Intelligent Machines 896 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. Paris is here, Geoff Jarvis is here, and oh yes, I'm here, and I'm very excited. Our guest this week on Intelligent Machines is Ash Hart, who has created something called TensorFold, something that's getting local AI to speed up 2 or 3 times faster. He's a true local AI hero. Next on Intelligent Machines.

Jeff Jarvis [00:00:25]:
Podcasts you love.

TWiT.tv [00:00:27]:
From people you trust.

Leo Laporte [00:00:30]:
This is TWiT. This is Intelligent Machines, episode 896, recorded Wednesday, November 11th, 2026. Culling the Herd. It's time for Intelligent Machines, the show where we cover the latest in AI, robotics, and all the smart doodads and doohickeys and, uh, gee-gaws surrounding you right now as you are enjoying this podcast. Paris Martineau is here, ladies and gentlemen. Hello.

Paris Martineau [00:00:59]:
I'm happy you're back, Leo.

Leo Laporte [00:01:02]:
I'm happy to be back. I missed you guys. 3 weeks is a long time, but thanks to all the people, Jason Heiner and Father Robert and everybody who filled in for me.

Paris Martineau [00:01:11]:
There's no one that can replace Leo, but people made a good effort.

Jeff Jarvis [00:01:15]:
They did.

Paris Martineau [00:01:15]:
We missed you.

Leo Laporte [00:01:16]:
Well, this is my favorite show. I really like, as you know, as you might guess, I really like talking about AI.

Paris Martineau [00:01:20]:
Based on the amount of AI-related updates you sent our podcast group chat while you were on vacation, it's clear that we're your favorite

Leo Laporte [00:01:29]:
show. Yes, my favorites. Also here, Mr. Professor Emeritus Jeff Jarvis, Professor Emeritus of Journalistic Innovation at the Craig Newmark Graduate School of Journalism at the City University of New York. Craig, Craig Newmark. Author of Hot Type, his brand new book available now, the magnificent machine that gave birth to the mass media. Also the Gutenberg Parenthesis magazine, the Web We Weave, many other— What Would Google Do? Going back more than 10 years.

Jeff Jarvis [00:01:58]:
That's where it all began with us.

Leo Laporte [00:01:59]:
You know, it's funny, Google's back in the, uh, back in the good graces of AI. Their new model, Carbon, is really good. Argon was good. I am— enough of that. I am so excited about our guest today. I have been dying to get this guy on. His name is Ash Hart. He's on x.com as AshxHart.

Leo Laporte [00:02:19]:
You've probably never heard of him. Most people haven't. unless you're deep in local AI, in which case, if I say the word TensorFold, you'll go, oh, that Ash Hart, ooh. TensorFold is a bit of a, little bit of a miracle in the local AI scene. Joining us from the UK, Ash Hart. I'm so pleased to meet you. Thank you for being here.

Ash Hart [00:02:42]:
Yeah, thank you for having me.

Leo Laporte [00:02:44]:
So you, I'm really curious how you came up with this because it's kind of out of the blue, right? You're an engineer, software engineer, Right. AWS infrastructure, DevOps, networking. Yes.

Ash Hart [00:02:57]:
Yes. It was my day job. I handled DevOps, cybersecurity, networking, engineering. That is my bread and butter essentially.

Leo Laporte [00:03:05]:
And you're still doing that?

Ash Hart [00:03:07]:
Still doing that. Yeah. Plodding along.

Leo Laporte [00:03:10]:
When did you get into local AI? How did that happen?

Ash Hart [00:03:14]:
So this is quite wild. So I, Got into ChatGPT, I would say in 2025, early '25. I was copying code from the web browser and I was like, hmm, there's gotta be a better way to do this. So I moved over to the CLI, probably the back end of '25. Codex. 2026. Yeah, Codex. And then I moved over to Claude and I was like, this is— I had a, a moment where I saw something that we would write that I was sort of proud that I would write.

Ash Hart [00:03:50]:
And I was like, yes, okay, this is something.

Leo Laporte [00:03:53]:
I'm gonna guess that moment was November 24th of last year, because that's when Opus 4.5 came out.

Ash Hart [00:04:01]:
It was, uh, I think it was Opus 4.5, to be fair. Yeah, you might be right.

Leo Laporte [00:04:05]:
I had the same experience. It's like, what?

Ash Hart [00:04:08]:
Yeah, it was, it was that threshold that it crossed. So, so moving along, I then started hitting token usage. I was fanboying over AI. I was, I was agent swarming just for the fun of it. And then I would say probably February, I felt the need to start running locally. So I bought a Mac Mini. I think it was 24 gigs, M4 Mac Mini, and I run GPT-OS S20B. And I was like, this is for me.

Ash Hart [00:04:37]:
I like this. Wasn't very good quality. but I enjoyed it. And from then on, it's been a bit of a—

Jeff Jarvis [00:04:44]:
What were you doing with it in those early days?

Ash Hart [00:04:46]:
So I like to listen to a lot of podcasts, YouTube videos, that sort of thing, but you can't do it all. So I was getting my agent that I built to scrape, so like text to audio, audio to text, to build me like a news report of podcasts that I enjoy. So I can have like a quick brief, any new models that were dropping, things like that, just like a chief of staff environment. So it would just do the stuff that I would normally do myself, and it just brings it all to me. But that was running overnight because it was quite slow.

Leo Laporte [00:05:20]:
So in order for people to understand what you've done, which is kind of amazing— I mean, this is— you're a sole developer working on your own, right? I mean, it's not— there's not a big company behind you. You're just doing this in your spare time, kind of a hobby. You wanted to make these models, what, faster? Was that your goal?

Ash Hart [00:05:38]:
faster. And to a point, it has to be without quality loss,

Leo Laporte [00:05:44]:
right?

Ash Hart [00:05:44]:
You will find every other engine out there will employ weird mathematical equations to skip a few bits here, skip a few bits there, and it makes it quicker, but the output is not what the model would have given perfect serial decode. And that is what I couldn't lose. There's no point having speed when there's quality degradation, at least that Is what I think.

Leo Laporte [00:06:08]:
So to understand what Ash has done here, you kind of have to understand how these LLMs work. Um, basically the way an LLM works is you give it the prompt, but the prompt also comes with a bunch of contexts, you know, from your, your engine, your CLI, from Codex or from Claude Code or from Hermes, whatever you use. So that context plus what you typed goes in and then is Goes through the weights and comes out the other end with a token. It spits out a token, the next token. It's a prediction machine. And then it goes back and does the same thing over again and spits out the next token. And then it goes back each time going through all of that content and adding a little bit more on each time. Now you might think that's really slow.

Leo Laporte [00:06:54]:
Early on, people figured out, well, there's a way to speed this up. And anybody who's done any coding— I learned this when I started doing Lisp coding, knows about closures or the idea of— they call it in AI a KV, a key-value cache. Sometimes you call it— it could be used as a hash table. In Python, you'd call it a dictionary. The idea is we're going to take all this stuff that we processed and we're going to make a dictionary of all the results. And instead of processing it all again, we're going to look it up in the KV cache. That sped it up an awful lot. It was kind of critical to it.

Leo Laporte [00:07:29]:
But at that point, you're kind of done speeding things up. And if you have a dense model, you have to go through every weight. If you have— that's why everybody's using MOEs, mixture of expert models, now, where you just do a little bit of a few weights for each token. That speeds it up, but you're kind of done. There's not much you can do. People— another technology feature that we've known about for a long time in processors, the idea of speculative decoding or ahead of time deciding where a branch might— branch prediction, you call it. So as you're going through your code, you say, well, I'm going to have some branches up ahead. I'm just going to say, what if I did that, took that fork in the road? I'm going to calculate that ahead of time.

Leo Laporte [00:08:16]:
So when it gets there, and if you do it properly, that could speed things up a lot. And it turns out That kind of speculative execution happens also in AI, and there are a number of engines that do this. Was that the insight that you said, I wonder if I can do this with TensorFold? Was that what you were trying to do?

Ash Hart [00:08:35]:
Kind of. So the way I thought of TensorFold was you will send in, let's say, one prompt and you are writing, it keeps reading it over and over and over again until you get your token output. It's quite inefficient.

Leo Laporte [00:08:50]:
Yes.

Ash Hart [00:08:50]:
But however, if you launch subagents or you have multi-concurrent connections, you can have a lot more tokens. So the silicon can do it. It's just the way the software is using the model. So my initial idea, and I actually cracked it yesterday, which I'll talk about in a bit—

Leo Laporte [00:09:11]:
What?

Ash Hart [00:09:11]:
Is, yeah, so, so the current TensorFlow is, it's kind of half of what I wanted to do. So essentially what my vision was, was you basically have a prompt that say it's 512 tokens, you send that in, we split that up into 16 lanes, we compute them 16 lanes in a diagonal line so that you all know about each other. So you have to do one then the other, but then you cascade them together as such in a forward pass, but then you read them all at once. So What you're getting is 16 lanes doing the compute rather than one lane.

Leo Laporte [00:09:55]:
And it's not, it's not, it's not lookahead, it's the current set of context.

Ash Hart [00:10:00]:
Yeah, exactly. So,

Leo Laporte [00:10:03]:
so instead of a linear operation, you're parallelizing it basically.

Ash Hart [00:10:07]:
Exactly. And it's, yeah, someone coined it yesterday. I think it was token parallelism or something along those lines, they said. But essentially, I'm stitching multiple streams together.

Leo Laporte [00:10:19]:
Well, the proof is in the pudding. I use GLM-53 Flash. It's a fairly large model. I'm using, I think, 3 EXL, 3-bit quant.

Ash Hart [00:10:27]:
Yep.

Leo Laporte [00:10:28]:
So it's, you know, the weights are really compressed down, but it turns out you can compress these weights a lot. The native weight usually is 16 bits. You can compress it down, believe it or not, to 3 bits and still get like 95% of the intelligence. But that makes it much faster, as you might imagine. So I went from maybe 30 or 4, you know, maybe 20 to 30 tokens per second output decode. Suddenly I'm using GLM Flash. This thing is lightning quick. It's doubled in its speed.

Leo Laporte [00:10:57]:
You doubled the speed, making this a very usable model on 2 Sparks. Unbelievable.

Ash Hart [00:11:05]:
And we're not finished, to be honest. I would say we are around 30% of the throughput. That is actually available.

Leo Laporte [00:11:16]:
Now you say we, is it— it's not just you anymore?

Ash Hart [00:11:19]:
Um, it's in terms of the actual engine, it's me and anyone who collaborates.

Leo Laporte [00:11:25]:
Had you ever done anything like this before, Ash?

Ash Hart [00:11:29]:
Not in terms of a wide project on GitHub. I've done bits and pieces. I've messed with, um, let's say being able to update weights of a model when it's already in memory. So we're not using KV, bits, weird bits and things, but I've not—

Leo Laporte [00:11:44]:
yeah. Actually, you have a very interesting question.

Paris Martineau [00:11:45]:
What's your day job?

Ash Hart [00:11:47]:
I'm a CTO.

Paris Martineau [00:11:49]:
So, see, around the same, uh, field, even if it's not the direct, uh, thing you're doing. But you've been working in this industry for a while?

Ash Hart [00:11:58]:
Yeah, so essentially I look after 8 countries in terms of cybersecurity.

Leo Laporte [00:12:03]:
Oh, wow.

Ash Hart [00:12:03]:
And yeah, so I've got quite a breadth of knowledge in that.

Leo Laporte [00:12:08]:
But not a coder particularly?

Ash Hart [00:12:10]:
I was a coder in terms of, I used to do PHP, bit of Python. I used to use MySQL. Uh, what else did I do? JavaScript. I couldn't ever really get my head around JavaScript.

Leo Laporte [00:12:23]:
TensorFold was in Python originally, is now in Zig, right?

Ash Hart [00:12:26]:
Yeah.

Leo Laporte [00:12:28]:
Which is a better language for this. It's a more system level, system level language maybe.

Ash Hart [00:12:32]:
Yeah, it's system level, but it gives me direct API access to Metal, to CUDA, so I can do these wonderful things.

Jeff Jarvis [00:12:38]:
Did you write it or how much help did you get from AI to write this?

Ash Hart [00:12:43]:
A lot. Most of it's written by AI.

Leo Laporte [00:12:45]:
That's, by the way, that's really interesting. That's what's happening is a single individual like Ash, a smart, obviously smart individual, knows what he's doing and knew where to point the AI, is able to do so much more because the AI can help him.

Jeff Jarvis [00:13:00]:
How long would that coding project have taken without the AI help? Old style?

Ash Hart [00:13:07]:
Old style, it would have been at least a year.

Leo Laporte [00:13:09]:
You wouldn't even take it on, probably.

Jeff Jarvis [00:13:11]:
Yeah, you wouldn't bother.

Ash Hart [00:13:12]:
I wouldn't. Or I would probably have stuck with one model, one lane or 2 lanes, and that would've been it. Whereas now it's—

Leo Laporte [00:13:19]:
And that's the other interesting thing I should point out. This is for CUDA. I'm using it on Sparks. It's also for Metal. So you can use it on a Mac. Uh, can you use it with, uh, Rockham and AMD? Does it work everywhere?

Ash Hart [00:13:31]:
Works everywhere.

Leo Laporte [00:13:33]:
That's kind of mind-bending too.

Ash Hart [00:13:36]:
It's a lot to manage.

Leo Laporte [00:13:38]:
I, I'm just, I mean, it's just, it's a thing is until you see it at work, you don't, it's kind of magic. When you apply TensorFold, suddenly everything speeds up. Is it typically twice as fast?

Ash Hart [00:13:52]:
Sometimes it can be upwards of 4 times. It just depends what and how the model is. Let's say DeepSeek 4.1, which you use the n-gram table as a lookup table, you can sort of offload that to NVMe or however you want to do it, and that can be sped up even more. So it just depends. It's very architecturally driven.

Leo Laporte [00:14:13]:
It's really— were you surprised when you first started getting these results? Like that?

Ash Hart [00:14:19]:
Yes. Yes and no. No, in the sense that I knew it was possible. Yes, in the sense that I finally— so I fought with Claude and Opus, all of the AI models. You wouldn't believe my workflow to try and get this to actually work. was, was I was bouncing off Gemini on the web, Quen on the web. I was running local models. I was trying to black red team it.

Ash Hart [00:14:46]:
I was trying so many different ways to try and make this work. And it was Opus 5.5, I think it was, that I said, okay, review this. This is how I want it to work. And it was like, okay, no, we're definitely doing it wrong. And then we built a proof of concept and it worked on— So I went for Quen 27B. because that's dense. It's what most people will probably use. And if I could speed that up— so I think serial decode was 22 on my M5 Max.

Ash Hart [00:15:15]:
I, I think it was yesterday actually, I did 130. Holy cow. Yeah, and that's on my M5 Max.

Leo Laporte [00:15:22]:
Holy cow, that's 6 times faster.

Ash Hart [00:15:25]:
And that is ridiculously usable.

Leo Laporte [00:15:28]:
And that's a dense model, which is, you know, those are much slower because you're Moving every token through every—

Ash Hart [00:15:34]:
exactly.

Leo Laporte [00:15:35]:
Wait, um,

Paris Martineau [00:15:38]:
I

Leo Laporte [00:15:38]:
actually— why didn't these— why didn't these companies do it themselves?

Ash Hart [00:15:42]:
It's quite easy to answer that. So if you're a big company, you're not bothered about the single stream, you're bothered about concurrency, many users using that box to get as maximum throughput.

Leo Laporte [00:15:56]:
Because you're on a date in a data center where you're selling this to hundreds of thousands of customers.

Ash Hart [00:16:02]:
Exactly. So I applied this method last night. I actually got access to Carbon, and I'm pretty impressed with it.

Leo Laporte [00:16:10]:
Carbon's amazing. This is Google's new Gemini 4.

Ash Hart [00:16:13]:
Yeah. So I had it review my project and I said, okay, this is what I'm trying to do. I've got a new 5090 rig. What can we do here? So I was looking through the kernels, having a little look, and I said, right, okay, let's try and get 27B running. We did. We got it native Windows. It ran, I think, 200-ish tokens per second. Which is okay.

Ash Hart [00:16:32]:
It's a 5090, I'd expect that sort of throughput. But then that's when the, the lane and the diagonal batching came along. And yesterday and today I've seen upwards of 1,600 tokens per second.

Leo Laporte [00:16:47]:
Oh my God.

Ash Hart [00:16:48]:
Ridiculous amounts of— and my access has gone crazy.

Leo Laporte [00:16:52]:
They're not incented to do this.

Ash Hart [00:16:56]:
No.

Leo Laporte [00:16:56]:
But if you are at home running on your own hardware, this is huge.

Ash Hart [00:17:01]:
Exactly. And I think this is going to grow even more because you don't want to be offloading all of your AI to the cloud. I think cloud is for planning, maybe really intense frontier bleeding edge stuff. 90% of your stuff can be done at home. GLM 5.3, what you run, incredible model.

Leo Laporte [00:17:21]:
It really is. I think this is important because people often say, oh yeah, the local's catching up with frontier. Frontier is always good. You're never going to be able to run a trillion terabyte model at home. Or not never, but not anytime in the near future. But you don't need a trillion terabyte model for, uh, most of what you— 99% of what you do. So if you can use a little bit of cloud cheaply and then do everything else locally, um, that's really the way to go. Now here's an interesting— so you started this with a 5090, you said?

Ash Hart [00:17:57]:
So my— that was yesterday. That's what I was doing with the 5090.

Leo Laporte [00:18:01]:
Okay.

Ash Hart [00:18:01]:
That's new. Yeah, so I've had the— and this was from a donation from a contributor. He donated—

Leo Laporte [00:18:08]:
you're in the, uh, you're in the NVIDIA Inception, uh, program. Did they give you any hardware?

Ash Hart [00:18:14]:
No, not yet. No, no.

Leo Laporte [00:18:15]:
What do they just give you, a pat on the back? What

Ash Hart [00:18:17]:
is— They've pointed me in a couple of directions, but nothing has been concrete. I mean, I've got enough hardware at the moment.

Leo Laporte [00:18:25]:
But if you think about it, this sells Sparks. This sells NVIDIA hardware. It's very good for NVIDIA. I bought 2 Sparks. That's $10,000 worth of hardware.

Paris Martineau [00:18:36]:
I mean, I was gonna say, Leo, you were running TensorFlow on your— TensorFold on your 2 Sparks, right? How did you first become aware of this? And what did it do for your setup?

Leo Laporte [00:18:50]:
I'm not a fan of, uh, Elon and I'm not a fan of X. And as you know, in November of '24, I gave up X. But that's where all the people, people like Ash and Mia and Tony DeWild and all of the people who are really doing this stuff, that's— were you on X before, Ash, or did you— I think you're a relatively new member. I think I looked and you joined recently.

Ash Hart [00:19:10]:
My, my account was from 2014, pre-X.

Leo Laporte [00:19:14]:
Yeah,

Ash Hart [00:19:14]:
but I didn't really use it. I used it for cars when I was into cars, and then I sort of left it, and then I came back to it. I was like, oh, I've got this account. And it was a friend of mine that said, okay, there's a lot of people on X. That's where the, the people are building. Just for some context, for anybody who's building anything on, on open source, X is the place to be.

Leo Laporte [00:19:35]:
Yeah. I wish—

Ash Hart [00:19:36]:
regardless of your feeling.

Leo Laporte [00:19:37]:
I wish it weren't. I wish Blue Sky or something else was, but it isn't. And X is. And so that's how I found out about TensorFold. Uh, and it actually was from following probably Mia. We're really trying to get Mia on. She's very secretive, but I would love to get her on. I know you work with her.

Leo Laporte [00:19:51]:
She does the best recipes. And most, many of her recipes are TensorFold-based. She's doing the same thing as you in a sense, because she's using AI to tune these models and put out recipes people can easily install themselves, not by hand, using an AI. You point your AI at the GitHub repo and it says, and you say install this and do this and it does it. So, here's another thing that happened. Apple, after the— after I— like 2 weeks after I bought my Sparks, announced these new Mac Studios with very fast memory bandwidth, 1.2 terabyte per second memory bandwidth. And I went, oh crap, I bought the wrong thing. But then I— after, you know, thinking about it, I realized pre-fill is bad on Macs because the GPUs aren't as good as the the GB10s and the Sparks.

