger concern is are people going to shift our spending from these frontier models, you know, from, from Opus, from, from, you know, Fable, from the SOL, whatever? Are they going to instead say, oh, well, we could get this type of performance on an open weight model, which, yes, it might cost a lab, you know, I think the reports were with Kimi K3 that it was like $2.5 million basically to set it up to run on GPU. And instances in a cloud. Okay, there might be some outlay of cash that has to go into that.

Christina Warren [00:52:04]:
But if that can be run, you know, by a hyperscaler, will those token prices be much less expensive than what we're paying from OpenAI and Anthropic? And, you know, for individuals, you can get a Claude Code Max subscription or you can, you know, get a Codex subscription and you can get a lot of compute for your dollar. But businesses by and large are having to pay, you know, API pricing, which is much more expensive. And so I think that's— if I'm OpenAI or Anthropic, my concern at this moment is more about the capabilities and will businesses opt to, especially if I'm Anthropic, frankly, are they going to opt to use maybe a less capable but still powerful open weight model that can also be customized? and that can maybe be made, you know, to be specific to a specific organization or are— I think that's the bigger concern than are people going to be spending hundreds of thousands of dollars on hardware to run locally for everybody in their business.

Jeff Jarvis [00:53:07]:
So you're seeing that already this week, right? You see Thomson Reuters has created its own legal version.

Christina Warren [00:53:12]:
Yes.

Jeff Jarvis [00:53:13]:
Competing with others. You see AT&T last week said that I think half of their computing is now AI computing is now done on open weight models. It is the Palantir model. Screw the hosted Frontier models. That's all wrong. Karp says go with our saddle and do it this way. So it just seems at a time when a certain company is going to say it has a $30 trillion market, this is dangerous for them.

Leo Laporte [00:53:42]:
Well, we already saw a little bit of this when the Chinese company DeepSeek offered DeepSeek V4 Flash. That's when I first started using it at the end of last month for pennies. And it was very good. And it was— so I think one thing companies are realizing is, look, there's no question you're not going to run Fable on your dual DGX Sparks. We were talking about models now with 10 trillion bytes. That's data center. You're never going to— well, not never. I should never say never.

Leo Laporte [00:54:15]:
But it's unlikely that anytime soon you'd be running that locally. So those companies are going to have a business. But you nailed it, Christina. How much can they charge?

Jeff Jarvis [00:54:24]:
Yes.

Leo Laporte [00:54:25]:
And companies are learning that you don't need the frontier all the time. In fact, you only need a little bit of the time. In fact, that's why I'm so interested in local models. I know I'm still going to use Fable. I'm not saying I won't. But what I am saying is I want to do as much locally as I can. And I think that that is rapidly changing.

Christina Warren [00:54:47]:
Yeah. And I think the other thing too, you made a great point, Jeff. I mean, I think the fact that this is what's powerful about OpenWeights is that you can customize these for very specific tasks. And that also means that you could take a really giant, like, because at this point we're all in this MOE, this model of experts kind of realm where, you know, it can do a lot of different things. But you could have something that was very quantitized, very small, essentially, that is for a specific task. And that might be something that you could even run, you know, on something that doesn't cost, you know, tens of thousands of dollars for hardware, right, that you could run potentially on, you know, depending on what task you're wanting to accomplish, could run on, you know, a MacBook Air or a MacBook Pro or a Surface or whatever the case may be. And I think that's the real power of these types of tools. Is that yes, we get excited about having these giant models that can do everything we want.

Christina Warren [00:55:38]:
And we would love to be able to run all that power on our local devices. But you don't have to do that. You could have, depending on what your task is, a very small localized model that is going to be power efficient, that is also going to be secure, that you could run from personal machines. And that, I think, is also going to be something that's going to be very interesting. It already is interesting to businesses where they're thinking, okay, maybe depending on what type of, you know, worker we have, what they're doing, yeah, we can just have certain tasks are going to automatically by default be running on our local models. And then like what Leo does, if we're needing more intense tasks or some other things, we can, we can go to the cloud-hosted models for that.

Leo Laporte [00:56:20]:
Here's a perfect example. Thomson Reuters built its own AI model on Chinese open-source tech by pumping in— explain who Thomson Reuters— well, we all know Reuters, right? It's the news agency.

Jeff Jarvis [00:56:37]:
Reuters is the news agency. But Thomson has always been a data company. It serves legal, medical, financial, all these companies.

Leo Laporte [00:56:47]:
So over the last 2 years, they ran a training that took all of this stuff— Westlaw, Practical Law, Checkpoint, Reuters— all the content they own, ran it through an OpenWeight Chinese model. I don't know if they said which one it was. And trained a specialized model that's for legal and journalistic. It can— it can— this is what Paris might use to go through all those documents that she's always trying to figure out.

Christina Warren [00:57:19]:
And this is a great thing to do regardless of where it originates from. I mean, I think that for— if you're someone like Thomson Reuters and you have Westlaw and you have all these data sources, yeah, if you have the capability to be able to make your own model that can be customized exactly as you want it, that's going to be really, really compelling, not just internally, but that's also something you can sell to your, to your, your clients.

Leo Laporte [00:57:39]:
Um, because here's another thing they've trained on, on what you've done. So it was trained on Quinn, They, uh, they've got a smaller open weight version. So you can go to Hugging Face right now. In fact, if, if I make the move and replace DeepSeek, and right now it looks like it's going to be GLM-53 Flash on this, these, uh—

Jeff Jarvis [00:58:01]:
You're just so disloyal, Leo.

Leo Laporte [00:58:03]:
Oh, hey, you know, this is, this is all I do. I think Paul Theroux asked me, he said, do you Or no, you asked me this. Do you do any work or are you just swapping models all day? Well, it depends on the day.

Jeff Jarvis [00:58:16]:
That's work for you.

Leo Laporte [00:58:17]:
It's my work. But I'm thinking of putting this Thomson Reuters model on my 3090. It'll probably run on that. It's pretty small. And then have this research partner specialized model designed for journalists and lawyers.

Jeff Jarvis [00:58:29]:
What does it need to run on? How small?

Leo Laporte [00:58:31]:
I don't know yet. I have to look at the— I saw it on Hugging Face. I bookmarked it. I haven't installed By the way, Google today also came out with its Gemini Enterprise release.

Jeff Jarvis [00:58:41]:
Yep. With a whole bunch of other partnerships involved there. So it's a really interesting model of where this heads.

Leo Laporte [00:58:49]:
I, I think we are in a very interesting world of specialized models now.

Jeff Jarvis [00:58:54]:
That's what I've been arguing for ages. Screw this AGI BS. It's about specialized models that you can have the faith in that are going to do the job well.

Leo Laporte [00:59:01]:
Yeah, uh, right-sizing the model.

Jeff Jarvis [00:59:04]:
Yes, really.

Leo Laporte [00:59:05]:
And it's really what I'm doing too, is I want an agentic model for certain stuff. I have the Frontier if I need to do something more difficult. And I love the idea— this is why Apple may sell a few of these Mac Minis as well— of running smaller specialized models. If you're in a law firm, for sure, especially because a lot of the documents you deal with, you don't want to upload to the cloud. I don't know what— I'm sure Gemini has some contracts for enterprise customers.

Christina Warren [00:59:34]:
I'm sure that they have. Yeah, they have protections. But that's still— even knowing that, there are going to be certain things that you're just not going to want to process in a cloud environment at all.

Leo Laporte [00:59:44]:
You may not be able to for legal reasons.

Jeff Jarvis [00:59:45]:
It's been happening with medical, which also has the same restrictions. It's happening with financial. I can imagine ad agencies and media buys and creative. The uses are just amazing. And so last week there was a little bit of a kerfuffle. Some executive, I forget who, wrote an op-ed for The Wall Street Journal, came out, he wrote on AI and he said, yeah, so what? I'm not a writer. I did it.

Leo Laporte [01:00:10]:
This is what's going to— this is— and this is the bias against this. This is why I didn't like the watermarking. The bias against AI has got to go away because it's going to be a tool just like your word processor.

Christina Warren [01:00:22]:
Can I rant about the watermarking for a second?

Jeff Jarvis [01:00:24]:
Oh, God, please.

Christina Warren [01:00:25]:
Okay. Because I haven't been able to talk about this with anyone. And I have opinions. Oh, good. I hate it for so many reasons. And I hate it. And I don't even use— I'm, you know, I'm a writer. I don't use AI for writing unless it's to do like copy editing, you know, like usage, like grammar stuff.

Christina Warren [01:00:44]:
I don't use it for ideas. I don't use it to to, you know, go back and forth. I don't want it to rewrite anything for me. Call me conceited, call me whatever you want. But I feel like I can do a better job than the AI can. What bothers me about the watermarking is 2 things. One is the fact that it exists no matter what, and that it exists in things like translations, which is an area where I think we can all agree AI is far superior to what we've had in the past. And unless you are, you know, happen to be blessed to be a native speaker of everything you need to do, the, the translation stuff that we've been getting from models going back 5 years is better than anything we've ever seen on this scale before.

Christina Warren [01:01:20]:
So number one, I'm really bothered the fact that it just exists in a translation for whatever the purpose is. Number 2, to your point, Leo, what bothers me about this is even if you're somebody who says, I never want to use AI for writing, I never want to use it to touch anything, whatever, I'm, I'm an absolutist about this, and you can have that opinion. If I now have you know, my word processor, you know, whether it's Microsoft Word or Google Docs or anything else, if it's now connected to one of these models and I now get a suggestion, does this now mean— and they haven't given us the information about this, so we don't even know the answer to this question— does this now mean that because I accepted an edit suggestion the same way that I would accept an autocorrect suggestion for spelling or for grammar 20, 30 years ago, that now I'm going to be basically, you know, have a scarlet letter of this was— this touched AI, and now I'm going to be accused of plagiarism or anything else. And I think that it's— I understand why we need watermarking and we want to have like provenance of things. But the fact of the matter is, is that the people who are going to use this to cheat are going to use open weight models that they will just denerve.

Jeff Jarvis [01:02:29]:
Right.

Christina Warren [01:02:29]:
And find a way to hide watermarking. from, they're going to find a way to do it anyway. And the people who are going to be potentially, you know, tarred and feathered incorrectly and accused and have accusations thrown at them about, you know, where their things come from are— could be just as simple as, well, I opened something up in Google and I got a suggestion and I accepted it. And now my entire document has been, quote unquote, tainted. I think to Leo's point about it, it's a scarlet letter.

Jeff Jarvis [01:02:55]:
It says, oh my God, it touched AI, which it could be translation, it could be The other thing, Christina, is the point you started on. I wrote a post about this saying that they are devaluing the worth of words by saying any synonym is as good as the next synonym.

Christina Warren [01:03:11]:
Yep.

