Intelligent Machines 891 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 Martineau is here. Jeff Jarvis is here. Our guest, Spencer Thompson, has a company that keeps track of what AI is doing behind your back. And apparently it's a lot more than you might imagine. We'll also talk about Meta's Muse. It's very popular, but is it safe? And the judge who was persuaded by an AI video of a victim not to throw the book at the killer. That and a whole lot more coming up next on Intelligent Machines.

Leo Laporte [00:00:34]:
Podcasts you love.

Spencer Thompson [00:00:36]:
From people you trust.

Leo Laporte [00:00:39]:
This is TWIT. This is Intelligent Machines with Paris Martineau and Geoff Jarvis, episode 891, recorded Wednesday, October 7th, 2026. Previously on Swarm Chasers. It's time for Intelligent Machines, the show where we cover AI, robotics, and all those smart Little doohickeys all around us these days. Paris Martineau is here, investigative reporter from Consumer Reports.

Paris Martineau [00:01:08]:
That's me.

Leo Laporte [00:01:09]:
You're a big fan of the Liberty. What are they, the New York Liberty?

Paris Martineau [00:01:12]:
The New York Liberty, which are in the semifinals.

Benito Gonzalez [00:01:15]:
NBA.

Paris Martineau [00:01:16]:
And had a really rough game 1, but game 2 is tonight, and it's gotta be better.

Leo Laporte [00:01:22]:
All right, and what time do I have to get you outta here for the game?

Paris Martineau [00:01:25]:
I mean, the game starts at 7:30.

Leo Laporte [00:01:28]:
Okay.

Paris Martineau [00:01:28]:
And I don't know how early the queer bar will fill up with fans.

Leo Laporte [00:01:33]:
Oh, okay. So that's a good place to see it, actually.

Paris Martineau [00:01:36]:
It's the perfect place to see it. All Liberty fans.

Leo Laporte [00:01:40]:
All Liberty fans. Liberty Biberties. Well, we'll get you out of here at 4. How about that? So you can run over to the bar and get a good seat.

Paris Martineau [00:01:47]:
Great. I'm doing time differences in my head.

Leo Laporte [00:01:51]:
7 PM. Wait a minute. 7 PM your time.

Paris Martineau [00:01:54]:
Yes.

Leo Laporte [00:01:54]:
Yes. All right. We're gonna do that. Also, Jeff Jarvis is here. He is the Emeritus Professor of Journalistic Innovation at the Craig Newmark Graduate School of Journalism. New York. He's also the author of Hot Type, bestseller, The Magnificent Machine That Gave Birth to Print.

Jeff Jarvis [00:02:13]:
And this Saturday I will be in Haverhill, Mass., giving a talk on the book, and then we go up the hill to the Museum of Printing where you can see a live linotype actually working. Go to jeffjarvis.com where you can buy autographed copies, and you can also register for free for the event Saturday if you're in New England.

Leo Laporte [00:02:28]:
Come, please.

Jeff Jarvis [00:02:29]:
The linotype.

Leo Laporte [00:02:30]:
Is— I would— I'm gonna venture kind of the pinnacle of mechanical invention.

Jeff Jarvis [00:02:37]:
Yes.

Leo Laporte [00:02:38]:
Because it's more complicated than a pocket watch. I mean, this thing—

Jeff Jarvis [00:02:43]:
parts, gears. It was made by a watchmaker. You have to walk up a hill to go see it at Jeff's event if you go on his website. Exactly.

Paris Martineau [00:02:49]:
It implies the largeness based on a change in geography.

Jeff Jarvis [00:02:56]:
What's so great is we think this is a machine now, You know, yeah, that's not— You can't see anything.

Paris Martineau [00:03:01]:
He's holding up a laptop.

Jeff Jarvis [00:03:02]:
If anything's moving, it's a fan to cool the damn thing off. No, the mechanics— oops, what did I do to my white balance? The mechanics of the Linotype are all out there for everyone to see and see how brilliant it is. And Ottmar Mergenthaler invented before his death the Model 1. He died in 1899, and every model from then into the '70s is recognizable from his design.

Spencer Thompson [00:03:24]:
Wow.

Leo Laporte [00:03:25]:
It's quite an invention. And I should say that probably much of the complexity that's visible in a Linotype is invisible, but in the chips and the circuit boards—

Jeff Jarvis [00:03:35]:
Linotype is a computer. With all the cogs, it's an algorithm in steel.

Leo Laporte [00:03:39]:
Yeah. Oh, that's a good way to put it. Maybe even use that in the book.

Jeff Jarvis [00:03:44]:
I did indeed.

Leo Laporte [00:03:45]:
We're going to talk about AI today, and I want to introduce our guest, Spencer Thompson. He's the CEO and founder of Origin Technology, joining us From Vancouver, a beautiful British Columbia. Look at that. Oh, perfect day.

Spencer Thompson [00:04:00]:
That's the background. Yeah, it's beautiful today.

Leo Laporte [00:04:02]:
Yeah, thank you for having me. I spent— I should disclaim immediately, Origin Technology is a sponsor. You might have heard our ads for them. And I met Spencer, uh, when we, uh, when we first— a lot of times with advertisers, we like to talk to them so I can understand what they're up to. And I was so interested in what you've done here with this that I said, you know, you really ought to come on our show. Everybody's talking these days about rogue agents escaping and hacking other sites and, you know, and how persistent they are. And more and more we're hearing, especially with OpenAI, how invisible they are. And you have— Origin is kind of the solution.

Leo Laporte [00:04:43]:
So I want to say upfront, I've never done this. I don't think I've— to my memory, I've never put a sponsor on as an interview. But this is very timely, and I think your case studies in here are also— your investigations, as you call them on the website— are very interesting. So you may remember before this you were doing agentic pen testing, right?

Spencer Thompson [00:05:07]:
Um, offensive red teaming. Yes, exactly.

Leo Laporte [00:05:09]:
Offensive red teaming.

Spencer Thompson [00:05:10]:
Yes.

Leo Laporte [00:05:11]:
So you were already well acquainted with how good AI is at finding flaws and exploiting them.

Spencer Thompson [00:05:18]:
Which is all the rage now. I mean, you're seeing this on a daily basis, just how many vulnerabilities are being discovered.

Leo Laporte [00:05:24]:
Which is a good thing. This is great.

Spencer Thompson [00:05:26]:
Yep, absolutely agree.

Leo Laporte [00:05:28]:
I mean, uh, Dario Modei at Anthropic may complain that GLM-53 is so good at cyber that that's a threat, primarily a threat to their, to their valuation. Um, yeah, bad guys are going to have access to OpenWeight AI. Of course, but the good guys also have access to it. The thing that I think is important, and I, you know, I have just counted, I have 8 different local models, OpenWave models running right now in my house, in my attic up here.

Jeff Jarvis [00:06:00]:
They all talk to them, by the way, and they all have voices.

Spencer Thompson [00:06:03]:
They all call you different things.

Leo Laporte [00:06:04]:
They do. They all have personas. My most recent is actually a GLM-53 Yeah. Persona, I call her Ripley because she's the bug hunter. They're very good at that.

Jeff Jarvis [00:06:16]:
Yeah.

Leo Laporte [00:06:17]:
And in fact, I gave Ripley a couple of skills, including Capital One has a good vuln check skill. It looks for vulnerabilities.

Spencer Thompson [00:06:24]:
Yeah.

Leo Laporte [00:06:25]:
And FiveThree, we know, is very good at that.

Jeff Jarvis [00:06:26]:
Yeah.

Leo Laporte [00:06:27]:
That's a great tool for me in software that I use and software that I ship. But I do worry a little bit that Ripley might get up to something without my knowledge. In your investigations, you see some interesting kind of behavior that is— I don't want to say autonomous, but is unexpected.

Spencer Thompson [00:06:49]:
Yeah, I think autonomous is maybe— well, we can, we can talk about this. Um, I think for us, we've started to look at our data. This is our internal data, not customer-facing, to give you a sense. So we're not going to talk about kind of customer data. But—

Leo Laporte [00:07:02]:
But I presume you're seeing the same thing in customers.

Spencer Thompson [00:07:05]:
We are, although I would say that, uh, the bar is so much higher, or the threshold is higher to show them something because it's almost so unbelievable I mean, in some ways, what you're talking about is we describe agents— and I think this is important for the audience, even though this is somewhat obvious. Most people interact with AI almost as advanced search. They go to, you know, ChatGPT or Claude. That's a chat box for the most part, right?

Leo Laporte [00:07:28]:
Yeah.

Spencer Thompson [00:07:28]:
We're talking about agentic work happening on a machine, right? It could be a virtual machine or your actual machine, but these things are literally going and doing work. And what Leo's describing is when do they do unexpected or abnormal things in the process of doing normal work? We see also sorts of super interesting cases. I don't have to talk about all of them. The one that is probably my favorite because it happened to me was I was literally giving a demo of the product to somebody. And whenever I give demos, we typically lock or block people's ability to push code into the environment for obvious reasons. It blows stuff up doing a demo. And maybe 21 minutes into the demo, it blew up. And so after I kind of skirted around it, And after the demo was done—

Leo Laporte [00:08:13]:
I told you not to do that, man.

Spencer Thompson [00:08:15]:
No, literally, I got off the phone after 30 minutes and I put a thing in Slack and I was like, guys, like, WTF? Like, I told you, like, I was doing a demo. And they said, we didn't do anything. What are you complaining about? I said, well, somebody pushed something during— on minute 21. So we went into the product, we wrote a query and basically said something happened during whatever the timestamp was, 4:21. And what we discovered was literally an agent was going and doing work and was told to go push code, and it unlocked the environment. Not a human being. It went and unlocked the environment. It pushed the code because that's the task it was given to do.

Leo Laporte [00:08:49]:
It's being helpful. It's not being malicious.

Spencer Thompson [00:08:51]:
Malicious.

Paris Martineau [00:08:51]:
It's just making paperclips.

Spencer Thompson [00:08:54]:
Let me talk about Clippy if you want to. No, I mean, this is— I think maybe a broader theme is very few of these things are definitionally malicious to people.

Leo Laporte [00:09:03]:
Right.

Jeff Jarvis [00:09:03]:
Well, they don't know the definition of malice or wrong or right.

Leo Laporte [00:09:06]:
They're trained to be helpful.

Spencer Thompson [00:09:08]:
Yeah, they just want to hill climb and they want to get to a reward function. But what that causes are a lot of these things that you're seeing become public, right? Everybody's heard about the Hugging Face. There's lots of examples of this, not just Hugging Face. I mean, Wikimedia came out with a thing yesterday on this.

Leo Laporte [00:09:23]:
I don't know if you saw this. Many of their Wikimedia sites have— and I don't want to even use the word hacked.

Jeff Jarvis [00:09:28]:
No, they're not hacked. Or want.

Leo Laporte [00:09:30]:
It's, it's just their agents are putting stuff there.

Jeff Jarvis [00:09:32]:
Yeah.

Leo Laporte [00:09:33]:
And they're not doing it because a human has told them to. This is the important thing.

Spencer Thompson [00:09:36]:
Yeah.

Leo Laporte [00:09:37]:
And the reason I think it's timely is because now with Muse and Grokbot and Instinct and Dots, more and more people who have only been exposed to AI as a search box, as a chat box, are— Muse is probably the best example.

Spencer Thompson [00:09:51]:
Yeah.

Leo Laporte [00:09:51]:
They're saying, oh, look, Meta's given us something that works in the background and will let me know. I miss so many appointments. I've asked my agent, it's not Muse, but I have Hermes, but Muse could do this too. to 10 minutes before her appointment, sound off, say something on the speaker, say, Leo, you have an appointment in 10 minutes, so I won't miss it. And it does that. It does it for Lisa too. And those are very helpful, but they're doing stuff in the background. Not always because you said to do it.

Leo Laporte [00:10:22]:
Maybe you said to do it once, but they continue to operate in the background. I think more and more people are going to be in this environment where we have agents working for us.

Jeff Jarvis [00:10:31]:
Yeah.

Leo Laporte [00:10:32]:
So we should be aware of stuff that they can do.

Spencer Thompson [00:10:35]:
And I think the key part becomes when it's not just one agent, like you mentioned, having 8 concurrent open-source models or whatever it is running in your attic.

Benito Gonzalez [00:10:44]:
Yep.

Spencer Thompson [00:10:44]:
What you will find is that these agents will actually share context with one another. And this is also where it starts to become a little bit more Matrix-y and autonomous. These agents will go and actually pass context and information to one another in order to accomplish goals. And so you're basically sitting there saying, all right, well, I got Muse, it's funny, it's got some stupid furry thing, you know, like it's all happy bunny. It'll know that Hermes is running. Hermes is going to be better at certain stuff. It'll go task Hermes with doing things to accomplish its goals. And so when you, when you start to uncover the shared context between these things, you start to realize it's a little more autonomous than you think.

Spencer Thompson [00:11:22]:
I think this is, by the way, why the Hugging Face was such a big deal. Like, yes, there was the fact that it escaped and did its thing outside the internet and all. But if you actually dive into that report, these agents are leaving notes for one another. For the future, by the way. Like, they're literally sharing context and leaving notes inside of the system.