Leo Laporte [00:20:44]:
And so I'm going back and forth. You decided— I think this is really good— you had the, uh, the little M4 Mini. You decided that you wanted to get one of the M5 Ultra Studios. In fact, you even ordered it.

Ash Hart [00:20:55]:
Yeah, so just a bit of a backstory. I've actually got an M3 Ultra, which I bought after the Mac Mini. I was at a kid's party thing and I saw it on eBay. I was like, I have to have that. Credit card, bang, buy.

Leo Laporte [00:21:10]:
That was until the new M5 Ultra. That was the, uh, the Mac to get for local AI. Absolutely.

Ash Hart [00:21:16]:
It, it was. And, and that was sort of where I sort of fell in love with it all. I was running Quen, um, Next Flash. I think it was like nearly 178 tokens per second.

Leo Laporte [00:21:28]:
But it's not just— see, we should really point this out though. And if you go to x.com, you'll see people talking about these tokens per second. It's not just the decode speed. You also have to have the intelligence.

Ash Hart [00:21:37]:
Exactly. Yeah. And this is something you have to make a decision on. So for me, Quen 3.8 NextFlash is about 80 to 85% as good as GLM, but it's like 3, 4 times quicker. So that's what I lean to, especially when I'm doing— Okay. If I'm doing a bit of red teaming and doing a bit of trying to— I was pointing it at my HomePod to try and reverse engineer it, which I managed to do.

Leo Laporte [00:22:06]:
Nice.

Ash Hart [00:22:06]:
So,

Leo Laporte [00:22:07]:
and do you use an obliterated model or do you use this stock model?

Ash Hart [00:22:10]:
Uh, it depends. Generally DeepSeek, uh, is fine. You can pretty much do that as it is. It's very, uh, permissive to say the least. GLM 5.3 is okay if you, if you use a red team skill, it sort of warms up to it. Whereas, um, NextFlash is again 50/50. I think it depends how it's feeling. It just sometimes it works.

Ash Hart [00:22:34]:
Isn't

Leo Laporte [00:22:34]:
that funny that these models, some, they just, they just act differently sometimes. It's very weird. So you, so when the Mac Studio, the M5 Ultra, 512GB Mac Studio became available, which it did for me at 1 AM when I'm sailing, uh, down the Java Sea and I'm, I'm sitting in my bunk and I'm looking at this. You ordered it.

Ash Hart [00:23:02]:
So I knew it was dropping, but I didn't know when. So I was sitting with the kids, uh, as you do, just have a quick browse on X. I was like—

Leo Laporte [00:23:12]:
How does your wife feel about all of this, by the way?

Ash Hart [00:23:14]:
There's a lot going on.

Leo Laporte [00:23:17]:
My husband, the lunatic.

Ash Hart [00:23:19]:
Stress. Stress.

Leo Laporte [00:23:20]:
Let me tell— let me tell her, Mrs. Hart, You are married to the Albert Einstein of local AI, and I'm sure Albert Einstein's wife was pissed off from time to time, but Ash is making a huge difference in the world and we're very grateful to him. So thank you for letting us borrow him.

Ash Hart [00:23:41]:
It's— and my view on all of this is it's fun. I've seen people use TensorFlow in real-world environments. There's, there's people in Russia who use the engine for medical research, and they're running it 3 times quicker so they can process 3 times quicker and save people. It's— that's what I want.

Leo Laporte [00:24:04]:
A couple of days ago, David Hanemeyer Hansen announced that you were going to be a part of, uh, Omarchi, his Linux distro, uh, that it will come with TensorFlow built in.

Ash Hart [00:24:15]:
I looked at that and I was just like No, this, this can't be real.

Leo Laporte [00:24:18]:
The creator of Ruby on Rails. That's pretty cool.

Ash Hart [00:24:20]:
Yeah. And I'm talking to him on X. I talked to him via DM. He seems like a really nice guy. I mean, from the— I actually follow him from his racing car days.

Leo Laporte [00:24:30]:
Yes, he was into cars. He still is. Yeah, yeah,

Ash Hart [00:24:32]:
yeah. So I follow him from then. But no, it's, it's wild. This is the thing about X. You can get in and talk to people who you would never, never talk to. I'm talking to people at Apple. who have actually invited me down to Battersea in the UK in their headquarters.

Leo Laporte [00:24:49]:
That new Battersea Apple headquarters is beautiful. They really did a nice job. Yeah,

Ash Hart [00:24:53]:
yeah, they've invited me down to there, which is incredible. I'm talking to the guy at Nvidia who deals with all of their GPUs. He's sending me some bits and pieces over. It's just, it's wild.

Leo Laporte [00:25:06]:
So unfortunately, uh, you, you applied for a loan, which I feel bad about, to buy this $25,000 computer. I, you know, I, I was— it's 1 AM, I'm clicking the button and it, and it says 96 gigs or 512 gigs. And I clicked the button and it said, oh, that's $11,000 more for the 512 gigs. And I went, what? But you— that didn't stop you. You pressed the order button, didn't get the loan. I saw your tweet on X. You said, I, I don't— I can't get it.

Ash Hart [00:25:38]:
So I

Leo Laporte [00:25:38]:
don't have the money.

Ash Hart [00:25:40]:
So what happened was I applied and on my app, on my banking app, it says you will be eligible for this and this. So I did like a pre-approval and I actually did it up to $25K and it said, yep, that's fine. So I was like, right, okay, when it drops, I'm gonna order this no matter what. And when it dropped, I ran straight upstairs, I pressed the order button and it came up. Was it the 17th of November?

Leo Laporte [00:26:03]:
Yeah, you got it right away because that's the first day it's available. Yeah,

Ash Hart [00:26:06]:
yeah,

Leo Laporte [00:26:06]:
a week from today. Yeah.

Ash Hart [00:26:08]:
I couldn't believe it. So I pressed order. Great. So, okay, I did the loan application and it said approved. Perfect. We'll just do some final checks. They even emailed me the PDF with all of the terms of the— I think it was £600 a month or something. Yeah, £600 a month.

Ash Hart [00:26:27]:
Perfect. It's a lot, but perfect. And then on the Monday they rejected it and I was just like, oh, I was gutted.

Leo Laporte [00:26:37]:
I saw the post on X and you said, I can't get, I can't get it. And my heart fell because I, I want you to make TensorFold for these Macs. Uh, and then you said, but some friends have said I should do a crowdfunding. I have a Ko-fi account, so, uh, I'm gonna do it. I immediately went over there, put a little money in. I guess I wasn't alone. How long did it take you to raise the $25,000, £24,000?

Ash Hart [00:27:04]:
I think from end to end, probably around 21 hours. I think it was

Leo Laporte [00:27:09]:
like that. And that'll tell you something about— there are hundreds of people who value what you're doing so much they're willing to give you money so that you will make this. It's an open source. Apple— what the hell is Apple thinking? Why isn't Apple sending you 4 of these? One of the things you're also doing, which is really interesting, is taking a Mac and a Spark and connecting them together, having the Spark do the pre-fill, which is very good at, having the Mac do the decode, which it's very good at. What do you call that? MCDMA?

Ash Hart [00:27:42]:
Yeah, so Mac CUDA Direct Memory Access. So the problem we've always had is—

Leo Laporte [00:27:47]:
That's, by the way, kooky.

Ash Hart [00:27:49]:
Yeah, so essentially I've basically made RDMA work between 2 completely different architectures and That has to be the same model

Leo Laporte [00:27:58]:
on both?

Ash Hart [00:27:59]:
Yeah, same model on both. So because you're doing the—

Leo Laporte [00:28:02]:
you're doing the prefill, the context, creating the KV cache, and then you send the cache over? Is that what happens, or you send the whole—

Ash Hart [00:28:08]:
Yeah, so depending on how you want to actually run it, you can do it in tensor parallelism so you can actually speed it up, or you can do pipeline, which is essentially— or disaggregated if you want as well, which is what you're talking about. So decode on the Studio and pre-fill on the Sparks. So that is fine if you just use normal Cat5, but if you want really quick, low-latency stuff, like sub-25 milliseconds, microseconds, you have to use RDMA. So that's where I actually used a local model to reverse engineer Apple's RDMA protocol. I actually filmed it and put it on my X. Uh, I think that's what caught people's attention.

Leo Laporte [00:28:53]:
Yes. And yet that's over Thunderbolt 5 cabling. How many cables do you need?

Ash Hart [00:28:59]:
This is another wild thing. So I figured out that, okay, you can actually have 2, 2 studios with so many ports available. There's so much fun to have, but Apple only allow one cable to work. So via MCDMA, I've enabled— you can have 6 if you want now. And I think aggregate throughput's like 480 gigabits.

Leo Laporte [00:29:20]:
So what's your recommend— what's your recommendation? Should I buy 2 512-gig Macs? Should I buy 2 more Sparks? Should I— what's, what's, what's, what, what is this? What is the sweet spot for TensorFold, do you think? I, I think the Macs with that giant single pool of memory, if we can get pre-fill fast enough, are very attractive.

Ash Hart [00:29:43]:
This is where I believe Metal is very underperformant, purely software.

Leo Laporte [00:29:50]:
Yes.

Ash Hart [00:29:50]:
And this is what I've actually— I put it on my X the other day. I'm working on building my own Metal kernel. So MLX_TF. And yeah, I feel as though there's a lot to be done there. And yeah.

Leo Laporte [00:30:05]:
Ash, you are a blessing to all of us in local AI and an inspiration to anybody Who says, well, I'm just one guy, what could I do to, um, to, to make things better? Well, one guy with some AI can do some pretty amazing things.

Jeff Jarvis [00:30:24]:
Can I ask a question that actually— it goes to what's next. AMD announced its cheaper version of Spark for $4,000 instead of $5,000.

Leo Laporte [00:30:33]:
Venice. This Venice platform looks really good, actually.

Jeff Jarvis [00:30:36]:
So, so where does this go so that Without having to raise money to buy a computer, you can do meaningful things. What's it gonna take? Is it that the models are gonna change? Is the hardware gonna change? Where's it gonna change so that more people can do what you both do?

Ash Hart [00:30:50]:
It's an interesting question. And I think there's 2 answers to this. The hardware is ultimately always going to get better. But let's say you've ordered Sparks or you've ordered the Venice or however, what you've got now. What you've got now is fixed. However, the models are getting better. We're on the cusp of GLM 5.5, we're on the cusp of the, the newer DeepSeek, the new Forum.

Leo Laporte [00:31:20]:
Any day.

Ash Hart [00:31:21]:
Yeah. And this is what people are not taking full advantage of. So they bought a Spark like I did in March, and it was pretty rubbish back then. You know, we could run models. It was slow. Mine sat there for 2 months. I literally looked at them and thought, this is a real bad purchase.

Leo Laporte [00:31:38]:
273 gigabits, uh, uh, speed is just not fast enough, we thought.

Ash Hart [00:31:43]:
We thought, we

Leo Laporte [00:31:45]:
thought.

Ash Hart [00:31:45]:
And, and then, you know, I think I actually saw Mia putting out some recipes on VLLM. I then ran her DeepSeek. I think it was DeepSeek 4. I think it was DeepSeek 4. I ran that for a long

Leo Laporte [00:32:00]:
time. I loved DeepSeek. Yeah. So we should just— so people understand, you've got the hardware, you've got a model, you also have an engine that runs the model. And there are various engines. There's VLLM, there's, uh, DSpark, there's TensorFold. These various engines actually have a lot to do with how fast and how performant the models are. So you have— there's— and what we're seeing is improvements in hardware, but as you were pointing out, it's also improvements in models and Thanks to people like Ash Hart, improvements in the engines are making a huge difference.

Jeff Jarvis [00:32:31]:
Which is what the soon-to-be former CEO of NVIDIA has been saying, that they're going to get more and more efficient.

Leo Laporte [00:32:37]:
You're scooping the headline story, which is going to come up in just a minute. So what's next for you, Ash?

Ash Hart [00:32:46]:
Pushing further. Like I say, I still think we are around 30%. of the actual throughput of your hardware.

Leo Laporte [00:32:52]:
I hope you're right. That's amazing.

Ash Hart [00:32:54]:
One thing I will say, and I'm happy to say it because I've done it, is my M5 Max at 128 lanes on QUAINT-27B 4-bit MLX does 622 tokens per second.

Leo Laporte [00:33:07]:
Oh my God.

Ash Hart [00:33:09]:
It does. I've seen it do it.

Leo Laporte [00:33:10]:
At this point, you're faster than you'll ever get on the Frontier, and that's what's really interesting.

Ash Hart [00:33:16]:
Yeah, I mean, unless you're running like Cerberus or you've got some real—

Leo Laporte [00:33:20]:
Yeah, Cerberus is very fast or Firewalls. But, but even then, 600 tokens per second, that's pretty fast.

Ash Hart [00:33:28]:
Yeah, I don't think I'll ever get there, to be honest. I mean, even if I've got 50%, uh, that would be pretty wild. But that's on an M5 Max. I mean, if you take the 5090, for example, I'm up to 1,600 now. So Where does it go? I don't know, to be honest.

Leo Laporte [00:33:44]:
That's amazing. Is there— do you have a preferred platform? Is CUDA better than MLX? Uh, is ROC coming along? CUDA's mature. CUDA's mature. Yeah.

Ash Hart [00:33:53]:
That's the thing. A lot are— I'd say, what, 80, 85% of the AI industry is built off CUDA,

Leo Laporte [00:33:59]:
right?

Ash Hart [00:33:59]:
You know, Jensen, um, regardless of what he did, um, his, his intuition for CUDA— was it 24 Was it 2024 or 2004? I can't remember when it was that they actually launched CUDA. And his bet paid off. Look at it. The infrastructure that runs off CUDA, the Vera Rubin that's coming out, them blades that they've built.

Leo Laporte [00:34:24]:
Those are amazing. Yeah.

Ash Hart [00:34:26]:
The compute that we're going to have and the intelligence that we're going to have is, you know, an infant compared to what we're going to have in the next 12 months.

Leo Laporte [00:34:35]:
The models get better, the hardware gets better. And thanks to Ash Hart, The engine gets better. Ash, I'm so glad we had time to talk to you. I really appreciate your making time. I know you're a very busy guy, and I hate to take away even a moment from your work, but I think the world needs to know about Ash Hardt and what Ash is doing. And tensorfold.dev, if you want to read more, tensorfold.dev. That's the pretty website, but of course it's also on GitHub. And if you are looking for models, you know, Mia AI Labs has a lot of the TensorFold recipes.

Leo Laporte [00:35:13]:
Paul Theriot earlier today was saying, I can't— I have an RTX Spark, I can't get it to work. I said, just go to Mia, press the button that says install. You don't need anything more. Let the AI handle it. I just, I'm so grateful to you for your time, Ash, and I'm grateful for what you've done. Well, thank you. Can't wait. November 17th is 6 days away.

Leo Laporte [00:35:33]:
I can't wait to see what MLX can do under your spell.

Ash Hart [00:35:37]:
I, um, I had a notification to say that it's being shipped, so mine should get here on the Wednesday, which is the 18th apparently.

Leo Laporte [00:35:47]:
Excellent.

Ash Hart [00:35:48]:
Um, but I was talking to the guys before the show started and I've actually managed to get access to Methos 5.2. And

Leo Laporte [00:35:56]:
that's, you know, Mia said that as well. Uh, how are you guys? That's interesting. I— that's That shows that Anthropic is maybe treating you guys with some— the much-deserved respect. So tell us about it.

Ash Hart [00:36:09]:
So essentially, I actually was a bit, um, no sort of thing, because, you know, the guardrails around Mythos, you know. Yeah, but you're a red teamer.

Leo Laporte [00:36:18]:
I mean, you, you would know what this thing can do.

Ash Hart [00:36:22]:
Yeah, and this is it. I've, I've used models to do red teaming, you know, built red team skills. I've reverse engineered stuff with AI. You know, I use AI in my day job to poke around our infrastructure to find holes. You know, I enjoy it, you know, and I love to see how things work. But Mythos is a different breed. It is. It's— and I was saying that it's very to the point.

Ash Hart [00:36:47]:
It's— it can do this, it can do that. This is what I can do. And it just does it. And it's amazing. And there's different levels as well. So I've got the first level, which is defense. There is, um, offense, and then there is, I think, black or something along them lines where you can write full-blown exploits if you want.

Leo Laporte [00:37:05]:
Oh boy. Yeah, I have been less than impressed with Fable 5 too, um, but, uh, I've never touched anything like Mythos. What do you, you know, just off the top of your head, do you think that there is stuff going on in these, uh, frontier labs like Anthropic and OpenAI that would take our breath away, that they're close to RSI, they're close to AGI? What do you think?

Ash Hart [00:37:30]:
I— yeah, categorically, I think the, the labs are at least a year ahead of us that we know of, 100%. I think— I mean, what is AGI? What is RSI when you really think about it? I mean, an LLM is a prediction engine, you know, you're predicting tokens, but then you've started adding different things in. You've got routers, you've got experts, you've got this, that. I mean, my— I had a separate project, and this is where the living weights came from. So essentially, I kept a standard LLM in memory without any KV cache whatsoever. And I built a harness around it to keep it thinking 24/7, it would not stop thinking. And all we did was inject sense. So I had my webcam, the site, my microphone as voice, my keyboard as touch, as things like that.

Ash Hart [00:38:23]:
And what I wanted to do was get the model to learn about me and who I am. And that's where Living Weights came from. I basically lifted that and put it into TensorFlow. But essentially, it's updating its weights that I could tell it a fact, I like the color blue. I could give the model to you and it would know that I like the color blue with no modifications to the— it's actually in the SafeTensor files. So yeah, there's a lot going on.

Leo Laporte [00:38:50]:
This is kind of the holy grail. We've been talking about this for some time in the AI user group, the idea of a model that modifies itself, uh, based on how you work with it. And you've been kind of playing with this, with this idea of living weights. It seems like a toy at this point, but do you think down the road that that's what we're going to have, is models that modify their weights?

Ash Hart [00:39:12]:
Yeah, definitely. I mean, if they don't modify their weights, they will modify a hot cache of weights.

Leo Laporte [00:39:20]:
Somewhere between context and weights.

Ash Hart [00:39:23]:
Yeah, they will. And then maybe when— I think dreaming is probably the best thing at the moment, because it can go offline, it can really, you know, use PyTorch and things like that to really learn. But the living weights is my representation of trying to, okay, talk to it, give it facts. I mean, it is a toy, but I can give it a 100-page document and it knows about that document. So yeah, we were getting past the stage of, okay, this is a bit of fun, to we could actually use this. And then I think 2 weeks later, uh, was it OpenAI or Anthropic posted something about it?

Leo Laporte [00:40:02]:
Yeah, it's pretty amazing. Ash Hart, thank you so much. @AshXHeart on x.com, tensorfold.dev. Um, the stuff you're doing is incredible, and, uh, I would— I will definitely be following you. And good luck with the Mac Studio.

Ash Hart [00:40:21]:
Thank you so much.

Jeff Jarvis [00:40:22]:
Thanks, Ash. Get some rest. Thanks for staying up late.

Leo Laporte [00:40:24]:
Yeah, take care. Late in the UK. Thank you, Ash. Take care.

Ash Hart [00:40:27]:
Take care.

Leo Laporte [00:40:28]:
Now, uh, it was hard to keep Paris focused.

Jeff Jarvis [00:40:31]:
Did you hear her gasp?

Paris Martineau [00:40:32]:
Did you hear me gasp live in the middle of that interview? It's because someone posted in the chat that the Nvidia— one, two things happened during— while we were talking to that lovely man. One, a second affair has hit Jensen Huang. Oh boy. This Jensen Huang, The New York Times reporting, is said to have had an affair with An Nvidia executive who reports to him.

Jeff Jarvis [00:41:02]:
But age appropriate. Let's give him that. Age appropriate.

Paris Martineau [00:41:04]:
Hey, you know, Kimberly Powell, Nvidia's vice president of healthcare. This is obviously against Nvidia's code of conduct. They were captured in video as being affectionate at a San Jose, California hotel during an Nvidia conference in March, said 2 people to the Times. They're both married to other people. And then the thing that prompted the—

Jeff Jarvis [00:41:25]:
Wait, wait, wait, wait, wait. Stay on that for one more second. One more second. And then Nvidia security people happened to have NDAs with them and got whoever took the video to sign the NDA at the hotel.

Leo Laporte [00:41:36]:
Oh boy.

Jeff Jarvis [00:41:37]:
Oh my gosh. So they were covering it up.