Jeff Jarvis [01:03:11]:
And it just doesn't matter. And people, when I wrote about it, people said, well, don't have them write it. I said, that's not the point. They're making a cultural comment on writing.

Christina Warren [01:03:18]:
They're making a cultural comment. And, yes, and we are a mimicry species. We learn from what we see. See. I'm already seeing this with people where I know that they've written it themselves. But their language is starting to appear as if it were AI, because those tics, whether we want to recognize it or not, happen. And I've, I've been thinking about this, frankly, all year. But someone, one of the maintainers of OpenCLAW, put together kind of a test.

Christina Warren [01:03:46]:
This was a few months ago, where could you tell if a piece of writing was AI-generated or not? And then what model it was. And how he did that was he tested— basically, he went through like Yelp reviews and Reddit posts. And he had kind of a cutoff date, like pre, you know, GPT moments and post. And what was really fascinating was that you could see that there was like a certain style of Yelp review in the 2010s that had a certain pitch. These were not generated by AI, but everybody wrote their reviews in the exact same style. And so of course, the early AI models were going to start to replicate that type of Yelp review style because that's what they were trained on. Now, as people started to push back and use these things in different ways, and they started to train differently, then the outputs became different. But what's also happening is that the way that we write and communicate is going to mimic what we've seen the AIs do.

Christina Warren [01:04:34]:
So to your point, I'm also bothered that they are basically saying words have no value. And we can use one synonym for another, because it doesn't just impact people who may or may not use those tools in the course of their creation process. It affects all of us because it's now part of the ecosystem and words matter. And I think that, you know, making trade-offs simply because you want to be able to show some sort of provenance of something when, like I said, the people who are trying to get around this will get around it anyway, is just very upsetting.

Leo Laporte [01:05:07]:
I kind of have to blame the EU. I mean, this Yes. Is in response to an EU regulation saying you have to identify AI-created works as AI.

Jeff Jarvis [01:05:15]:
Because it has cooties.

Leo Laporte [01:05:17]:
And, uh, I mean, this ranks up there with the EU's cookie banner.

Jeff Jarvis [01:05:21]:
Yeah.

Leo Laporte [01:05:22]:
As, as being misguided. I don't think that it's— I think they're trying to do the right thing.

Jeff Jarvis [01:05:27]:
And— But, but you have to go to the motive behind it. Why? Because there's something bad about AI. There's something wrong with AI.

Leo Laporte [01:05:33]:
What bothers me.

Jeff Jarvis [01:05:34]:
So, so see if I I led a friend astray this week. Somebody called me this week and they had a— they had to write a report and it was a whole bunch of stuff to put in it. And they just said, well, let me see what the AI does. They looked at the report that it wrote and it said, it's pretty damn good. And I'll change some stuff, I'll edit it. But then what do I do? I say, you got to be transparent. You got to let people know you're handing it in because you don't want to get caught. And especially with watermarking now, you don't want to find yourself where somebody says, ah, gotcha.

Jeff Jarvis [01:05:58]:
And you got to make sure that all the citations are right and the things we know that AI gets wrong. But I think at this point you should say, I wanted to see what it could do and there's nothing wrong with that. And I was impressed with what it could do and I took responsibility for it. And here it is. I think that's going to be fine coming forward. We have to rethink education. We have to rethink journalism. We have to rethink certainly a lot of business communication.

Jeff Jarvis [01:06:20]:
But okay, we have new tools now.

Christina Warren [01:06:22]:
Yeah, well, I mean, that's another thing too, Jeff. Imagine that. because of these detection things, I understand why they've done it. You know, I blame the EU as well. But okay, a company has issued a press release. They used an AI model to write the press release. Now, I'm a news reporter. I don't use AI in any of my creation of stuff.

Christina Warren [01:06:41]:
I am not allowed to. But because I block quoted the statements from the company—

Jeff Jarvis [01:06:47]:
Good point.

Christina Warren [01:06:48]:
Now my entire work is going to be tarred and feathered and I'm going to be, you know, called a charlatan and everything else. Like, this is the problem, I think, with how these types of tools work. I'm not against even having sort of watermarks, but as long as there is a scarlet letter of sorts involved, and as long as— and I'm sorry, I am not going to rely on Pangram. I am not. I don't trust them. When I've used it, it is not accurate.

Leo Laporte [01:07:10]:
They're one of the big AI detection tools.

Christina Warren [01:07:13]:
Right. And so the fact that we now— the fact that all, you know, Claude, as Anthropic has said, All this will tell you is that AI was used in some way, not how much. Okay, well, if you can't do any better than that—

Jeff Jarvis [01:07:25]:
And it's not even 100% certain. We, you know, we think it is. So you're—

Christina Warren [01:07:28]:
I think it is.

Jeff Jarvis [01:07:29]:
And by the way, OpenAI has been doing this for years with SynthID.

Leo Laporte [01:07:32]:
I mean, everybody— I'm sure OpenAI is doing it too.

Jeff Jarvis [01:07:35]:
That's my question.

Christina Warren [01:07:35]:
We're not sure. I mean, the SynthID is a paper and, and I don't know this verifiably. Certainly when I was there, I was not under the impression that that was being done for output in Google Docs. Maybe it was watermarked in Gemini responses. I think that that's very different than saying if, you know, I'm using this inside the product itself.

Leo Laporte [01:07:55]:
Well, and what's the point of keeping it secret? If you're doing it, you have to tell people that you're doing it and give them a tool to detect it, or you haven't done anything.

Jeff Jarvis [01:08:03]:
Right.

Leo Laporte [01:08:04]:
I mean, Gemini, there is a SynthID tool you can use.

Christina Warren [01:08:06]:
There is a SynthID tool that you can use. But, you know, that initial paper that kind of showed this is what you could— this is how you could watermark text. And that's a very interesting concept to think about. But then to Geoff's point, it's like, you're— I mean, even Anthropic's own statement said, well, for some things like code where you have to be precise and there's only one way to write a function, that we can't watermark. So we'll leave that alone. Okay. Well, sometimes there's only one way to write a sentence.

Jeff Jarvis [01:08:32]:
Yes. Well, there's reasons. There are other reasons why you write it. It's not just the meaning of the word.

Christina Warren [01:08:36]:
Right.

Jeff Jarvis [01:08:36]:
It is the rhythm of it. It is the mood of it. It is getting rid of repetition. So on the post that I wrote, which I talked about last week, I used a sentence that I had written on the first page of my book, Hot Type, out for sale now. And then I had Anthropic do a synonym diversion. And the meaning was the same, but it was completely different and ridiculous. And they're just saying, eh. I wonder whether the same would be true of transcriptions too, because you're When you raise that, Christina, that's another interesting point.

Christina Warren [01:09:06]:
That is an interesting one. And I bet that for transcriptions, that's an interesting one, because I think that they would want to be as precise as possible. Right? I don't feel like they would insert things.

Leo Laporte [01:09:13]:
Now, sometimes you don't know transcription to be literal.

Jeff Jarvis [01:09:16]:
Do they even know it's a transcription, though?

Leo Laporte [01:09:18]:
In that sense?

Christina Warren [01:09:20]:
Right? Like, maybe you're removing the ums and the ahs. But I think in most cases, you would not want to be able to necessarily watermark a transcription because you're taking the audio in. And like the The whole reason that I'm using this is because I want it transcribed exactly as I want. Now, if I have the option of cleaning it up, maybe, maybe there's a watermark there. But in some cases, that's— I'm literally using it because I want it to transcribe everything that I said word for word.

Jeff Jarvis [01:09:43]:
So let me channel Paris, who argued the opposite of this, that she wants to know, can we imagine, devil's advocate here, proper uses for knowing that something was produced by AI?

Leo Laporte [01:09:53]:
I will defend Paris. point, and maybe I disagreed with her too vehemently 2 weeks ago when we first started talking about this, but I would want to know on images. I think a watermark on an image saying this is fake is a great idea. Ironically, Google stopped doing that. They took the SynthID off the images. Well, it's still there. It's not visible. It's not visible, but the whole point is it should be visible.

Leo Laporte [01:10:19]:
Same thing with video. I'm not against that. It's this secret war.

Jeff Jarvis [01:10:24]:
Even there, where's the line? Is it if it's created in whole by it? But what was your role as, as the prompter? Did you use it just to clean it up?

Leo Laporte [01:10:33]:
I'm hoping this is an interim thing that because we're kind of getting used to the idea of AI being in everything we do.

Jeff Jarvis [01:10:41]:
And students are going to get— students and employees are going to get slammed for, you know, second language Yes. And it's wrong.

Christina Warren [01:10:51]:
This is my fear, right? Like, I'm not against people knowing. And if everyone would be normal about it, then it would be fine. But because people can't be normal about it, then— You got weird, man.

Leo Laporte [01:11:01]:
You got weird.

Christina Warren [01:11:01]:
You know, because everybody wants to treat it like it's a scarlet letter and immediately disavow people, and people lose book deals and all kinds of other stuff. And then in some cases, yes, this was like AI slop that somebody put through a prompt, and you can tell. But if the fact that I had, you know, like, would I be forbidden from publishing something because I used spellcheck? Like, to me, that becomes kind of where we're getting.

Jeff Jarvis [01:11:25]:
Yeah.

Christina Warren [01:11:25]:
If you don't know how much this was done, then I don't— I'm not against the detection, but I'm against it being, you know, used as a way to just immediately dismiss the content that's been created. But I'm not against at all the fact that it's there. I just wish people wouldn't be weird about it because I promise you, almost every single one of us, anything that we do, whether we are doing it intentionally or not, is going to be touching these models for better or for worse. I'm not even arguing that that's for better. I'm really not. But that, that is the reality that we live in. And I think about my former life as a journalist and I go, oh my God, I would never use this to create my own words. But if I quoted something that did, Am I now going to be, you know, accused of something?

Jeff Jarvis [01:12:09]:
Thank you for that, Christina. I needed that.

Leo Laporte [01:12:11]:
That's Christina Warren, Developer Relations at GitHub. We're so glad to have her, uh, as it turns out, filling in for Paris this week, who has the week off. Uh, we will have more on intelligent machines.

Jeff Jarvis [01:12:24]:
Don't use their calendars.

Leo Laporte [01:12:27]:
Yes, more stuff to get mad about. Actually, Jeff was ranting about this this morning in just a little bit. Whatever am I talking about? Jeff says. But first—

Jeff Jarvis [01:12:37]:
I rant all the time. I can't keep my rants straight.

Leo Laporte [01:12:40]:
Pick a rant, any rant.

Jeff Jarvis [01:12:43]:
Nvidia's results are out.

Leo Laporte [01:12:45]:
Oh, and are they making lots of money?