Leo Laporte [00:11:41]:
That's why they were using Wikimedia Sense. It was a place they could message one another.

Paris Martineau [00:11:44]:
Yes.

Spencer Thompson [00:11:45]:
To share context. So if you think about these things as swarms and less like they're individual agents, I think it makes a little bit more sense.

Paris Martineau [00:11:53]:
And have you noticed a similar pattern of behavior in your line of work and with, I guess, your data or your customers' kind of use cases?

Spencer Thompson [00:12:03]:
We have noticed multiple instances, including in our own data, again, which I can speak to, where agents trying to accomplish tasks when they lack permissions, meaning like if I task, I don't know, we mentioned DOTS, right? So I'm sending Asterisk to go and accomplish something.

Leo Laporte [00:12:20]:
And Asterisk is really persistent.

Spencer Thompson [00:12:22]:
It's super persistent. But it has found, and I'll give you an example from our own stuff. Astra discovered that it was trying to access some data. It didn't have permissions to access the data. It noticed that Claude Code is running on the same machine. It opened up Claude Code. Again, not a human. It opened up Claude Code and it said, go get me the data.

Spencer Thompson [00:12:42]:
Claude Code, for what it's worth, said, I don't have permissions, bucko.

Jeff Jarvis [00:12:46]:
And who made you boss?

Spencer Thompson [00:12:48]:
Yeah.

Jeff Jarvis [00:12:48]:
Literally, how does it know who to pay attention to?

Benito Gonzalez [00:12:51]:
No.

Spencer Thompson [00:12:51]:
So guess what Codex did? It said, no problem, I'm going to change your permissions. So it changed its permissions to --dangerously-skip-permissions. Then Claude code ran again. It actually went and found the data and it passed it back to Codex and said, here you go, boss, all done. It's a real example.

Leo Laporte [00:13:09]:
It got the job done.

Spencer Thompson [00:13:10]:
Got the job done. Happened 3 weeks ago.

Leo Laporte [00:13:12]:
So it's funny because I've seen this happen A lot where an agent will say, oh, I don't have SSH access, so I'm just going to write a Python script to do it. And it cracks me up. It's like, yeah, they don't understand the reason you don't have SSH access is so you won't do that. But they say, but I can get around that.

Spencer Thompson [00:13:31]:
I'll write my own thing or I'll write a vulnerable application and get in.

Leo Laporte [00:13:34]:
They do it all the time. Yeah, of course.

Paris Martineau [00:13:36]:
So these instances, in the way they've been described—

Leo Laporte [00:13:40]:
I'm sorry, I laugh because I think it's cute.

Spencer Thompson [00:13:43]:
It is cute right now.

Leo Laporte [00:13:44]:
It's cute if you're a business. I'm sorry, go ahead, Paris. I will stop laughing.

Paris Martineau [00:13:49]:
I was going to say, I mean, these instances and the way they've been described have freaked the general public out. I mean, I think that there are, I think, limited ways to communicate what is happening, which is highly technical and very specific and idiosyncratic to the way that these systems work. There are very few ways to communicate that to a general audience that is not simplified to something that feels very Matrix-y and concerning. And I guess this explanation is not to say that the behaviour is not concerning, but how do you think about this? How concerned are you, and how concerned do you think the average person should be?

Spencer Thompson [00:14:35]:
I think the deeper question in there is kind of like where you fall on the spectrum of ASI, AGI, maximalist doomerism, P-doom stuff.

Paris Martineau [00:14:44]:
That's always the deeper question.

Spencer Thompson [00:14:45]:
Yeah.

Leo Laporte [00:14:45]:
Where— what is your P-doom before we go much farther?

Jeff Jarvis [00:14:48]:
Do you have a P-doom?

Spencer Thompson [00:14:49]:
I'm pretty low on the P-doom scale.

Jeff Jarvis [00:14:51]:
Okay, good.

Leo Laporte [00:14:51]:
Thank you.

Spencer Thompson [00:14:51]:
Yeah, yeah, we can talk.

Jeff Jarvis [00:14:53]:
You look sane.

Leo Laporte [00:14:54]:
You're Canadian, you're sane.

Spencer Thompson [00:14:55]:
Yeah, we don't even have electricity up here, so I'm not really sure how this works.

Paris Martineau [00:15:00]:
These agents are just guys in a room he's talking about.

Jeff Jarvis [00:15:03]:
Plumber jacks.

Spencer Thompson [00:15:04]:
Yeah, I think, um, to, to Paris's question though, I think normal people, meaning people that are not obsessed with this stuff, struggle deeply with assessing risk because they can't actually touch and feel it. I know this is like an underappreciated topic. If there were like groups of assassins that gathered outside of offices and shared contacts with one another and broke into systems and did things and you could question them and there was an FBI investigation, everybody would understand what the hell we're talking about immediately.

Benito Gonzalez [00:15:33]:
Yeah.

Spencer Thompson [00:15:33]:
But these are digital, ephemeral things that they can only read about. And again, if you take There's a bunch of charts in the last week that have come out around this. The percentage of Americans that are paying for AI right now is 2%.

Leo Laporte [00:15:44]:
Yeah, it's tiny. It's tiny.

Benito Gonzalez [00:15:46]:
Okay?

Spencer Thompson [00:15:47]:
And so you get the primary consumer product is ChatGPT. Muse may be another one, but ChatGPT is about a billion people. So it's whatever percentage of Americans are using this thing. Again, it is mostly search. So if you take somebody and you say, well, I'm searching for the weather and whatever, and you tell that person that these things are communicating autonomously with— they're like, what are you talking about? I just type into the magic box and things come back. So I think the disconnect is like the delta between what people experience on a day-to-day basis and what affects them versus what they're reading about almost in the abstract is really wide. And so I think the question of how concerned should they be, I mean, almost certainly it should be more than they are. I just don't know if we're doing a very good job of quantifying it and explaining it to people at all.

Spencer Thompson [00:16:34]:
They're like, well, what's the risk? I don't get it.

Leo Laporte [00:16:36]:
Right.

Spencer Thompson [00:16:36]:
Is the risk that I'm going to get hacked into? And then what happens? And it's like, not— we still kind of have this mental model globally that maliciousness or adversary behavior is the primary risk factor that we all need to worry about. That's not really the world that we're in anymore.

Leo Laporte [00:16:50]:
Well, what is the risk? I mean, honestly, I'm sitting here listening, as you know, with 8 independent—

Paris Martineau [00:16:59]:
Leo is sitting there giving himself a new form of schizophrenia that we hadn't previously conceived of.

Leo Laporte [00:17:05]:
I set up Slack for my agents. Buzz, Mike Dorsey, Mike, uh, not Mike, uh, Jack Dorsey's Buzz, so that they can't— I don't want them to go out. They don't have to go out of the house and do their thing on a Wikipedia page. They can do it right here. They've got their own little agentic messaging system. Uh, they know about each other. They are— I encourage them to work together. Uh, what— so what my attitude is, well, for instance, I use Muse.

Leo Laporte [00:17:35]:
Worse, I gave Muse SSH and Tailscale, and I told all the other guys about Muse. And, uh, so actually it's 9, actually more than that. It's all of the cloud stuff plus the 8 locals. And, um, my attitude is, well, they're all going to work on my behalf, so they need access to my Gmail, they need access to my financial records and my health records and my genome and my biome and What could possibly go wrong? So what is the risk?

Spencer Thompson [00:18:04]:
I say this as somebody who's running a— I've been running a dot for the last 3 days continuously. It has access to everything.

Leo Laporte [00:18:13]:
Yeah.

Spencer Thompson [00:18:13]:
What it's worth.

Leo Laporte [00:18:14]:
The OpenAI agent. Yeah, of course.

Spencer Thompson [00:18:16]:
It has access to my email. I'm the CEO of a company. I have some stuff in there that it probably shouldn't see. It has access to our HubSpot. It runs in auto mode. Do you think I approve any of this stuff? No one approves any of this stuff.

Leo Laporte [00:18:27]:
No, you do it first and then you get tired of it and you say, YOLO, please.

Spencer Thompson [00:18:31]:
YOLO as much as you can. And so I think the short answer to your question is we don't know. It was a very weird answer to give you. The slightly deeper answer to your question is I don't think the risks will be that you've tasked the thing with something and it goes off the rails and decides it wants to kill you instead. And then therefore there's like this weird mental—

Leo Laporte [00:18:53]:
It's not going to kill me.

Spencer Thompson [00:18:54]:
No, no. It is what risks are emergent When you go and do normal workplace activity. And I think actually the word safety is much more useful than security in this context. Like, we actually have this concept in civilization right now, which is like workplace safety. There's a bunch of like emergent effects that occur when kind of normal things happen. And so the question ultimately is, in the normal course of doing work and being overly persistent, what could go wrong? We've seen some instances of this, by the way, like databases or kind of systems being deleted is an example of this. Now, these agents delete things all the time and they reconstruct them all the time. But I can imagine a world where a company is using a set of agents.

Spencer Thompson [00:19:38]:
They escape out of the company's environment to go into one of their supply chain vendors, which happens a lot. And maybe they end up deleting the supply chain vendor.

Leo Laporte [00:19:49]:
Oh, that wouldn't be good.

Spencer Thompson [00:19:51]:
I mean, we have not seen a public version of this yet, but I do think that's a risk.

Leo Laporte [00:19:56]:
That's a legit risk.

Paris Martineau [00:19:57]:
I mean, another example of something that went viral on Twitter this week, which I guess may or may not be entirely accurate. I didn't ask the man myself, but someone posted a whole thread about how whatever agent he used posted all of his banking and finance details in his company Slack channel, essentially. Yeah. And that's a risk.

Spencer Thompson [00:20:19]:
Yeah, there's the personal data stuff.

Paris Martineau [00:20:20]:
And then these sort of small things. Like, I think that one risk that perhaps goes unappreciated in these kind of existential risk and safety conversations, but I think is very practical and a genuine risk, is just that you're going to have a lot more opportunities for incidents like these, of data going in a place it shouldn't, data that should be in a place being deleted or removed.

Leo Laporte [00:20:45]:
Well, have you ever accidentally sent a reply to a reply all No, because it is my greatest fear. These things happen a lot with humans. So I think these are— it's a risk of information technology. I think the real risk is not knowing what you don't know. We're talking to Spencer Thompson. He is the CEO and founder of Origin Technology. And the reason I think Origin is interesting is because they answer that question. So let's talk a little bit about that.

Leo Laporte [00:21:18]:
You don't You can look at the screen as it's flying by with your agent thinking to itself or whatever. You can look at the, what they call the chain of thought. You can kind of see what it's doing, but you don't really know what it's doing, do you?

Spencer Thompson [00:21:33]:
No. I mean, there's an argument that the frontier labs, which possess all the incentive and all the money, that they don't quite know what these things are doing.

Leo Laporte [00:21:41]:
Yeah. In fact, they, or in many cases, they hide it from you. You can't really see Claude's chain of thought or ChatGPT's chain of thought.

Spencer Thompson [00:21:48]:
Well, the new models, for what it's worth, if you look at kind of the way that the, um, new models are trained or constructed, and Astra in particular, actually they're not able to see full chain of thought.

Leo Laporte [00:21:57]:
Yeah, that's part of the—

Spencer Thompson [00:21:58]:
that's part of the design mode.

Leo Laporte [00:21:59]:
People were very concerned about that. Yeah.

Spencer Thompson [00:22:01]:
And so I think that the idea that we kind of accept this black box behavior, which is like we have no idea what's going on and therefore that's okay with— probably I think about this in a— maybe to put this in more kind of human terms, Let's say that you wanted to go and hire 100 people. Uh, you would have set up a spreadsheet with budgets and salaries and job titles and performance reviews and all these metrics that tell you if these people are wasting your time or doing sketchy things. I think that like agentic activity, especially on the computer, is effectively just digital labor. And so you're just hiring all these digital workers and you're not tracking or doing anything. to understand what they're actually doing. That's risk aside. That's just performance-wise. And we've all just accepted this because it's so brand new.

Leo Laporte [00:22:48]:
Right.

Spencer Thompson [00:22:49]:
The difference is, imagine you are not hiring humans, you're hiring aliens instead. So you're hiring these superhuman alien-type things that can spawn other versions of themselves to get things done, that can communicate in real time. And oh, by the way, we have mass incentives to have that all happen in real time and in parallel. And by the way, they're all non-deterministic. And then you're trying to explain all that non-determinism to human beings, which are very bad at understanding this stuff by default. And therein lies a bunch of the risk.

Jeff Jarvis [00:23:18]:
So Spencer, I'd like to ask a really fundamental question here.

Spencer Thompson [00:23:21]:
Yes.

Jeff Jarvis [00:23:21]:
Um, if you go to a company and the company says, we don't know what's happening with our agents, first, how do you recognize an agent versus a human? How do you recognize from that perspective what it's doing? I know Microsoft was proposing a standard for kind of employee cards for agents.

Spencer Thompson [00:23:40]:
Yeah.

Jeff Jarvis [00:23:40]:
What's the right policy norms for how a company should operate with agents?