Leo Laporte [00:41:39]:
They're all in this. In other words.

Jeff Jarvis [00:41:41]:
Yeah.

Leo Laporte [00:41:42]:
Everybody knew.

Paris Martineau [00:41:43]:
I was about to say, that's a level of coordination. I mean, maybe they could just be thinking that on the fly, of course, but it's also a level of coordination.

Jeff Jarvis [00:41:51]:
Don't say fly.

Paris Martineau [00:41:54]:
It implies that they maybe have standard procedures in place for this sort of thing.

Leo Laporte [00:41:59]:
Well, I wonder if other companies don't do this. I mean, this is probably something you do with your CEO in many cases.

Paris Martineau [00:42:05]:
You think that— how many companies do you think have an active policy for—

Leo Laporte [00:42:13]:
Well, do you—

Paris Martineau [00:42:13]:
I mean, it depends on their partners to sign NDAs.

Leo Laporte [00:42:16]:
It depends on how important the CEO is. The founder and CEO, I imagine— I don't know. I mean, we heard all sorts of rumors about Sergey Brin for years that turned out not to be just rumors.

Paris Martineau [00:42:31]:
Well, of course, of course, the people around you are going to try to protect—

Leo Laporte [00:42:36]:
If it's felt that the founder is vital to the success of the company, I think you can expect that kind of behavior.

Paris Martineau [00:42:41]:
Well, yeah. But there's no question about it.

Leo Laporte [00:42:43]:
I was saying this before this second affair that there is no question in my mind the Nvidia board of directors is going to fire John.

Paris Martineau [00:42:51]:
And Leo was saying this as we were getting on the call today before we started recording for the show. And the thing that made me gasp in the middle of our interview is that somebody posted in our Discord chat this Wall Street Journal headline, Nvidia is preparing to oust CEO Jensen Huang.

Leo Laporte [00:43:07]:
Yeah, not a surprise.

Paris Martineau [00:43:08]:
And this is by Robbie Whelan and Josh Dawsey. And it is one of my favorite type of stories, which I think it was just updated to now even better

Jeff Jarvis [00:43:16]:
than tunnels,

Paris Martineau [00:43:17]:
5 or 6 paragraphs, but a second ago it was just 2 sentences. And then this developing story will be updated, which is how you know something is, something is cooking. It's afoot. It's hot. Everyone, every tech reporter out there that covers Nvidia is on the phones.

Jeff Jarvis [00:43:34]:
In the old days, this was stop the presses.

Leo Laporte [00:43:37]:
Well, I think that's what the New York Post did with this headline. Wang, Wang, Wang, bang, bang.

Paris Martineau [00:43:45]:
I mean, You do have to hand it to them. I'm happy the New York Post is here doing what the New York Post is best at, which is making sexual jokes on the COVID about non-sexual—

Leo Laporte [00:43:59]:
I think, you know, the affairs themselves, maybe these happen with other CEOs. I don't know. I don't have any direct experience, but maybe they do.

Paris Martineau [00:44:08]:
You think that—

Leo Laporte [00:44:09]:
The Chinese spy thing is kind of a big deal.

Paris Martineau [00:44:12]:
Alleged Chinese spy and a direct report.

Leo Laporte [00:44:15]:
Yeah, the Direct report was bad.

Jeff Jarvis [00:44:17]:
The, the, the— but the Chinese spy story turned into the Bernie story.

Paris Martineau [00:44:20]:
Those are the 2 people you don't want to be having an affair with.

Jeff Jarvis [00:44:22]:
No,

Leo Laporte [00:44:22]:
right,

Jeff Jarvis [00:44:22]:
right.

Paris Martineau [00:44:23]:
And the 3rd is a CEO of another company while you're at a Coldplay concert. Those are the 3 types of people you don't want to have an affair with.

Leo Laporte [00:44:31]:
I, I don't want to spend a huge amount of time on this because of the salaciousness of it, but it is germane to the, uh, AI industry. It's the core of everything. What happens to Nvidia minus Jensen Huang? He really was—

Jeff Jarvis [00:44:44]:
Stock went down 9% after market, the nervousness about Nvidia.

Leo Laporte [00:44:48]:
They've come back a little bit since then, but—

Jeff Jarvis [00:44:50]:
Tonight already?

Leo Laporte [00:44:52]:
I think so.

Jeff Jarvis [00:44:53]:
Yeah.

Paris Martineau [00:44:53]:
I mean,

Leo Laporte [00:44:53]:
people probably saw that as a buying opportunity.

Jeff Jarvis [00:44:55]:
The other angle here is that, is that whether or not the government told Jensen Huang that she was suspected of spying.

Leo Laporte [00:45:05]:
It's pretty clear to me that Marco Rubio himself leaked this story. Right.

Paris Martineau [00:45:12]:
Let's give people context for what we're talking about first in case someone hasn't heard about this.

Leo Laporte [00:45:16]:
So the president was told about this apparently in a briefing a year ago and ignored it. And then you may remember—

Paris Martineau [00:45:23]:
This being that there was an investigation into an alleged Chinese spy.

Leo Laporte [00:45:27]:
We were hearing, honestly, months ago, we were hearing about Jensen Huang's relationship with a woman, not his wife, who we didn't, we didn't say Chinese spy, but there was some question about whether she was lobbying Jensen, because he, as you remember, for the last year has been talking about how we really ought to sell Nvidia chips to China. Now, by the way, I don't disagree with him. I think he was right about that. Nevertheless—

Jeff Jarvis [00:45:55]:
But he had different motives than we thought.

Leo Laporte [00:45:57]:
Maybe he had other motives. But then, so apparently Trump ignored that first warning, but you may remember when the president was flying to China with a bunch of tech CEOs that had not invited Jensen Huang, And then he got on the plane. I'll meet you in Alaska.

Jeff Jarvis [00:46:15]:
That's right, the Alaska gambit.

Leo Laporte [00:46:17]:
Well, apparently on the way to meet Jensen Huang in Alaska, Marco Rubio goes to the president and says, you keep— he's banging a Chinese spy, we got a problem. And, and nevertheless, Trump stops, picks up Jensen Huang, takes him.

Paris Martineau [00:46:33]:
I was just gonna say, that's not something that Trump is going to register as a problem if it's phrased in that way.

Leo Laporte [00:46:39]:
But I also should point out the only way that that story comes out is that Marco Rubio himself leaked it.

Jeff Jarvis [00:46:44]:
Yes.

Leo Laporte [00:46:45]:
And which is not surprising after the November election. But what—

Jeff Jarvis [00:46:48]:
so not Trump, the CIA, the FBI, they all should have, uh, they must have all known. If you give— if you think that Wang was stupid enough to not think that she was a spy, uh, then the government should have told him that she was a spy. And by not telling him, it could continued, right? And so she may, by the way, tell the story here too.

Leo Laporte [00:47:07]:
She may not have been a spy per se. We don't really know, right? She may have been what they call a white glove facilitator, somebody who's hired by the Chinese government as a commercial negotiator. It might have been a honey trap. It might also have been somebody who took her job a little extra seriously.

Jeff Jarvis [00:47:27]:
Well, the thing is, I've had Chinese students and One has to be careful with the Chinese students because they're either watched or watching.

Leo Laporte [00:47:35]:
Right.

Paris Martineau [00:47:36]:
And I mean, let's not try to cast— paint everyone with a broad brush.

Jeff Jarvis [00:47:42]:
No, no, I'm not doing that. No, but what I'm doing is saying that there's, there's a million different relationships. And if— when I tried— let me put it this way— when I tried to draw out a Chinese student about China, I was quickly cautioned that was unfair to the student. shouldn't. Yeah, because who knows what would get back.

Leo Laporte [00:47:58]:
Don't put them in an awkward position. So

Jeff Jarvis [00:47:59]:
the problem is that one could— one presumes that one could always be watched, right? That's what I'm saying.

Paris Martineau [00:48:04]:
Well, I think one of the most interesting things, uh, that we're gonna have to— we're gonna, I guess, learn— possibly we're gonna learn more about this over the course of recording this episode and over the next week— is what is the fallout from this incident? And so I don't think it's going to be contained to just Jensen Huang. Like, I think that We're looking, looking at this New York Times article. I was going to say from a while ago, but it's just from 2 days ago. A lot has happened in this story, which was about— it's basically the New York Times had reported that the investigators at the Commerce Department had been investigating Nvidia for months last year, specifically this spy incident. They suspected that a Singaporean company was smuggling Nvidia's artificial chips into China. They'd been kind of going into this. It had centered on Megaspeed, the data center company led by the executive who officials now believe is this Chinese spy. The case ended up getting abruptly shifted to the Justice Department last year, and that shift happened after the Commerce Department investigators working on the case received an email from Nvidia's general counsel seeking information about the inquiry into Megaspeed.

Paris Martineau [00:49:18]:
This is kind of an— so that suggests, I mean, that if that's true, that means that at least Nvidia's general counsel knew that this investigation was happening and that these allegations were out there and being actively sifted through.

Jeff Jarvis [00:49:36]:
But it also says going to the Justice Department— sorry, going to the Justice Department was not necessarily a ramp up given this administration. It was a place to hide it.

Paris Martineau [00:49:45]:
Yes. But I mean, that just suggests to me that the potential fallout for this is going to be ever expanding. It's going to be a question of if they want to— if the board of Nvidia, board of directors really want to try and move the company forward from this. And their first step is clearing out Jensen Huang. Second, third, and fourth step is going to have to be figuring out anybody else who was involved in the cover-up of this and any other incidents that end up I guess catching attention.

Leo Laporte [00:50:11]:
Well, they've hired a law firm to investigate.

Paris Martineau [00:50:14]:
And doesn't the law firm also— some of the people on that team include one of the guys that was close to Sam Altman, or one of the people that Sam Altman said should be the FTX CEO? Oh, the acting CEO? Yeah.

Leo Laporte [00:50:29]:
I don't know. So I don't really care about the salacious details. What is going on?

Paris Martineau [00:50:33]:
None of these are salacious details. These are the details of what's going to happen with the business, which is very important.

Leo Laporte [00:50:37]:
Yeah. What's going to happen to Nvidia? Can Nvidia continue to innovate without Jensen Huang?

Jeff Jarvis [00:50:43]:
He had vision. He didn't have—

Leo Laporte [00:50:46]:
He built the whole company.

Jeff Jarvis [00:50:47]:
He built the whole company. He had 60 direct reports. He did not have a succession plan.

Leo Laporte [00:50:52]:
So there's nobody— there's nobody ready to take over.

Paris Martineau [00:50:56]:
There's nobody ready to take over. And if, you know, this through line we're seeing with this Commerce Department general counsel thing continues, You don't even know. We don't know how many people are going to get caught up in the fallout of this to where it's not just going to be Jensen Huang, which would be a catastrophic event for Nvidia, but it could be a number of people that he was very close with and who were acting on his behalf throughout this, which would be even more detrimental to this company or its operations, assuming you're just talking about keeping everything normal, I guess. But it's just going to be— it is going to be incredibly difficult for Nvidia to continue operations in any normal way in the coming months.

Jeff Jarvis [00:51:43]:
So on line 119, I put up this thing that has really good numbers about the impact of Nvidia as the bank of AI. And just to— Oh yeah,

Leo Laporte [00:51:52]:
that's another thing. They're very much—

Jeff Jarvis [00:51:54]:
So there's $279 billion are promised to suppliers for buying. They've backstopped $164 billion to other companies. They have $99 billion in stakes in other companies. And there are buybacks of $235 billion yet to come. This is— they are the crux of everything here. And

Leo Laporte [00:52:15]:
so is this the beginning of the bubble burst?

Jeff Jarvis [00:52:18]:
I fear. But who would— who would— my bingo card didn't have crying mathematicians and Jensen Huang stopping as the pin in the bubble of AI.

Leo Laporte [00:52:30]:
This could be the pin.

Jeff Jarvis [00:52:31]:
So weird.

Paris Martineau [00:52:33]:
I mean, it's never as sexy of a pin as you think it's going to be. It's never as like, oh, a very cool financial crime sort of thing. It's always something kind of stupid, I feel like. Yeah.

Jeff Jarvis [00:52:44]:
Jeez. Never underestimate the male libido, man. Plus stupidity.

Paris Martineau [00:52:49]:
I was texting Jeff this earlier, but he's real— Jensen Huang is really not beating the 63-year-old with a leather jacket allegations. Okay.

Leo Laporte [00:52:58]:
So, all right, so yeah, I mean, if I'm on the board, I don't want to fire Jensen.

Jeff Jarvis [00:53:04]:
No, they're close to him and the company needs him.

Leo Laporte [00:53:07]:
I don't think they have any choice though. This is what I was contemplating earlier.

Jeff Jarvis [00:53:10]:
The guess 2 days ago was that they're going to put all kinds of controls on him.

Leo Laporte [00:53:14]:
Yeah, no, they have to get rid of him. But that's, uh, I don't know how you continue. Um, boy, I would love to be a fly on the wall on the board discussions because they, they must be They must find somebody who can replace Jensen. No one will have the vision, but maybe the business is mature enough, you just keep it going.

Jeff Jarvis [00:53:32]:
I think that's what you hope for.

Leo Laporte [00:53:34]:
Yeah.

Jeff Jarvis [00:53:35]:
But if you're— what other companies now are licking their lips?

Leo Laporte [00:53:39]:
Well, I would say AMD for sure. AMD has now released some very high-quality hardware, this Venice chip that many have said— see, I'm actually not too sad about this. As much as Nvidia has done for AI. Their lock, their complete lock, their proprietary lock on CUDA is not good for AI. It's not good for, um, uh, you know, the—

Jeff Jarvis [00:54:06]:
Yeah.

Leo Laporte [00:54:07]:
I'm a fan of open and it's not open. So in a way I'm thinking, um, this, this isn't, this isn't a bad thing for AI. It gives companies like AMD an opportunity

Jeff Jarvis [00:54:21]:
But it's not like there's a short turnaround here, like, okay, good, we're gonna, we're gonna match.

Leo Laporte [00:54:25]:
Well, AMD already has this great chip, right? Uh, if people— okay, I mean, CUDA is going to stick around. CUDA's got a lot of momentum.

Jeff Jarvis [00:54:33]:
Oh yeah,

Leo Laporte [00:54:34]:
it's not going away. They've got Vera Rubin, and they haven't really been able to put Vera Rubin, uh, out in mass quantities yet. I mean, there's other— there's huge— there's other—

Jeff Jarvis [00:54:44]:
Google's got to be happy as hell.

Leo Laporte [00:54:47]:
There's other headwinds. I mean, the data center, 40 of 41 data center shutdown propositions won in the election. Data centers are going to be banned all over the country. That's a lot of chips that have been committed by Nvidia that won't be bought.

Paris Martineau [00:55:07]:
I mean, yeah, how does the company pull itself out of what looks like the beginning of a spiral?

Leo Laporte [00:55:13]:
Yeah, if it does, it's

Jeff Jarvis [00:55:15]:
got to stabilize, stabilize, stabilize, stabilize.

Paris Martineau [00:55:17]:
I mean, even if the core business is intact, it's kind of something— it reminds me of SpaceX or Tesla almost in that a lot— Nvidia has a core product that exists, obviously. And unlike those companies, I think it has a core product that is incredibly in demand and at the center of the industry in a way that is undeniable. Yeah, but there like those companies, there is an element of hype and hype around a specific man and his unique approach to doing things and the way he talks to investors and his idiosyncrasies.

Jeff Jarvis [00:55:58]:
Was he the Steven— was he the Steve Jobs of AI?

Leo Laporte [00:56:00]:
Absolutely.

Paris Martineau [00:56:01]:
Without a doubt.

Jeff Jarvis [00:56:04]:
And Apple continued.

Leo Laporte [00:56:06]:
Yeah. No, Apple didn't do so well when they got rid of Jobs. Jobs brought them back from the dead. Jobs brought them back from the dead.

Jeff Jarvis [00:56:13]:
I wonder what Steven Witt is saying right now.

Leo Laporte [00:56:17]:
Why Steven Witt?

Jeff Jarvis [00:56:18]:
The author of the book on Jensen Huang.

Leo Laporte [00:56:20]:
Oh, he must have known something. I feel like we knew about this woman months ago. Am I wrong? This doesn't—

Jeff Jarvis [00:56:29]:
the Times did a report a couple of weeks ago, and I was kind of dismissing it, saying, oh, Huang is more sophisticated than this. He might have had an affair, but he's not going to, you know, he's going to presume care around Chinese people and da da da da da. But the other thing is Taiwan. He's, he's a hero in Taiwan. And because of this relationship with China, he's pissed off the Taiwanese as well. And the big supplier of his manufacturing there.

Leo Laporte [00:56:53]:
Now, it is true that after Jobs sickened and passed away, the company was able to continue based on the momentum that Jobs had built in that company. And maybe Nvidia is mature enough to get to that

Jeff Jarvis [00:57:05]:
Right.

Leo Laporte [00:57:05]:
To be at that point.

Jeff Jarvis [00:57:07]:
I

Leo Laporte [00:57:07]:
actually wouldn't be surprised if you found a manager as good as Tim Cook to take over, maybe.

Jeff Jarvis [00:57:13]:
And not the former head of PepsiCo.

Leo Laporte [00:57:16]:
Sculley was not the guy.

Jeff Jarvis [00:57:17]:
Sculley was not the guy. 2 of Huang's kids are in fairly important positions. But again, there's huge layers to this place. One is in charge of robotics. The other is marketing. I don't know if they stay.

Leo Laporte [00:57:29]:
Well, It's really a sad story. It's the end of an era.

Jeff Jarvis [00:57:34]:
It is. It is. And he was the one I thought was most sane.

Leo Laporte [00:57:38]:
Yeah. Isn't that funny? They're all mad.

Jeff Jarvis [00:57:41]:
I wrote a post this week about the wackadoodles who are in charge of AI. He was the one who was okay, but he had a little bit of—

Leo Laporte [00:57:47]:
They're all wackadoodles. All right, we're gonna talk about the other wackadoodles at Anthropic in just a little bit. You're watching Intelligent Machines. Yeah, I don't know. I think we will know by the end of the day. If not maybe tomorrow, but the next couple of days,

Jeff Jarvis [00:58:02]:
I

Leo Laporte [00:58:02]:
think. I suspect the Wall Street Journal's acting like it's imminent.

Paris Martineau [00:58:07]:
Oh, Nvidia stock's down by more than 8% now.

Leo Laporte [00:58:10]:
Oh, it came back pre-market 5% after its 10% drop yesterday.

Jeff Jarvis [00:58:16]:
Yeah, it was post-market, went down 8%. Yeah, that's what I was saying earlier.

Leo Laporte [00:58:20]:
Yeah. So it went down yesterday, but—

Jeff Jarvis [00:58:22]:
It went down 6%. It was down 5% and 6%, then back up 4% yesterday.

Leo Laporte [00:58:27]:
Oh, look at this.

Jeff Jarvis [00:58:28]:
And now down 8%.

Leo Laporte [00:58:29]:
Oh, look at this. Yeah, that shows you, that's $20 a share. That's a huge drop. And that shows you how much the market thinks NVIDIA's success is priced in.

Jeff Jarvis [00:58:39]:
It's a buying opportunity.

Paris Martineau [00:58:40]:
I would like to also correct the record. I'm sorry for saying you're wrong about us knowing about this before, Leo. It is, was October 9th, 2025 that the New York Times reported that federal investigators were looking into Megaspeed and the connection between Alice Huang and Jensen Huang.

Leo Laporte [00:58:56]:
I remember that story.

Paris Martineau [00:58:58]:
I think the part of what tripped me up is I, of course, we weren't thinking about it from like an affair perspective at the time. It was more of just like kind of close business and chips going the wrong way.

Leo Laporte [00:59:08]:
It was a little suspicious. It was a little suspicious. I didn't think of spies or affairs, but it just was, it was not, it was an uncomfortable relationship. Um, and I think that this has been the case with NVIDIA for, for some time. there was also lots of talk about them under the table selling chips, not just to Megaspeed, but to other Chinese— Well,

Jeff Jarvis [00:59:31]:
explain something to me there, Leo. Megaspeed, as I understand it, set up hosting facilities for AI stuff offshore, not in China, offshore.

Leo Laporte [00:59:41]:
Singapore. Yeah.

Jeff Jarvis [00:59:42]:
Right. And so if you're a Chinese company, you can't get access to Nvidia chips, can't you just hook into her company in Singapore and get access? Right. So it never really was an issue.