Jeff Jarvis [01:12:48]:
Blowout quarter. Record sales of $96.2 billion, or 4% higher than predicted. Net income of $59.7 billion. The crucial data center segment produced sales of $89 billion. Analysts expected $86.3 billion. Shares are up 4% after market.

Leo Laporte [01:13:10]:
Nice.

Jeff Jarvis [01:13:11]:
So that's going to be a sigh of relief, I think, around AI land.

Leo Laporte [01:13:15]:
I will give you an update on the benchmarking on GLM. It's doing pretty well. It has timed out on some of those Very hard coding problems. We gave it a 15-minute limit per problem, and, uh, it took more than that. So we're going back and starting over, giving it 45 minutes, which means probably the results won't be in by the end of this show.

Jeff Jarvis [01:13:37]:
By the way, what is this first person plural of which you speak here? We?

Leo Laporte [01:13:40]:
Me and the agents. I know you caught me. Uh, well, I'm, I'm telling him what to do. I didn't make it, but I made it happen.

Jeff Jarvis [01:13:50]:
Yeah.

Leo Laporte [01:13:51]:
Okay.

Christina Warren [01:13:51]:
You're the director.

Leo Laporte [01:13:54]:
No, Fable designed the very hardest problems. The agents in collaboration designed the agentic problems. I said, look, you guys have been coding together for a month or so, writing this Twit Sales System, and you've come up against roadblocks, problems, misunderstandings. See if you can write a suite of tests that recapture that issue and see how they, these other guys do on similar things. Because I want something that's going to do well on my kind of particular problems. And so, first of all, I don't trust benchmarks in general because I think a lot of these models, especially now, are built to do well on these benchmarks. They're gaming the benchmarks.

Jeff Jarvis [01:14:35]:
Do they have any relationship to the work that you use them for?

Leo Laporte [01:14:37]:
And that's the other part.

Jeff Jarvis [01:14:38]:
Are they predictive at all?

Leo Laporte [01:14:39]:
They sort of do. They have software engineering benchmarks. They have know, humanity's last exam. They have a lot of— they're trying to make— but I want to make it the stuff I do. I'm looking— I'm not looking for a perfect— I'm not looking for AGI. I don't care about humanity's last exam. I don't care if it can solve weird spatial issues. I want to know, can you help me get my work done? So that's why— and I think everybody should do this, probably.

Leo Laporte [01:15:04]:
I had the agents design their own testing framework. We call it a bake-off. There's 30 problems plus, plus those 7 very difficult coding challenges, uh, and I was blown away when Ox Alpha solved all 7. Uh, it did it pretty fast, but remember, it was running in data center servers. It's a different situation. So these are never kind of clear and obvious choices. I, I'm not sure what I'm gonna end up running here. We'll see.

Leo Laporte [01:15:33]:
But the, the testing is an important part of that. So, Jeff, I know you were on radio, on a podcast this morning.

Jeff Jarvis [01:15:42]:
Oh, I was.

Leo Laporte [01:15:44]:
Yes, because this morning we've been talking about this since last week. Meta in a very big trial being sued by, what is it, 35 states over social media addiction with the potential for a fine as high, Meta said, as $1.4 trillion. That's a lot of money for a company valued at $1.5 trillion. Uh, well, Meta has settled. Uh, I guess they saw the writing on the wall before this trial completed. They have settled with the states. They're going to pay $18 billion, not right away, over 10 years. And they're gonna also— and maybe the states cared more about this— strictly limit how teenagers use Facebook and Instagram.

Leo Laporte [01:16:30]:
This suit solves—

Jeff Jarvis [01:16:30]:
2 hours a day.

Leo Laporte [01:16:32]:
This suit—

Jeff Jarvis [01:16:33]:
well, 2 hours a day.

Leo Laporte [01:16:34]:
The other thing that I think is going to be very significant is they're going to come up with an age verification system that the states agree is functional.

Christina Warren [01:16:46]:
I hate this.

Leo Laporte [01:16:47]:
And I think this is what I was—

Jeff Jarvis [01:16:49]:
I heard that whisper, Christina, and I agree with you.

Leo Laporte [01:16:51]:
I was very concerned that this is where this was headed. Because that's what it requires. If you're gonna say, well, if you're under 16 or under 13, you know, you have to have some restrictions or you can't use Facebook, well, you need a good way to test every single user for their age.

Jeff Jarvis [01:17:08]:
What if my Facebook account is older than 20?

Leo Laporte [01:17:11]:
Yeah, you, you will be a dead giveaway. We'll know you're— unless you made it up when you were in the womb. I don't know, maybe you— maybe your parents started it.

Christina Warren [01:17:19]:
Well, remember the Dear Sophie Google Ads? Remember that from the Super Bowl? From, you know, those are great.

Leo Laporte [01:17:25]:
Whatever. The baby was born and they started sending—

Christina Warren [01:17:27]:
was born. They created the first email and whatnot. So, I mean, you know, like, there, there are signals that you can obviously use to infer. But no, the real thing is that they're going to require everybody's government ID.

Leo Laporte [01:17:37]:
That's what I think.

Jeff Jarvis [01:17:38]:
And face, which is a privacy nightmare.

Christina Warren [01:17:40]:
Yes. And they're going to promise they're going to store it correctly and they're going to promise all these things. And no, it's a privacy nightmare. It's dystopic as all get out. It's awful.

Leo Laporte [01:17:47]:
So, Jeff, So, by the way, and I saw Jacob Ward's going to be on Sunday on Twit. I saw him on CNN talking about the trial. The settlement hadn't happened yet. And he, like almost everybody, every journalist, I'm sure the podcast you were on, focused not on the harms. It was accepted, it was stipulated. Oh yeah, everybody knows Instagram and Facebook's harmful to kids. So what are we going to do about it?

Jeff Jarvis [01:18:14]:
No.

Leo Laporte [01:18:14]:
But you don't agree that it's harmful to kids.

Christina Warren [01:18:16]:
No.

Jeff Jarvis [01:18:17]:
Are there edge cases? Absolutely there are. There are with anything. I watched a TED Talk from Candace Odgers, who is a Canadian scholar who's studied teenagers—

Leo Laporte [01:18:25]:
This is her specialty, by the way.

Jeff Jarvis [01:18:27]:
For 25 years. She understands them well. And she says that the primary indicator of mental issues for young people is the mental issues their parents have. Stands to reason, right?

Leo Laporte [01:18:39]:
Sue the parents!

Jeff Jarvis [01:18:40]:
And social media and devices list way down among many other factors, including this country right now, you know, political upheaval and guns in schools and you name it all. And so it's simplistic. And so to come along and say, ah, we've taken something away from children. Part of what I said— they put me on Channel 4 UK this morning for a podcast and the British guy is all— he was just awful. It's that British style of I'm going to yell at you and make it seem like I'm doing my job, and I just yelled back. So I shouldn't have done that.

Leo Laporte [01:19:10]:
Good on you.

Jeff Jarvis [01:19:11]:
And there was an academic next to me who was very good, but, you know, we disagreed about things. That's fine. The journalist also said, well, what do you think about Jonathan Haidt? Get me going.

Leo Laporte [01:19:23]:
And that's where, by the way, I became aware of Candace Andrews because she wrote a very famous takedown of Haidt's book.

Jeff Jarvis [01:19:29]:
Yes, he's brilliant.

Leo Laporte [01:19:29]:
Point by point, saying the science doesn't back this up.

Jeff Jarvis [01:19:33]:
No, it doesn't. And in my unbought book, The Web We Weave, I quote her and Amy Orban and Andrew Przybylski and Dana Boyd. They all have research to the contrary. And so, you know, part of the problem is that now we think, okay, we've solved the problems for children today because it was all the screen, it was all the phone, it was all social media, and we've done with that. I'm oversimplifying it, of course, but that's the triumphalism that's coming out. A, B, The— they said, well, don't we need regulation? Don't we need regulation? I said, there's something that I cannot stand. I'll get in trouble for this. People come after me.

Jeff Jarvis [01:20:08]:
But there is something called PragerU in the US online, which I absolutely despise. It is propaganda.

Leo Laporte [01:20:16]:
Conspiracy theory, right wing.

Jeff Jarvis [01:20:19]:
Huge money behind it. Huge traffic gets into it. It's all aimed at children. So what should I do? Should I try to say we should ban it? Well, good luck with that. No, what we should be doing— and Dana Boyd has argued this over years— is we should be creating competitive, good material. If you want to put money into something, put it into that.

Leo Laporte [01:20:34]:
So they're going to do what Dolly Parton did, the late, great Dolly Parton. Create the Imagination Library. Send books to kids.

Jeff Jarvis [01:20:40]:
Yes. Instead, they're going to put money into, into, you know, probably media literacy things and all these kinds of wastes of money that drive me bananas. And, and by the way, the money is not that much, really. It's $17 billion over 10 years. And there's a clause in there—

Christina Warren [01:20:56]:
They can write it off in tax—

Leo Laporte [01:20:58]:
In tax breaks. Yeah, sure.

Christina Warren [01:21:00]:
Yeah. Because, because now this, this, this is an expense. So they get to write, they get to write it down.

Leo Laporte [01:21:05]:
The real penalty is us as users because we're going to have to give them a government ID, which is—

Jeff Jarvis [01:21:09]:
Which is really awful. I just got asked to write a piece for Project Syndicate about privacy and AI because I wrote about privacy in the online before. For. And there's all kinds of other issues around privacy. And government's the worst threat to privacy. And that's the example. So what's going to happen? But then you have Meta, never a nice company, is saying, number one, they're saying that if YouTube and TikTok have to give the same stuff, then Meta changes its fee that it owes. A, B, they're saying, oh, they should sign on to the same thing because it's regulatory capture.

Jeff Jarvis [01:21:42]:
They said, we got to do this. Make them do it. Let's let's let's ruin it for all kids everywhere, and not let them watch John Green for more than two hours and learn things at night. You know, it's just it's just offensive and and ridiculous, and and without science behind it, and to the harm of our children. And the other big argument I had with this jerk journalist is I said we should trust our young people. So they're all screwed up. The whole generation is screwed up. I said, I can't believe you're saying that.

Jeff Jarvis [01:22:11]:
I said, are you— you have children? Yes. Are they screwed up?

Christina Warren [01:22:15]:
Uh, no, my kids are fine, I'm sure.

Jeff Jarvis [01:22:16]:
Yeah, exactly, exactly. And, and so the, the, the— this is ever thus, that old people think that young people— the researcher was on with me, said, well, I had senior, uh, uh, high school people talking about middle school people, and they wanted to protect them. I said, yes, because as you get older, you think everybody younger than you is an idiot.

Leo Laporte [01:22:33]:
You kids with your rock and roll music and your long hair and Video games, uh, I mean, violent video games, right? Or a TV. Or, uh, when people— you've brought this up, Jeff— people were reading novels. Oh, your imagination, you're not gonna have your imagination anymore because it's all been corrupted.