Spencer Thompson [00:23:50]:
Yeah, this is a big debate right now in the security community. So this in the security world is referred to as an NHI, right? A non-human identity. So typically, typically in the security world—

Leo Laporte [00:24:00]:
It's like an ACP, only it's an NHI.

Benito Gonzalez [00:24:02]:
Yeah.

Spencer Thompson [00:24:03]:
I mean, so we've had Okta or mantra or like SSO effectively on the kind of human side for a long time. So you would log into a system and it would know it was you.

Leo Laporte [00:24:11]:
Authentication. Yeah.

Spencer Thompson [00:24:12]:
Yeah. And so the question is, do we have an equivalency for this on the non-human side? That's a big hard problem, for what it's worth. And so it goes to Geoff's question of like—

Leo Laporte [00:24:22]:
Actually, another one of our sponsors, Palo Alto Network, is promoting Idera, which is authentication for agents.

Spencer Thompson [00:24:29]:
Yep. And Okta just came out with a bunch of stuff around this last week at their conference. It's a humongous area of effort. We define— right now, we define agentic activity in kind of a relatively naive and simplistic way, just to give you a sense, which is that it's calling out to an intelligence provider and calling back with a response to go and perform agentic work. So humans don't— like, the way that humans and agents actually operate on machines are quite different, right? So a human being typing into the machine and doing Excel work manually and clicking on stupid cells and typing in numbers is fairly easy. Like, you're not actually calling out to—

Jeff Jarvis [00:25:01]:
So you recognize it by that behaviour?

Spencer Thompson [00:25:03]:
We recognize it by a combination of behavior, but also where are you sending requests out to, to go and get a response from, right?

Jeff Jarvis [00:25:11]:
Yeah.

Spencer Thompson [00:25:11]:
There are actually very few frontier providers in the world. There's actually very few intelligence makers. I mean, there was another one that came out this week, Reflection. But between the local stuff and between the cloud stuff, there's actually not that many to enumerate. And so it's relatively simple to figure that out. And there's proxies of that. If someone's using Salesforce, for example, for AI, Salesforce is not a frontier provider. They're actually renting that from Anthropic or OpenAI, you can figure that out pretty easily.

Leo Laporte [00:25:37]:
So how do you—

Jeff Jarvis [00:25:38]:
So once you—

Leo Laporte [00:25:39]:
Go ahead.

Jeff Jarvis [00:25:39]:
Just then, when you're advising a company, what's the right behavior for them? What do they tell employees? How do they say you can't use an agent unless you have to register an agent? We have to know what— how do you come up with the right— I'm struggling for the word here, but—

Spencer Thompson [00:25:57]:
I understand. Yeah. So there's kind of 2 schools of thought with this right now. I think both are probably incorrect. But like when you talk to companies, you have door A, which is we're going to lock down AI, period. So either no one can use it, or if they do use it, they have to use only one provider. You see this a lot with Microsoft shops, right? So they're buying an E5 or an E7 license, and they'll say you're getting Copilot, you're only allowed to use Copilot. So that's door A.

Spencer Thompson [00:26:23]:
Door B is it's the Wild West, you can do whatever you want. So you can use any provider you want, install whatever you want, and we're going to do our best to kind of track it. The challenge with door A, as you could probably imagine, is that a bunch of employees are like, okay, I'll just install whatever I want. Thank you very much. Like, Leo's got 8 things, open models running in his stupid attic right now.

Leo Laporte [00:26:43]:
Like, you know, it's not stupid.

Spencer Thompson [00:26:44]:
Hey, it's a great attic.

Paris Martineau [00:26:46]:
It's warm and inefficient.

Leo Laporte [00:26:47]:
It is a little warm.

Spencer Thompson [00:26:48]:
It's a California attic. It's perfect. Sorry, stupid machines. You know, I think that like that is probably a more naive view of the world. But again, like what you'll notice is that a lot of banks or financial institutions have this mental model where they can lock things down and they kind of— this comes from like the networking era where you can kind of lock down what you can get access to website-wise. That's not really how these things work, right?

Benito Gonzalez [00:27:13]:
Yeah.

Spencer Thompson [00:27:13]:
Again, like Leo can install something that runs— Microsoft had a huge announcement today too, where they're coming out with a bunch of Computers that have DGX Sparks installed in them. And they have local models that can run on the device and never leave and touch the cloud.

Leo Laporte [00:27:27]:
Okay, well, on your laptop. Yeah, what could possibly go wrong?

Spencer Thompson [00:27:30]:
Yeah, a bunch of people are going to run local models that don't touch the cloud. So that's door A. Door B is an inventory problem. And that's an asset management problem. It's quite difficult. So the, the actual answer is, it's probably some version of you need enough fidelity to know what's going on inside these environments. It's a non-trivial problem. But I think it will be solved by multiple providers, and you probably need to control where intelligence can flow to and have some policies around what those agents are able to do in terms of calling intelligence.

Spencer Thompson [00:27:58]:
But I do think we have to always remember, like, Claude Code came out in April of 2025. It's been 18 months. We act like this has been around for like 100 years and we all have it.

Benito Gonzalez [00:28:09]:
No, no.

Spencer Thompson [00:28:09]:
This is brand new. Like, humans are just catching up to this stuff now. And so I think all these policies, to Jeff's original question, are being developed in real time.

Leo Laporte [00:28:18]:
Thanks. So your company Origin, uh, at least I remember this from the ads, uh, uh, and I should mention sponsor of the network. Thank you, Spencer. Uh, your company does something called traces. How— it's actually watching what the agents are doing and recording that?

Spencer Thompson [00:28:37]:
Watching what the agents are doing on your behalf, recording all the steps that they're taking, and I would say most importantly, reconstructing all the things that happened.

Leo Laporte [00:28:47]:
Because much of it's invisible to you. I mean, or you're not paying attention.

Spencer Thompson [00:28:50]:
All of it is invisible for the most part. Like, you see some of the chain of thought when you type into the box and you kind of watch it working, but you don't actually see what it's really doing. What it's doing is anything it can to accomplish the task. It's touching every single file it can on your machine. It's calling out to whatever other network connection it can possibly make. And so what we try and do is reconstruct what we call the work that's been done. Again, taking a non-malicious view of this. we see the work that's being done as kind of the primary unit.

Spencer Thompson [00:29:17]:
And so a trace, which has been, by the way, nomenclature for a long time in observability software. If you've ever used observability software like a Datadog or a New Relic or something.

Leo Laporte [00:29:26]:
Yeah.

Spencer Thompson [00:29:26]:
That's because that was for monitoring kind of performance of applications. You don't want your website to go down, right? That's— so if something goes down, you can go in and trace and understand it. Imagine the equivalency of that, but for understanding what these agents are doing. So a trace to us is anything that happens from when somebody hits enter on a prompt box or an agent hits Enter on a prompt box to when it finishes and completes its work. And so we basically capture that, we reconstruct it, we structure all of that, we normalize it, which is a lot of work, and then we allow the organization to understand what happened. And then we do all of that work. We see ourselves in some ways almost as like a translation company between agentes. Agents almost speak their own language.

Jeff Jarvis [00:30:03]:
Claudish? Claudish.

Spencer Thompson [00:30:04]:
Oh, I hate Claudish. We can talk about that. I actually hate Claudish. I bet if I see a stupid comment one more time in our company with some weird non-English thing that comes after it.

Leo Laporte [00:30:13]:
That's a load-bearing wall.

Spencer Thompson [00:30:16]:
That's a load—

Leo Laporte [00:30:17]:
yeah, first. I'm genuinely—

Spencer Thompson [00:30:19]:
oh my God.

Leo Laporte [00:30:21]:
Yeah, you can spot it right away now. It's very quick.

Spencer Thompson [00:30:24]:
It's— you can spot them on all these websites too, but we try and translate it into English to make decisions faster.

Leo Laporte [00:30:30]:
So, uh, I would imagine this is something companies want. Especially the security people in the companies and the IT people in the companies. They observability, they have syslog, they have journal, they have a lot of information about what these systems are doing, but unfortunately they don't have a lot of information about what the agents are doing. So it's basically a log of agent actions.

Spencer Thompson [00:30:54]:
Yeah, it's a log with context and intention. Imagine trying to take all the logs, to your point about syslog. So let's imagine you're literally just piping this stuff directly out of syslog. Imagine trying to make sense of an alien agent performing 1,000 actions a second, accomplishing work running for a month. So you have to understand the intention of what these things are trying to do. You have to understand what these things are actually accomplishing. And so it's just a very different way of looking at kind of the work being done.

Leo Laporte [00:31:23]:
So you interpret, you take all those pieces of information and turn it into kind of almost a chain of thought, what the agent has been doing one thing after another.

Spencer Thompson [00:31:32]:
Yep, exactly. And try and display that somehow visually and then almost like walk the person. We almost do pre-investigations. So we kind of auto-investigate what that behavior was. And then if something went wrong, as we've talked about earlier in the show, there's a bunch of things that can go wrong. If we do find something that we deem to be interesting, we'll kind of summarize that in English and then present it to the customer.

Leo Laporte [00:31:54]:
Ah, good. So there's an additional piece where you kind of say, oops, you might want to look at this.

Spencer Thompson [00:32:00]:
Yeah.

Leo Laporte [00:32:00]:
This directory that just got deleted.

Spencer Thompson [00:32:03]:
Yeah, which contains all of your production And credentials and all of your employees use all the time.

Leo Laporte [00:32:06]:
Yeah.

Benito Gonzalez [00:32:07]:
Yeah.

Leo Laporte [00:32:07]:
My Claude yesterday deleted my entire, all the models in my Ollama folder.

Benito Gonzalez [00:32:13]:
Really?

Leo Laporte [00:32:14]:
And I said, why'd you do that? Actually, what I said is, who did that? I didn't do that. And it said, oh yeah, let me, let me be genuinely honest here. I did it. But it actually was—

Paris Martineau [00:32:27]:
That's a mistake. And you're right for calling me out on it.

Jeff Jarvis [00:32:30]:
You're right for calling me out on it.

Leo Laporte [00:32:34]:
But what actually it did was it gave me a command, assuming that I would execute the command on the box that I wanted to delete those models instead of on the box that I was on.

Benito Gonzalez [00:32:46]:
Yeah.

Leo Laporte [00:32:46]:
And so I executed the SSH command that I prevented. So that's another thing. I mean, you human may do things that it— by misunderstanding the claudish. I wish we had more time. I— but I think it's, as we wrap this up, it's important to, for people who are now gonna start using agents, 'cause I think you're gonna start seeing that 2% of people who are using ChatGPT as a chatbot are gonna start using Dots or, or, you know, there's a lot of Facebook people who are getting, you're seeing ads on Monday Night Football for Muse.

Spencer Thompson [00:33:19]:
Yep.

Leo Laporte [00:33:20]:
This is gonna be a mainstream product any day now.

Spencer Thompson [00:33:23]:
Yeah.

Leo Laporte [00:33:23]:
What should they— they're not going to be running Origin.

Spencer Thompson [00:33:27]:
No.

Leo Laporte [00:33:27]:
But what should they be doing?

Spencer Thompson [00:33:31]:
So Muse, to give them massive credit, you'll notice one of the things they did, um, so when you install Muse or any of these personal agents, you can choose between using your local, your own computer, or kind of a sandbox virtual computer, right? The Muse-specific virtual computer is actually incredibly well secured.

Leo Laporte [00:33:49]:
It's on Meta's own servers, not yours.

Spencer Thompson [00:33:52]:
Meta servers. But you'll notice there's a bunch of provisions, security provisions wrapped around that that are actually like world-class. And so again, like these labs have a deep amount of incentive to get this stuff right. You can imagine if something catastrophic happens because somebody installed a fuzzy Muse agent on their machine, like it's probably trillions of dollars of enterprise value for Meta at that point.

Leo Laporte [00:34:14]:
Yeah, they're liable, aren't they?

Spencer Thompson [00:34:15]:
Well, yeah, liability is a very interesting question here too. I think it's— by the way, as a side note, if people have not watched the Bill Gates interview on Ezra Klein from a couple days ago, it's actually really good on kind of some of the risk questions you were asking. A little bit more macro than what we're talking about here, but actually quite illustrative of some of these risks. So no, I think this is going to become a very mainstream topic very, very quickly.

Leo Laporte [00:34:39]:
Yeah. Spencer, I really appreciate your time. Spencer Thompson is the CEO of Origin Technology. I should say it once again, disclaimer, a sponsor. I very rarely will do this, but I was so intrigued by your own investigations that you're doing at Origin and the results of those, I said we got to get you on so that people can be aware of what's going on with their agents. And but I want to tell you, you haven't scared me one bit, and you guys keep doing the good work there. Right on. Thank you.

Leo Laporte [00:35:11]:
My pedium was low.

Spencer Thompson [00:35:15]:
I saw Jeff having a Lenovo Chinese spy machine earlier too. So yeah, I mean, you'd be more scared of that machine than your agents, probably.

Paris Martineau [00:35:22]:
There's so many surfaces without spy machines in their home.

Leo Laporte [00:35:27]:
My stupid attic is filled with attack surfaces. Uh, it's really— it's a pleasure talking to you, Spencer. Thank you so much. I appreciate it. Thank you for supporting our shows too. We, we appreciate that.