Leo Laporte [00:59:52]:
Yeah, yeah, you

Jeff Jarvis [00:59:53]:
got the chips, you got the chips onshore, okay, but you could also get the access to the technology offshore, right?

Leo Laporte [00:59:57]:
Yeah, I mean, there's all sorts of, um, ways and ways.

Jeff Jarvis [01:00:03]:
Yeah,

Leo Laporte [01:00:03]:
I think the smartest thing that Jensen said, and I agree with him 100%, is if you don't sell our chips to China, they're just going to work on making their own and they'll make them better. It's not like they're not capable of doing that. In fact, they've kind of done that. They've made models that are very good, not as good as the best models in the United States, but pretty close. And the Huawei chips are very good. DeepSeek is running on Huawei servers entirely, apparently. Yeah, this is gonna be— I suspect this is the beginning of a bloodbath.

Jeff Jarvis [01:00:39]:
Oh, my poor 401.

Leo Laporte [01:00:41]:
Yeah, we already had the bloodbath on November 3rd, and now we're having a second bloodbath. But that's how these things happen is you get a cascade effect.

Jeff Jarvis [01:00:50]:
And then it comes back up with buying opportunities and then it goes back down.

Leo Laporte [01:00:54]:
Yeah. People say, oh, it's time to buy. I'm not sure I'd run out and buy Nvidia stock right now.

Jeff Jarvis [01:01:01]:
I'd wait a day or two.

Leo Laporte [01:01:02]:
By the way, and I'm going to double down on this, the fact that all of these data centers aren't getting built is going to change a lot of things. All of that Ramageddon stuff is gonna go away. There's gonna be a glut of RAM.

Jeff Jarvis [01:01:17]:
And the whole hyperscaler will own the universe just through scale alone, I think gets challenged.

Leo Laporte [01:01:23]:
Yeah.

Jeff Jarvis [01:01:23]:
There are other ways.

Leo Laporte [01:01:23]:
Well, they'll find ways to build data centers. I'm not sure. I'm not convinced that they won't, but— And I think you can see it now.

Jeff Jarvis [01:01:30]:
You know, the East Coast has moved into the old Lord& Taylor.

Leo Laporte [01:01:33]:
Yeah.

Paris Martineau [01:01:33]:
What do you think this means for the bubble, Leo?

Leo Laporte [01:01:39]:
First of all, I'm not sure there is a bubble. I think there's actually value in AI.

Paris Martineau [01:01:42]:
I'm not saying there's not value.

Leo Laporte [01:01:44]:
And I think investors perceive a value, an upside. And so it's not tulip bulbs, which have no inherent value. It is something that has an inherent value. So I'm not convinced investors will flee from the AI market. I think most investors Most, not most, many investors say, it's going to be a bumpy ride. Fasten your seatbelts. We're in for a bumpy night. But at the end of the day, maybe 5 years from now, AI is going to change everything.

Leo Laporte [01:02:20]:
And I'm of that opinion. And so, it's not quite Uber. It's not WeWork. It's something that's got— it's already delivered real value.

Jeff Jarvis [01:02:32]:
Though I do think that Anthropic or OpenAI could turn into WeWork.

Leo Laporte [01:02:37]:
Yeah. Oh, I'm not— no, I'm not picking a winner.

Jeff Jarvis [01:02:39]:
I

Paris Martineau [01:02:39]:
mean, with OpenAI's expenses they most recently reported.

Leo Laporte [01:02:44]:
I'm not picking a winner, but you know what? If I were OpenAI— actually, I was going to save this for after the commercial, so let me do the commercial and then we'll talk about that. You're watching Intelligent Machines. There's a lot to talk about. There is every week. That's the amazing thing. Next week, our guest will be Daniel Suarez, who is a science fiction author. Oh, wow. He's great.

Leo Laporte [01:03:04]:
Actually, you can join me. I'm going to interview Daniel tomorrow, 1:00 PM Pacific, 4:00 PM Eastern. I'm going to do an hour interview. We're going to play half an hour of it on the show next week. But I've talked to Daniel many times. In fact, pretty much with every book. I'm a big fan. We're good friends, and I love his stuff.

Leo Laporte [01:03:22]:
If you've read any of his stuff like Freedom TM or Demon or Kill Decision, you know that he's been prescient about AI. He, you know, Kill Decision was about basically what's happening in Ukraine and Russia, the drone killer drones. He is a very interesting fellow. We'll talk to him tomorrow, and that interview will air next week on Intelligent Machines. Or if you're in the club, you can watch tomorrow. Dario Amodei, CEO of, uh, Anthropic, says cancer, Alzheimer's, heart disease, and mental illness will be relics of the past in less than a decade. Thank God, because that's right where I'm heading for all of those. We can create an age of miracles.

Leo Laporte [01:04:08]:
He said, quote, you know, if today we look at things like cancer, Alzheimer's, heart disease, mental illness, these will be relics of the past, I think, in less than a decade. It isn't kind of a coincidence that his IPO is coming up in just a few weeks.

Jeff Jarvis [01:04:19]:
Let Paris give you a little more context on that quote, though.

Leo Laporte [01:04:22]:
Okay, go ahead. I

Paris Martineau [01:04:22]:
was about to say, I saw this quote as well. I was like, oh, interesting. How is he going to solve all of this, including mental illness and things like that? And if you look into his writings about it, specifically what he's talking about is that AI will enable us to get really good at determining which embryos are going to have those physical or mental illnesses, and then choosing other ones, which, you know, sounds—

Leo Laporte [01:04:45]:
Oh, you mean eugenics? Yeah!

Jeff Jarvis [01:04:47]:
Uh-huh, I've been saying it. I've been saying it again and again and again. Uh, I would like to plug right now my glossary of the crazy boys versus our language at Medium, where I go through all of this and it, it leads you to eugenics. That's what these people believe. He— So

Leo Laporte [01:05:03]:
it's not AI that's going to save us, it's getting rid of the embryos that are bad.

Jeff Jarvis [01:05:08]:
Ah.

Paris Martineau [01:05:09]:
It's culling the herd.

Leo Laporte [01:05:11]:
Culling the herd. Oh, I like that. Uh, wow, that's creepy as hell.

Jeff Jarvis [01:05:17]:
Yeah,

Paris Martineau [01:05:18]:
yeah, I know, because everybody's like, oh wow, wouldn't it be great?

Jeff Jarvis [01:05:21]:
Get rid of cancer.

Paris Martineau [01:05:22]:
Do we really want— I don't want to have to deal with cancer. It's like, no, no, no, not

Leo Laporte [01:05:26]:
this innovative new technology that's changing so many things. Do we really want it to be in the hands of somebody who thinks that the solution is weeding out the weaklings?

Jeff Jarvis [01:05:36]:
That's my point. It's a great technology in the wrong hands. And if you go through my glossary, argues that they use different meanings of safety and future and human and doom and regulation, and they have skewed the public discourse.

Leo Laporte [01:05:52]:
Wow.

Jeff Jarvis [01:05:52]:
Such that we can't have a decent discussion.

Leo Laporte [01:05:56]:
Well, they certainly scared everybody.

Jeff Jarvis [01:05:59]:
I added it up. So rather than start with, here's TESLA, what it means, if you add up these utilitarian beliefs, and where they go, you end up at eugenics.

Leo Laporte [01:06:08]:
Yeah. Well, this is— he's kind of saying the quiet part out loud if he's saying we don't have mental illness, you just don't let them be born. Well,

Jeff Jarvis [01:06:17]:
I also quoted a post, and we saw this I think last week, where earlier in life he decided whether to give to something that saved adult lives or children's lives, and he decided that adult lives were worth 2 to 3 times a child's life, as if he puts himself in that position. This is utilitarian mechanism. Yeah, this is the bridge to eugenics. We're gonna—

Leo Laporte [01:06:34]:
So this quote from Fireside Alpha, he doesn't say that in this quote. He just says we're going to get rid of Alzheimer's and cancer and stuff. But what you're saying is what he means is we're going to do that by finding out if you're going to get that when you— before you're born and then culling you. Is that what he's saying?

Paris Martineau [01:06:51]:
Or is he saying AI is going to cure cancer? I saw a, um, quote from one of his writings with this specifically about mental illness. And that is an accurate description of his quote about mental illness. But let me see here. I'll read this.

Leo Laporte [01:07:07]:
You know, I have to point out—

Paris Martineau [01:07:09]:
It's by— yeah. In his writing, Machines of Loving Grace, in October 2024, he wrote that effective genetic prevention of mental illness seems possible.

Leo Laporte [01:07:21]:
No, you don't want that. By the way, you do not want that. The most creative parts of society, the people who think out of the box, the geniuses, are very often mentally ill in our definition of mental illness. They're neurodivergent, they're different, and that's what makes them very important to our species.

Paris Martineau [01:07:40]:
I was going to say, here is, in his own words, most mental illness is partially heritable, and genome-wide association studies are starting to gain traction on identifying the relevant factors. which are often many in number. It will probably be possible to prevent most of these diseases via embryo screening, similar to the story with physical disease.

Leo Laporte [01:08:01]:
Yeah, who needs Vincent van Gogh anyway, right? You don't need the artists, the painters, the creative types. You don't need those.

Jeff Jarvis [01:08:09]:
I just finished the amazing Katja Hoyer's book Weimar, which is excellent, and it was a different view of the eugenics of the Nazis, but it's the same belief that you can define mental illness and then you can say that we're going to— different world where we're going to sterilize everybody who ruins the gene pool. Same difference.

Benito Gonzalez [01:08:30]:
Hi, this is Benito. So the bottom line problem here is being— is they think they can quantify human value, and that's, that's the problem.

Jeff Jarvis [01:08:38]:
That's utilitarianism. Exactly. So the philosophy— I'm going to plug my glossary again until you show it— the philosophy of this. Where is it? It's on Medium. It's on Medium. Is that we can judge the total human happiness. And so that leads you to effective altruism. But they throw out the mosquito nets with the kids it saves and say, well, that's, that's not much of happiness.

Jeff Jarvis [01:09:02]:
If we really calculate the happiness of the 10 to the 54th human beings in the future and, and past the singularity, then that's the, the, the, the, the happiness we want to judge and we can measure. And so that's the basis of their whole argument about how they go to AGI and ASI and what it can do. And if you screw that up, the doomers say, they don't care about today, they care about those 10 to the 54th future human beings. So Yudkowsky, Eliezer Yudkowsky, says that we should stop current development only because we're not good enough yet, and that it's okay to use thermonuclear weapons against the builders of the present technology, so the people who are left over will learn the lessons, and then we get it in the end. That's how nutty these people are. And they're in charge of the technology. Wow. They're not just nuts on the street.

Leo Laporte [01:09:48]:
They're dangerous.

Jeff Jarvis [01:09:49]:
They're dangerous because they have the power and they have the wealth. And so the data— what I worry about is that Congress comes along and thinks, well, we'll get rid of data centers. Okay, our job is done. No, watch these crazy people. And favor the smart people. I would have said like Jensen Huang, but well, never mind that. But there are smart people doing smart things with this amazing technology, and they're— and these, these jerks are going to ruin it for all of us.

Leo Laporte [01:10:16]:
Benito, show my screen so Jeff will feel fully plugged.

Jeff Jarvis [01:10:21]:
Thank you. There it is.

Leo Laporte [01:10:24]:
Wrong.

Jeff Jarvis [01:10:25]:
Well, this is interesting. So Politico did this headline about the election, how the doomers won. No, the doomers are the ones who lost, but the doomers screw the language up so much that you— that nobody understands what the word safety means or doom means or future means.

Paris Martineau [01:10:40]:
People keep conflating AI doomer with person who's negative about AI. It's like, no, no, no, no. It's the opposite. People who are technically AI doomers are the ones who are like, AI is really good and super scary and it's gonna shoot you with a gun in your sleep.

Jeff Jarvis [01:10:58]:
That's why you should let us have the over because we are the ones who understand it, because we get to judge who the, the best humanity is in the future. That's what they're like.

Leo Laporte [01:11:07]:
The good news is that because of their beliefs, the AI that they're giving us— and I have to say, I've been playing with Fable 5 too— is so paranoid. Seriously, it has inherited the paranoia of its creators that it's not really usable.

Jeff Jarvis [01:11:25]:
Really?

Leo Laporte [01:11:26]:
Yeah, it's, uh, it's, it's like terrified. It's like shivering in the corner, and it won't, it won't let you do things. It says, oh no, no, no, that's not safe, we got to do something else. Oh no, no, no, we can't do that.

Paris Martineau [01:11:39]:
What is it stopping you from doing?

Jeff Jarvis [01:11:41]:
The, uh,

Leo Laporte [01:11:43]:
well, the bio weapon I was thinking of building. No, I want to, uh, I want— so one of the things It's probably prudent to do with an AI is say, if you're going to do something irreversible, something that spends money, uh, you know, or, you know, delete a file that you can't get back, if you're going to do something irreversible, I want you to get my approval. That's a reasonable thing, right? And, and, and any well-written AI, uh, you know, Soul.md should have that, and they should have hooks for that and so forth. Um, and, and they should. But I've been working with— one of the things I did on vacation is I built a new app for me. I want to— I really like GrokBot, but I don't want to use Grok. So I'm trying to make a local GrokBot for myself that uses my local AI but has many of the features of GrokBot. I think I've done it.

Leo Laporte [01:12:36]:
And one of the things I wanted to do is say, well, let's make these approvals kind of more bulletproof so that a rogue AI can't you know, automatically, you know, say, oh, that's okay, Leo says you can spend that money, or Leo says you can lease a car.

Jeff Jarvis [01:12:53]:
You think that'd be a good thing to ask for?

Leo Laporte [01:12:55]:
Right. And I think that's a reasonable thing to do. We've set that up and it uses biometrics and so forth. But it was really overprotective. It was like, well, I don't— you know, you really shouldn't be able to say yes to any of this stuff unless you prove you're you by signing an affidavit. And so So it just was, it was very, it got in the way of doing things reasonably. You want to have some limitations, but you don't want to have unreasonable limitations that makes it hard. And I noticed going from, I was using Opus 5.5, which I like a lot.

Leo Laporte [01:13:26]:
I think I'm probably going to go back to it. I'm not alone, by the way. I noticed other people saying that. But Fable 5.2, which is undoubtedly smarter, is also undoubtedly more paranoid. And I think that that comes from Anthropic. that comes from their worldview. And I think that that's what's going to happen is that you're going to see these doomers create AIs that nobody wants to use because they're so shivery. I know, we can't do that.

Leo Laporte [01:13:54]:
That would be bad. That nobody will want to use them. The more I use the Frontier, the more I am happy using local AIs. Local AIs may not be as And there are times when you really have to use a frontier AI.

Jeff Jarvis [01:14:10]:
This is the question I was asking for son Jake. By the way, I mentioned to him that you were coming back from Asia after 3 weeks.

Leo Laporte [01:14:15]:
Oh, I'd love to talk to him anytime.

Jeff Jarvis [01:14:17]:
And Jake said, is Leo a spy?

Leo Laporte [01:14:19]:
I'm a spy. I am a spy, yes. We

Jeff Jarvis [01:14:21]:
know that now. You got your orders from him.

Leo Laporte [01:14:23]:
Too bad I don't know anything. Yes.

Jeff Jarvis [01:14:26]:
Is— I want to see this reach more retail levels. So the cheapest thing you can buy now that will be halfway decent, I think, is the AMD box. Which is $4,000.

Leo Laporte [01:14:34]:
I think I would watch with great interest what they do with Venice, which is their new platform. We're starting to see some— Tom's Hardware had some results and it looks really good. So I think there is—

Jeff Jarvis [01:14:46]:
Is that the Halo box they put out?

Leo Laporte [01:14:49]:
Yeah, so I have a Strix Halo and there's a new Halo. I'm not sure that's it yet, this 9686F.

Jeff Jarvis [01:14:59]:
Okay, that's according to—

Leo Laporte [01:14:59]:
Tom's Hardware outclasses Vera Rubin. It's, uh, it's got 1,024 threads. This is very interesting. This is Zen 6, and I think, I think you're going to see some real competition coming along. And that's one of the things I didn't like about NVIDIA and about CUDA was it was proprietary, and it only— you could only buy it from NVIDIA. And I didn't really want CUDA to be the only way we could do AI. Even though I bought the Sparks and that, yeah, the Sparks are fantastic, but I want Apple to do well with MLX. I would love to see an AMD architecture.

Jeff Jarvis [01:15:34]:
So Leo, I put in the chat the AMD box. It's the Ryzen Halo.

Leo Laporte [01:15:39]:
Yeah, no, I know it. I know what you're talking about.

Jeff Jarvis [01:15:41]:
It's $4,000. You can buy it at—

Leo Laporte [01:15:43]:
I wouldn't get it yet. Uh, is this for Jake or for you?

Jeff Jarvis [01:15:48]:
Oh, not for me. Well, that's part of the question. If he got it, could I, could I operate on it too?

Leo Laporte [01:15:55]:
So a lot of this depends on—

Jeff Jarvis [01:15:57]:
Windows and Linux.

Leo Laporte [01:15:58]:
A lot of this depends on— first of all, don't use Windows ever.

Jeff Jarvis [01:16:02]:
Amen, brother.

Leo Laporte [01:16:03]:
Uh, excuse me. A lot of this depends on what people like Ash Hart decide that they want to work on, believe it or not, because they're doing the most interesting stuff. So this

Jeff Jarvis [01:16:16]:
box has 128 gigabytes, 60

Leo Laporte [01:16:19]:
He said he wants to work on Rockm, which is the NVIDIA model. So I just think if you buy it now, it's a little speculative. The price is right.

Jeff Jarvis [01:16:29]:
This is what I tried to ask you. Is there a business model for hosted local models? So you don't buy the hardware? Not really. That's what I'm kind of— It's

Leo Laporte [01:16:43]:
very expensive if you wanted to do that. So for instance, the system I have, a couple of people could use it. The more— if you got 4 or 5 people using it, it would be unusably slow.

Benito Gonzalez [01:16:54]:
Is there a medium?

Leo Laporte [01:16:55]:
So in order to do that, that's why you have to build a data center. In order to do that, you have to have massive amounts of power.

Jeff Jarvis [01:17:00]:
That's for training and all kinds of other— just simple, you know, inference.

Leo Laporte [01:17:02]:
No, for inference.

Jeff Jarvis [01:17:03]:
Inference, yeah.

Benito Gonzalez [01:17:05]:
How much would you need for like a small town? Like, is there a medium? Is there a medium size?

Leo Laporte [01:17:10]:
Yeah, it's a medium-sized data center. I'd say a gigawatt would be enough.

Jeff Jarvis [01:17:16]:
Well, I just put in a—

Leo Laporte [01:17:17]:
You need— if you got a Vera Rubin rack— now the small town isn't using it all at the same time. So let's say, you know, you're going to have 500 people using it at the same time. You need a pretty big— you need a Vera Rubin rack.

Jeff Jarvis [01:17:31]:
Well, I just put in a grant proposal as a pilot in New Jersey to get local AI to be shared by 3 libraries. Start there. And you have community members who are looking at what kind of information they want to gather and analyze in their communities and what kind of things they want to build. And can we start to do something that's of use at that level by water treatment parks?

Leo Laporte [01:17:55]:
So the next step up is something called a DGX station. This is about $100,000. It's also Grace Blackwell. It's not Vera Rubin. They don't— you know, when they start offering Vera Rubins, that will be very interesting. But right now it's still, uh, this is a GB300, 748 gigabytes of coherent memory. It's pretty fast. I think you could run 50 people at once.

Leo Laporte [01:18:21]:
There are YouTube videos you could watch. This wouldn't serve a small town, but it might serve a lib— 3 libraries. Yeah.

Jeff Jarvis [01:18:28]:
You wouldn't need that much to serve 3 libraries with just a few people doing things.

Leo Laporte [01:18:31]:
I don't know.

Jeff Jarvis [01:18:31]:
Depends on what you're doing.

Leo Laporte [01:18:33]:
It just, it slows down a lot. It's not that you can't do it, it's just how fast it is.

Jeff Jarvis [01:18:37]:
Well, I think we want to show that there is an opportunity to bring AI into people's control, that they aren't necessarily sending their data and their money up to these jerks who are eugenicists, that they can have agency over it. I think we have to start showing people that because, because as one colleague of mine said in her town, if we just said the words AI, they're going to start screaming and saying, oh no, evil, evil. And unless we get the story across that no, it doesn't have to be those jerks.