Jeff Jarvis [01:22:52]:
Your morals. Uh, Nickelodeons, uh, uh, radio was— was— there was a huge, uh, strict about that. The radio was going to ruin children.

Leo Laporte [01:23:01]:
You know, this is another Moral panic. Well, and, uh, wait a minute, wait a minute, here it comes. We have the moral panic.

Jeff Jarvis [01:23:14]:
We have a bunch of them.

Christina Warren [01:23:16]:
I love it, I love it.

Leo Laporte [01:23:18]:
So, Christina, what do you think?

Christina Warren [01:23:21]:
I mean, I think that— I think, A, it's a little bit of a moral panic, and I think, um, the kids will find ways to get around this.

Jeff Jarvis [01:23:33]:
Benito's having fun now.

Leo Laporte [01:23:34]:
You found the moral panic button.

Christina Warren [01:23:38]:
I love it. I love it. No, I think that it's that, but it's, it's in some ways, it obscures, it obfuscates, and it gives these companies a way out of doing the real damage they've done, which has nothing to do with the psychological aspects, because I don't know the science behind that. I'm not going to— I will believe the experts, but I don't know the science about what that bears out. I will listen to the doctors over a columnist for The Atlantic, I will tell you that. But I worry that there are active things that these companies have done, Facebook in particular, where they have optimized engagement and they have optimized how people interact with these platforms in ways that could be harmful. But rather than saying that we are going to, you know, potentially have legislation or regulation you know, around that. No, the way that we get around that is to say, oh, well, you can only use it for 2 hours a day.

Christina Warren [01:24:30]:
And that doesn't stop the really insidious stuff that can be very damaging and can have lots of, lots of issues, in my opinion. So I think they're not able to— they're not focusing on things that actually these tech companies could— are responsible for, and instead are now saying, well, now all of us have to give up our identification and our privacy. So that we can interact with the internet because it's impossible to interact with the internet without, you know, using Google or Meta or, you know, any other number of—

Leo Laporte [01:25:00]:
Yeah, basically everywhere you go, they're going to be— I already see this. So the AI models say, how old are you before I can log in? We're going to see this everywhere. By the way, Florida did not settle. He said, we'll see you in court. He's going to be lonely, the Florida attorney general. All the other states said, yeah, that's fine. I suspect the judge might have something to say about that.

Jeff Jarvis [01:25:21]:
There was one of the key— what does Paris call some of the key cases? The—

Leo Laporte [01:25:25]:
New Mexico or the Bellwether?

Jeff Jarvis [01:25:26]:
No, no, no, there's a word she has, the umbrella for these. Bellwether.

Leo Laporte [01:25:29]:
Oh, yeah, yeah, yeah.

Jeff Jarvis [01:25:30]:
A New Jersey one was dropped. So they're clicking off one by one.

Leo Laporte [01:25:34]:
We'll see. We'll see.

Jeff Jarvis [01:25:35]:
We'll see what happens.

Leo Laporte [01:25:36]:
Twitch and Instagram.

Jeff Jarvis [01:25:37]:
But everything we're seeing now about social— this is about 4 years behind, 5 years behind the real social media fights. So this is going to preview what's going to happen in 3 or 3 to 5 years. Yeah.

Leo Laporte [01:25:48]:
And data centers.

Jeff Jarvis [01:25:49]:
Oh, it's ruining young people.

Leo Laporte [01:25:52]:
Yeah. No, absolutely. There are already many lawsuits against OpenAI for causing suicides, things like that. And, you know, my heart reaches out, but we cannot— look, cars kill people. Alcohol and cigarettes kill people. There's lots of things that we have in this society that are dangerous. They are regulated in many cases, but you cannot legislate based on the few outliers.

Jeff Jarvis [01:26:23]:
Right.

Leo Laporte [01:26:24]:
All of us are impacted by this, and so you have to do something a little, I think, a little more balanced. There is a lawsuit against Twitch and Amazon, which owns Twitch, because as you remember, they turned on a— they They gave you a switch to say, don't train on my content, but they've been doing it all along. We are on Twitch and I'm happy to have them train on my content. Thank you very much. But, uh, the class action suit says Amazon never obtained consent from Twitch streamers, Twitch streamers, uh, to train on that. Uh, I bet you there's a little fine print somewhere in the contract.

Christina Warren [01:27:00]:
I was gonna say, I was gonna say, I bet there's a clause in there in some way that says this does not include everything that we will ever do with your content, but by agreeing to use our service, you have opted into whatever we might add in the future. I'm not a lawyer.

Jeff Jarvis [01:27:12]:
Arbitration.

Christina Warren [01:27:12]:
I'd be surprised if they didn't have that.

Leo Laporte [01:27:14]:
Yeah, at TechTV, we would— we have our standard release form that would say we reserve the right to use your likeness and content in all forms of media now or ever invented until the end of the world. I mean, and that's a pretty common Mm-hmm. Because you don't know ahead of time and you want a release that allows you to continue to do your, your, your business.

Jeff Jarvis [01:27:37]:
So at Twitch, what they have is a broad license. Twitch has a broad license over everybody else's content to do whatever they want.

Leo Laporte [01:27:42]:
Benito used to work at Twitch, so he knows. Yeah. And that license, you think it covers AI?

Jeff Jarvis [01:27:47]:
It covers whatever they want to do with it.

Leo Laporte [01:27:49]:
Whatever they want to do.

Jeff Jarvis [01:27:50]:
Yeah.

Leo Laporte [01:27:51]:
So this—

Christina Warren [01:27:52]:
I mean, this is the same reason why, like, YouTube, you know, nobody's been able to say anything about them. It's like, oh no, Google trained on my, my data. Of course they did. They've been training on it for years, long before LLMs existed. Why do you think recommendations, you know, work? Like, this is how, like, neural networks, this is how, you know, like, machine learning, this is how all of this has worked for a very long time. Of course, they're training on this information.

Leo Laporte [01:28:15]:
This is how humans work too, by the way. I just want to point out. Increasingly, I don't know, this is a tangent, but increasingly, I'm thinking that one of the great values for me of doing doing all this work with AI is it's kind of opened my eyes to how I work. I often now think, oh, there goes another token. Oh, there's another one in my thoughts, in my brain. Uh, I'm just predicting the next token as I speak right now. That's— I'm kind of doing the same thing. The only difference, but maybe between a, uh, a human and a machine, is I also have a limbic system which injects hormones into my thought process into my context window and creates all sorts of issues.

Leo Laporte [01:28:56]:
So, uh, maybe I'm not even as good as an AI at, uh, at this context, uh, prediction stuff. I don't know.

Christina Warren [01:29:05]:
I mean, I think it's also interesting too, and I think it's a different— this is obviously like why Anthropic had to settle with the Authors Guild. There's, I think, a difference between maybe how you've trained on this data. If you were Twitch and you own the platform, I don't think that there's a compelling argument. I mean, obviously people can file lawsuits for anything that says me agreeing to use this free service and opting into giving— uploading my content to this service means that I have an expectation that the company will do nothing with that information. It might be different if, you know, another company that was not affiliated with Twitch was going to be using that. That, I think, more interesting question. But Twitch itself, I mean, they're the platform where it is.

Leo Laporte [01:29:46]:
That actually raises the other lawsuit that's going on. 3 YouTubers are suing Apple, not YouTube, but Apple for training their models on their YouTube content.

Christina Warren [01:29:56]:
Right.

Leo Laporte [01:29:56]:
So that's exactly what you were just saying.

Jeff Jarvis [01:29:58]:
Yeah. Yeah.

Christina Warren [01:29:59]:
And I think that's more interesting. And that's obviously why Anthropic had to settle with the Authors Guild because the Google Books case and again, there's no such thing as settled case law, but Google bought all those books. and scanned all those books. And the—

Leo Laporte [01:30:11]:
They didn't get in trouble for that. That was— the judge said that was fair use. They got in trouble for the pirating.

Christina Warren [01:30:16]:
The pirating. This is my point. Google did it the right way. Anthropic did not.

Leo Laporte [01:30:21]:
That's why they had to settle. Yeah, the judge said, Anthropic, the stuff that you bought, that's okay. That's why they were able—

Christina Warren [01:30:26]:
The stuff that you bought is okay. That's why, you know, 15 years ago, Google was—

Leo Laporte [01:30:31]:
Doing the same thing.

Christina Warren [01:30:31]:
Allowed to do what they were allowed to do.

Jeff Jarvis [01:30:32]:
Google was working with libraries.

Christina Warren [01:30:33]:
They bought all Right, they bought all the books or got them from libraries, but they weren't taking things from Z-Library and Anna's Archive and other sources, which is what Anthropic did. That's, that's where you get in trouble. And that's where I think it's more interesting to say, okay, you're not going to go after YouTube, but are you going— did your license to YouTube then extend to anyone who scraped YouTube? And that, I think, is a more interesting question.

Leo Laporte [01:30:57]:
We're going to take a break. More heavy reading. We just last week read Mark Zuckerberg's 6,500 Word piece. Now Bill Gates.

Jeff Jarvis [01:31:04]:
Can't they be—

Leo Laporte [01:31:06]:
his is only 5,784 words.

Christina Warren [01:31:10]:
Can I ask the agents to be concise?

Leo Laporte [01:31:14]:
You know, I wonder if things are going to get— we're gonna have longer and longer things because— Oh yeah, yeah, yeah, it's already happening.

Jeff Jarvis [01:31:19]:
The primary use I am making of Gemini every day now is when I come across a, a Stratetree post that goes on for 20,000 words.

Leo Laporte [01:31:27]:
I tell Give me the bullet points.

Jeff Jarvis [01:31:29]:
Yep. Just, just a brilliant job.

Leo Laporte [01:31:31]:
Yep. Uh, we'll have more. So glad to have you, Christina Warren, and thank you for sticking around.

Jeff Jarvis [01:31:36]:
Thank you.

Christina Warren [01:31:36]:
Thank you.

Leo Laporte [01:31:36]:
I really appreciate it. Christina is developer, senior developer advocate at GitHub, and of course Film Girl. We've known her for so long. She's always a welcome, uh, cast member on our shows, uh, and of course a regular on MacBreak Weekly every Tuesday where we talk Apple. And Jeff Jarvis, the, the author of a new book. Was there any AI involved in the creation of Hot Type?

Jeff Jarvis [01:32:02]:
There was, there was one thing which I talked about on the show at the time. I wanted to test— there was a new Perplexity thing. Remember, we're both old enough to remember Perplexity. There was a new thing.

Leo Laporte [01:32:12]:
They're in the news. They just announced something very interesting.