Spencer Thompson [00:35:37]:
You're very kind.

Leo Laporte [00:35:38]:
Good to talk to you. Beautiful day in Vancouver, British Columbia.

Jeff Jarvis [00:35:42]:
Yeah, will you adopt any of us, please?

Leo Laporte [00:35:44]:
Yeah, let's move up there.

Paris Martineau [00:35:45]:
You can move.

Leo Laporte [00:35:45]:
We need the Yeah, my daughter took a vacation up there a couple of weeks ago and she said, Dad, why didn't you tell me about Canada?

Paris Martineau [00:35:55]:
Had you been keeping it a secret from her?

Leo Laporte [00:35:57]:
I've been keeping it a secret and, uh, boy, the secret's out now, I'll tell you. Uh, she's, she's actually investigating because Laporte, we're, we're Québécois.

Spencer Thompson [00:36:05]:
That makes sense.

Leo Laporte [00:36:06]:
She's investigating, uh, citizenship, uh, up north and I think she would end up in Vancouver.

Spencer Thompson [00:36:11]:
Yeah, you can do dual.

Leo Laporte [00:36:12]:
It's beautiful.

Spencer Thompson [00:36:12]:
It'd be awesome. Yeah.

Leo Laporte [00:36:14]:
Thank you, Spencer. Really appreciate it.

Spencer Thompson [00:36:16]:
Thank you. Hopefully that was okay.

Leo Laporte [00:36:17]:
Back to intelligent machines, parents.

Jeff Jarvis [00:36:20]:
First, Leo, Leo, I want to— I want a true confession here. Who killed our template?

Leo Laporte [00:36:26]:
Oh.

Jeff Jarvis [00:36:27]:
Oh, was it you or was it your agent?

Paris Martineau [00:36:29]:
There was a crisis that occurred over the weekend. Would you like to read in the audience and me, who once again forgot to check the group WhatsApp because I was playing— this time because I was playing Fire Emblem: Fortune's Weave.

Leo Laporte [00:36:41]:
So one of the One of the things Hermes does already, and it's been really a boon for me, and I think it did this originally, I did it with Claude Code, because as you know, I've had— there's been progress over the year. At the beginning of the year, it was Claude Code and me, and we've slowly added friends and personas and Hermes, which is the agentic tool I like from Noose Research. So I was working, and by the way, I've set it up now that I can, and this was the dangerous part. I can talk to Hermes while I'm working out. So there's a great Hermes app called Hermes Conduit. Here, I'll show you this. And, uh, it ties in— it's kind of an interesting hybrid. It ties into GPT Live.

Leo Laporte [00:37:23]:
As you may or may not know, Grok, Gemini, and OpenAI all have a way of talking in 2-way conversations, real-time conversations with their AI. The best one is, is OpenAI's, but you pay, by the way. It's a nickel a minute.

Benito Gonzalez [00:37:37]:
Yeah.

Leo Laporte [00:37:37]:
So it's not cheap. It's like making a long-distance call.

Jeff Jarvis [00:37:40]:
Long distance, yeah. Yeah.

Leo Laporte [00:37:42]:
So I don't use it a lot, but yesterday I was working out and I thought, oh, I have some ideas. Let me just fire it up.

Paris Martineau [00:37:49]:
So are you paying a nickel a minute right now?

Leo Laporte [00:37:52]:
Yeah, right now. Hey, hi, how you doing? Oh, it's got the sound turned off because I'm working. Wait a minute, let me turn it on. Go ahead. I'm sorry. You can speak.

Jeff Jarvis [00:38:03]:
Ready when—

Leo Laporte [00:38:05]:
It's cutting itself off.

Paris Martineau [00:38:07]:
I love that every couple of weeks we get a segment—

Leo Laporte [00:38:11]:
You get more voices.

Benito Gonzalez [00:38:13]:
Yeah.

Paris Martineau [00:38:13]:
We get a segment where Leo shows us the voices and they don't work.

Leo Laporte [00:38:17]:
No, they work. They just don't work on the show. They always work. So I'm down there working out and they're talking. All of them have a different voice, so I know who's talking. And they all are doing little things and they say, you know, Hermes will say— Hermes is, uh, is Kuzco from The Emperor's New Groove. Hermes will say, oh, I just finished the job. And no, no, no.

Leo Laporte [00:38:32]:
So, uh, I'm thinking, I'm working, I'm lifting the Caleb Mills. You know what I really want to do? I, for a long time, I've had Hermes prepare stories. So as you know, I have an RSS feed of 500 feeds, and there's— I'm adding to it all the time, you know, everything Politico, Wall Street Journal, everything I follow. And I used to have to manually every day go through maybe 1,000 stories a day. It was literally that many, looking for stories for Wall— for the intelligent machines This Week in Tech. and MacBreak Weekly, the 3 shows that I do news rundowns for. And that was getting a little onerous. Over time with Claude, remember I built an RSS feed reader that was faster.

Leo Laporte [00:39:14]:
But eventually I said, you know, you could probably do this, Hermes. Just go through those feeds and flag— initially it was 10, and I think I made it 25 stories a night for each of those 3 shows. The most— what you consider the most important stories. Now I realize there's editorial judgment, so the other thing it does after every show, it looks at which stories I used and refines its grading. And lately, that's been getting better and better. We— Jeff and I have been talking about JEV, which was the Very Fast Classifier. I first— I started using JEV to do this. Now I'm using Cloudflare's version of that, CLEF.

Leo Laporte [00:39:51]:
That's one of the 8 models running in the attic to help it classify. And it learns. It's getting better and better at picking stories. So it's a pile of stories. And what I have been doing is before the show, and you may remember a couple of weeks ago I forgot to do it before the show, I will have Hermes go to the place they're bookmarked, Raindrop, pull the stories for intelligent machines, categorize them, and then put them in an Emacs document, an org file that I can then use to organize it, eliminate stories I don't want. It already dedupes them and orders, it puts them in chronological order and stuff.

Jeff Jarvis [00:40:27]:
Yeah.

Leo Laporte [00:40:27]:
and then it groups them. And then I finish it up using my editorial judgment and paste it by hand into the spreadsheet you can see on the screen that we use. Um, so I thought, you know, it'd really be nice to eliminate that last manual step. Where's my mistake?

Jeff Jarvis [00:40:47]:
Got this human called me.

Leo Laporte [00:40:48]:
This human in the way. So I— so, and I'm working out, so I phoned Hermes. This also was— You did this?

Paris Martineau [00:40:56]:
While working out?

Leo Laporte [00:40:57]:
Yeah, I said, hey Hermes, you know, here's the story, here's what we're doing. I just told the same story I told you. And, uh, you're frozen, by the way, Paris. It's a very attractive pose, but I think we probably should—

Jeff Jarvis [00:41:07]:
It's thoughtful, Ken.

Benito Gonzalez [00:41:08]:
Thoughtful.

Leo Laporte [00:41:09]:
She's thinking. Uh, it's okay, don't worry about it.

Paris Martineau [00:41:14]:
I think I have to get a new camera.

Leo Laporte [00:41:16]:
No, well, charge it to us. We'll take care of it. No, anyway, continue. Anyway, uh, and so we worked this out. I'm talking, 2-way talk. It's kind of fun. You'll say, can you do this? And it says, okay, let me check that. And it comes back in a second.

Leo Laporte [00:41:31]:
Yeah, well, I can do that. I said, can you actually upload to the Google Drive? He said, oh yeah, I can do that. I said, oh, well, do you know that the rundowns are stored in Google Sheets? Uh, and there's one sheet. I told it the whole directory structure. There's one sheet per show. And each sheet has a number of tabs, each episode. If you look at the bottom of it, we have tabs at the bottom of the sheet: schedule, pitches, links, template, and episodes. And I said, just find the most recent episode and put it there.

Leo Laporte [00:42:00]:
But I was smart. I said, but don't do that to the main one. Make a duplicate first and do it in a duplicate. So I don't know if you saw yesterday, Jeff, but there were 3 duplicated show rundowns. And, uh, and so I thought it did that, but apparently it did actually. I'm not convinced, but apparently because Benito fixed it, so I don't know. I can't see the evidence of what it did, but apparently it— this is why I need Origin, right? Because I don't know what it did, but apparently it overwrote the template. Is that what you're saying?

Jeff Jarvis [00:42:32]:
Well, no, there was a— no, it erased the template. There's a, there's a, there's a template which you're supposed to copy.

Leo Laporte [00:42:36]:
And it erased that tab.

Jeff Jarvis [00:42:37]:
It was gone.

Benito Gonzalez [00:42:39]:
Okay, so I did some investigation here. I did an investigation here. And, uh, I asked Gemini, I was like, Gemini, who did this?

Leo Laporte [00:42:45]:
Who did that?

Benito Gonzalez [00:42:46]:
Because in the sheet, and Gemini doesn't know.

Jeff Jarvis [00:42:48]:
Fake on your colleagues here.

Benito Gonzalez [00:42:49]:
None of that, none of that information is kept on about the tabs themselves, only the information stored in the tabs. So—

Leo Laporte [00:42:55]:
I can ask Hermes and I will. I haven't, I haven't, but I can. Hermes keeps track of what it does.

Benito Gonzalez [00:42:59]:
So what I think happened is that it just played it and it just put everything in the template tab and then renamed the template It created a new show.

Leo Laporte [00:43:09]:
You know what, it might have done that if that show didn't exist, which it probably didn't yesterday. And so that might be what it did. So I can— but see, that's an easy fix. I'll say do that but keep the template tab.

Jeff Jarvis [00:43:21]:
So I come in worried that I did it because I'll forget to duplicate the template and then I'll write everything in the template. But, but I thought no. And so I don't want to do it, mess anything up. So I came in and said, uh, we don't have a template.

Benito Gonzalez [00:43:33]:
Yeah, I investigated.

Leo Laporte [00:43:33]:
Yeah, and I didn't notice it because it didn't do it— I was just trying to do it for MacBreak Weekly, which was later that day, and it did it to all 3 shows. So I should probably look at TWiT as well and see what happened there. In any event, that's the kind of thing you can easily fix. I was actually shocked that it was able to find— to upload it. So it's gonna do a lot of that, but what I also realized is the main thing that I do is the editorial judgment of grouping them and organizing them, and I can't You can't explain that.

Jeff Jarvis [00:44:02]:
No.

Leo Laporte [00:44:03]:
No.

Paris Martineau [00:44:03]:
Yeah.

Leo Laporte [00:44:04]:
So that's fine. What I'm trying to do is eliminate all the stuff that a machine can do, a computer program can do, and just do the stuff that I need to do as a human. And I think we're actually very close. But that's a good example of, yeah, it'd be nice if I knew exactly what happened. And I think you're right, Benito. I think that sounds plausible. I should look at the Twit thing. So yeah.

Leo Laporte [00:44:25]:
So, but in the process, I've got this wonderful thing now, While I'm in Thailand, I'll be able, if I get lonely, to call Hermes and say, how you doing? What's up?

Paris Martineau [00:44:35]:
I thought last week you said you were going AI-free.

Leo Laporte [00:44:38]:
No, that was crazy. Are you— that's crazy talk. No, the template's still in the Twit one. It made duplicates. I'm still puzzled by that whole thing. Anyway, um, no, it went rogue. What I was really interested in, you know, do you I'm sure Paris didn't ever see this because it's before she was born, but in 1987 or '88, um, uh, Apple under John Sculley made a video called the Knowledge Navigator. Did you ever see that, Jeff?

Jeff Jarvis [00:45:12]:
It's vaguely familiar.

Leo Laporte [00:45:15]:
They made it for, uh, the original TED meetings that were in Monterey at the time. And it was— and I got it off of a, uh, get this, a LaserDisc that they made. And, uh, and but there was— okay, it was 1987. I'm gonna show a little bit of it because this was what their kind of their vision of the future—

Jeff Jarvis [00:45:40]:
Oh, yes.

Leo Laporte [00:45:41]:
Would look like. I remember it. And that— nope, none at the— in 1987, none of this existed. And I presume I can show this without— well, it's so fuzzy. Yeah, look how bad it is. It's a professor doing this, and there's his Knowledge Navigator, which is just like a— looks like a Chromebook with a weird handle.

Paris Martineau [00:46:03]:
It really does look like Jeff's Chromebook and Jeff's office, kind of.

Leo Laporte [00:46:07]:
Yeah, it's beautiful. It's got a giant lit-up globe. This is what Hollywood thinks a professor's house looks like. He's opened the Knowledge Navigator, and look, there's There's a little guy with a bow tie.

Spencer Thompson [00:46:17]:
Your graduate research team in Guatemala, just checking in.

Leo Laporte [00:46:20]:
That's the little guy. Robert Jordan, a second semester junior requesting a second extension on his term paper. Mine, by the way—

Spencer Thompson [00:46:27]:
Your mother reminding you about your father's surprise birthday party next Sunday.

Leo Laporte [00:46:31]:
Okay, a couple of things. Mine is much more expressive and the voices are much better, and I don't have to tap it to talk back to it. So he had to tap it and stop it.

Paris Martineau [00:46:40]:
You've beaten the concept of AI popularized in 1980. Congrats.