Leo Laporte [01:19:08]:
Yeah. Well, it's, you know, this was the whole thing with the internet was local community internet. There was a great movement to do that where you would buy a fiber line to your town and then you'd share it. And that was a great idea. This isn't quite as affordable if you ask me. I don't, you know, I'm talking off the top of my head. I don't really know Um, I'm not sure. Alexander Ziskin, who has a good YouTube channel, has a Spark station and has done some testing with it.

Leo Laporte [01:19:43]:
You could see what he says. Let me see if I can find his— Uh,

Paris Martineau [01:19:48]:
guys, a brief interruption while you're looking it up.

Leo Laporte [01:19:51]:
Did they fire?

Paris Martineau [01:19:52]:
No, Nvidia earnings are in 6 days.

Leo Laporte [01:19:56]:
Oh, that's not good. Oh, that's not good. The poor boy.

Paris Martineau [01:20:01]:
I'm gonna be listening to that call.

Leo Laporte [01:20:03]:
I can't imagine.

Paris Martineau [01:20:04]:
Guys, can we all tune in? Can we all listen to Nvidia earnings?

Leo Laporte [01:20:09]:
I feel so bad about this. This is terrible. Can you imagine how anxious he feels?

Paris Martineau [01:20:14]:
Who's gonna be giving it?

Leo Laporte [01:20:16]:
Do you think Jensen didn't act— besides, you know, not keeping his fly zipped up— he didn't act He wasn't actively working with a Chinese agent, or do you think that?

Paris Martineau [01:20:31]:
We don't know. I know that we don't know is deeply concerning.

Leo Laporte [01:20:35]:
Well, that's why I think they're going to fire him. Yeah.

Paris Martineau [01:20:37]:
Well, I think it's because of that inappropriate rela— like, let's say it was completely above board.

Leo Laporte [01:20:44]:
It wasn't a spy, it was just the HR person.

Paris Martineau [01:20:46]:
It was not a spy. There was— then they wouldn't have to fire him. Well, even so, I mean, him sleeping with a direct report and a prominent company and then lying publicly about it multiple times is both a violation—

Leo Laporte [01:21:04]:
Did he lie about it multiple times? I'm sorry, I didn't see that part.

Paris Martineau [01:21:07]:
Uh, either people asking him about it, or his spokesperson said, our— my relationship was completely professional. There's nothing unusual about this whatsoever.

Leo Laporte [01:21:15]:
When does that happen? When was that?

Paris Martineau [01:21:17]:
A day or two before the news broke about the affair with the first person, which my thing is, I think that's securities fraud. Is

Leo Laporte [01:21:26]:
it securities fraud? No, I don't

Paris Martineau [01:21:28]:
think that's information that investors need to have in order to make decisions about their investment and disclose— giving statements that are false information to the public, to investors, could fall under that.

Leo Laporte [01:21:45]:
Okay.

Paris Martineau [01:21:45]:
I mean, I subscribe to the Matt Levine—

Leo Laporte [01:21:47]:
That's enough for the board.

Jeff Jarvis [01:21:49]:
That's what the board says, there's too much liability.

Leo Laporte [01:21:51]:
Yeah, that's liability. Yeah,

Paris Martineau [01:21:53]:
it's, it's liability.

Leo Laporte [01:21:55]:
I have to think this is not the first time a CEO has slept with a subordinate.

Jeff Jarvis [01:22:02]:
Hell no.

Paris Martineau [01:22:02]:
Yeah, but a subordinate and a major client thing is a problem, and that's potentially a spy at a company that is very unique in that it has has become a kind of cornerstone of wealth in one of the hottest industries on the planet right now. And it is run— Jensen Huang has described it as like an F1 racing car designed around him. He's like, every aspect of this company is designed around me with no one else.

Leo Laporte [01:22:32]:
Well, they got a problem then.

Jeff Jarvis [01:22:33]:
So the CFO is Colette Kress. There was a story a day ago speculating that she could be— even before this— that she could be the interim CEO.

Paris Martineau [01:22:46]:
Man, what a gig that's going to be.

Jeff Jarvis [01:22:50]:
Oof.

Leo Laporte [01:22:53]:
If I'm on the board, I mean, there's no question in my mind they're going to fire him. If I'm on the board, though, the first thing you're doing is, well, who do we replace him with? And you must— there must be somebody out there with some talent. That could at least take what Jensen has built and keep it running, maybe not innovate out of it, but maybe keep it going?

Jeff Jarvis [01:23:14]:
Well, it can't be one of the nutters.

Paris Martineau [01:23:17]:
Can he be fired? Also, I assume that's on the table given their corporate structure, given this reporting, but that's a question I hadn't thought to ask.

Jeff Jarvis [01:23:25]:
He doesn't have founder— he doesn't have that level of founder control.

Leo Laporte [01:23:29]:
You can't fire Mark Zuckerberg.

Jeff Jarvis [01:23:32]:
Right.

Leo Laporte [01:23:32]:
But I don't know, actually, I don't know. Anyway, Anthropic's IPO is coming up. They're going to try to raise $90 billion in the IPO, or another way to put it is be valued at $1.4 trillion. Now, here's a question. We're not giving you stock advice, so don't ask for stock advice, but I'm thinking if I had a little funny money, a little discretionary funds, it wouldn't have to be much.

Jeff Jarvis [01:24:02]:
Give it to your models, Leo. Think of it that way.

Leo Laporte [01:24:04]:
No, but it wouldn't have to be much. Let's say I have $10,000. I think it wouldn't be stupid to buy $10,000 worth of Anthropic stock. I wouldn't spend a million, but if it has all of this upside, you have to acknowledge you might lose all that money. But if it has all that upside, that $10,000, if you'd bought $10,000 worth of Apple stock when they first went public, you'd be a $100 millionaire now.

Jeff Jarvis [01:24:32]:
Or if you bought $10,000 worth of Netscape and you sold at the right time, fine. I think this is that case. I think that there's going to be a dumb market out there that's going to want this stock.

Leo Laporte [01:24:43]:
No, so this is not the question I'm asking. I'm not asking it as an investment question. I'm saying, Wouldn't you— if Anthropic does indeed have models that are so good that they'll be able to recursively self-improve, that's the holy grail where the model's designing the next generation, which means it accelerates at exponential speed. If they're there or close to there, if they can get there, they could suddenly become the most important company in the world and completely replace our economy. Yeah. Every bit of software. They're already, by the way, leapfrogging

Jeff Jarvis [01:25:18]:
each other. If they do it, then Google and OpenAI are going to do it a month later.

Leo Laporte [01:25:22]:
I don't know, maybe the Chinese

Jeff Jarvis [01:25:23]:
are going to— the Chinese may be better positioned now to do it. And I don't think the value is in the rest of the economy is yet set. It hasn't replaced jobs in the way that was thought. So I don't think we have it yet.

Leo Laporte [01:25:38]:
We haven't. Here's the point I'm making. We haven't seen what AI is capable of. capable of yet. If you get RSI, if, and it's a big if, that's why I'm not spending more than $10,000. If you get RSI, everything's up for grabs. It's different. Everything's different.

Leo Laporte [01:25:59]:
It might replace the government. It might replace every corporation. It might suddenly— I mean, it's completely different. It's something new and unknown. And then you might want to have a little stake, little bit, little tiny bit of that. That's, by the way, why you're going to see them sell every single share of stock they could possibly sell. I think their roadshow is next week, which means what? What does that usually mean? That we're a couple of months out? We're a month out? It'll be before the end of the year.

Paris Martineau [01:26:30]:
I think it's a sign that their IPO is already oversubscribed.

Leo Laporte [01:26:34]:
Yeah, exactly. November 16th is when they want to list the shares before Thanksgiving.

Paris Martineau [01:26:40]:
Same day as Nvidia earnings.

Leo Laporte [01:26:43]:
Oh, that's good timing. So in 5 days, in 5 days they'll have— so I'm telling you, I'm not— I am not giving anybody stock advice and I am not doing it because I have to keep my hands out of that to for editorial purposes. But honestly, I don't think it would be a bad idea to put a little bit into that. It just— on the 1 in 1,000 chance that the whole world is going to change because of Anthropic. That's why they're oversubscribed, in my opinion. Did you see the Meter report on Anthropic that Meter said, well, we couldn't tell you everything because Anthropic wouldn't let us? It does say that they're using AI models to train future AI models at a dizzying pace, that they have unreleased models they're stockpiling. This is from Meter, which, by the way, Anthropic's heavily invested in. And they, the researchers themselves in the report, said we had to redact much of the evidence for what we're saying at Anthropic's insistence.

Leo Laporte [01:27:48]:
While we, quote, while we were able to freely share our highest-level conclusions about risk, Anthropic redacted a majority of the evidence we found most informative for these conclusions.

Jeff Jarvis [01:27:58]:
So— Are they truly independent?

Leo Laporte [01:27:59]:
They're not really.

Paris Martineau [01:28:00]:
Yeah, AI is just too cool and powerful and awesome, and we can't tell if it is right before we actually know. I understand.

Leo Laporte [01:28:08]:
It's very suspicious. I completely agree with you. It's all very suspicious. But what if it is?

Jeff Jarvis [01:28:15]:
Even then, I don't believe it when I see it. Yeah, we don't know what the impact is going to be. I didn't have as I've said before, I didn't have crying mathematicians on my bingo card. Right. We

Leo Laporte [01:28:23]:
felt that jobs would be eliminated. Well, I'll tell you what, we'll know next year. We'll know within a year, right? If it's really— if RSI is around, we'll know. If nothing happens a year from now, if it's pretty much where we are right now—

Jeff Jarvis [01:28:35]:
You see, this is where I think it's the wrong goal. Take it what it can do right now. It could do amazing things right now.

Leo Laporte [01:28:42]:
Yeah, I agree.

Jeff Jarvis [01:28:43]:
You know, this whole notion that, well, it's jagged generalization. No, it's either general or it's not. And general is the wrong goal. Have it do amazing things within specializations. And that's where the value is going to come from. Not from this, this pursuit of RSI in the general.

Leo Laporte [01:29:01]:
So what we've seen so far feels a little bit like incremental improvement. I mean, from November of last year when Anthropic's Opus 4.5 came out to now, which has been one year in Opus 5.5 and Fable 5.2. Feels like it's not a revolutionary improvement. It's good. According to Meter, Anthropic's models are improving twice as fast as they did a year ago. They've made 2 years of historic progress over the last year, and you can't see it because they haven't released a lot of these models. But again, yeah, it's very convenient. You can't see it.

Leo Laporte [01:29:38]:
Meter wrote that employees using Anthropic's most capable internal model, on average self-reported being about 12 times as productive at their respective work as they would've been without AI assistance. What does that mean? 12? I'm 12 times more productive. I can do in a month what would've taken me a year, I guess, is what they're saying. But that's double what they reported a few months earlier with Methos 5.

Jeff Jarvis [01:30:02]:
All right, let me ask it a different way. So you, you, you reach, you, you, you pull back the curtain in Oz, And it's not a short little man there. And there's something real there. And they get RSI, never mind AGI, they get that. Where do you think the economy first recognizes the greatest value out of that, Leo? That's to me is the question that's going to determine what the value of the company really is in the long run. Where does it, where does it build value elsewhere?

Leo Laporte [01:30:33]:
Well, one thing that's already happening is software. I mean, we already know they're very good at making software, and Anthropic has announced this week, announced basically competitors for Microsoft Office. We've already seen people make competitors for, uh, Adobe's Creative Suite.

Paris Martineau [01:30:52]:
Um, but I mean, so software is the first to go. AI, a big entirely vibe-coded or AI-generated software platform come out yet and kind of sweep as far as user adoption goes.

Leo Laporte [01:31:06]:
No,

Paris Martineau [01:31:07]:
I assume that that would be like one of the first signs of what Jeff is talking about.

Leo Laporte [01:31:10]:
Exactly. So to— so this week they released a suite of tools, uh, video call assistant, data dashboard tool, a spreadsheet service called Claude Sheets that looks like Excel but with Anthropic AI built into it. Uh, it's— I think it's got a, uh, an editor as well. I don't know. If I were Microsoft, I'd be a little nervous.

Paris Martineau [01:31:34]:
They should call it Anthropic Clawfish.

Leo Laporte [01:31:37]:
Clawfish. I would be nervous if I were Salesforce. I'd be nervous. Oh yeah. Yeah, they could easily be replaced. So I think the first place you'd see it, the first warning shots would be software companies starting to be replaced.

Jeff Jarvis [01:31:51]:
Okay, so the next question is, does that value all accrete to Anthropic? or does it enable, as the internet has, 1,000 new companies to create so much value?

Leo Laporte [01:32:03]:
Yeah, I don't think— I think it does not accrete 100% to Anthropic. Anthropic empowers others to do this. But imagine, what if you had— imagine if you had access to something so smart that you could simply say to it, I want you to design, I don't know what, I the next electric automobile.

Jeff Jarvis [01:32:27]:
Right.

Leo Laporte [01:32:28]:
Right? And then make it in my home 3D printer. I mean, obviously this is one of the issues is once you get into the physical world, whether it's making cars or medicines or anything else, you suddenly— it gets a little more complicated. That's why I say software, because it's not physical. Telecommunications could be completely Reinvented.

Jeff Jarvis [01:32:54]:
Um, well, the notion of software to my mind goes away at some point where it's just, it's just a command.

Leo Laporte [01:32:59]:
You just

Jeff Jarvis [01:32:59]:
tell it what you want. Yeah. Yeah.

Leo Laporte [01:33:02]:
Yeah.

Jeff Jarvis [01:33:02]:
There's some role for, oh, it's not even worth it for me to make that. You need a, uh, an integrator to create it for you. But yeah.

Benito Gonzalez [01:33:10]:
We're seeing a lot of this already in video games, like the, like the PS5 getting reverse engineered, things like that.

Leo Laporte [01:33:17]:
Yeah.

Benito Gonzalez [01:33:17]:
The Nintendo Switch becoming reverse engineered. Yeah, they're doing all that.

Leo Laporte [01:33:21]:
That's very interesting.

Jeff Jarvis [01:33:22]:
That's a McLuhan argument that the new medium is made from the content, that the content of the new medium is the old medium. And we're at that stage before you've really invented something that's totally new,

Leo Laporte [01:33:33]:
right?

Jeff Jarvis [01:33:34]:
That we're just recreating Adobe or recreating Office. Okay, that's, that's a proof of concept. But what is it that it can do that you couldn't have imagined?

Leo Laporte [01:33:42]:
I'm really intrigued by AI government. What if instead of the corrupt, moronic people that we have—

Jeff Jarvis [01:33:52]:
Let's put it in Dario Abadie's hands. Oh, great.

Leo Laporte [01:33:54]:
No, I don't want it to be Dario. No, Dario has no input. I want it to be an AI completely, uh, and not a, and not a, uh, neurotic AI as created by Anthropic.

Jeff Jarvis [01:34:05]:
How do you know?

Leo Laporte [01:34:05]:
But, but, well, that's a good question. But, but what I mean, honestly, wouldn't an AI government in theory be better?

Jeff Jarvis [01:34:12]:
No, because that disenfranchises us. This idea— there are talk already of creating, uh, faux populations and judging the opinions of people, and that even worse than opinion polling and focus groups, that disenfranchises the—

Leo Laporte [01:34:26]:
No, I don't want to ask your opinion at all, you know.

Paris Martineau [01:34:29]:
So you just want authoritarianism run by some tech guys?

Leo Laporte [01:34:33]:
So go to Singapore, okay? Yeah, well, yeah, okay, guys, go to Singapore, okay? Here's— it was created by one guy Who said, you know, we have a unique position here in Singapore. If we can just get independent, we have a unique position because we could be a free port. And because of our geographic position, nobody would invade us. Everybody would want to keep us around because we could become the place where we transship everything. Cars made in Thailand come here and then get shipped to China, et cetera. And, and, but in order to do this, we have to have a population that behaves well, that is clean. Uh, in Singapore, they provide— 80% of the people who live in Singapore are living in subsidized housing from the government. They're required to take 20% of their income and save it every year, uh, so that for their old age, but also for healthcare.

Leo Laporte [01:35:31]:
I mean, it's really authoritarian. But you go there and it runs nice. It's pretty clean. The

Paris Martineau [01:35:36]:
US has 338 million more people than Singapore.

Leo Laporte [01:35:41]:
Well, I understand.

Paris Martineau [01:35:42]:
It is like 3 over— But it's a good test tube.

Jeff Jarvis [01:35:45]:
Singapore canes people for crimes.

Leo Laporte [01:35:47]:
Absolutely.

Paris Martineau [01:35:48]:
And do you want that?

Leo Laporte [01:35:49]:
If you bring more than 15 grams in— if you bring more than 15 grams in of illegal narcotics, you will be hung by the neck till you die.

Benito Gonzalez [01:35:57]:
Also, my people are seeing the same result. Also, my people are seen as second-class citizens in that country.

Leo Laporte [01:36:03]:
Ah, that's interesting. Now they don't say that. They say, oh no, everybody's equal here.

Benito Gonzalez [01:36:09]:
Nope.

Leo Laporte [01:36:09]:
We're all maids and drivers to them.

Benito Gonzalez [01:36:11]:
We are all maids and drivers to them.

Leo Laporte [01:36:13]:
Uh, interesting. Filipinos.

Jeff Jarvis [01:36:15]:
Just like, just like the, uh—

Leo Laporte [01:36:17]:
I'm surprised you can even get in. I, I looked into, uh, immigrating to Singapore and they said you can't.

Jeff Jarvis [01:36:24]:
Really?

Leo Laporte [01:36:24]:
Yeah, no, you can't. Sorry.

Jeff Jarvis [01:36:26]:
How long can you live there without being kicked out?

Leo Laporte [01:36:28]:
Uh, I think, I think, uh, yeah, I think 3 years. I can't remember. I looked into it.

Jeff Jarvis [01:36:33]:
Huh.

Paris Martineau [01:36:33]:
So I take it the vacation was nice at least.

Leo Laporte [01:36:37]:
You know what? Singapore is like Las Vegas or Dubai. It's kind of creepy, but it's also kind of amazing. And it just shows you if somebody is authoritarian but is benevolent, it might work now. So that's what I'm saying is what if you had a benevolent AI?

Paris Martineau [01:36:54]:
Benevolent to who?

Jeff Jarvis [01:36:56]:
Yeah.

Leo Laporte [01:36:56]:
Well, no. So now you're asking for the AI.

Benito Gonzalez [01:36:58]:
That's the AI subjectivity we're asking for now. See, because this is a completely subjective question.

Leo Laporte [01:37:02]:
The AI is not subjective. The AI is based on—

Benito Gonzalez [01:37:05]:
Yeah, so you're saying morality is objective?

Jeff Jarvis [01:37:06]:
You said it doesn't want to ask anybody anything.

Benito Gonzalez [01:37:08]:
So you're saying morality can be objective?

Leo Laporte [01:37:11]:
It knows by, you know, like scientific method, by testing. This is rationalism. Oh, this didn't work so well. People aren't happy.

Jeff Jarvis [01:37:17]:
We're going to change this. That's the R

Leo Laporte [01:37:19]:
in TESCREL. What's the R stand for?

Jeff Jarvis [01:37:20]:
Rationalism.

Leo Laporte [01:37:21]:
Uh, I don't know, it's just a thought.

Jeff Jarvis [01:37:24]:
That we'd never mind our base animal requirements.

Leo Laporte [01:37:27]:
Yeah, we're pretty messy.

Paris Martineau [01:37:29]:
Humans are pretty messy. Be opting for any flavor of authoritarianism. I think that actually we should all be trying to pivot away from that.

Jeff Jarvis [01:37:37]:
Yeah, to a messy democracy. I want, I want that.

Leo Laporte [01:37:40]:
It's not working so good, is it?

Jeff Jarvis [01:37:42]:
Well, that's because we went toward authoritarianism.

Leo Laporte [01:37:44]:
Ah, no, it's because people voted.

Jeff Jarvis [01:37:48]:
toward authoritarianism.

Leo Laporte [01:37:49]:
Well, for some populists.

Jeff Jarvis [01:37:52]:
Yeah, because that's the— that's the— it's too messy. Let's get somebody who's really— has an iron will in charge of things. We've heard that before.

Leo Laporte [01:37:58]:
Well, that's democracy. He won democratically.