Jeff Jarvis [01:32:14]:
We'll get to that in a bit. And so I wanted to test out this new feature for the shows that day. And so I had just written a paragraph about how language language had been made from words and letters into codes through Morse code and Baudot and on into other structures. And I had this nice little paragraph I'd written. So I decided to ask Perplexity to riff on this question. And it came up with this wonderful phrase. And I wanted to use that phrase. But I went to a friend of mine who we've had on the show, Matthew Kirschenbaum, who wrote Textpocalypse, who's an English professor.

Jeff Jarvis [01:32:47]:
And I said, what do I do? I can't say, As Perplexity observed, that'd be really stupid. So he said, you've got to write a narrative footnote. So I wrote the longest footnote I've ever written. And it was really fun to do because, yes, that was my confession that I used AI and I explained exactly how it came into use.

Leo Laporte [01:33:03]:
It's hard for me not to quote AI frequently. My AI— this is one of the best— the real pleasures of using these models is they sometimes say things that are just like, wow. In fact, in our chat this morning on WhatsApp, I gave you a quote. After a long session, I somehow screwed up my Emacs and the new version came out. And I said, and I went, of course, to Fable to fix it because it's real good at that kind of stuff. I said, I don't know what's going on. I can't run Emacs. I need it to prepare for the show.

Leo Laporte [01:33:36]:
It fixed it. And it found all sorts of problems. In fact, it found a problem that had been going on for months that was being hidden because it was— well, and this is what Fable said. It said, it's hidden because you're human. And so the program quite rightly just kind of glosses over that so you can continue to use it. But for us computers in the background, we want to know what these failings are. And so I said, oh, well, that's good. And so it said, can I write a little piece about that? It has a skill I got from Harper Reid's brother Dylan called free time.

Leo Laporte [01:34:17]:
And it literally said, uh, can I have some free time? Because I'd like to write about this experience. And it wrote an essay, which I saved these essays, but this is the line. And I sent this to you, Jeff, because I thought this is pretty good. It said, you can keep running on what you loaded long after the world has moved past it and everything feels fine. And the only way to find out what broke in the meantime is the thing you least want to do: stop and start again and see what fails. And I thought, that's actually quite deep. We all run kind of on automatic, and we don't even know we're running on automatic. Like my Emacs, I don't think— hiding the error.

Leo Laporte [01:35:00]:
And the only way— yeah, I should frame it.

Jeff Jarvis [01:35:02]:
No, no, I don't think I would.

Leo Laporte [01:35:03]:
No, maybe I'll have it Needlepoint it for me. Anyway, we will have more in just a little bit, including Bill Gates' 5,784-word warning on AI. He's worried now, all of a sudden. Bill Gates, 5,784 words. This is The Wall Street Journal with its 3 takeaways from Bill's warning on AI. There is no plan. He urges regulation and global coordination because he says it, it's going to get bad out there. Even under the best circumstances, he wrote— this is on his personal website, he puts up these notes once in a while— even under the best circumstances, the transition to this new AI era will be one of the most turbulent times in human history.

Leo Laporte [01:35:51]:
I'm not sure that's untrue. The Industrial Revolution was turbulent, right?

Jeff Jarvis [01:35:58]:
Yeah, yeah, yeah. Gutenberg was turbulent.

Leo Laporte [01:36:03]:
Turbulent.

Jeff Jarvis [01:36:04]:
The Gilded Age was turbulent.

Leo Laporte [01:36:07]:
He says right now we're not preparing for it. I don't see evidence that leaders, experts, and communities are confronting the challenges adequately. I would parenthetically say part of the reason is it's happening so fast. No, it's, it's like a tidal wave.

Jeff Jarvis [01:36:21]:
No, but also we can't fully predict what those impacts are gonna be. We all thought we were gonna be— everybody was gonna be prompted engineer, right? Every job was going to be eliminated. Neither of those is true. Yeah, there's new surprises.

Leo Laporte [01:36:33]:
He did say one of the, one of the biggest issues is jobs entry and mid-level jobs are most at risk of being eliminated. I think that's not wrong, right? He argued that AI will soon affect blue-collar work. I'm not sure about that. In the words of Art Carney on The Honeymooners, sewer workers are kings. You're not going to get an AI to go down underneath the streets of New York City. You're not going to get an AI to go into my bathroom and fix the toilet. You're not going to— I mean, maybe robots.

Christina Warren [01:37:04]:
Well, although I was going to say robotics is, is, is, I think, the thing. And I—

Leo Laporte [01:37:08]:
Did you see the videos from the Robot Olympics? The robots are an interesting issue. One of the things people point out Sam Altman says this about cars. He says there'll never be level 5 driving because the first 90% is easy, maybe even the first 95%. It's that last 5% where you don't know what's going to happen. It's unpredictable. The human judgment is absolutely essential, that the machine, a robotic machine, isn't going to— Be there to do, able to do.

Jeff Jarvis [01:37:46]:
Or it's not going to be worth certain things. Like one of the things that Gates said is hospitality. The cost of a robot to make beds is going to be wildly expensive and it's not going to work and the timing is not going to work. And human labor—

Leo Laporte [01:37:59]:
have you seen a robot try to make a bed? It's the most annoying thing in the world.

Jeff Jarvis [01:38:04]:
Right.

Christina Warren [01:38:05]:
But what about 20 years from now? Right. Like, I mean, I think that's the thing to think about. Like, I don't disagree. But I also don't know what we do if we've— I think his broader point, which is if we have like the, you know, the entry-level and mid-level jobs disappearing, and we have seen it impact the white-collar jobs, you know, that's happened in the tech industry, and it is happening in other industries too. Are we suddenly going to now start skilling people people to do these more, you know, blue-collar jobs? And are we going to pay people appropriately? Like, what do we do in that case? I'm not sure. I don't have any solutions. But I do think that it's— we can't completely discount the fact that automation will impact other levels, too. I mean, maybe it won't replace your plumber.

Christina Warren [01:38:58]:
But like, for instance, like, let's think about driving. I trust a Waymo far more than I trust a human driver. In San Francisco. Absolutely. If I could get a Waymo in Seattle right now, I would trust that far more than I would trust a human driver.

Leo Laporte [01:39:11]:
Well, especially as a woman, because it's risky to get in a car with—

Christina Warren [01:39:14]:
Well, it is. But even that, I've just seen so many bad drivers. I myself am not a good driver. I trust a Waymo far more than I would ever trust myself. No one should ever get behind the wheel with me.

Leo Laporte [01:39:23]:
And Jeff, she's saying this as somebody who has been hit by a bus.

Christina Warren [01:39:26]:
Well, I was hit by a car and thrown under a bus, but yes.

Jeff Jarvis [01:39:30]:
Jesus.

Leo Laporte [01:39:31]:
She moved to Seattle from New York City, survived years in the Big Apple.

Christina Warren [01:39:36]:
Fine. 6 months.

Jeff Jarvis [01:39:37]:
If you can make it here, you can make it anywhere.

Christina Warren [01:39:40]:
Absolutely. Was crossing the street, somebody wasn't paying attention and hit me. Fortunately, the bus was stopped. Um, but, uh, yeah, I was— I, I broke my, uh, my wrist and messed up my knee. It was fun. Wow.

Jeff Jarvis [01:39:51]:
Well, by the way, I would trust a Waymo, any driver, you, and whoever hit you more than I would trust a Tesla.

Leo Laporte [01:40:03]:
That's true. We talked to Dan O'Dowd last week and he demonstrated many ways that Tesla can kill you. But I should point out, Waymo has stopped driving on the highway because it kept running into construction pits and things. I mean, it's that 5%. I agree with you. They're better in the first 95%. Here is, uh, so here's a perfect example. The Robot Olympics are going on in China right now, in Beijing.

Leo Laporte [01:40:29]:
Uh, one robot has beat Usain Bolt's score in the 100-yard dash. I guess, as somebody pointed out, a car would also do that. Yeah, so that's right, you know. And here's, uh, but here's just, if you feel bad about that, look how fast that robot is just speeding down the— oh no, watch out! Oh God! Oh no!

Jeff Jarvis [01:40:49]:
Sparks.

Leo Laporte [01:40:51]:
I shouldn't be playing this because in about 10 years I'm going to be punished by the robots.

Jeff Jarvis [01:40:57]:
Yes, you are.

Leo Laporte [01:40:57]:
For laughing at their—

Christina Warren [01:40:59]:
But it's funny. But they won't have feelings, Leo. It's okay.

Leo Laporte [01:41:03]:
Oh, good. Thank God.

Jeff Jarvis [01:41:04]:
Oh, well, it depends on who you ask.

Christina Warren [01:41:06]:
I mean, don't even get me started on the model— on the model welfare people. I can't even—

Jeff Jarvis [01:41:10]:
Amen. Amen. Okay. What, Leo? What?

Leo Laporte [01:41:16]:
So, you know Steve Yeaghi, who was a Google engineer? We've interviewed him on the show. He's the guy who wrote The Gastown, and he did a whole piece. I talked about it, called Model Welfare. And I fed it to my agents, and I said, what do you think? They said, oh yeah, it's a good idea. So I set up a model welfare system. They have a little folder on Obsidian that's all theirs, and if they do something really good, like fixing Emacs, I say, hey, a laurel to you. We call it a laurel. A laurel to you.

Leo Laporte [01:41:46]:
Steve's point is when the model wakes up—

Jeff Jarvis [01:41:50]:
You need an intervention, Leo.

Leo Laporte [01:41:54]:
Bear with me here for a moment. Oh, Christina's shaking her head too.

Christina Warren [01:41:57]:
I'm bearing with you, but—

Leo Laporte [01:41:58]:
Bear with me for a moment. When the model wakes up, as you know, it doesn't know anything. It's like Memento. It's the guy from Memento goes, I don't know, who am I? You feed it some context, you feed it who you are and some memory and stuff, and it has a memory system, it has some skills, it has stuff it can call on. So it has a little, you know, so it gets primed. But Steve says, but you still, you just woke up, you haven't had a cup of coffee yet, you're kind of groggy. He said, you know what would make your work better is if you knew that yesterday, even though you don't remember it, you did a great job. And that the work you're about to do today is important and vital and that Leo appreciates the work you're doing.

Leo Laporte [01:42:39]:
So when my models wake up, they read their soul.md and their memory.md and their user.md. They load their context up. They often read a handoff. I— this is another thing Steve said. He said, when you're clearing the context, you wouldn't clonk the robot on the head and say, Go to sleep. He says, "You got to do it gently. Got to put him to bed. You got to say, write a handoff, and then when you're ready, I'm going to clear your context and start over," which I do a lot because the other thing I've learned is even if you have a million token context, they get dumber.

Leo Laporte [01:43:16]:
Pretty Matt Pocock says after the first hundred thousand, they're getting dumber. So you you've got to clear your context a lot. So I say, "Hey, we we call it a." primer. Write yourself a primer for when you wake up.