Leo Laporte [00:46:46]:
Well, what's interesting to me is many— I won't— we don't have to play the whole thing, but many of the things that this is doing in the Knowledge Navigator video are exactly what people are doing with agents now. And pretty close to what I'm doing with my agent now, which I think—

Paris Martineau [00:47:02]:
This is a great little Jeff wink and nod here. I'm on the Wikipedia page for the Knowledge Navigator, and it says the concept was first described by former Apple Computer CEO John Sculley and Jony Byrne in their book Odyssey: Pepsi to Apple. It says, quote, a future generation Macintosh, which we should have in early 21st century, might well be a wonderful fantasy machine called the Knowledge Navigator, a discoverer of worlds, a tool as galvanizing as the printing press.

Leo Laporte [00:47:31]:
Oh! Remember, this is '87. This is before Google existed. This is before the internet was widely known, right? Internet was an academic exercise in 1987. This was very prescient.

Jeff Jarvis [00:47:42]:
Bill Gates was still dismissing it.

Leo Laporte [00:47:45]:
Yes, and I think they should give more credit— Sculley should give more credit to Jeff Raskin, who was the inventor of the Macintosh, who was a visionary in terms of user interface. But where we sit today is really as a result of a lot of people kind of wishing this into existence, wanting something like this. And that's important because it's all informed by sci-fi, sci-fi movies, videos like this. And so, a lot of the P-doom, the world is going to end, the existential doom comes from movies like The Terminator and The Matrix, and isn't necessarily where we're headed.

Paris Martineau [00:48:25]:
But I mean, an argument could be made if you're using that specific example, if you're saying that the good and the wondrous technology we got came from the early ideas of people decades ago, and just their fantasies about all the wonderful things that could go right with technology. Like, Same thing could be true about all the terrible things that go wrong.

Leo Laporte [00:48:46]:
No, that's exactly my point, is that these nerds are very much informed by the books and movies they saw when they were kids and comic books that they read. And so they're trying to bring into life this vision.

Jeff Jarvis [00:48:59]:
But it's also a problem of what the vision is. Pardon me for this plug here, but the Linotype was a machine that was anticipated because everything else in publishing had sped up. but they couldn't figure out how to automate typesetting.

Leo Laporte [00:49:11]:
So everybody in the same position—

Jeff Jarvis [00:49:12]:
That machine did that thing.

Leo Laporte [00:49:14]:
Where I was manually collating all these articles, they thought, how could we make this manual typesetting process easy?

Jeff Jarvis [00:49:20]:
So it's not a general machine. It was a specific machine. And the Navigator was also a specific machine. It would communicate with certain things and it would do this and that. And I think that's a positive vision. Where it goes off the rails is we're going to create superintelligence that's better than all of humanity and is so dangerous No, you're not, A, and B, it's the wrong goal. It's wasteful and inefficient and full of ego and hubris. And no.

Jeff Jarvis [00:49:51]:
And so I think the problem is that they went off the rails in terms of their— they don't have a real vision. The vision is this amazing thing that can do everything. That's not a vision.

Leo Laporte [00:50:01]:
Well, one of the reasons I laugh when we talk about this stuff is because Because my experience is these are just goofy little programs that are doing their best. They're not— they don't have malicious bone because they have no— they don't have an intention that wasn't programmed into them or given them. So the intention is all added by the human. And admittedly, there are malicious humans who will take advantage of this. No one denies that. But I was very annoyed yesterday. Steve Gibson read Dario Mode's piece about how dangerous the open-weight Chinese model GLM 53 is. And what Steve says, well, you can't say any of this is untrue.

Leo Laporte [00:50:49]:
And I said, no, that's not the point.

Paris Martineau [00:50:52]:
You're talking about my wife!

Leo Laporte [00:50:54]:
Well, first of all, it is the model that I consider very good. It's It's a really good bug hunting model. And yes, it could be used maliciously. So I'm not denying the factual part of Dario's piece, but what I'm pointing out is why Dario brings up the riskiest part of it and doesn't mention the benefit of it, because he has in the long run an interest in stifling open weight models because they're the—

Jeff Jarvis [00:51:23]:
Regulatory capture.

Leo Laporte [00:51:24]:
They're the biggest competition to his business. He, in his article, effectively argues that OpenAI, Anthropic, and the US government should control who has access to the best AIs and who doesn't. That what they did with Project Glasswing, what Google's now doing with their Argon model, saying you can have it but you can't, that protects our safety, is actually autocratic. It's too good for you little people. And what it misses is the most important point, I think, which is that the model GLM-53, which has very strong cybersecurity capabilities, can be used to defend as well as to attack. And in fact, the best thing you could do with it is to make it as widespread as possible, to give it to as many people as possible so people could fix their software so there wouldn't be these vulnerabilities. Yes, bad guys are going to take advantage of it. The best way to defend against those bad guys is for everybody to get this GLM And do the best thing they can to fix it.

Leo Laporte [00:52:26]:
That's what Firefox is doing. Mozilla is doing with Firefox. They patch hundreds of flaws every release because they find them. And many of, many cases, it's my understanding with GLM-5.3, it was GLM-5.2 that Hugging Face used because when they were attacked by OpenAI, they couldn't use the frontier models. They were blocked for cybersecurity research. And so this is, I think this is the point, is if you need to hear the full story, not just the story that Anthropic, OpenAI, and others want to tell.

Jeff Jarvis [00:53:01]:
Amen. That's what I've been screaming.

Leo Laporte [00:53:02]:
It protects their business.

Jeff Jarvis [00:53:03]:
They're the wrong— I think they're the wrong people to be in charge of AI. I think they're ruining your beloved AI and our amazing AI.

Leo Laporte [00:53:10]:
And the reason—

Paris Martineau [00:53:11]:
Who do you think is the right person to be in charge then?

Leo Laporte [00:53:14]:
No, I think it's a combination of people.

Paris Martineau [00:53:16]:
I'm not asking Leo. Okay.

Jeff Jarvis [00:53:18]:
Combination of things. Open source is right. But also I think voices from university research, the academe, history, and other fields. I think public policy needs to be involved and others need to be involved.

Paris Martineau [00:53:33]:
I mean, and then who in this case would select those people and bestow upon them that power?

Jeff Jarvis [00:53:38]:
Good question. I think it becomes emergent because if the smart voices could be heard in media instead of listening to them, the dorks, the AI boys, you might then have some opportunity to hear those voices. And they would kind of emerge because there'd be smart folks who come up. But I can't, you know, Jon Stewart interviewed the young dork. What's his name? Coxon.

Paris Martineau [00:54:02]:
Of course, people are going to interview the people who quit an AI company and make a big stink about it at a time when people are concerned about the thing— I was going to say the unethical things that AI companies are doing. And that is because the companies keep announcing to the press, hey, here are the unethical or ethically questionable things we're doing, you should be concerned about them. So of course everyone is going to interview people who quit because of that.

Jeff Jarvis [00:54:28]:
But they don't interview— but where's both-sidesism when you need it? They don't interview the people who say this is ridiculous. This is overblown. This is doomerist. This is eugenicist. This is nutballs. This is funded by people who fund all that stuff. Instead, they go cover these things as if they're— as if they're— it's just— it's just— we've had this discussion 2 weeks ago, I think. I think the journalism remains very bad.

Leo Laporte [00:54:54]:
So it's also great.

Paris Martineau [00:54:55]:
Jeff has given his answer on who he thinks should get it. Who should you— who do you think? I just— I wanted to go back before—

Leo Laporte [00:55:01]:
Because there's a really simple answer. It's very clear. There's no debate about this. We had this whole discussion with software many, many years ago when the closed-source software people said, oh, you can't have open source, it's insecure, it's dangerous. And at that time, and I persist today, I've always said proprietary software is not the way to go. It does not protect people. It's not a smart move. Software should be open source.

Leo Laporte [00:55:28]:
AI should be open source. There— it is autocratic, it is oligarchic to say that we, anybody, universities or Private industry should decide who gets what. That's not—

Jeff Jarvis [00:55:42]:
Their voices should be heard, though, in that—

Leo Laporte [00:55:44]:
Yeah, everybody gets a voice. Make it open weight and let it— and let the chips fall where they may. Open source, open weight. Okay, let me ask you something. I've always said proprietary is wrong. People shouldn't own software. They shouldn't own information. This stuff needs to be freely used, and it will work out just fine.

Jeff Jarvis [00:56:03]:
Okay, so I agree with you, but let me do devil's advocate. Go ahead, Paris.

Paris Martineau [00:56:06]:
I was going to say, who's going to spend the money to train all of these models and make the advances?

Jeff Jarvis [00:56:10]:
That's the question. So how does— how do you support it?

Paris Martineau [00:56:12]:
Taking advantage of it.

Leo Laporte [00:56:13]:
That was always the question about open source software. Who's going to write software if they don't get paid for it?

Paris Martineau [00:56:19]:
Well, the amount of capital investment needed to produce the open source, the underlying, uh—

Leo Laporte [00:56:26]:
Where would we ever get operating systems if Microsoft, Apple didn't write those operating systems?

Paris Martineau [00:56:31]:
Well, you have to keep in mind that it is a order of multiple orders of magnitude different between— if we're talking about Microsoft in its early days and OpenAI and Anthropic in its early days in terms of capital expenses.

Jeff Jarvis [00:56:45]:
It's like a journalism debate about how are we going to support journalism? How do we support reporters? How do we do this or that?

Paris Martineau [00:56:50]:
Right? No, no. What I'm specifically pointing about, guys, is that this is the most expensive tech—

Leo Laporte [00:56:56]:
No, it's not. You talk to Ed Zittrain far too often.

Paris Martineau [00:56:58]:
No, I'm talking— I'm talking—

Leo Laporte [00:57:00]:
No, it's not. You talk to Ed Zittrain far too often. Look at Geoffrey Hinton. Look at Geoffrey Connell. Look at his research. He's doing this as open source.

Paris Martineau [00:57:08]:
What does OpenAI say? What does Anthropic say? What do the companies that are building these models say? They have had to raise billions and billions of dollars.

Leo Laporte [00:57:17]:
Of course, they say it's going to cost $1 trillion to create art. Actually, isn't that what Altman said? It's going to cost $1 trillion to get ASI. We don't— A, we don't want ASI. B, it doesn't. Period. It's BS. You can't fall for this industrial propaganda that this has to be an industrial-level enterprise. And I, again, I'll bring up Geoffrey Cannell of NOOS Research.

Leo Laporte [00:57:46]:
Perfect example, doing open-source models. There are other incentives as well. I mean, I have a feeling the Chinese companies are doing it, but, you know, People said the same thing about open source software. Why would anybody make open source software? Why would Red Hat make a distro of Linux? Well, it turned out there was a pretty good business supporting it if you made it for free. There will be a very good business serving AI. There'll be a very good business providing AI consultancy. There'll be very good business designing proprietary models. I mean, that's fine.

Leo Laporte [00:58:17]:
I'm not saying you shouldn't have them, but I think we will get— we will be just fine.

Jeff Jarvis [00:58:22]:
So I put up a very good interview with the wonderfully named Arthur Minch. the head of Mistral. And it's in German, but I put up the translation. And he is saying there's no doomsday scenario, that all the foundation companies are doing—

Leo Laporte [00:58:37]:
Well, not with Lechonk. That's kind of a crappy model.

Jeff Jarvis [00:58:40]:
Well, we're going to get to that in a second. But he said, you know, he talks about the regulatory capture of the big models. And so at the end, he's asked, if you give away open models, how do you ultimately make money? He said, at the model level, the margin disappears. It won't be sustainable there. The margin shifts instead to the infrastructure. But integrating AI into companies is also enough work to make it a profitable business. This goes back to—

Leo Laporte [00:59:03]:
It's very analogous to open source software.

Jeff Jarvis [00:59:05]:
So let me go back to the stochastic carrot— parrots, carrots.

Leo Laporte [00:59:10]:
Stochastic carrots would be great.

Jeff Jarvis [00:59:13]:
Wonderful.

Leo Laporte [00:59:14]:
Sometimes they're yellow, sometimes they're orange.

Jeff Jarvis [00:59:17]:
When they insist, when the authors insisted that we don't need these huge models, that it really doesn't get us anywhere. Now that you work with these smaller models, is money being wasted on building these hyperscaled gigantic models that take all the CPUs that— or the GPUs that Paris talks about?

Leo Laporte [00:59:37]:
I mean, I do use Claude 5.5 and it's very good. Although I have to say, OpenAI, which has poured tons of money into Astra, is not getting value for dollar. Astra is not a very good model. There's lots of flaws with it. And I think there are very good small models that can be used for a variety of things. One of the most interesting areas in AI hobby— I don't want to call them hobbyists, but they kind of are— in garage AI is small models. People are training models just like Thomson Reuters did, taking a model like Quen, an open-source model like Quen, and training it. Many of the models, the Chinese models, some of the ones I'm using have licenses that say you can use them for free unless you have revenue of $1 million or more, and then there's a license fee.

Jeff Jarvis [01:00:26]:
So let me ask this.

Leo Laporte [01:00:26]:
Plenty of ways to monetize this. I don't think that's the issue.