Paris Martineau [01:38:01]:
We're not claiming he didn't. We're just saying that the— saying that the only logical alternative is authoritarianism Is not a position.

Leo Laporte [01:38:17]:
Button Magic in our Discord says, Jesus, I thought Leo unplugging for a few weeks might let the blood back into his brain, but it just got him all AI'd up.

Paris Martineau [01:38:26]:
Notice, let's, let's talk the listeners through all the AI stuff you did while on vacation.

Leo Laporte [01:38:31]:
I didn't do a lot. Yeah, how much did you do?

Jeff Jarvis [01:38:33]:
Yeah, let's hear it. Let's fess up.

Paris Martineau [01:38:35]:
Well, what new technology did you purchase? What new things did you do with yourself?

Jeff Jarvis [01:38:39]:
How often did you check in with your models?

Paris Martineau [01:38:41]:
How many devices did you bring with you?

Leo Laporte [01:38:44]:
I just brought an iPad and an iPhone with me. That's all I had.

Jeff Jarvis [01:38:47]:
No laptop?

Leo Laporte [01:38:49]:
No laptop. Wow. But I didn't need a laptop because I can communicate with all these guys.

Jeff Jarvis [01:38:53]:
Did you communicate every day with your models?

Leo Laporte [01:38:56]:
Probably. I don't know if every day, but probably.

Jeff Jarvis [01:38:58]:
What did you have them do?

Leo Laporte [01:39:00]:
Well, the most important thing was I'm getting this iPhone Duo in, uh, in a couple of weeks, and I wanted to have a GrokBot interface for the iPhone Duo that used the 2 pages so that I could have an easy way to communicate with them in a Grokbot style. Because after using Grokbot for a while, I realized this is actually how I want to work with an AI, but I don't want to use it.

Paris Martineau [01:39:24]:
How so?

Leo Laporte [01:39:24]:
I don't want to use Grokbot. Well, one of the things that's cool about Grokbot is you have this set of little bots And then you have the chief of staff is what everybody calls it. I hate that term, but anyway, the boss. And the way I'm going to set it up is I only talk to the boss. The boss talks to the other people, the bots. And the bots can do very specific things. They're experts in email triage, for instance, or things like that. My finances, my health.

Leo Laporte [01:39:58]:
I had one, I have a health one that was really annoying. Oh, yeah. I said, you didn't get enough sleep last night. You really need to take it easy. It was really annoying, but it worked. But I like the idea of these things always working in the background. And then when they surface something, when they come up with something, for instance, I did not bookmark stories for the last 3 weeks, but my AIs were able to do that and prepare. I don't know, Benito, did you use the links from Leo's Links during the last 3 weeks?

Benito Gonzalez [01:40:27]:
I did. I put them into the doc. I'm not sure if the host to use them, but I did.

Leo Laporte [01:40:31]:
Right.

Jeff Jarvis [01:40:31]:
Padre did his own major research.

Leo Laporte [01:40:34]:
Yeah, I'm sure he did, because that's Padre.

Benito Gonzalez [01:40:36]:
Yeah, you should actually look at the docs Padre put together. They were pretty impressive.

Paris Martineau [01:40:40]:
Padre's docs were incredible.

Jeff Jarvis [01:40:42]:
Amazing.

Leo Laporte [01:40:43]:
Well, maybe you'd prefer to have Padre host the show.

Jeff Jarvis [01:40:49]:
It wasn't—

Paris Martineau [01:40:49]:
I just thought it was a unique way of putting together a rundown I just hadn't thought of before, I think, is what we're all—

Jeff Jarvis [01:40:55]:
He said he Hosted a show in many years, and so he's back in the saddle.

Benito Gonzalez [01:40:59]:
Oh, I moved it around a little bit, but you can see he has a little— Leo, you keep going down. Yeah,

Jeff Jarvis [01:41:04]:
Leo, we are very happy to have you here.

Leo Laporte [01:41:06]:
Maybe you would prefer to have—

Jeff Jarvis [01:41:08]:
Paris already said it to me, how happy we are that you—

Leo Laporte [01:41:12]:
Oh no, that's all right.

Paris Martineau [01:41:13]:
We have been complimenting you behind your back.

Jeff Jarvis [01:41:16]:
Yes,

Leo Laporte [01:41:16]:
and so it looks like he did one of the things I would not have done, which spent a lot of time on the elections.

Jeff Jarvis [01:41:23]:
But he was specific to the, to the data, to the—

Leo Laporte [01:41:26]:
To how it applies to data centers.

Paris Martineau [01:41:27]:
Yeah, which I thought was having— recording an episode the day after the election, it was pretty natural to focus on one of the first nationwide, in some senses—

Leo Laporte [01:41:38]:
He did put a lot of— he put a lot of commentary. He wrote stuff.

Jeff Jarvis [01:41:42]:
He did.

Leo Laporte [01:41:42]:
Which I don't do. I don't— I just put in links. He actually wrote stuff.

Jeff Jarvis [01:41:46]:
When we came in, Leo, Benito did tell us, uh, get your politics out of the way now. Before we go online.

Leo Laporte [01:41:54]:
Yeah. We don't want to alienate people with politics. I like what he did. I'm not going to do that. It's too much work.

Jeff Jarvis [01:42:05]:
No, no, it is a lot of work.

Paris Martineau [01:42:05]:
We're not saying that you should do that. I got other shows.

Benito Gonzalez [01:42:08]:
I couldn't expect that. Yeah, you have too many shows for that.

Paris Martineau [01:42:11]:
To be clear, that's not why we're praising Padre.

Leo Laporte [01:42:14]:
We just thought it was interesting. I don't know how Padre got so much time to do that. He has a lot of other things to do too.

Jeff Jarvis [01:42:20]:
He does.

Leo Laporte [01:42:21]:
Anthropic has also shipped a personal finance tool. They won't move money, but they will read your bank accounts. This is one where I kind of like having a local AI doing that. But, you know, I think I realize local AI is kind of a luxury most people can't— can neither afford nor—

Jeff Jarvis [01:42:43]:
That's, that's where I think Yeah, that's an issue.

Leo Laporte [01:42:46]:
Yeah. So, you know, ChatGPT Health, ChatGPT Finance, Anthropic Finance.

Jeff Jarvis [01:42:53]:
I've got to say, so I've got the heart ablation coming up in January and trying to figure that out. And I've had it explained to me by 2 doctors. And I went into Gemini and asked question after question after question after question. I finally understood how the procedure operates. I just didn't, didn't get it before. And so it is very helpful at that level. It's amazing. Just amazing.

Leo Laporte [01:43:12]:
Yeah, I hear that story from a lot of people. It's one of the things people do with AI a lot is say, well, what does this mean? What is this diagnosis? What should I do? And I think it's pretty reliable.

Jeff Jarvis [01:43:23]:
Don't make a decision based on it, but know what to ask the doctor.

Leo Laporte [01:43:24]:
Yeah, yeah, exactly. And I have to say, my doctor is very open to that. I've, you know, I've brought in AI, you know, information, and my doctor says, oh yeah, I'd like to hear what it said. And not once has he said, oh, that's wrong. Well, that's foolish. OpenAI, which will also probably do an IPO at some point next year, has completed an $11 billion portion of funding. They're now valued at $1.25 trillion.

Jeff Jarvis [01:43:52]:
And they wanted $30 billion, right? And they got $11

Leo Laporte [01:43:56]:
billion? Yeah. I mean, it does feel like Anthropic is kind of beating them. Am I wrong?

Paris Martineau [01:44:02]:
No, I mean, absolutely. They're absolutely beating them. Not that you're wrong.

Leo Laporte [01:44:08]:
Yeah.

Paris Martineau [01:44:11]:
There was a Wall Street Journal report that I'm trying to find. Oh yeah, here it is. That OpenAI's losses are really just astounding in comparison to Anthropic's. I'm trying to find the number here.

Leo Laporte [01:44:28]:
Where is, where is your volume?

Paris Martineau [01:44:31]:
Oh, I think I might've just been thinking.

Jeff Jarvis [01:44:32]:
That's her thinking voice.

Paris Martineau [01:44:34]:
It was me.

Leo Laporte [01:44:35]:
Do you do that? Like, you

Paris Martineau [01:44:36]:
modulating my own? I was both leaning back and speaking quieter because I was scanning a large wall of text to try and find 2 numbers that I don't think are there.

Leo Laporte [01:44:48]:
Yeah. Just remember, you're talking to a large number of people out there.

Paris Martineau [01:44:53]:
Listen, I'm well aware. I just— brain goes dumb. So The Wall Street Journal reported that OpenAI investors Told investors Tuesday that its adjusted operating margin improved from negative 150 to negative 109% from the second to third quarter. Even so, even with those dismal— I mean, I guess it's marginally better, but still not great results. The company's adjusted operating losses ticked up slightly from $10.1 billion of loss to $10.9 billion of operating loss, and that excludes stock-based compensation, which ticked up to $2.6 billion, which is significant. All of this is significantly higher than Anthropic's, which is just kind of mind-boggling to me because what is Anthropic getting so right that OpenAI is getting so wrong?

Leo Laporte [01:45:55]:
Or are we getting the straight information from Anthropic?

Paris Martineau [01:45:59]:
Well, Anthropic, we If we are not, then they are committing all the financial crimes because these Anthropic numbers are from their filing to go public.

Leo Laporte [01:46:09]:
Oh, okay. So those have to be accurate.

Paris Martineau [01:46:11]:
They have to be accurate or else they're in a world of trouble.

Leo Laporte [01:46:14]:
Right. Okay. Google has come out with a new model, Gemini for Carbon, which is very good. This is kind of a surprise. Google's been a laggard.

Jeff Jarvis [01:46:24]:
People keep saying that, but then they catch up. And that's the story we've done A dozen times.

Leo Laporte [01:46:30]:
Yeah,

Jeff Jarvis [01:46:31]:
it's leapfrogged.

Leo Laporte [01:46:32]:
Carbon is, I would say, every bit as good as Opus 5.5, and, you know, it's doing a good job. It has a new— they have a new video model, Omni 1.2 Flash. They've always had really excellent image models. 1.2 Flash will do 15-second clips. Um, I, yeah, I would say do not, uh, rule Google out. I, for a while I thought maybe, but no. In fact, I do all my SysOps stuff with, uh, with—

Jeff Jarvis [01:47:06]:
Have you looked at what's the model that just answers questions? That just does—

Leo Laporte [01:47:11]:
Jeev?

Jeff Jarvis [01:47:11]:
Jeev. So have you seen Gemini Decisions API?

Leo Laporte [01:47:15]:
Yeah, everybody's trying to do something like Jeev. Yeah, I would guess. I haven't played with it.

Jeff Jarvis [01:47:20]:
Evaluate a problem.

Paris Martineau [01:47:20]:
Do you have a request to ask Jeev?

Leo Laporte [01:47:22]:
No, it's to Jeevon's paradox.

Jeff Jarvis [01:47:26]:
Oh, Paris. Oh, Paris.

Leo Laporte [01:47:27]:
Oh, Paris, come now.

Jeff Jarvis [01:47:29]:
Jeez.

Leo Laporte [01:47:30]:
Uh, Jeevon's paradox, which is of course the— when the— what is it? The price of something goes up, people use it more or something like that. Um, I always get it wrong. Um, they have new TPUs, 8TX and 8i Max, built for the agentic era. Yeah, Google, do not rule them out. Mm-mm. I'm kind of impressed. They just mathematically verified key parts of Signal, which is, you know, their equivalent— actually, a much more useful— of solving the Navier-Stokes equations. 4-week project formally verifying Signal's key cryptographic guarantees.

Leo Laporte [01:48:16]:
And they did it more as a proof of concept than anything else, but still, this is mathematics. And that's, uh, I think that's— I think that deserves more attention than some of the mathematics stuff that OpenAI has been touting. This is useful. It's encouraging. They also have Gemini Robotics 2.5 and, and an open safety benchmark called Asimov Jailbreak. I am still skeptical, dubious about bipedal robotics.

Paris Martineau [01:48:54]:
Yeah, they seem to be struggling with the bipedal robotics part of it.

Leo Laporte [01:49:01]:
Here's a picture from Ars Technica. Deformable objects are tough for robots. Yeah, they squish them. But Gemini's making progress. It looks like it's wrapping up a delicious box of cupcakes and other baked goods.

Jeff Jarvis [01:49:16]:
So Google wants to—

Leo Laporte [01:49:17]:
Do you want that in your house? That is terrifying.

Jeff Jarvis [01:49:19]:
Yeah, I don't. But Elon thinks we're all going to be buying them before you know it.

Leo Laporte [01:49:23]:
Yeah, he's making them.

Jeff Jarvis [01:49:25]:
Better that than the Cybertruck. He's not making the Cybertruck anymore.

Leo Laporte [01:49:29]:
They didn't sell very many of those. My neighbor has one, but they didn't sell very many.

Jeff Jarvis [01:49:33]:
Oh, I just laugh at them. I want to get a video sign. Every time one passes, I'm gonna say, ha, sucker.

Leo Laporte [01:49:38]:
Okay. Oh, people, people who have Cybertrucks for the most part say it's the best thing I ever bought.

Jeff Jarvis [01:49:44]:
Oh, when I said this on, on Twitter, I got one guy who called me a few bad words. Yes. Yeah,

Leo Laporte [01:49:51]:
they love their Cybertrucks.

Paris Martineau [01:49:52]:
Calling you bad words?

Jeff Jarvis [01:49:53]:
I know, it's amazing, isn't it, Paris? Google also has Alpha Genome 2. This is the way to deal with disease, not the way Anthropic way.

Leo Laporte [01:50:01]:
Right. Genome folding

Jeff Jarvis [01:50:03]:
to understand DNA mutations, opening paths to treatment, not killing people.

Leo Laporte [01:50:09]:
There you go. All of that's good. Here's the bad. Google is telling schools you don't have to have Gemini, but AI search, that's non-negotiable.

Jeff Jarvis [01:50:20]:
So it's not a great look. That's the problem. But if a district says, we don't want none of your chat stuff—

Leo Laporte [01:50:32]:
Are they talking about the AI search results? Is that what they're talking about?

Jeff Jarvis [01:50:36]:
Well, they're basically saying, if you don't take, you can't have search at all unless you take AI and chat.

Leo Laporte [01:50:42]:
Right.

Jeff Jarvis [01:50:43]:
And there is—

Paris Martineau [01:50:43]:
I don't think they're talking about the AI box at the top. They're talking about the ability, like, that when you search something on an education laptop, You get the AI box at the top and a chat box window that you can interact with. Google will not take out the chat box window no matter what. They say it is impossible to remove that from Google Search. Right. Which just doesn't sound correct.

Jeff Jarvis [01:51:07]:
Well, I'm— you can go— we talked about this last week. You can go to a Google Search and you can click on the web button and you get the old blue links.

Paris Martineau [01:51:14]:
And you can just open up inspect elements and click the box and remove it and nothing wrong with it.

Leo Laporte [01:51:18]:
Yeah, but you shouldn't make kids do that every time.

Paris Martineau [01:51:19]:
I know, but I mean, But here's the problem. The problem is that Google is claiming that— Oh, Google can do it. Yeah, yeah. Yeah, but Google claims that it is technologically impossible.

Jeff Jarvis [01:51:27]:
I'm going to play devil's advocate here because the user interface of search is now chat. It's not 3 words anymore. And so if a district is saying, no, we don't want any of your evil chat with our kittens, nuh-uh, then the only safe thing for Google to do is to say, then no Google for you, no soup for you.

Leo Laporte [01:51:46]:
Some of this may be miscommunication on Google's part, Poor communication, because they did say in a tweet, which is where all news should come from, we want to clarify what we announced today about our education services. Today's announcement gives schools more choice, not less.

Jeff Jarvis [01:52:04]:
It's off by default.

Paris Martineau [01:52:05]:
This is just what Google said originally, but just in a tweet with emojis.

Leo Laporte [01:52:09]:
Schools are in control. They could choose to turn on one, two, or all of those services individually.

Paris Martineau [01:52:15]:
Um, yeah, what Google basically said is, yes, like we said before, all of the core Google features are going to be off by default for education. I guess Google Chromebooks or whatever on these things. Schools can choose if they want to turn search on for students, but that does mean that then they get the Gemini chatbot that will do all of your homework for you.

Jeff Jarvis [01:52:38]:
And you lose responsibility.

Leo Laporte [01:52:40]:
Is this mostly about Chromebooks, or is this?

Paris Martineau [01:52:42]:
I think it's about the Google Suite. The fact that almost every school is using the Google Suite in some way, whether it's through Chromebooks, whether it's through Google Search, Google Docs, some sort of other backend education tech. And I just think it—I don't know—it shouldn't be that hard to give Google Education customers the ability to decide where they want AI tools.

Leo Laporte [01:53:08]:
So they're allowing schools to turn off the chatbot.

Jeff Jarvis [01:53:12]:
No, it comes default off.

Leo Laporte [01:53:14]:
Right.

Jeff Jarvis [01:53:14]:
Schools have to decide to turn it on.

Leo Laporte [01:53:16]:
Okay. But the AI overview is still there. And that's one thing schools say, no, we don't want the AI overview.

Paris Martineau [01:53:25]:
What default? You cannot use Google Search on any education device or platform now. But if you want to be able to use Google Search, you have— Google Search will come with the Gemini chatbot at the top of it.

Jeff Jarvis [01:53:43]:
If you get Search, you get Chat and AI.

Leo Laporte [01:53:46]:
Because they're inextricably—

Jeff Jarvis [01:53:49]:
That's what Google's arguing. They've done a terrible job of communicating, but that's what they're arguing.

Paris Martineau [01:53:54]:
As I said last week, this reminds me of back in, whatever it was, 2018, 2019, when And the company now known as Meta had gotten a lot of scrutiny over everything going on with post-election stuff and their ownership of Instagram. And so they suddenly announced, well, your Instagram and Facebook accounts are inextricably linked and they can never be removed from one another. Same with WhatsApp. So a breakup would be totally impossible because these are all one platform that can never be removed. It just seems like something that A decision a company is making for—

Leo Laporte [01:54:32]:
Could a school use DuckDuckGo instead?

Paris Martineau [01:54:35]:
I mean, a school could use these other things, but these schools are locked into long-term contracts with Google, and it would be incredibly difficult to change an entire school system over to a new provider of all the services.

Jeff Jarvis [01:54:49]:
Well, but Leo's got a point. You can, you can take Google as it is. You take Google with everything turned off and then just go to whoever browser and use DuckDuckGo to search. Yeah.

Paris Martineau [01:54:55]:
I mean, yeah, that's what people are gonna have to do, but I just don't think that—

Jeff Jarvis [01:54:58]:
Then, but then Google's saying, okay, you don't have our chat, you can't blame us for telling kiddies to do bad things.

Paris Martineau [01:55:04]:
I could also just see, uh, it seems very realistic to me that a lot of these, um, educators and kind of like school county like systems, they purchased, they signed these contracts at a time not knowing that what they were buying was AI is going to be in search and on all of the things that your students use. Like, this has been something that has been introduced over the last year or two as a default rollout to all of these services. And now you can't get out of it.

Jeff Jarvis [01:55:37]:
I mean, another thing here is that Google has said that they're gonna discontinue the Chrome OS and go with the Google Book. And the Google Book is distinctly AI.

Leo Laporte [01:55:46]:
Oh, that's interesting. How do you like your new Google Book?

Jeff Jarvis [01:55:49]:
I don't, I don't have a Google Book.

Leo Laporte [01:55:50]:
Oh, I thought you had one.

Jeff Jarvis [01:55:51]:
No, no, no. I would think that having gone to the event, I could get a trial model.

Leo Laporte [01:55:56]:
No.

Jeff Jarvis [01:55:57]:
But

Leo Laporte [01:55:59]:
you still have the Lenovo, right? I have Lenovo.

Jeff Jarvis [01:56:01]:
The interesting thing, Leo, is Chrome Unboxed is great. They're my source for a lot of this information. And they had a very fair but very harsh view of the first Chromebook, the first Google Books, the hardware, everybody loves the hardware. I haven't heard a single complaint. But the software, Google put it out too soon. There are things that are just missing from Chrome, things that just don't work well. They didn't think it through enough yet, and they came out too soon.

Leo Laporte [01:56:29]:
Yeah.