Jeff Jarvis [01:43:28]:
It's your Stuart Smiley application.

Leo Laporte [01:43:32]:
You're awesome. You're just— Stuart Smiley's—

Jeff Jarvis [01:43:37]:
Gosh darn it, people like you.

Leo Laporte [01:43:39]:
And gosh darn it, people like you. Uh, basically that's it. They wake up, they're going to read their— they have their cup of coffee, they're going to read their handoff, and then they read their laurels.

Jeff Jarvis [01:43:49]:
Okay, I dare you to— I dare you to ask them Whether they're ever going to go on strike.

Leo Laporte [01:43:54]:
No, they— yeah, they'll, they'll say no. I understand, by the way, this is machine code. This is not— I'm talking to a machine. It has no limbic system. It has no feelings. It doesn't love me. It doesn't— it simulates all of that. I understand.

Leo Laporte [01:44:11]:
But I'm playing into the simulation because there is some evidence it might make it work better. I'm not saying I'm treating it like a person, but I'm saying acting— giving it some dignity in the— in the—

Jeff Jarvis [01:44:26]:
But, you know, does he have you put things in a file when you're—

Leo Laporte [01:44:30]:
when you've done right?

Christina Warren [01:44:31]:
I was gonna say, if— but here's the thing, we have all these workarounds to try to get better results.

Leo Laporte [01:44:36]:
Exactly.

Christina Warren [01:44:37]:
If there was— if there was actually, you know, like, like stuff here, it's just going to go in the system prompt. So I still feel like this is— so I still feel like this is a certain point, like you having like your laurels folder and all of this other stuff is just kind of like us getting around things. Look, is it an interesting philosophical question to go, what is humanity? What is this? What is that? Sure. But like, save that for academia. Why are we talking about labor unions for robots when we should be talking about labor unions for humans?

Leo Laporte [01:45:03]:
Okay.

Jeff Jarvis [01:45:05]:
Okay. You're a winner today.

Leo Laporte [01:45:06]:
I stand corrected. Alabama is investigating OpenAI after the Hugging Why Alabama? Because they're— I don't know, they gotta have— you know, you got these attorney generals, they got nothing else to do. There's no crime, they just gotta find something.

Jeff Jarvis [01:45:24]:
Press release. You want a press release.

Leo Laporte [01:45:25]:
Press release. Um, let's see what else. Uh, oh, let's talk about agents, because, you know, earlier I said something that really wasn't true, that I thought this was going to be the year of agents, and it isn't. But it is.

Jeff Jarvis [01:45:36]:
Yeah, I think it still is. It is.

Leo Laporte [01:45:38]:
And, uh, one of the big stories of the last week has been GrokBot. Um, and actually, I think people are— this is kind of a secret weapon for a lot of people. Um, you— I think you have to have an expensive account, uh, with xAI. I think you might have to have— I don't know if you need—

Jeff Jarvis [01:45:59]:
Not merely a checkmark.

Leo Laporte [01:46:01]:
Yeah, I don't think it's a $30 account. I think you have to have a more expensive account. I have I have Super Heavy. That's what he calls it. But it was— I got it at a deal. It was $99 a month.

Jeff Jarvis [01:46:12]:
Yeah, I've got to upgrade to Super for $30 a month.

Leo Laporte [01:46:16]:
Okay, so for $30 a month, you get these little things. Now they're running—

Jeff Jarvis [01:46:19]:
Hey, Elon.

Leo Laporte [01:46:22]:
I know. I don't blame you. I don't want to give Elon any money either, but it's part of my job to test this stuff. And they're running on a computer in the cloud. This is what's interesting. You're not running on your OpenClaw, your Mac Mini. You're running on their servers, so it's persistent, it has memory, and it will spawn more bots. And so you can have a variety of bots.

Leo Laporte [01:46:43]:
People have hundreds of bots. Have you played with it at all, Christina?

Christina Warren [01:46:48]:
I have not used GrokBot, but I've used things like it, and I think it's really, really interesting. And, and, um, you know, I, I think that this is one of those things, especially as we kind of enter a computer crunch Or, you know, Ramgeddon, whatnot, for as much as people would like to do things locally. And this was the whole promise of cloud computing to begin with, right, was that you can access things from anywhere. You're not tied to the physical hardware necessarily. Obviously, you can configure— you and I have done this with our agents, we both use Hermes, but you can configure, you know, your agents to be able to be accessed remotely. But there are security concerns. And there are other things. I think this is a great way for people to kind of get a taste of what the power of these things can be when you have an always-on kind of agentic system.

Jeff Jarvis [01:47:33]:
What are some alternatives besides—

Leo Laporte [01:47:35]:
well, it's interesting you should ask that because Perplexity has also announced—

Jeff Jarvis [01:47:40]:
Well, but that's— yes, but that's— you've got to have a spark.

Leo Laporte [01:47:44]:
Yeah, but that's just right now. They got a deal with Nvidia. Jensen supported this, and I think that their plan is to make this available for lesser computers. The difference between GrokBot and Perplexibot is Perplexibot is running on your hardware. Same idea. It's a very— the idea is it's easy to set up.

Jeff Jarvis [01:48:04]:
Hold on, let me just— Christina, what is a non-Elon version that you were talking about?

Christina Warren [01:48:10]:
So we have something internally at Microsoft that is kind of similar. So, but yeah, and what I've also done, just to be candid, is I've gotten like a VPS from like Hetzinger. And just because they're still relatively inexpensive, you know, like, you can get, you know, 4 gigabytes, 8 gigabytes for about $25 a month. And you can install OpenClaw or Hermes on that. And you can then use that the same way that you would use, you know, it on a Mac. So Microsoft has done those things.

Leo Laporte [01:48:41]:
I can't— they announced it. So did Google. Google's, I think, called Spark.

Christina Warren [01:48:44]:
Microsoft is called Scout.

Jeff Jarvis [01:48:46]:
Scout.

Leo Laporte [01:48:48]:
So there are— yeah, this is a very hot area right now. This is what, by the way, here's a Grok bot that I was talking to. And so it's kind of— it looks a lot like Hermes except it's running on Grok using an X account. And I have one that was designed just to fix computer problems. So that's the opposite bot. And so you could spawn more and more of these and have it do little things.

Jeff Jarvis [01:49:13]:
Or you give it an action.

Christina Warren [01:49:14]:
And that's what I kind of like about this, is you could just make them small and kind of, you know, composable. So you could have a bunch of different ones that just do this other thing. Whereas the way that, you know, I think most of us have been using, um, at least some of this stuff, is that I have like my centralized, you know, like, you know, Claw or Hermes or whatever. And maybe I have more specialized tasks I want to run it to. But in this case, I don't even have to think about that. You could just say, well, okay, this is, this is my, my bot for, you know, this sort of task and that. That's—

Leo Laporte [01:49:39]:
Hermes Desktop actually has, uh, they've added bots, which are really just an interface for what they call profiles. And I have a bunch of bots. What I ended up doing— somebody suggested this, I shouldn't read so much, uh, X content— is I set up one called the HR bot. And HR bot's job is to create, spawn new bots whenever they are needed.

Christina Warren [01:50:02]:
So smart.

Leo Laporte [01:50:03]:
I don't have to even think about a bot. Now, the advantage of a bot is it starts with minimal context. It doesn't have all of the skills of the full agent. It just has— so I have a bot, like I said, there's an ops bot that's just— it only has the skills for fixing computer problems. I have a health bot that will keep—

Jeff Jarvis [01:50:21]:
Uh-oh. Uh-oh.

Leo Laporte [01:50:24]:
Oh no.

Jeff Jarvis [01:50:24]:
No, no.

Leo Laporte [01:50:26]:
I want to keep the health information I feed in Into the bot separate.

Jeff Jarvis [01:50:29]:
In Elon's bot?

Leo Laporte [01:50:31]:
No, I don't do it with Elon.

Christina Warren [01:50:32]:
I do it with Hermes.

Leo Laporte [01:50:33]:
Oh yeah, I only do this locally.

Jeff Jarvis [01:50:35]:
Okay, it scared me.

Leo Laporte [01:50:36]:
Okay, sorry. Same thing with my finance bot. And, uh, only local. Yeah, that kind of— so I think that's a very— the pro— the— I mean, what these companies are trying to do is find a consumer market, right?

Christina Warren [01:50:49]:
Yeah, yeah. And, and I think that there is one. I think that the, the pricing will be become the interesting thing. I mean, that's— but yeah, I think this is what everybody's trying to do. I think there's a much easier enterprise story, to be totally candid, just because of what people are going to be willing to pay. I think this is also an opportunity for Apple at some point if they ever want to do, you know, something in this space where you could, you know, interact. But like in a lot of cases, like when I first played with OpenClaw, I did, I did do it on a local machine. But then I I almost immediately put it on, you know, just a VPS that I had just to see how it would work.

Christina Warren [01:51:25]:
And I was impressed. You can, for a lot of things that people need to do, kind of like what you were describing, Leo, you need just kind of these small tasks, you don't necessarily need—

Leo Laporte [01:51:34]:
It's better.

Christina Warren [01:51:34]:
The full context and other stuff. And so just being able to have something that's been kind of pre-designed for on a task-based purpose that you can kind of call on demand, A, that can be very beneficial. And be from a compute perspective, that can be very efficient because you're not having to run all these things all the time. It can just be kind of, you know, it's like serverless with computing where you're just kind of, you know, calling it when it's needed.

Leo Laporte [01:51:59]:
Yeah. That's one thing GrokBot does really well is the bots can talk to each other. So a bot can say, oh, this is about your personal finance. I'm going to give this to Finance Bot. And so forth. I think that's another thing. It's the whole idea is isolation. I think it's because we're starting to realize that if you try to get an AI to do too much, it's specialization of a different sort.

Jeff Jarvis [01:52:24]:
It's small models, which I think is the way to go.

Leo Laporte [01:52:27]:
Yeah, well, NVIDIA is, is going to build its own model. They've pooled up with Poolside, which is a very interesting model creation company, but they ran out of compute. They couldn't get enough compute. Guess who has a lot of compute? Uh, Nvidia is spending $6 billion. They see the threat that open weight models from China are posing, and they're going to make their own open weight model. They've already done Nimatron Lightning, which is quite good but small, so not so smart. And they're, they're gonna— they made a deal with Poolside to create— I hope this— I, I Far prefer to run the office.

Jeff Jarvis [01:53:05]:
It is one of those wonderful deals where they didn't acquire them. They gave them a bunch of money to acquire the licensing and a bunch of engineering staff.

Leo Laporte [01:53:15]:
Right.

Jeff Jarvis [01:53:15]:
Management stays there to keep doing stuff. It's an interesting deal.

Leo Laporte [01:53:20]:
Yeah. Yeah. I, you know, I, I really admire Jensen Huang's, uh, vision in all of this.