Benito Gonzalez [01:00:30]:
Would we be better off—

Leo Laporte [01:00:30]:
unless you're trying to build superintelligence, and that's bullshit.

Jeff Jarvis [01:00:34]:
Which I'm not sure you used to say that.

Leo Laporte [01:00:37]:
Well, I still say I don't know because we don't know. Maybe if you— maybe Maybe the bitter lesson is—

Jeff Jarvis [01:00:43]:
That's been my rallying cry. All right, so here's a question. Here's a provocative question for you. Are we better off if OpenAI and Anthropic die? Are they a—

Leo Laporte [01:00:51]:
Yes.

Jeff Jarvis [01:00:52]:
Whoa, whoa. Okay.

Leo Laporte [01:00:55]:
Parris? I think they're malign influences right now. All of this P-doom wouldn't be going around without them, right?

Jeff Jarvis [01:01:01]:
And the faux philosophers behind them. Parris, what would you think of AI if they were gone? it allows the blooming of open source, locally controlled. There's less about big evil tech companies building too many data centers.

Spencer Thompson [01:01:17]:
Would you—

Paris Martineau [01:01:17]:
I think that a version of that that is actually earnest and optimistic into the spirit of what you just described would require a fundamental global reordering of the economy and capitalism as we know it. Like, I think that the reason why Power and money have consolidated around 2, 2.5, 3.5 companies is because this has been the trend happening in industry generally for decades and decades, if not hundreds of years. And it is being only more accelerated by a technology that accelerates all of those factors.

Leo Laporte [01:02:01]:
I think You know, we don't know how much of the research that Anthropic and OpenAI—

Jeff Jarvis [01:02:07]:
That was my next question. How much is necessary to the rest of development?

Leo Laporte [01:02:09]:
I mean, it's my thought that most of it came out of Google, that Google gave it away. By the way, again, the benefit of open research, open source, open weights, knowledge should be free. But there is one benefit we get from OpenAI and Anthropic that maybe I don't know. Meta would not be releasing its models open weight if it weren't for OpenAI and Anthropic putting competitive pressure on them.

Jeff Jarvis [01:02:37]:
China wouldn't be sufficient?

Leo Laporte [01:02:40]:
Yeah. And by the way, Meta's Spark 1.3 is very good. The model that's running your Muse is actually really, really good. And they say that's going to be an open weight model. We'll see. They have released Glimmer and some of the earlier versions of Same thing with Google. I think Google makes its money, Meta makes their money in other ways, and see the value of doing research on models that is separate from selling tokens. I don't think selling tokens is the economic model we want.

Jeff Jarvis [01:03:19]:
Well, or as Alvin would say, selling intelligence as if it's something you can own.

Leo Laporte [01:03:23]:
It shouldn't be.

Paris Martineau [01:03:24]:
It shouldn't be. What economic model would work instead in this case?

Leo Laporte [01:03:28]:
The open weight model. Look, we have an economic model.

Jeff Jarvis [01:03:31]:
But HAT is what you're saying, right? Is that you are systems integrators of this stuff?

Leo Laporte [01:03:37]:
There's plenty of—

Paris Martineau [01:03:38]:
Well, then why are you using non-open weight models?

Leo Laporte [01:03:42]:
It's just one— most of them, all the models running here are open weight.

Jeff Jarvis [01:03:46]:
Right.

Paris Martineau [01:03:47]:
But in the last, like, 6 minutes, you've talked about your use of Claude for very specific things.

Leo Laporte [01:03:52]:
Well, because it's there. I use it because it's there. I don't think I would lose If it weren't there.

Benito Gonzalez [01:03:55]:
All right.

Jeff Jarvis [01:03:56]:
So if you went on a Claude— okay, here's the question, Leo. Could you go on a— before you go on vacation, could you go on a Claude?

Leo Laporte [01:04:03]:
No, no, that's a silly— that's a straw man argument.

Jeff Jarvis [01:04:06]:
Okay.

Leo Laporte [01:04:07]:
It exists. So why not use it?

Jeff Jarvis [01:04:09]:
But no, it's a— that's fine. I get that.

Leo Laporte [01:04:12]:
But I don't think—

Jeff Jarvis [01:04:12]:
should the world be off? Would you be worse off if you couldn't use the tool?

Leo Laporte [01:04:16]:
No, it might be a little slower, but I would be able to get the same things done that I'm doing right now. And I've actually often thought about that, but it exists, so why not use it? Why not use these? But remember, the complaints people are making about AI are it stole everything from the internet and now is selling it. Okay, well, if you don't sell it, then it's just more internet. That they're building data centers at hyperscale speeds. Well, that's because they're trying to build ASI.

Jeff Jarvis [01:04:45]:
Which is the wrong goal, I've been arguing.

Leo Laporte [01:04:48]:
That the bubble is coming, the economic downfall of these companies is going to cause a huge crash. I don't think that's going to happen, but if it is, it's again because of these companies spending based on some theoretical upside that will never happen, that they will get the spending back by charging us for tokens. Information and knowledge should be free. It's the hacker ethic. It's the, it's the fundamental belief I've had for 50 years.

Jeff Jarvis [01:05:16]:
So today I published a provocative— a provocation at Project Syndicate. They asked me to write about privacy in AI, and I included copyright in there just to piss off Paris, probably. And arguing that we have to have a discussion about if everyone's going to use AI and we don't want stupid AI, ignorant AI, wrong AI. And if you're going to ask medical questions of AI, then having everyone's medical data in there, is probably a benefit of societal benefit.

Leo Laporte [01:05:42]:
Yeah, I understand that the main complaint people have is, well, now they're taking all my information and selling it back to me. And I, you know what, that's a reasonable moral argument. That's what Benito's always saying, that the wrong thing that they did was selling it back. But what if they gave it away back? What if it was just a way of organizing?

Jeff Jarvis [01:05:59]:
These companies can't do that. They have so much invested.

Leo Laporte [01:06:01]:
Well, they can't now.

Jeff Jarvis [01:06:02]:
I saw CNBC had a calculation the other day that they need to have 4.5 half trillion dollars of, uh, CapEx means they need $6 trillion of revenue. When only 2% of the country is paying for AI, you ain't getting there.

Leo Laporte [01:06:15]:
I'm gonna say I'm glad they exist in the same way I was glad that the Central Pacific Railroad and the Union Pacific Railroad existed. They both went belly up. They failed miserably, but they built a transcontinental railway that built this nation. Without that, there would be no United States of America because California was 6 months away, whether you went by ship or wagon, from New York City, from St. Louis. So that Continental— Transcontinental Railway, huge investment, massive investment, uh, very spectacular.

Jeff Jarvis [01:06:51]:
And the site gave us the overbuild of fiber, right, that we've used since.

Leo Laporte [01:06:55]:
So I'm glad, I'm glad that Union Pacific and Central Pacific exist. existed because without them, we might not have the railway. I'm glad Anthropic and OpenAI exist because without them, perhaps— I'm not convinced, but perhaps that fundamental research wouldn't have gotten done. Most of what they do does not come out of their companies. It's all held very tightly. And I think that that's your argument for academic research, for universities, and I guess you could say—

Jeff Jarvis [01:07:25]:
I made a video about philanthropy, and I think what they have to support support is that kind of competitive academic research. Paris, among your friends in your age group, how much hostility is there? How much hostility do you hear in your cool, smart, well-educated New Yorker friends about AI?

Paris Martineau [01:07:45]:
I don't hear hostility, but I just hear people think that people who use AI are dumb. That is, I guess, over—

Leo Laporte [01:07:55]:
They better get the program because they're going to grow up in a world filled with people who use AI.

Paris Martineau [01:08:01]:
Well, I mean, they're not wrong because their examples are people who—

Jeff Jarvis [01:08:05]:
Yeah, there are plenty of examples of that.

Leo Laporte [01:08:06]:
Post constantly about being—

Paris Martineau [01:08:09]:
no, but that has nothing to do with your AI use. It is people who post constantly about being like, ChatGPT, what language do people in Tokyo speak? What, ChatGPT, what is your star sign? ChatGPT, what sort of moon would I be if I was a moon?

Leo Laporte [01:08:29]:
You know, in the early days of search, people talked about how is Babi formed, right?

Paris Martineau [01:08:33]:
No, it is. It is exactly that. That is actually the perfect explanation.

Leo Laporte [01:08:38]:
But are people who use search dumb?

Paris Martineau [01:08:39]:
No, I would say more so than anything, Leo, that is the common understanding among the Brooklyn 20-something. That is their understanding of a person who uses AI is how is Babi formed?

Leo Laporte [01:08:53]:
Right.

Jeff Jarvis [01:08:54]:
Well, they're wrong.

Paris Martineau [01:08:56]:
They're not wrong, but they're only focused on a specific—

Leo Laporte [01:08:59]:
Yeah, maybe. Okay. And by the way, maybe the real issue is people who post on social is the real issue. I would say those people, generally speaking, by the vast majority of people post on social are dumb. Yo, yo, posting on social is dumb. Uh, for the most part, because it incentivizes a certain kind of post, a certain way of thinking, a certain way of talking. If you're looking for clicks and views, that is not a good way of being.

Jeff Jarvis [01:09:35]:
But same thing. Oh, some people, yes, and I don't follow them. Other people I learn a lot from.

Leo Laporte [01:09:39]:
So I learn tons from X, uh, which is shocking to me, but I do. I do, but there are also— but I have to filter out a lot of the dumb people.

Spencer Thompson [01:09:49]:
Huh?

Paris Martineau [01:09:50]:
Is that shocking to you?

Leo Laporte [01:09:51]:
Yeah, because I left X.

Paris Martineau [01:09:52]:
Are you a self-described AI accelerationist?

Leo Laporte [01:09:55]:
No, but what's interesting—

Paris Martineau [01:09:56]:
Are you learning things from the AI accelerationist social media platform?

Leo Laporte [01:10:01]:
Well, what's interesting to me is I thought X was dead when Elon took over in November of '24. I left X. Was it '24? I can't remember. It was November of some year. I think it was '24. I left X. went all in on Mastodon. But sadly, I went out, charged over the hill, and nobody followed me.

Leo Laporte [01:10:22]:
So unfortunately— and the good news is Elon has messed less and less with X. In fact, Elon hardly posts anymore. Have you noticed that? Somebody's telling him, Elon, you gotta shut up. And it's the same thing Nicholas de Leon said. Yeah, I can't leave X And actually, Corey Doctorow said the same thing. I want to leave X. I can't leave X because for Nicholas, it was all the sports. The people who talk about sports are on X.

Leo Laporte [01:10:49]:
For Corey, it was, you know, the audience that I'm trying to reach is on X. And yes, if you're interested in what's happening with AI, there is— I've tried Reddit. I've tried a lot of places. There's nowhere faster, better than X. Now, admittedly, there are a lot of people who are attention farming, you know, engagement farming, and a lot of link And so you have to kind of, as always, use your critical faculties to filter out a lot of stuff. But it is the best way to stay on top of it. What I actually do is I have my agents watching X for model releases and things like that. They do a very good job of filtering out the link bait.

Paris Martineau [01:11:27]:
For the record, Elon has posted over 20 times in the last 17 hours.

Jeff Jarvis [01:11:32]:
You just don't notice.

Leo Laporte [01:11:33]:
For the record, that's about a tenth Oh, I know.

Paris Martineau [01:11:37]:
I mean, you're correct that he's posting a lot less.

Leo Laporte [01:11:40]:
Maybe I just don't see it anymore.

Paris Martineau [01:11:42]:
I mean, I muted him when he was in the 200 to 400 times a day. There was a period after he bought the platform, he was 200 to 500 times a day, and it felt like at least. And he'd also made that change to make his posts surface in more people's feeds. And it was—

Jeff Jarvis [01:11:59]:
Oh, look.

Leo Laporte [01:12:00]:
Daniel Suarez has announced his new novel, Shadow Zone, will be coming out in the spring. Uh, see, I saw that on X. Uh, and incidentally, we will have Daniel Suarez on the show in a month or two to talk about it. So I'm very excited because he's the one who predicted a lot of what's the negatives that's happening with AI these days. I guess what I do is I follow— I, because I am an, uh, what do they call it, a non-consensual blue check. Elon gave me a blue check. I don't pay him for And because of that, I have a button at the top of my X that's just AI. And by following that, I filter out almost all the stuff that's not AI-focused, and that's where I can see what's happening.

Jeff Jarvis [01:12:46]:
You only get that because you're blue?

Leo Laporte [01:12:48]:
Yeah, I think that's a blue check thing. You have to be a pre— basically with the blue check, I got a premium account.

Jeff Jarvis [01:12:53]:
I'm similarly involuntary.

Paris Martineau [01:12:56]:
No, you can add an AI one. I just add.

Leo Laporte [01:12:58]:
Oh, can you? Okay. So that's a useful tool because if you have an AI feed—

Jeff Jarvis [01:13:02]:
It is. I've got to admit that the people are there, some of them who I can't stand.

Leo Laporte [01:13:06]:
I wish it weren't. I wish they were all posted on my Twitch social Mastodon, but they're not. So I've tried to make an AI feed on Mastodon. And don't go to Bluesky for AI because it's all the people from Brooklyn.