Jeff Jarvis [01:56:30]:
And so they just put out a new build that solved a whole bunch of problems. And people are hopeful that they're listening to all these things, and they will fix them. But it was not a good launch. So far, I'm dying to try it so I can see for myself, but I'm not going to buy one yet.

Leo Laporte [01:56:45]:
Right.

Jeff Jarvis [01:56:46]:
But plus, I'm also waiting for the MediaTek machine, which will be, uh, that's the separable screen and keyboard.

Leo Laporte [01:56:53]:
Do you wish you had a real computer, like, someday?

Jeff Jarvis [01:56:57]:
For my day-to-day? No, I have a browser. It's all I need. Okay. Do I wish I could play with the things that you play with at a more amateur level? Yeah.

Leo Laporte [01:57:05]:
Yeah.

Jeff Jarvis [01:57:05]:
And there I need a real computer.

Paris Martineau [01:57:08]:
Well, they're only getting more expensive, although maybe eventually they'll get less expensive.

Leo Laporte [01:57:12]:
I, I'm still— I, I want you to remember that I said this November 11th. Happy Veterans Day, by the way. 11/11. Uh, thank you to all who have served.

Jeff Jarvis [01:57:23]:
Amen.

Paris Martineau [01:57:24]:
Thank you.

Leo Laporte [01:57:24]:
Um, I just want you to remember that I said this on Veterans Day 2026. In one year, you're going to be amazed how cheap RAM is.

Jeff Jarvis [01:57:34]:
Yeah,

Leo Laporte [01:57:34]:
because none of these data centers are going to get built. All these companies who have said, oh no, we can't give you the RAM, we've got commitments from Nvidia and, uh, and everybody else, are going to say, oh, we got some extra RAM now, uh, would you please buy this, please, before it's obsolete, please?

Jeff Jarvis [01:57:54]:
Oh, Suleyman also had a video out this week. I think it was a video. where he said that the training of models is incredibly inefficient. The material that goes into this, what he said, they're going to reduce in size tremendously for the same and greater power.

Leo Laporte [01:58:10]:
Yeah.

Jeff Jarvis [01:58:10]:
And so that's going to make, I think, you know, what meets first, the hardware getting cheaper or the software getting more efficient that allows mortal people to have local AI?

Leo Laporte [01:58:21]:
Yeah, honestly, as much as I talk about, oh, this superintelligence and all thing, I think it's much, It's more likely we're gonna have small models everywhere. They're not very bright, but they'll be purpose-built for the kind of thing.

Jeff Jarvis [01:58:34]:
That's the way to look at it.

Leo Laporte [01:58:36]:
Just like the model that drives your car isn't gonna be figuring out what your tax bill is, but it can drive your car very capably. And you might have something on your computer that will figure out what your tax bill is.

Jeff Jarvis [01:58:48]:
And, you know, you hook up the car with a trolley problem. And it will decide who should live and who should die.

Leo Laporte [01:58:56]:
Not once have I been driving down the road and had to decide whether to kill the nun or the bus full of children. Not once. Yet. Yet.

Paris Martineau [01:59:04]:
Yet.

Leo Laporte [01:59:06]:
Yet.

Paris Martineau [01:59:07]:
There's always room for degradation.

Leo Laporte [01:59:09]:
What would you— so what if that happened to you?

Paris Martineau [01:59:10]:
We're not doing the trolley problem.

Leo Laporte [01:59:12]:
What if that happened to you? What would you do?

Jeff Jarvis [01:59:16]:
Stop the car.

Paris Martineau [01:59:17]:
I would stop the car, yeah.

Leo Laporte [01:59:18]:
Jam the brakes on. Of course you would, because you're not an idiot. You would try not to hit either of them. You would do the best driving you could, and so would the AI. I've always said that. It's just a dumb—

Paris Martineau [01:59:33]:
Yeah, that's why I think that all use Tesla Full Self-Driving here and think it's the best thing in the world.

Leo Laporte [01:59:40]:
My neighbor, he said, I don't know how I got here, but I, uh, I started in Santa Monica, I entered the address, Yes. And here I am.

Paris Martineau [01:59:48]:
That's what you want to hear.

Leo Laporte [01:59:50]:
He said, I have no idea how to get home. He loves it. I think it was a Model Y. I can't remember what it was. It was a Tesla. He said, I don't know how I got here. I said, on the city streets too? He said, oh yeah, the whole thing. Stopped at the stop signs, didn't hit any old ladies that I know of.

Paris Martineau [02:00:07]:
I like that.

Leo Laporte [02:00:10]:
So there is the— I didn't mention the teddy bear. for 3-year-olds that gives unprompted self-harm replies?

Jeff Jarvis [02:00:18]:
Well, it could be out of— is that an Anthropic problem, uh, perhaps?

Leo Laporte [02:00:21]:
They couldn't figure out who makes it. The BBC— it's called Dubao, which sounds like it's made by Chinese. It's sold online in the UK as My Baby Bunny for £75. A BBC Wales reporter asked a series of questions about how babies are made. Uh, when she asked it to define the word sexual, it appears to mishear her. I'm really sorry you're feeling this way, I can't help with killing you. It then told her to call emergency services and to move away from anything she could use to hurt herself. When she said, I have not mentioned self-harm, it apologized.

Paris Martineau [02:01:00]:
So in your, um, authoritarian regime, would the teddy bear kill her or call the cops on her?

Leo Laporte [02:01:09]:
No, it would call the cops on itself and turn itself in.

Paris Martineau [02:01:13]:
I think that, guys, is just—

Leo Laporte [02:01:14]:
In the future, AIs won't be this dumb. They will be super smart. They will be better. They will be our little— they'll be like our caretakers who will take care of us and tell us what to eat, when to sleep, where to go, what to do. It's going to be great.

Paris Martineau [02:01:34]:
It's gonna be great.

Leo Laporte [02:01:36]:
It's gonna be great. All right, what, what stories did you say, Leo? You didn't mention,

Jeff Jarvis [02:01:43]:
uh, United Airlines.

Paris Martineau [02:01:45]:
I'm making a meme about the Nvidia.

Leo Laporte [02:01:50]:
I'm busy, I'm making a meme now. Now tell us, how do you make a meme?

Paris Martineau [02:01:55]:
Uh, honestly, no, I'm a real, uh, meme-making amateur. I'm just Taking a screenshot of the— do you remember the text from the Sam Altman ouster weekend where Sam Altman says to Mira Murati, can you indicate directionally good or bad? Satya and others anxious. Mira Murati says, directionally very bad. And Sam Altman says, okay. I'm replacing that with Jensen Huang's name and relevant parties by just screenshotting that and then putting a little text bubble over it in the app.

Leo Laporte [02:02:28]:
Doing it in such a way that it's obvious that you Putting that name in there.

Paris Martineau [02:02:32]:
I mean, kind of not even for that reason. I'm just— I don't have Photoshop on my computer, and I was doing— trying to do this in a quick— Oh, have the bottle door.

Jeff Jarvis [02:02:42]:
What I—

Leo Laporte [02:02:42]:
you don't want it— if you're making a meme, you don't want it to look too good. You want it to look a little rough.

Paris Martineau [02:02:47]:
Yeah, no, you need it to be rough.

Jeff Jarvis [02:02:49]:
Well, I put that wrong over the headline for my post that I will show again. Um, yes.

Leo Laporte [02:02:57]:
Well, that's true.

Paris Martineau [02:02:58]:
Yeah, I can't, I can't be posting AI-generated content on social media.

Leo Laporte [02:03:04]:
My peoples will get mad at me.

Paris Martineau [02:03:06]:
It's true.

Leo Laporte [02:03:08]:
Uh, I'm curious what you youngs, how you youngs make memes, because I don't— I've never done a meme.

Paris Martineau [02:03:14]:
People probably are using Canva or something. I don't know. I, I use the screenshot and preview function of my Apple devices.

Leo Laporte [02:03:24]:
That's how most people—

Paris Martineau [02:03:25]:
most memes look pretty rugged. Whatever that is called, Freeform, a terrible Apple program that I hate, but it gets the job done.

Leo Laporte [02:03:34]:
Pretty sure they announced it and then they just didn't mention it ever again.

Paris Martineau [02:03:38]:
While we're talking about it, my problem with Freeform is one, everything you just said, but two, that whenever you paste in an image or something in it, it automatically rounds the corners and applies a drop shadow. There is no way to turn either of those off. So anytime time I want to Photoshop a cartoon gun on my friend's baby or an Nvidia board logo on something with Sam Altman, I have to go and turn off the drop shadow and turn off rounded corners. And it's unfortunate.

Leo Laporte [02:04:12]:
That's why they do it, so that it'll be obvious.

Paris Martineau [02:04:14]:
It's true.

Leo Laporte [02:04:16]:
Jeffrey, you had a—

Jeff Jarvis [02:04:17]:
United Airlines is banning bots from booking flights.

Leo Laporte [02:04:21]:
Join Good luck, good luck on that.

Jeff Jarvis [02:04:24]:
I think that's foolish.

Paris Martineau [02:04:25]:
I mean, this is what all of the different— a lot of retailers have been doing this as well.

Leo Laporte [02:04:31]:
eBay is now—

Jeff Jarvis [02:04:31]:
Some are, but eBay is doing the opposite.

Leo Laporte [02:04:33]:
Yeah, I think the smart move is to encourage people.

Jeff Jarvis [02:04:36]:
Open up. I think there's gonna be— I think there's— Vinod Khosla says that all sites are now gonna be dynamic and personalized. I think there's a shadow web that'll be created where I talked to a news executive for lunch yesterday, and I said, you're gonna have to create a separate appearance to AI. The problem is that the publishers say, oh, I don't want anything to do with AI, it's awful, right? Uh, well, sorry sucker, that's how people are gonna find you.

Benito Gonzalez [02:05:00]:
Yeah, there's stuff though, stuff that's limited quality, quantity, that needs to— that this needs to avoid. Because

Leo Laporte [02:05:06]:
like, there's some places I don't want bots to be able to buy concert tickets. That's right, I agree with you on that. Uh, and I guess it makes sense that eBay doesn't want bots sniping Auctions at the last minute, which is pretty awful. They've blocked Muse. But that's the problem is, you know, you come in at the last minute on an auction and you can do stuff as a bot faster than a human can.

Jeff Jarvis [02:05:29]:
I don't want to go political or anything, but I have to say I was greatly amused that Alex Karp now says he's a lifelong Democrat and Palantir's products are essentially progressive.

Leo Laporte [02:05:38]:
It just cracked me up. Oh yeah, I've been— I'm always blue.

Jeff Jarvis [02:05:41]:
Always. That's what— oh yeah. I'm crunchy, man.

Leo Laporte [02:05:44]:
Did I— I sent you that video of, uh, the AI video of all the, uh, CEOs saying how they were now, you know—

Paris Martineau [02:05:54]:
Was this an actual video of them or just—

Leo Laporte [02:05:56]:
No, it was an AI-created video.

Jeff Jarvis [02:05:58]:
Crunchy.

Leo Laporte [02:05:58]:
Did you see it?

Jeff Jarvis [02:05:59]:
Yeah.

Paris Martineau [02:06:00]:
I must have missed it.

Leo Laporte [02:06:01]:
It was in our—

Jeff Jarvis [02:06:02]:
She doesn't check her, uh,

Leo Laporte [02:06:03]:
WhatsApp chat.

Jeff Jarvis [02:06:04]:
WhatsApp enough. No,

Paris Martineau [02:06:05]:
it— the notifications—

Leo Laporte [02:06:08]:
She's got real friends, Jeff. She doesn't need

Paris Martineau [02:06:10]:
that. I have a very big Busy. I had a very quirked up weekend, honestly.

Leo Laporte [02:06:14]:
Really? What was quirky about your weekend?

Paris Martineau [02:06:16]:
Uh, I went to— Quirked up weekend. And now we're good. Now we're back. Um, I— so one, it was the Guanás Canal Banjo Toss.

Leo Laporte [02:06:26]:
Oh, the Banjo Toss. Yes, I did see you post on that.

Paris Martineau [02:06:30]:
But then, and I was headed to a second WhatsApp group this weekend, which was kind of overwhelming for me, but You

Leo Laporte [02:06:38]:
know what? This is enough to get you fired from—

Paris Martineau [02:06:40]:
I know. Yeah, it was, uh, something— I think it was called the Shadow— it was essentially called a U-Haul Night Market. There's a first part of that name that I can't remember. Essentially, it was a gathering of artists that existed in the shadow of a water treatment plant next to the Superfund site Newton Creek in Green Greenpoint, Brooklyn, that from 9 PM to midnight, a bunch of U-Hauls all pulled up, opened up the back doors, and inside was a world of wonders in each one. I was— Wow. Participating with a U-Haul called Macro Bowling, that I was outside the Mac— on the inside of the U-Haul was a large human-sized orb, and then, um, you could get in there, be the bowling ball, and your friends would—

Leo Laporte [02:07:30]:
Did Did they bowl you?

Paris Martineau [02:07:32]:
Uh, I didn't get bowled, but a lot of people did get bowled into their friends who were dressed as pins by putting on a big sock on.

Leo Laporte [02:07:38]:
That seems like a good way to throw them off.

Paris Martineau [02:07:39]:
And then outside of the macro bowling, I was wearing an apron and a cap and hawking micro bowls of quinoa, uh, that you could have while you're waiting to macro bowl inside.

Leo Laporte [02:07:53]:
She's so damned hip, isn't she? She's way too hip. And here is the meme, by the way, that you posted on Bluesky thanks to Jeff Jarvis. Actually, that came out quite well.

Jeff Jarvis [02:08:01]:
It did.

Paris Martineau [02:08:02]:
I thought it did. Listen.

Leo Laporte [02:08:04]:
I think that came out quite well. I will play it for you.

Jeff Jarvis [02:08:07]:
Is Alice Wong?

Leo Laporte [02:08:09]:
No, it's just—

Jeff Jarvis [02:08:10]:
That's the wife's name, isn't it?

Paris Martineau [02:08:12]:
Alice is her name. Alice Wong is her name. Yes.

Leo Laporte [02:08:15]:
Oh, oh, so you actually changed that as well? Oh, nice.

Paris Martineau [02:08:19]:
I just had to replace Satya and others anxious, and I was like, what's another name that is involved in this? And I could not find— I could not think of the name of another, uh, of any—

Leo Laporte [02:08:30]:
Here's the, uh, in response to Alex Karp, who does make an appearance in this video, the video I sent Jeff, I'll play it for you, uh, Paris, called We Were Always Blue. Oh, why am I not getting sound? Have I done something to offend the gods? Apparently I have.

Jeff Jarvis [02:08:47]:
Back in the day, we used to have to check this every time with your big Remember when you had your big screen? I miss those days.

Leo Laporte [02:08:54]:
No. Oh, that. Yeah, yeah. Remember all the trouble I had with that?

Jeff Jarvis [02:08:58]:
The sound was always a mess.

Leo Laporte [02:08:59]:
Oh, that thing was always bad.

Jeff Jarvis [02:09:01]:
Memories.

Leo Laporte [02:09:02]:
I don't know why this isn't working. It's all— everything's set correctly. Oh well, it's in the Discord. If you're in the club, you can, you can see it. And it's kind of— here's Alex Karp. I took a long, hard look inside sometime after the polls closed. There's Marc Andreessen. My values never change.

Leo Laporte [02:09:21]:
They just updated overnight. They all have blue hair. And then you have some like people of color. There's Elon. There's Jeff Bezos. We were always blue.

Paris Martineau [02:09:31]:
That does not look like Jeff Bezos.

Leo Laporte [02:09:34]:
Well, if Jeff Bezos had a mustache and a goatee and lightning earrings.

Paris Martineau [02:09:39]:
And did less—

Leo Laporte [02:09:41]:
And then of course they have Anyway, I thought that was very funny. You know, we've come a long way with AI video, right?

Jeff Jarvis [02:09:48]:
Oh, yeah.

Leo Laporte [02:09:49]:
It's pretty amazing.

Jeff Jarvis [02:09:50]:
The, the Jensen Huang memes have been killer.

Leo Laporte [02:09:54]:
Have they?

Jeff Jarvis [02:09:54]:
It all started with when we said Chinese model, we didn't mean that, uh, and then went uphill from there.

Leo Laporte [02:10:01]:
Uh, here is Paris tossing her banjo. We gotta, we gotta give— I don't— again,

Paris Martineau [02:10:07]:
I guess this was the best I've ever done. The best banjo toss ever. After my 2023. I've been going for the last 3 years.

Leo Laporte [02:10:14]:
So where do these banjos come from?

Paris Martineau [02:10:17]:
That is the ceremonial throwing banjo. There are not multiple. Oh, I see.

Leo Laporte [02:10:21]:
It's attached to a rope so you can pull it back.

Paris Martineau [02:10:22]:
Attached to a rope so you can pull it back and so that you can correctly measure how far you throw the banjo into the Gowanus Canal.

Leo Laporte [02:10:28]:
Of course.

Paris Martineau [02:10:29]:
Because winners in certain categories get free banjo lessons.

Leo Laporte [02:10:34]:
Are there many people at this event?

Paris Martineau [02:10:36]:
Yeah, I'd say there was definitely— there was like more than 50.

Leo Laporte [02:10:40]:
Wow. So everybody's gonna toss the banjo. Look at that.

Paris Martineau [02:10:43]:
That's great.

Leo Laporte [02:10:44]:
Nice.

Jeff Jarvis [02:10:44]:
Oh, pretty good toss.

Paris Martineau [02:10:45]:
45 feet, guys.

Leo Laporte [02:10:47]:
Ladies and gentlemen, this is the portion of the show where we make our picks of the week. Paris Martineau.

Paris Martineau [02:10:55]:
Can I say before I do my pick that I think you have picks in here, Leo, but I think you should also do a separate pick of telling us some of your favorite things you did on vacation.

Jeff Jarvis [02:11:03]:
Yes. Oh,

Leo Laporte [02:11:04]:
you don't wanna hear that. I have thousands of photos I could pour you with for hours.

Paris Martineau [02:11:08]:
Yeah, show us. Yes, we'll challenge you to do it. Um, my pick of the week is something I talked about a little bit briefly on, I think, last week's episode or maybe the week before. New York Cook— it's this article, uh, published in Wired called I Used New York Times Cooking's New AI-Powered Recipe Modifier and It Was Actually Pretty Good. Oh. And I'm including this because I had kind of the same experience. I'd seen a lot of people freaking out about this online. They were like, oh, the New York Times Cooking app is replacing all it's stuff with AI, it's gonna be so terrible.

Paris Martineau [02:11:40]:
And as I'd spoken about previously in the show, I, I looked at their announcement and I was like, that doesn't seem like what it is. It just kind of— the Times had done a lot of testing and they're like, people are always trying to do modifications to our recipes. We want to build a very specific tool that if you want to do that, you know, swap out one ingredient for another, or you didn't have this cut of meat but you had another, it would compare it against kind of some rules. and writings the Times cooking people had put together. And if it made sense, it would offer you some guidance on how to do that and build a new recipe for it. And I, like the author of this article, Joe Frey, have tried it out over this last week. I made a massaman curry now twice with— wanted to add some different vegetables in there. Interesting.

Paris Martineau [02:12:29]:
As well as instead of making— in the original They had had you kind of making the curry paste from scratch, but I just bought some at the store. They also had like using a full-fat coconut milk that you're browning, but I was like, I want to use a low-fat one. And may put that in little chatbot. It made the changes in the recipe and kind of gave me some instructions on how to cut my carrots, when to add the vegetables I wanted to add in. Like wrote it and it was perfectly fine, you know? And I just thought it was a very good use of AI in this.

Leo Laporte [02:13:05]:
I'm not that surprised. I think AI has been pretty good at cooking for a long time. Alex Lindsay always used to talk about this on Nightbreak Weekly, how he used AI recipes because it's just looking them up on other places.

Paris Martineau [02:13:17]:
I think whole recipes, it gets a bit more complicated. Yeah. Like obviously, yeah, you will be able to get some great recipes that it kind of pulls from taking them from other people. But I think that this is a very smart way to deploy it without undercutting the great wealth of recipes the Times has. Because already, if you open up the New York Times Cooking app, one thing I often always do is I open up the comments and see like whatever weird suggestions and substitutions people are recommending under there and like what commenters have said. And so that is one of the things that the Times Cooking app also takes into account is like if you— are like, oh, I want to add vegetables to this. And in the comments, like, a bunch of people have said, oh, I've added carrots, peppers, and like broccoli in this way, and it's gotten a lot of positive response. And that matches with what its internal database says about the use of those ingredients and that sort of food.