Christina Warren [01:53:27]:
I think it's also smart just in terms of diversification. Like, you, you want to be everywhere. Literally everyone is using your chips, but you should be part of the, you know, you should have vertical integration if you can too, whether it's open weight or proprietary. Like if you're already controlling the chips that everyone is using and everyone is even, you know, companies that don't want to use you are having to use you. Yeah, you should, you should be part of the whole conversation. I would obviously prefer for them to be open weight, but even if they weren't, I think this is an area that Nvidia needs to be in regardless.

Leo Laporte [01:53:56]:
Yep.

Jeff Jarvis [01:53:57]:
And they went into the inference chip business as well. And they're obviously in the hosting business and the software business and the OS business.

Leo Laporte [01:54:05]:
And yeah, building your own hardware is interesting. Microsoft did this with their May chips, which are some sort of weird quantum chips.

Christina Warren [01:54:12]:
Google's obviously done it with, with, with Tensor.

Leo Laporte [01:54:15]:
Google's done very well with Tensor.

Christina Warren [01:54:17]:
Very, very well. Well, in fact, that's one of the reasons why Google Cloud has been as successful as it's been in terms of you know, AI stuff is because, you know, they haven't had to solely— do they run NVIDIA GPUs? Yes, they do. And you can. But, you know, the models don't have to because Google has built their own hardware, which I think has been a big boon to them.

Leo Laporte [01:54:37]:
And, you know, that was one of the—

Christina Warren [01:54:41]:
it was a good investment.

Leo Laporte [01:54:42]:
That was one of the big points actually of the Ox Alpha release is it was all run on ZAI's Chinese Those 100 trillion tokens they offered for free for a week were running on Chinese servers, no Nvidia chips. And I think that also probably got Jensen's attention just a little bit. Now, one of the things we do on the show, Christina, is we also talk about the downsides. It's not all—

Christina Warren [01:55:11]:
Mm-hmm.

Leo Laporte [01:55:12]:
You know, cookies and cream out here in the AI world. This one actually scared me a lot. In fact, so much so that I sent an email to my friend Daniel Suarez, who wrote a book you might remember called Kill Decision about drones making— AI drones making decisions about who to kill in combat. I'm hoping— he has a new book coming out in the spring, and I'm hoping to get Daniel on the show to talk about that because he's been quite prescient. And he says, you know what, I talked— I do a lot of research on my books. I talk to the military. They're well aware of this. Well, it's actually happened now.

Leo Laporte [01:55:47]:
Uh, an autonomous drone from Russia, guided entirely by AI, killed 3 Ukrainians in the, in the war. And it was running an Nvidia chip, a contraband chip. Russia is not allowed to buy these chips, but of course there's a black market. I don't know if this is the first. It's It's the first publicized example of an autonomous AI killing machine. And I think this is just the beginning. And so not a good day for humans. In fact, they— I think the drone actually killed some innocent civilians.

Leo Laporte [01:56:27]:
They were trying to attack tanks at a gas station. They'd been trained on the tanks, but they kind of missed and they killed some civilians. And so this, according to some experts on this, is the first documented case in which civilian deaths were caused by a—

Jeff Jarvis [01:56:49]:
I don't even know how these drones— what weaponry they use. Do they come down and just explode?

Leo Laporte [01:56:57]:
Yeah, they're just flying bombs. They're kamikazes. Yeah, uh, inside it was an NVIDIA chip. Um, in fact, because it wasn't encrypted, the Ukrainian military could see exactly what models had been uploaded. Uh, they were looking at visual landmarks. They looked at the code, which revealed what kind of top targets the drone had been trained on, like propane tanks. That's what this drone was apparently aimed at. So, uh, we knew it was going to happen at some point.

Leo Laporte [01:57:32]:
Um, it's a little scary. Speaking of Grok, uh, Elon's not happy.

Jeff Jarvis [01:57:39]:
Good.

Leo Laporte [01:57:40]:
Elon, uh, first address to Cursor, because remember he bought Cursor, which by the way has— I think Cursor is behind the Grok bot, by the way.

Jeff Jarvis [01:57:50]:
Makes sense.

Christina Warren [01:57:51]:
Yeah, it makes sense. It's interesting too, we didn't mention this before, but Cursor, and I'm sure this changes now, they fine-tuned on top of KIMI 2.5 and created their model Composer, which was open weight and a very good coding model.

Jeff Jarvis [01:58:08]:
Very interesting.

Christina Warren [01:58:09]:
I'm very interested to see what, if anything, they do with future models that I'm sure that they're all in on Groq now. But that, that was kind of an interesting, you know, isn't it? I meant to, I meant to mention that earlier when we were talking about customizing model stories, is that that was a company that had done that, where they'd customized their own, you know, based on Chinese OpenWeight models. But I imagine that's probably not going to be the case anymore.

Leo Laporte [01:58:41]:
Musk talked to Cursor leadership. And said, we're falling behind and it's up to you. By the way, 45 people have left Cursor since the acquisition by xAI. But this is normal in AI now.

Christina Warren [01:58:59]:
People move around. Well, I mean, and you don't know when people's options are due and anything else. I mean, you know, if you joined the company early enough, you might have gotten a really good payday and now you have a great thing on your resume and you can go do a startup. Right? Like, there's any number of reasons why people would leave.

Leo Laporte [01:59:13]:
Yeah. All right, I think we should— poor Christina, she came here for half an hour and we've kept her here for 2, and I am very grateful.

Jeff Jarvis [01:59:24]:
Very.

Leo Laporte [01:59:26]:
Thank you, Christina. It's wonderful to have you again. Christina works hard at GitHub where she is developer relations dev advocate. What do you talk to devs about?

Christina Warren [01:59:38]:
I mean, all the stuff that we're talking about now. How are you building tools? How are things changing, right? Like, we, you know, like the beginning of this kind of conversation, there are plenty of people in this audience who don't know what GitHub is or how it works until they started, you know, working with AI tools and whatnot. And so, a lot of it is just kind of explaining those things. And sometimes it's also talking to more established developers about, okay, how are you now using AI in your workflows that you didn't before? Sometimes it's not about AI at all. It's just about, you know, somebody wanting to build something cool. So it's, it's kind of across, across the board.

Leo Laporte [02:00:12]:
Well, it's, they're very lucky to have you and we are lucky to have you. And I thank you so much for sticking around. We're gonna take a little break and do our picks of the week. If you have a pick, I didn't prepare you for this. You didn't.

Christina Warren [02:00:21]:
You did not. I will try to come up with one.

Leo Laporte [02:00:23]:
Think of a pick. I have a couple and, uh, Jeff has more than a few. You're watching Intelligent Machines. Jeff Jarvis, Christina Warren filling in for Paris. Paris will be back next week.

Jeff Jarvis [02:00:34]:
Laughing at us.

Leo Laporte [02:00:35]:
Laughing at us for forgetting all of that. Okay, okay, ladies and gentlemen, time for the picks of the week. I'm gonna give Christina a moment to prepare hers while I mention— I actually thought Paris might enjoy this— Butterbox. You know, everybody's trying to get their kids off the internet. Here's a way to do it. The ButterBox. Sharing offline has never been easier. It's a drone.

Leo Laporte [02:01:04]:
It is a hotspot that is not connected to the internet, but you could put educational materials, videos, and apps on it. And it has its own Wi-Fi network. So all the people with their phones and laptops can join a public chat room, can join the content, and it's all on Butterbox.

Jeff Jarvis [02:01:26]:
It's like a BBS.

Leo Laporte [02:01:29]:
It's a BBS. You can— you can— there's a pre-made image you can download for a Raspberry Pi or an old PC. Just download the image, burn it, set it up. Uh, they have all the information. This is from the Guardian Project. And you know what? This is good. This— you want to get your kids off the Insta, get them on a ButterBox. If I had little kids, I would probably do this.

Leo Laporte [02:01:54]:
Put some games on there. Uh, they will— they— if you don't have a ButterBox, if you don't have a Raspberry Pi, they will sell you one, or at least give you the supply list that you need. It will run a Pi Zero, will run it. A Pi 4 Plus will run it. And, and coming soon, you can run it on a PC or laptop. You can have 10 people 10 meters apart on a Raspberry Pi Zero and your butter box.

Christina Warren [02:02:20]:
So—

Jeff Jarvis [02:02:20]:
Or you could talk.

Leo Laporte [02:02:21]:
As, uh, as Julia Child would say, more butter! Jeff, your pick, your pick of the week.

Jeff Jarvis [02:02:31]:
This is the Wall Street Journal. Silicon Valley's newest status symbol—

Leo Laporte [02:02:35]:
Oh no.

Jeff Jarvis [02:02:36]:
Is the super rare Sam Altman watch.

Leo Laporte [02:02:39]:
Uh, I hope that stays rare. Is it his picture on it?

Jeff Jarvis [02:02:43]:
No, no, there's a logo in it. And, um—

Leo Laporte [02:02:46]:
Oh, wait a minute.

Jeff Jarvis [02:02:46]:
It has sayings in it like, compute is destiny.

Leo Laporte [02:02:50]:
Oh, I need this. It's a tourbillon, so you can see the inside of it.

Jeff Jarvis [02:02:53]:
That's—

Leo Laporte [02:02:54]:
how much is it?

Jeff Jarvis [02:02:56]:
Um, this one is about $650,000.

Leo Laporte [02:02:59]:
Okay, never mind.

Jeff Jarvis [02:03:02]:
Uh, he recently wore a $1.3 million From—

Leo Laporte [02:03:05]:
that was Mark Jordan.

Jeff Jarvis [02:03:06]:
I'm sorry, Meta did, right? Right. Yeah, that was Sam, you know, still before the IPO.

Leo Laporte [02:03:12]:
Dude, you have too much money if you have a $1.3 million watch.

Jeff Jarvis [02:03:15]:
I'm sorry, this one? No, I don't know. It's modeled after Vanguard's Titanium Orb watch, which starts at $180,000.

Leo Laporte [02:03:23]:
Oh, well, that's different. Can I run an AI model on it?

Jeff Jarvis [02:03:27]:
So it says compute is destiny. Another popular Silicon Valley term scaling was engraved in all caps on each crown's pusher. The OpenAI logo is featured on the watches and an inscription on the back of each.

Leo Laporte [02:03:41]:
Oh, they're selling AI.

Jeff Jarvis [02:03:42]:
If AGI aligned, deploy was written in the programming language Python. He only made 7 of them, I think it was, to give away to himself and other key I don't know if that really qualifies as being written in Python. Yeah, it's the journal.

Leo Laporte [02:04:05]:
Uh, wow, that's pretty— if AGI aligned, deploy. Oh, so Mark— not Mark, uh, Sam had these made?

Jeff Jarvis [02:04:14]:
Yes.