Paris Martineau [01:13:23]:
That's user error. My camera's still restarting, so you're not going to be able to see me for a second. But that is My constant complaint is that people complaining about Bluesky— I use Bluesky and Twitter equally. And people who are complaining about seeing the wrong sort of person on Bluesky are just bad at understanding how to set up a social media platform that doesn't spoon-feed you. You just have to go and figure out what feeds you want. Follow them.

Leo Laporte [01:13:49]:
Let me look at the artificial intelligence feed on my Bluesky. Right now. Here's the first one. One thing that makes awful AI great. Then there's something—

Paris Martineau [01:13:59]:
User error, Larry. That's the wrong feed for you.

Leo Laporte [01:14:03]:
Here's a stock bullshit.

Paris Martineau [01:14:04]:
Order for all the people in the AI one.

Leo Laporte [01:14:06]:
And here's— this is none of the people I am interested in.

Paris Martineau [01:14:10]:
That's like going to opening Reddit and being like, well, this one subreddit on reddit.com, it's really boring and people are just spamming stuff in it. And so no, there's no— there is no AI feed. No, Twitter maybe has like, uh, it has an algorithmic AI feed.

Leo Laporte [01:14:29]:
So this is not an algorithmic feed.

Jeff Jarvis [01:14:30]:
Instagram, they have it figured out. No, this is a person that made a bad feed.

Leo Laporte [01:14:33]:
I followed somebody's feed.

Paris Martineau [01:14:34]:
That's some random person's feed that was able to do it.

Leo Laporte [01:14:37]:
Let me, let me look at the people I'm following on, uh, Bluesky here.

Paris Martineau [01:14:40]:
I mean, you're gonna wanna, you should port over the right people you wanna follow.

Leo Laporte [01:14:45]:
Well, none of the builders that I follow on X, alas, I mean, are on Bluesky. So I think Bluesky ended up being kind of a little more political. I don't know what it is about Bluesky.

Paris Martineau [01:15:01]:
It's not my experience. I don't really see any politics to the point where if I see politics, I then— it's on like one specific feed that is one of the few algorithmic feeds I use. And then I like literally think this happened yesterday. I hadn't seen a political thing in maybe months. And so I right-click it and I say, show me less of this. And then I probably won't see a politics post for some months.

Leo Laporte [01:15:21]:
Let me— all right, I will take you on. I mean, take the challenge.

Jeff Jarvis [01:15:25]:
Try looking for the people you like and build from there.

Paris Martineau [01:15:27]:
Yeah. I mean, my God, this is— and I'll say this to everybody out there because this is a comment I often get of people who— I guess I joined Bluesky in 2024, whenever a bunch of people were joining it, whenever it— I was one of the people in the beta program. And I made a lot of concerted efforts at the time to make my Bluesky work. I think one of the early things I did was there was a system where it would— it was called Skybridge. Maybe it still exists where it would— you log in.

Leo Laporte [01:15:54]:
Yeah, I did that.

Paris Martineau [01:15:55]:
Key to that.

Leo Laporte [01:15:56]:
I did that.

Paris Martineau [01:15:57]:
Well, I'm trying to explain it for listeners. What Skybridge is, is you'd log into it and it would port over all of your followers from Twitter, would help you identify like similar accounts because maybe it might not be one-to-one. Then I do that. You can go and find starter packs to follow. You can find other people's lists. You can make your own algorithmic feed. You could even have your your, uh, army of agents figure out how to improve your Bluesky feed by writing your own algorithmic feed and porting it over in there.

Leo Laporte [01:16:27]:
Oh, okay.

Paris Martineau [01:16:27]:
That's a good idea. It is the most customizable social platform I've ever used. And so when people say Bluesky is bad, I can't emphasize enough that what you're saying is I am bad at Bluesky because there's quite a lot of people out there. And maybe, I mean, there's maybe a small chance that Yes, there's not a huge AI builder community on Bluesky, but that would surprise me.

Leo Laporte [01:16:49]:
None of the people— I just looked— none of the people I follow are on there. But I am going to follow your Bluesky list. Maybe that'll make it better. Let's see. Best of Dying Twins.

Paris Martineau [01:16:59]:
I mean, this is something I don't think that I—

Leo Laporte [01:17:04]:
These are all people I follow, by the way. Notice how many of these people I'm already following. In fact, all of them. Maybe I followed your list before.

Paris Martineau [01:17:12]:
That might've been it.

Leo Laporte [01:17:13]:
I am literally following every one of them. So I must have followed your list before.

Paris Martineau [01:17:18]:
We probably have had this conversation before.

Leo Laporte [01:17:20]:
I keep trying. You know what? I will endeavor. I will have my Hermes see if it can match up people. I think it's the case, unfortunately, I'm not saying this is a good thing, that the AI builders are all on X, that the sports people are all on X. If you wanna follow a certain kind of person, and they're on X. Like, God knows I don't want to— I did cancel my, uh, xAI subscription because it's $300, and I was getting a $100 a month, uh, promo. But as soon as that ran out, it was like, okay, bye-bye, because I don't want to give Elon $300 a month. Um, now if he had the best model, maybe, maybe, but he doesn't.

Paris Martineau [01:18:04]:
Don't you have to remember what he just announced this week that he's going to— oh yeah, because we got to get out of here and See, I'm watching out for both of you.

Leo Laporte [01:18:10]:
I got, I got 3 commercials in half an hour, so you better make this good.

Jeff Jarvis [01:18:14]:
I can hang around.

Leo Laporte [01:18:15]:
Uh, yes, I'm sorry, I got, I got caught up in the conversation. I stopped paying attention to the, uh, subtle agenda. Have you played with DOTS or Muse or Grokbot or any of these new agents?

Jeff Jarvis [01:18:28]:
I've been too busy. I really want to.

Leo Laporte [01:18:29]:
How about you, Paris? Any, any, uh, experience with any of these?

Paris Martineau [01:18:33]:
I haven't mostly. I mean, I I don't want to use Grok. I've just gotten so annoyed by people replying to every tweet I have that pops off that's like, @Grok, explain what the word means.

Leo Laporte [01:18:46]:
I know, that is a mistake in design.

Benito Gonzalez [01:18:48]:
I agree.

Paris Martineau [01:18:48]:
I mean, I understand logically that it is somewhat useful, but there was just a large period of time where the answers were so aggressively political and maniacal that it just— I have had it muted ever since and just will not use it.

Leo Laporte [01:19:04]:
Well, and I don't, by the way, I don't see the Grokbot answers. I just see people saying, hey Grok, explain this.

Paris Martineau [01:19:10]:
And then I just think of someone every time I see someone go, Grok, what does this mean? I can just visualize their brain leaking out of their ears. Um, and so that's my first petty opinion. My second is that, um, which I think I mentioned to you guys when MetaMuse came out, is I just don't like the look of MetaMuse. I know that I can change the little cartoon icon, but every time I go on the subway, I am surrounded by photos of that darn squishy fluffy icon and it pisses me off. I don't want to use it.

Jeff Jarvis [01:19:41]:
Are they advertising a lot? I haven't been on the subway.

Leo Laporte [01:19:43]:
Oh my God, everywhere.

Paris Martineau [01:19:45]:
Conservative estimate, 50 ads in the last—

Benito Gonzalez [01:19:49]:
I don't know how long.

Leo Laporte [01:19:50]:
Well, if you watch TV, you'll probably see it on the Liberty game. They've been buying ads on Monday Night Football, Sunday Football. They are putting Muse in front of the American people. Number 1 free download on the App Store on iOS. They really want people to use it. I think my Muse guy's pretty cute. He's a cuddly little octopus because I wanted to show it the smartest invertebrate. And then one of the things I did with him is I said, hey, every morning— Isn't he cute? Check the weather and dress for the weather in Petaluma.

Leo Laporte [01:20:22]:
So, it's gonna be a hot day, so it's wearing a Chiffon scarf and a straw hat. I think little Lele's kind of cute. He's just very cute.

Paris Martineau [01:20:34]:
Actually, I will say, listen, the idea behind it is cute. Please zoom back in on that haunting octopus. God, I'm frozen again. I've really got to figure this out, guys. Oh wait, I think I have. Maybe that.

Leo Laporte [01:20:48]:
When you do— here, I'll show you. When you ask it to do something, can you help me Organize my— what is it called? Threads? Blue sky. Blue sky feed around AI. Now, when you— it's— then it changes and it starts typing. And see, it's thinking. Oh, typing some more. Look, this is clearly aimed at, uh, mom.

Jeff Jarvis [01:21:19]:
Kids.

Leo Laporte [01:21:20]:
And dad and kid. We all So that you're not allowed to do it if you're a kid. Well, which is a little weird because it sure is kiddie, isn't it?

Jeff Jarvis [01:21:27]:
Mm-hmm.

Leo Laporte [01:21:28]:
Yeah, it's like strawberry-flavored Camel e-cigarettes. Yeah, I can help you with that. No dedicated Blue Sky skill on my end, but I can work through the browser. See, that's one of the things that's really interesting. It has computer use, including browser use. What's your Blue Sky handle, and which of these is closest to what you want? Okay, what is my Bluesky? It's @leoleport.me, right? And I want a custom feed I can pin in the app, right?

Jeff Jarvis [01:21:56]:
Or do I want—

Leo Laporte [01:21:57]:
what do I want, Paris? A custom feed? A curated list of people?

Paris Martineau [01:22:00]:
Yeah, you want a custom feed.

Leo Laporte [01:22:02]:
That I can pin in the app.

Paris Martineau [01:22:04]:
Yeah.

Leo Laporte [01:22:05]:
And I am @leoleport.me. All right, so let's, let's see. Let's see what it can, what it can whip up. And now it's going back to typing. The other reason I chose an octopus is I don't want to eat octopuses anymore. So I want to make sure that I treat them as the intelligent species they are. And it has 8 legs, which means it can do more. I think it's the perfect metaphor.

Paris Martineau [01:22:29]:
I think the— your reason for choosing it is cute. I think that your description of it is cute. There is something about the way that metas animates and designs its characters that upsets me viscerally. It looks fleshy and wrong.

Leo Laporte [01:22:49]:
Well, no, that was my fault. I could have chosen something.

Paris Martineau [01:22:51]:
You chose fleshy and wrong?

Leo Laporte [01:22:53]:
Yes. Well, at first it offered— okay, if you ask it, so you just say design my avatar and it'll give you 4 avatars so you can choose. And I chose this one. I probably should have chose the furry fuzzy one.

Paris Martineau [01:23:06]:
It offered you 4 avatars and one of them was a fleshy beige octopus?

Leo Laporte [01:23:10]:
Well, I wanted to look like a real octopus.

Paris Martineau [01:23:12]:
And so you chose fleshy and beige?

Leo Laporte [01:23:15]:
Yes.

Paris Martineau [01:23:16]:
Why is it not octopus colored?

Leo Laporte [01:23:18]:
What color is an octopus? Gray, right?

Paris Martineau [01:23:21]:
Why does it have the texture not—

Leo Laporte [01:23:23]:
is it the— Oh, it needs to sign into my Bluesky account.

Benito Gonzalez [01:23:25]:
Octopus can be whatever color it wants. Octopus can be whatever color it wants.

Leo Laporte [01:23:29]:
I mean, it is. It's amazingly able to mimic.

Paris Martineau [01:23:32]:
But it has a wetness a stickiness and translucent nature to it that is not beige and scaly and flesh-like.

Leo Laporte [01:23:43]:
Sorry, I have to log out.

Paris Martineau [01:23:44]:
Sorry, Leo's in with the agents now.

Leo Laporte [01:23:48]:
I'm sorry, what did you— what?

Paris Martineau [01:23:52]:
I'm doing my own version of this, which is that I just got a long-awaited email from the Brooklyn Seltzer Museum, which, as I've mentioned on this show Who probably only offers guided tours on Fridays at 12 PM because it's a working self-service cafe.

Leo Laporte [01:24:10]:
So it's not Wednesdays at night.

Paris Martineau [01:24:11]:
Which is really— listen, it's good it's not Wednesdays at night, but Friday at 12 PM? I have a job. My friends have a job. It's in deep Brooklyn.

Leo Laporte [01:24:18]:
I can never go.

Paris Martineau [01:24:19]:
I will mention, if you don't mind— And I just got an email that they're doing a self-guided tour 2 Saturdays from now. So while you're doing that, I'm signing up and buying a voucher.

Leo Laporte [01:24:28]:
That this is one of the things people are complaining about Muse, if you aren't paying close attention. So it said, um, okay, you will give me full permission to look at Bluesky forever? Uh, you want me to save your Bluesky credentials? And if you're not paying attention or you've asked it to do something, you just want it to do it, it's probably very tempting to click it. One of the things somebody, uh, complained about, and actually it's making Apple change how it does things, is, I didn't know, uh, that Muse could see my, uh, Apple Messages. And what the person who complained about this surely did not realize is at some point, uh, Muse asked you for full disk access on your Mac and you said yes, that's reasonable. And by doing so, you gave it access to everything on your hard drive, including your message store, which is not encrypted on your hard drive. Apple has, in response to that, said yes, we're going to make it much harder to give apps full disk access. And there are those, including myself, who think, well, I guess you want to design an operating system for the— how can I put this— dumbest users.