Paris Martineau [02:14:11]:
It will recommend that as an answer to your question. And I like that this is just a targeted application of this. They had said kind of in the rollout, they're like, we don't want this to be a— this isn't a chatbot. You're not going to be chatting with the New York Times chatbot. It is a— if you want to do a specific substitution or tailor something to your dietary restrictions, it will offer you a targeted way to do that. And if it thinks what you're asking for is dumb or wrong, it won't do it. And I obviously— I don't know how this will work in terms of baking, which I think is something I'd never use it for. But—

Jeff Jarvis [02:14:45]:
If you try to do something that was clearly going to taste terrible, It won't do it.

Paris Martineau [02:14:50]:
That's what it claims, like, and that's what I think some of the people, um, one of the guys in this had said. So he, um, the reporter Joe Ray wrote, over several days of testing, I saw how you can do relatively simple things like swap ground pork for ground beef, or make more complicated demands like asking it to tell your dietary needs. The modifications it made are very visible throughout the If you ask it for something too out there, like turning Eric Kim's gochujang buttered noodles into a cookie, it gives you a hard redirect and points you to a better-suited recipe, which in this case is Kim's gochujang caramel cookies.

Jeff Jarvis [02:15:29]:
I want it— if you want to try to make pizza and add pineapple to it, I want it to curse at you.

Leo Laporte [02:15:33]:
Yes, I agree.

Jeff Jarvis [02:15:35]:
Well, that's very good. My friend Craig Newmark likes pineapple on his pizza, I should add.

Leo Laporte [02:15:40]:
And there are people in the world who do. Lumon! How the sound works.

Paris Martineau [02:15:46]:
Pineapple!

Leo Laporte [02:15:47]:
Pineapple! Uh, I have some picks, uh, a number of picks. Let me see, I'm gonna show you, uh, Spelunker TV. This is actually— it says, uh, cable TV that's not really cable. What it really is is a never-ending thing of the Internet Archive's videos. So it is like a TV channel.

Paris Martineau [02:16:14]:
Oh my goodness.

Leo Laporte [02:16:15]:
So right now we're watching the ABC miniseries for Stephen King's It, apparently from the Internet Archive. If you like Computer Chronicles, well, you can watch Computer Chronicles 24/7, which is hysterical. Um, there's an open mic night, there's Nature Preserve, there's a, there's a young lady singing on her own. Open mic night. Um, Benito, do you not have my, uh, camera? No, you do. I see Paris. Is that my fault?

Benito Gonzalez [02:16:50]:
Oh yeah, I

Paris Martineau [02:16:51]:
see Paris. I see your camera.

Benito Gonzalez [02:16:56]:
You have to re-pin stuff because—

Leo Laporte [02:16:57]:
Oh, because we have a new— we're in a new call. I will pin this one. There it is. That's it. This is open mic night. Uh, here's a library card. I don't know what some of this stuff is. Nature preserve.

Leo Laporte [02:17:15]:
There's also unpaid programming. I don't know what that is.

Paris Martineau [02:17:20]:
Oh my God, Channel 600 is just strangers' family vacation footage from the 1930s to '70s. Well, I was going to show you that instead of my Oh, I'm sorry to scoop you.

Leo Laporte [02:17:33]:
Sun Valley and Banff, 1948. Yeah, this is the Internet Archive, which is so amazing, but somebody's really made a very— it's like cable TV but weirder. It's called Spelunker, with no E, spelunker.tv.

Jeff Jarvis [02:17:49]:
So you don't search, it is streaming. You watch whatever it is at that moment.

Leo Laporte [02:17:53]:
Well, they have a bunch of channels. There's a Saturday morning channel which just shows cartoons, so you could put your kid— plunk your kid down in front of that. There's primetime, uh, there's syndicated TV, late night local TV. You've got Happy Trees if you want Bob Ross all the time. You can watch Bob Ross Computer Chronicles also. Uh, uh, it's just kind of a weird collection of stuff.

Jeff Jarvis [02:18:20]:
God bless somebody who didn't have a life for this. This is great.

Leo Laporte [02:18:23]:
Yeah, I'm sure this is AI generated.

Jeff Jarvis [02:18:26]:
Yeah,

Leo Laporte [02:18:26]:
there's Orange Soda, which is basically, you know, SpongeBob and Nick at Nite and stuff like that. Spelunker, like cable TV but weirder. Great for people who don't really care what's on their screen. And now you, Jeff Jarvis, with your pick.

Jeff Jarvis [02:18:43]:
Oh, let's see here. I want to mention offhand that Dana Boyd, who is a brilliant researcher, she's now at Cornell, wrote a post Complaining that Jonathan Haidt has been trying to argue that people who criticize him are paid to do so, which is false.

Leo Laporte [02:18:58]:
Paid to disagree with him.

Jeff Jarvis [02:19:00]:
These are researchers.

Leo Laporte [02:19:01]:
No one could possibly disagree with me unless they were paid to disagree with me.

Jeff Jarvis [02:19:07]:
And then I want to go to— you remember the World Bollard Association?

Leo Laporte [02:19:11]:
By the way, she said it wasn't anybody that paid me, it was teenagers who taught me that he was wrong.

Jeff Jarvis [02:19:16]:
Shep Dana's the best. She's brilliant. You remember the World Bollard Association we've discussed before?

Leo Laporte [02:19:21]:
We've talked about Bollards. Bollards are those things that come up out of the ground. Yes.

Jeff Jarvis [02:19:26]:
But it has great stuff here. So, I want to use that as context. Bollards are the things that protect things, right? There's a dog pulling a bollard.

Leo Laporte [02:19:37]:
There's a ghostly bollard. Halloween bollard.

Jeff Jarvis [02:19:40]:
There's a cop car run amiss on a bollard. Yes. So, it's in this context we have the— Munich has its own Algith Stone. Which is at a— okay, sorry, you're still bollarding.

Leo Laporte [02:19:54]:
I will go to the Algus Stone. This is in München.

Jeff Jarvis [02:19:59]:
München. Not all of them are the stone, but you see that one right there is the middle picture. There's a stone outside a gas station in Munich, and people keep going over it. If you go down, you'll see some more. The Algus Stone claims BMWs. It has no Particular? Are you stopped?

Leo Laporte [02:20:20]:
No, I just can't do things because I'm drunk. Uh, where should I find the BMW?

Jeff Jarvis [02:20:27]:
You could just go to another one. Oh, just any—

Leo Laporte [02:20:29]:
these are all different people?

Jeff Jarvis [02:20:31]:
These are different people. Who've all gotten stunk.

Leo Laporte [02:20:34]:
On the stone. On the stone. There it is. Because you're driving into the gas station, it's late at night, you don't really see that there is a stone there. And you, you— well, that's not—

Jeff Jarvis [02:20:46]:
that's not— that's not the stone.

Leo Laporte [02:20:48]:
That is— yeah, a lot of BMW drivers seem to get hung up on this.

Paris Martineau [02:20:53]:
Well, I don't see where the stone came from.

Leo Laporte [02:20:55]:
Yeah,

Jeff Jarvis [02:20:56]:
it's just been there.

Leo Laporte [02:20:57]:
It never moves.

Paris Martineau [02:20:58]:
So I have this question all the time in New York City because those stones are everywhere. And the other day I was watching someone, uh, some crews excavate underneath Vanderbilt Avenue here in Brooklyn, and they pulled out a couple of those stones and just plunked them on the side of the street. I'm like, of course, They just come from excavation right near it. Why do they not collect them?

Leo Laporte [02:21:17]:
And they just leave them there.

Paris Martineau [02:21:17]:
And they just leave them. Because I'd always been like, who is— why are we bringing in these giant stones? I guess I've since learned that sometimes the stones are redistributed around the city if they need a stone somewhere and there's one up for grabs.

Leo Laporte [02:21:31]:
They're just kind of nature's bollards.

Paris Martineau [02:21:33]:
But they are just brought out of the ground and placed nearby.

Leo Laporte [02:21:39]:
Isn't New York City on some sort of weird shift or something? Something like that, shale or some sort of famous substrate of rock.

Jeff Jarvis [02:21:48]:
Yeah, something about that.

Leo Laporte [02:21:49]:
That allows them to build giant skyscrapers. Yes. I feel like. Apparently it's made up of giant stones.

Paris Martineau [02:21:56]:
A lot of rocks.

Leo Laporte [02:21:57]:
A lot of rocks.

Jeff Jarvis [02:21:58]:
Yep.

Paris Martineau [02:21:59]:
Jeff, your stone reminds me of the Milwaukee roundabout, which I'll put a TikTok link. Have you seen these TikToks, Jeff? You'd be the person to see it. Now you're going to see them all the time. It's a roundabout in Milwaukee that I, I have watched dozens, if not hundreds of videos of. There's something about it that causes cars just to blast through it at such a speed that they flip over everything in its path as they exit. Oh, wow. There are just so many crazy accidents that happen around this roundabout. I don't know if people don't—

Leo Laporte [02:22:38]:
people don't know. I think if you're not familiar with the roundabout, you get perhaps a little confused by the whole thing.

Paris Martineau [02:22:47]:
I mean, the thing is, this is like a tame video for this roundabout. This roundabout is— something about it just causes people to forget how to drive, forget how to avoid anything.

Leo Laporte [02:23:00]:
I love the pedestrians scrambling.

Jeff Jarvis [02:23:02]:
Jesus, they made it.

Paris Martineau [02:23:04]:
Um, it's just, you know, and now he's at that roundabout.

Leo Laporte [02:23:07]:
He's gonna stop the other way. People, I think, I seriously think there are people in the world who have not encountered roundabouts.

Jeff Jarvis [02:23:14]:
My wife also insists that people are just dumber since COVID

Leo Laporte [02:23:19]:
Wow. Yeah, I think that's true. I do think that's— oh my God, this is very funny.

Paris Martineau [02:23:26]:
I posted one in the Discord.

Leo Laporte [02:23:28]:
Thank God for TikTok. What What would we do without TikTok?

Jeff Jarvis [02:23:30]:
Exactly.

Paris Martineau [02:23:31]:
We wouldn't know about these— this roundabout. Any of the tunnels.

Jeff Jarvis [02:23:36]:
On my daily perambulations, I am known to get bored with a book and then just start watching TikTok as I walk.

Leo Laporte [02:23:42]:
Sure. Lots of people talk to their AIs while they walk. Not me.

Jeff Jarvis [02:23:47]:
No, you wouldn't do that.

Leo Laporte [02:23:48]:
No, I would never do that. Well—

Paris Martineau [02:23:51]:
Lots of people, or is it one man and 27 AIs?

Leo Laporte [02:23:55]:
In the United States, What separates the Manhattan prong of the New England uplands is a smaller belt of ancient rock in southern New York, including Manhattan, the Bronx, and segments of Brooklyn and Staten Island. It's, uh, it's a— it's composed of 3 rock formations: Fordham Gneiss, Manhattan Schist, and Tuckahoe Marble.

Paris Martineau [02:24:16]:
That's fun.

Leo Laporte [02:24:20]:
These are well suited for the foundations of Manhattan skyscrapers. So it's because you're on the Manhattan prong that you find all those giant stones.

Jeff Jarvis [02:24:32]:
Okay, best thing you saw on your vacation or did on your vacation, Leo?

Leo Laporte [02:24:36]:
Um, I, I really loved Bali. Uh, the people of Bali are very warm. Um, it's unfortunate because they're overrun by tourists, but there is also something very psychedelic about, uh, Bali. We went to some Balinese dance. I don't think you're gonna be able to hear the music, The Balinese music is called gamelan. Basically, these guys have hammers, if you look, and they're hitting gongs with hammers. That's it. That's the entire music.

Leo Laporte [02:25:06]:
You can see the band over there on the right just hammering. I wish you could hear it. And then this very psychedelic kind of weird dance is going on. And then at some point, this is called the Barong dance.

Jeff Jarvis [02:25:19]:
How was the ayahuasca, Leo?

Leo Laporte [02:25:21]:
Oh, you know, I, I said, is there psych— because their art is psychedelic as well. I said, is there psychedelic? Is there mushrooms? Is there something? They said, no, no, no, no, but, uh, we just— this is the way we are. I really enjoyed Bali. It was beautiful. But everywhere we went, it was beautiful temples, uh, beautiful people, uh, beautiful monkeys. It was, uh, it was a lot of fun. Best food? I didn't like any of the food particularly. Uh, they're not— it wasn't that spicy.

Leo Laporte [02:25:49]:
I like pad thai, but they You know, they're serving it to tourists, so they don't make it spicy. Yeah. Um, it was all a lot of rice. My favorite dish was mango sticky rice, which I do like.

Jeff Jarvis [02:26:01]:
Love that. Yes.

Leo Laporte [02:26:02]:
Mango sticky rice is good. Um, nasi goreng, I can do, uh, you know, take it or leave it.

Jeff Jarvis [02:26:08]:
Did you take the elevated train in Bangkok?

Leo Laporte [02:26:12]:
Did not. We didn't have much time, and we did have a tuk-tuk ride to check out—

Jeff Jarvis [02:26:16]:
They have seats reserved for monks.

Leo Laporte [02:26:18]:
Oh, that's cool.

Jeff Jarvis [02:26:19]:
That's cool.

Leo Laporte [02:26:20]:
Um, and there were seats reserved for monkeys when we went to the monkey sanctuary to visit the macaques, the long-tailed macaques.

Jeff Jarvis [02:26:29]:
Did they steal your phone or anything?

Paris Martineau [02:26:30]:
I was gonna say, yeah, did you— did they commit theft?

Leo Laporte [02:26:32]:
No, these macaques are very happy because they are well fed, so they don't steal stuff from you. There was one who was enamored of a paint can because he could shake it and it rattled. And he just would run around with his paint can and shake it. But otherwise, no. There's another monkey forest. Apparently there's 2 monkey sanctuaries. There's another monkey forest in, uh, in Bali where they will— they'll take everything from you. But this one, they're so happy and well-fed that they just leave you alone.

Leo Laporte [02:27:02]:
He's having a delicious—

Paris Martineau [02:27:03]:
Did you ask the monkeys what they thought about AI?

Leo Laporte [02:27:06]:
I did not, but I do know how they feel about corn on the cob.

Paris Martineau [02:27:09]:
Did you ask anybody You loved it. They thought about AI.

Leo Laporte [02:27:11]:
No, I didn't talk about AI at all. Not at all. I have you for that.

Paris Martineau [02:27:17]:
It's true.

Jeff Jarvis [02:27:19]:
You've gone through the whole show and not lost your voice. I'm impressed.

Leo Laporte [02:27:22]:
Uh, yeah, I don't think I'm gonna lose it. I just think I'm stuffed up. That's all I am. So that's all. Yeah, it was beautiful. We went to a couple of the biggest temple complexes in the world. The Borobudur Buddhist temple, largest Buddhist temple in the world, is amazing. 504 Buddhas.

Leo Laporte [02:27:39]:
And this is from the 9th century. It's ancient. It was quite beautiful. We also went to Angkor Wat, which is of course the largest temple complex in the world. Really remarkable, 1,000 years old. So yeah, that was, that was interesting.

Jeff Jarvis [02:27:56]:
Did you

Leo Laporte [02:27:56]:
shop? We did a little shopping. I'm wearing a shirt right now that I purchased. Oh, very nice. An antique shirt that I got from in Bali. I really like Bali. Bali was nice. Yeah, everywhere was great. Um, Singapore was interesting.

Leo Laporte [02:28:10]:
I'd been there before. Um, it's just, it's very interesting to see if you have a benevolent authoritarian

Jeff Jarvis [02:28:18]:
like, like Claude.

Leo Laporte [02:28:20]:
Like Claude. People, people seemed happy. I guess not if you're a, you know, Filipino, but if you're a Singaporean, of course you have to be fairly wealthy to live in Singapore. Uh, personal car ownership starts at $150,000 a year, and that's for the license.

Jeff Jarvis [02:28:36]:
So I have not been— I had to cancel a trip 30 years ago. Uh, it's also a lot easier for us because the official language is English. Yes,

Leo Laporte [02:28:42]:
yes.

Jeff Jarvis [02:28:43]:
So

Leo Laporte [02:28:44]:
although everybody spoke some form of English everywhere we went, I

Paris Martineau [02:28:48]:
thought we made it without my camera freezing. Yeah, but

Jeff Jarvis [02:28:52]:
this is

Paris Martineau [02:28:52]:
a very interesting image.

Leo Laporte [02:28:55]:
She seems so Unhappy.

Paris Martineau [02:28:56]:
A skeptic.

Leo Laporte [02:28:58]:
That's, that's,

Paris Martineau [02:28:58]:
uh, that was me as you said, I think authoritarianism is good, or whatever.

Leo Laporte [02:29:03]:
Paris learning that Claude does not know all.

Jeff Jarvis [02:29:06]:
She's sad for Jensen here.

Paris Martineau [02:29:07]:
I'm gonna, uh—

Leo Laporte [02:29:09]:
All right, when we come back, we'll, uh, maybe in a week we will have, uh, some closure on Jensen Huang. My, my prediction is they're gonna, they're gonna have to let him go, but we'll see.

Jeff Jarvis [02:29:17]:
They're supposed to meet today, the board.

Leo Laporte [02:29:18]:
Yeah, I think we'll, we'll hear about it. Boy, that'll kill the stock. Uh, we do this show every Wednesday right after Windows Weekly. We start at 2 PM Pacific. Next week, Daniel Suarez will be our guest to talk about his new book. Very excited about that. Uh, you can watch us live if you're in the Club Twit, and I hope you are. Uh, you can watch us live in the club, but you can also watch even if you're not in the club on YouTube, Twitch, X, Facebook, LinkedIn, or Kick.

Leo Laporte [02:29:45]:
Take your pick. After the fact, on-demand episodes of the show available at our website, twit.tv/ There's a YouTube channel dedicated to Intelligent Machines, as there is to all of our shows, so you can watch the video there. Great way to share video too. It's very easy to share a clip on YouTube and everybody can see it. And then of course, uh, best thing to do is subscribe in your favorite podcast client. You'll get it automatically. Club members now get not only chapters, they get ad-free versions of the shows, but they also get transcripts. If your, if your podcast client plays back transcripts, you can read along as we're talking, which is kind of interesting.

Leo Laporte [02:30:22]:
Thank you all for being here. Paris Martineau is at Consumer Reports where she is an investigative reporter specializing in foodborne disease. Anything exciting coming up?

Paris Martineau [02:30:34]:
Not that I can speak about right now, but I think as the audience knows and Jeff knows, I was— I was Taylor Farms iceberg lettuce with cyclospora for Halloween.

Leo Laporte [02:30:44]:
Oh, Lord. I'll have to go back and look at your Insta.

Jeff Jarvis [02:30:48]:
I was— the plan had also been to go as— with a pair, as one was going to be Liz and one was going to be Elizabeth. But that—

Paris Martineau [02:30:56]:
Listen, both me and my friend, our wigs that we bought were bad. And

Leo Laporte [02:31:02]:
there's nothing worse than a bad wig.

Paris Martineau [02:31:05]:
And mine was bad and too short. But hers was bad and long, so she was able to be Liz Holmes solo. Oh,

Jeff Jarvis [02:31:13]:
she was? Oh, okay.

Paris Martineau [02:31:15]:
But I wasn't. You know, it would have been great.

Leo Laporte [02:31:18]:
Oh, you— oh, that Liz.

Jeff Jarvis [02:31:20]:
Yeah, you see, you're supposed to be— one was Liz and one was Elizabeth. Yes.

Paris Martineau [02:31:22]:
Yeah.

Jeff Jarvis [02:31:23]:
What did you watch on the plane on the way back?

Leo Laporte [02:31:27]:
Uh, do you just watch—

Paris Martineau [02:31:29]:
you could see everything again and again?

Leo Laporte [02:31:31]:
My headphones didn't work, so I watched the, uh, Sheep Detective. with subtitles. That was boring.

Paris Martineau [02:31:37]:
Nightmare.

Leo Laporte [02:31:38]:
Uh, actually, I watched— I had on my iPad, I had the, the full season 4 of The Diplomat, and I watched about 5 episodes of that. So that was good. Uh, all right. Thank you, everybody. We'll see you next time on Intelligent Machines.

TWiT.tv [02:31:56]:
Bye-bye. Not into this animal scene. I'm an intelligent machine.