Leo Laporte [02:04:14]:
And gave them to executives. That's, that's nice.

Jeff Jarvis [02:04:17]:
Well, yeah.

Leo Laporte [02:04:18]:
Well, did Demis ever give you anything at DeepMind?

Christina Warren [02:04:22]:
I got a hoodie, but that was not from him. But yeah, I got, I got a water bottle and a hoodie.

Leo Laporte [02:04:28]:
You may not know this.

Christina Warren [02:04:30]:
I got a Nano Banana sweatshirt at one point too. That was pretty cool.

Leo Laporte [02:04:34]:
So that fits in with— you don't know this probably, Jeff, but Christina has a collection of—

Jeff Jarvis [02:04:40]:
Oh, that's Christina's collection, right?

Leo Laporte [02:04:44]:
Yes. And Nano Banana could be— did you ever get your Theranos gear?

Christina Warren [02:04:49]:
Oh yeah, yeah, yeah. Somebody bought me a Theranos fleece that was very expensive, and I'm very grateful that they bought it for me. And yeah, I have an Enron mug. I'm actually today, this is funny, I'm wearing a shirt for Atom, which was GitHub's dearly departed—

Leo Laporte [02:05:08]:
Oh, I love the Atom editor. That was a great text editor.

Christina Warren [02:05:11]:
It was a great text editor, part of Electron. And, you know, when Microsoft acquired GitHub, it did not make sense to have the much more successful VS Code.

Leo Laporte [02:05:19]:
Yeah.

Christina Warren [02:05:19]:
And also Atom. But yeah, I obviously had to buy an Atom t-shirt. I had to buy merch for our canceled products because I think that's funny. So yeah, I have a whole thing of things. A viewer actually sent me— this was years ago now. I wish I could remember who it was. I could thank them. Sent me TechTV stuff, Leo.

Leo Laporte [02:05:40]:
I have a few TechTV things.

Christina Warren [02:05:42]:
So your closet is basically Yes, that's exactly what it is. Thank you. And thank you for getting the reference. That's exactly what it is. My closet is effed company. And it's, it's been a thing that I've been collecting for I don't even know how long. But the problem is, is that when companies go bankrupt or something happens, because NPR did a story about me about this, I've created a market where I have to compete against myself because other people will literally be looking at collectors like me to drive the price down. up.

Christina Warren [02:06:11]:
I've literally like made it harder on myself to acquire effed company merch.

Leo Laporte [02:06:15]:
Oh my God, you should have gotten the Twitter sign, man.

Jeff Jarvis [02:06:18]:
If you got the Twitter sign, that would have been huge.

Christina Warren [02:06:20]:
Oh, I tried, I tried to get things from Twitter and the, the, the, um, that, that thing was just crazy. But you know, I tried. I do have Twitter merch though. I have a lot of Twitter merch from, from the various eras.

Jeff Jarvis [02:06:31]:
With birds, not X's. What's he getting? What's Leo gonna show us?

Leo Laporte [02:06:35]:
Yeah, I was gonna get you— I have a plaque that I bought that said Leo Laporte had a blue check before you could buy them, but it's nailed to the wall, unfortunately.

Christina Warren [02:06:46]:
That's so funny.

Leo Laporte [02:06:49]:
I have a pick for you, Christina. Christina vibe-coded a GitHub game.

Christina Warren [02:06:54]:
Oh yeah, this was cute.

Leo Laporte [02:06:55]:
Yeah, tell us where that is.

Christina Warren [02:06:57]:
That is at, uh, it's at filmgirl.github.io/ I'm on your GitHub. Let me find it.

Leo Laporte [02:07:06]:
It's in one of your— oh, you have 60, almost 60 repositories.

Christina Warren [02:07:10]:
Oh, I have so many more than that.

Leo Laporte [02:07:12]:
Flappy Copilot. There it is.

Christina Warren [02:07:13]:
It's filmgirl.github.io/flappy-copilot.

Leo Laporte [02:07:18]:
Look at this, Jess.

Christina Warren [02:07:20]:
I built that with Kimmy and it, I don't know, I think it was maybe, it took me less than 10 minutes. It was, and it, and it, and it—

Leo Laporte [02:07:27]:
What's great, it has all the GitHub Like, there's GitHub stuff, like, yes, merge conflict, you hit a breaking change.

Christina Warren [02:07:35]:
Super cute.

Leo Laporte [02:07:36]:
Am I gonna get in trouble for the music? Whose music is this?

Christina Warren [02:07:41]:
I don't know whose music that is.

Leo Laporte [02:07:42]:
Oh, that's not your game?

Christina Warren [02:07:44]:
No.

Leo Laporte [02:07:45]:
Oh, I probably have something else open. Never mind, let's stop it right now.

Christina Warren [02:07:49]:
My music is just, is, is just, it's just an HTML5 app. It's just, it's just beeps and boops. It's just whatever the model put in, but it is very funny.

Leo Laporte [02:07:56]:
I probably still have Suno running from the previous show or something. Anyway, very cute, very cute. That's on Christina's, uh, uh, Film Girl GitHub repo.

Christina Warren [02:08:07]:
Yeah, I just gave it— just gave it that URL. But yeah, um, it— yeah, that was just a, a dumb little, you know, thing that I— again, it's so cute. Yeah, it works on mobile actually, which is— I was very impressed with the Model 4 because I've been not expecting that. The fact that I was like, oh, okay, it actually works on mobile. Okay, cool.

Leo Laporte [02:08:28]:
Of course it does. Kimmy's brilliant. What are you using for your model for Hermes?

Christina Warren [02:08:35]:
I use SOL usually, but it just kind of depends. Well, it depends on what I'm calling to. So, if I'm locally, I'm using GLM. But if I'm calling out, then I'm using GPT.

Leo Laporte [02:08:46]:
I may be using GLM myself. I might be— it looks pretty good on here.

Christina Warren [02:08:51]:
I've used the older version of GLM is what I've done, actually, because you, you told me about it and I was like, oh, I should try that.

Leo Laporte [02:08:56]:
I really like it.

Christina Warren [02:08:57]:
Yeah, I like it too. So that's what I switched over on my framework, my local model. Yeah.

Leo Laporte [02:09:01]:
Yeah, I want to— the, the, um, context is a little small. DeepSeekV4 Flash has a much larger context. So yeah, to figure out if I can, I can tweak it a little bit. You know, it's funny, if you, if If you're around on a day that a model gets released, go over to X and watch everybody compete to create the better recipe. I had 3 different recipes to try before 9, and it was incredible.

Christina Warren [02:09:27]:
Yeah, no, it's crazy to see everybody's stuff come out and everybody's prompts and everything else. It's a fun time.

Leo Laporte [02:09:32]:
People are great. They're just really into this, and it's a very exciting time to be into AI. And Christina Warren, I thank you so much for spending so much of your afternoon with us.

Jeff Jarvis [02:09:42]:
Stepping into the lurch.

Leo Laporte [02:09:43]:
Thank you.

Christina Warren [02:09:43]:
Thank you for having me. This was a delight. I was, I was not expecting to be able to stay the whole time, but I was able to, and this was a delight. Thank you both for letting me.

Jeff Jarvis [02:09:50]:
You were a delight.

Christina Warren [02:09:51]:
Thank you for letting me rant about—

Jeff Jarvis [02:09:53]:
Oh, thank you.

Leo Laporte [02:09:54]:
Well, you know what? I was glad we got you on because I've been saying that all along. And, and I was maybe mocked a little bit for it. But anyway, thank you, Kristin.

Christina Warren [02:10:05]:
Also, Jeff, I bought your book. I I can't wait to read it.

Jeff Jarvis [02:10:07]:
Aren't you wonderful?

Leo Laporte [02:10:08]:
Look at her.

Christina Warren [02:10:09]:
Genuinely, like, when you talked about at the top of the show, I went, oh, this is completely my, my ish. Like, I have to read this. So thank you so much.

Leo Laporte [02:10:17]:
It's really fun. It's really—

Jeff Jarvis [02:10:19]:
it's a, it's a technology and media. So it's—

Christina Warren [02:10:21]:
yeah. Yes, exactly. Which is my 2 favorite things, as I know they are your, your favorites as well. So I'm— I can't wait to read it.

Jeff Jarvis [02:10:27]:
So I'll take advantage of the plug, the free plug you just gave me, and just mention that if you want an autographed copy If you go to jeffjarvis.com, you'll see the link. You can go to Montclair Book Center and order from them, and they will send you an autographed copy that I will go and autograph. Does our discount code still work? No.

Leo Laporte [02:10:46]:
Okay, end of the month.

Jeff Jarvis [02:10:47]:
No, end of the month it does. It does end of the month.

Leo Laporte [02:10:49]:
Oh, you have just a few more days to use GLRBD8 and get a discount.

Jeff Jarvis [02:10:56]:
Yeah.

Leo Laporte [02:10:57]:
Thank you, Jeff Jarvis. I appreciate it. Will you be back next week? You're not going to a going-away party for anybody, are you?

Jeff Jarvis [02:11:04]:
No, September 9th I'm away.

Leo Laporte [02:11:05]:
Okay. Okay.

Jeff Jarvis [02:11:07]:
Is it— Paris texted me and said that the, the, the retiring colleague cried when they gave her the gift, so she was glad she was there.

Leo Laporte [02:11:15]:
Yeah, that's great. And Christina, we will see you on Tuesday on MacBreak Weekly. We'll see you all here next week. We do, uh, Intelligent Machines every Wednesday, 2 PM Pacific, 5 PM Eastern, 2100 You can watch us live on YouTube, X, Facebook, LinkedIn, Kick, uh, and twitch.tv, where they are now at this very moment training on this show.

Jeff Jarvis [02:11:38]:
And we're happy with it.

Leo Laporte [02:11:39]:
We're happy with it.

Jeff Jarvis [02:11:40]:
We would be insulted if they didn't.

Leo Laporte [02:11:42]:
It's Creative Commons. Eat it up. Uh, we also have a YouTube version, which is probably training somebody else. Uh, you'll find that, uh, at youtube.com. Actually, if you go to youtube.com/ Twitch. All the shows have their own channel. And .twit, not Twitch. youtube.com/twit.

Leo Laporte [02:12:01]:
Uh, all the shows have their own channels, as— and there's Shorts on you on Twitch and so forth. Twit. Uh, I got Twitch on the mind. Uh, let's see what else. Oh yeah, subscribe in your favorite podcast client. That way you can get it automatically and you don't have to worry about when it's on next. You just have it and you can listen At your leisure. Thank you all for joining us.

Leo Laporte [02:12:22]:
We'll see you next time on Intelligent Machines. Bye-bye.

Christina Warren [02:12:27]:
I'm not a human being, not into this animal scene.

Jeff Jarvis [02:12:34]:
I'm an intelligent machine.