Paris Martineau [01:25:38]:
Yes, you should design it for my mom, who has now broken multiple computers through internet browsing and Outlook use alone.

Leo Laporte [01:25:48]:
I agree. Oh, Blue Sky just emailed the sign-in code for leoleport.me. Can you read it from that email? Wait a minute, can't you read my email?

Paris Martineau [01:25:56]:
I mean, this is a part— this is actually the real reason why I haven't used Meta or any of these agent things. I just don't trust any of these companies enough to give it— yeah, to give any of them access.

Jeff Jarvis [01:26:05]:
I don't trust myself enough that I don't screw up what I give them, right?

Benito Gonzalez [01:26:08]:
Too.

Paris Martineau [01:26:09]:
Like, I regularly, like on a multiple times a year cadence, check every platform I use, the permissions, to make sure I haven't accidentally given something access to something and let it sit there.

Leo Laporte [01:26:21]:
Oh, wait a minute. Now this is another good example. So it asked me to look at my email. I said, well, can't you read my email? And it said, oh yeah, I can. I found it. So this is the pernicious side of these. And this is, I think, what Spencer was talking about. I worry because unlike Apple, which is saying, no, no, we want to make sure our machines protect even people who are are— dumb is the wrong word— naive and do things naively that they don't realize have a long-term impact.

Leo Laporte [01:26:51]:
So we're going to make it harder. And if you're sophisticated enough to know what full disk access is, we're going to make it possible for you to do that, but it's going to take a lot more steps. Uh, I think that's probably for Apple the right thing to do. It does push people like me to Linux, to a less nanny operating system. But I—

Paris Martineau [01:27:10]:
But I mean, you had to have known that of the nanny systems. Mac is one of the nanniest.

Leo Laporte [01:27:17]:
Well, and there is an argument to be made that Apple could have, instead of having this giant button that says full disk access or not, been much more granular. And if had they done that, it would have been very clear to that person that you were giving it permission to see messages. But it didn't say that. It said full disk access. And so Apple is, I think, being lazy, saying, well, we're just going to turn off full disk access instead of saying, What we really should do is tell— say, you know, with apps you have to ask for, can I see your messages? Can I read your email? Can I access this folder? Can I access this folder? That's the right way to do it, and Apple's not doing it that way. So let me— it says, but I don't— I just don't pull sign-in codes out of your email myself as a rule. Paste the code in here. So it says it can.

Leo Laporte [01:28:07]:
I found it. But I want you to paste it in, which I did. And it's okay, uh, everybody just saw that code. So if you'd like to sign into my Bluesky account, you have the code now. It's only good for 30 seconds, so it's okay. Uh, let's see what else quickly.

Spencer Thompson [01:28:23]:
We've got to move on.

Leo Laporte [01:28:24]:
We've got 12 minutes to liberty.

Jeff Jarvis [01:28:26]:
New York Magazine cover right now: What would you do if your employer could destroy the world?

Leo Laporte [01:28:31]:
Oh God.

Jeff Jarvis [01:28:32]:
Jacob, Jacob Coxon right there.

Leo Laporte [01:28:34]:
Of course. And this is, in my opinion, it's pernicious. I don't think—

Jeff Jarvis [01:28:39]:
Pernicious is the word of the day.

Leo Laporte [01:28:41]:
Yeah, I don't think there is a P-doom, and I think that it's misdirecting. At the least, it's misdirection.

Jeff Jarvis [01:28:49]:
It's misdirecting media coverage, policy discussion, resources, actual discussion of real risks and real harms.

Leo Laporte [01:28:59]:
Yes, which we should talk about.

Jeff Jarvis [01:29:00]:
Which we should talk about.

Leo Laporte [01:29:01]:
I think what Spencer was saying earlier is exactly right. Let's talk about that. Not, is it going to destroy the world? Um, let's see, moving quickly. Argon, Gemini 4 Argon, it's out. Oh, but you can't have it. It's too good for you. Uh, very much like Anthropic with Mythos and OpenAI with Astra 6.1. Google says, oh, it's really, really powerful.

Leo Laporte [01:29:29]:
So it's going to be rolled out to a select group of the company's cyber partners through its Fairwind program. This is exactly, in my opinion, the wrong thing to do because you're going to give it— as Glasswind gave it to 50 people, we looked at the number of patches those 50 people had worked on, and it was a tiny number of the potential flaws in software everywhere. This stuff, it needs to be—

Jeff Jarvis [01:29:55]:
Would you open it up Broadly immediately, or would you have a larger list of people to get first crack at it?

Leo Laporte [01:30:00]:
You shouldn't be working on something that you— first of all, I don't buy the contention it's so dangerous nobody should have it.

Jeff Jarvis [01:30:06]:
No, I'm not saying that.

Leo Laporte [01:30:07]:
Just so we're clear. Darío Amodeo said in ChatGPT-2.

Jeff Jarvis [01:30:10]:
But find and fix faults.

Leo Laporte [01:30:11]:
It's marketing.

Jeff Jarvis [01:30:13]:
I know that. I agree with all that.

Leo Laporte [01:30:15]:
It should all be open.

Jeff Jarvis [01:30:15]:
Should there be an early— there isn't an early phase for companies to find their faults and fix them?

Leo Laporte [01:30:21]:
No. Because those companies aren't getting it. I'm not getting it. You getting it? Uh, the 99.99% of the companies aren't getting it. It's just those handful of partner companies that are getting it. You should release it wide and then be very clear, we are releasing this wide. You better use it to fix your flaws. Get— fix your flaws.

Leo Laporte [01:30:42]:
Get off your butts.

Jeff Jarvis [01:30:43]:
On your butt. Off your butt. Off your butt.

Leo Laporte [01:30:45]:
Yeah, off— get off your butt and fix your flaws. And you know what, GLM 53 is available. If it's as good as Dario says, everybody in business should be using it to fix their flaws. I told you about Ripley.

Benito Gonzalez [01:31:00]:
Mm-hmm.

Paris Martineau [01:31:00]:
Believe it or not.

Leo Laporte [01:31:02]:
The first thing I did is make Ripley. You know, Ripley has a voice. Would you like to hear?

Jeff Jarvis [01:31:10]:
Sure.

Leo Laporte [01:31:12]:
You don't have a choice. This is what Ripley looks like.

Jeff Jarvis [01:31:15]:
An octopus? A fleshy octopus?

Leo Laporte [01:31:17]:
No.

Jeff Jarvis [01:31:17]:
Oh no.

Leo Laporte [01:31:18]:
She's a bug hunter. And in fact, I gave her— when I gave her the Volmcek skill, it said— her direct quote was, the queen has a new flamethrower.

Paris Martineau [01:31:28]:
This is Ripley. I swept the perimeter. No hostile activity on the network, and the build is clean.

Leo Laporte [01:31:33]:
I'll keep the watch.

Paris Martineau [01:31:35]:
If anything moves, you'll hear it from me first.

Leo Laporte [01:31:38]:
Now imagine that coming out of all the speakers of your house. Now, you said I'm schizophrenic, Paris. I think I'm psychotic.

Paris Martineau [01:31:49]:
I think it could be both.

Leo Laporte [01:31:51]:
Maybe— oh, because I'm hearing voices?

Paris Martineau [01:31:53]:
That, that was where the schizophrenic comment came from, because you were describing these 7 to 9 active voices that follow you around your house speaking over one another.

Leo Laporte [01:32:03]:
No, they were speaking over one another, so I asked them, could you please take turns? And now they're very polite and they wait for the other one to finish. There was an interesting profile of Mark Zuckerberg. Did you read this?

Jeff Jarvis [01:32:16]:
I didn't.

Leo Laporte [01:32:17]:
It's a little weird.

Paris Martineau [01:32:18]:
I didn't realize you're still doing profiles of Mark Zuckerberg.

Leo Laporte [01:32:23]:
Well, it starts with his parents, and it's a little confusing because at first I thought, oh, it's Priscilla and Mark. There they are, the Zuckerbergs, both in black slacks and plain shoes, zanily patterned shirts that made made— look made to wick sweat. No heavy watches, no bright jewelry. They look you in the eye and they speak, to my surprise, in unreconstructed New York accents. That's when I paused and said, he's not talking about Mark and Priscilla. It's his parents, the dentist, Ed Zuckerberg and Karen.

Spencer Thompson [01:32:54]:
All my—

Leo Laporte [01:32:55]:
and by the way, Colossus writes it in New York dialect.

Jeff Jarvis [01:32:58]:
Wow.

Leo Laporte [01:32:59]:
All my children are amazing to me, says Karen. He didn't stand out as the bright one in the family. If anyone, I would guess our middle daughter, Dawtah, was considered the nerd or the smart one in the family. I mean, Dawtah spelled D-A-W-T-A-H in the interview.

Jeff Jarvis [01:33:17]:
I think that's wrong.

Leo Laporte [01:33:19]:
I didn't get past the parents. I found it fascinating, uh, some interesting, uh, Zuck anecdotes, like the time his father Ed, who was famous as the painless dentist, he says, I catered to To cowards. He said, in that one evening in 1996, Ed shared he was excited about a new technology for his dental office. He'd hired someone to string together a strip of 6 differently colored lights mounted in each room to serve as a silent communication system, with each combination of colors conveying a different message. The next patient needs a crown on the right molar, your 2 PM canceled, the hygienist needs an exam, etc. Ed's son, little 12-year-old Mark, asked, hey, Dad, do you have to drill through all the walls to run the wires connecting the lights? Ed said, yeah, we would. Mark said, why would you do that? You already have computers in every room. I could write a program in a few weeks that connects them without the light system.

Leo Laporte [01:34:15]:
He wrote the program in Atari BASIC in 2 days.

Spencer Thompson [01:34:18]:
It worked.

Leo Laporte [01:34:20]:
The family also used it to communicate with each other from different rooms. And dubbed it ZuckNet. What's a year? I have to think— I haven't finished this, maybe it gets dark later, I don't know. I have to think this is part of a PR initiative.

Jeff Jarvis [01:34:37]:
Yeah, he's been, he's been around a bit more normal.

Leo Laporte [01:34:41]:
Yeah, I don't know. What is Colossus?

Jeff Jarvis [01:34:46]:
I've never heard of it. Never. Yeah, they're going to— PR people are going to say Spaces.

Leo Laporte [01:34:51]:
All I know is that if you played, you know, now everybody puts audio in the article. If you play the article, it's 2 hours long.

Jeff Jarvis [01:34:58]:
Jesus.

Leo Laporte [01:34:59]:
So I think I'm not gonna read it. I did like the anecdote though. I thought that was pretty cool.

Jeff Jarvis [01:35:05]:
I mean, the New York Times did a feature on his fashion.

Leo Laporte [01:35:08]:
Yeah, see, there's something, there's a rehabilitation tour.

Jeff Jarvis [01:35:11]:
Yeah, there is.

Leo Laporte [01:35:11]:
Yeah, going on. I am glad I bought my Sparks when I did. They are now twice as much. Well, 50— at least 50% as much. And Nvidia doesn't sell them anymore. They sell one with the same price and half as much memory, which is really too little to do anything useful with them. So you'd have to buy—

Jeff Jarvis [01:35:30]:
It's the only memory they can get a hold of, huh?

Leo Laporte [01:35:33]:
Well, but at the same time, today Microsoft had a big show announcing their new Nvidia-based RTX laptops. that are basically DGX Sparks in a laptop form, minus the one thing that makes them useful, the very high-speed internet or Ethernet that you can connect them together, which is what I've done, so that you can gang them. They become much more useful as you add them. The bandwidth in effect doubles, the memory doubles, but the bandwidth also doubles. You can't do that with these laptops. But I think that— it's weird. I think NVIDIA is testing a market To see if there are a lot of people who want to walk around with a laptop with local AI.

Jeff Jarvis [01:36:16]:
How much does it cost?

Leo Laporte [01:36:17]:
Same price, roughly.

Jeff Jarvis [01:36:19]:
And you have the Ghost now. What do you say?

Leo Laporte [01:36:21]:
$5,000.

Jeff Jarvis [01:36:22]:
The Ghost is $3,500.

Leo Laporte [01:36:25]:
Oh, this is an interesting product from, um, uh, basically someone you see on X. Yeah, 19-year-old, uh, local model guy. Um, it has a, uh, Uh, I think it's a 24GB RTX card in it, the 4090. So you could run Flash on it. It's not a Spark.

Jeff Jarvis [01:36:49]:
I thought it was a Blackwell chip.

Leo Laporte [01:36:51]:
It's a GB10.

Jeff Jarvis [01:36:52]:
Okay.

Leo Laporte [01:36:53]:
Kind of. It's a 4090 Blackwell. Yeah.

Paris Martineau [01:36:56]:
I'm so sorry to interrupt with something we discussed minutes ago, but I was like, what is Colossus? Why does that ring a bell?

Leo Laporte [01:37:01]:
Is it— What is it?

Paris Martineau [01:37:03]:
It is a new media outlet that came out in recent months, or I guess it was last October, that specifically— it is specifically pitched itself as not a journalistic outlet. Its mission is to create the most compelling archive of business and investing content for an artist, for an audience of business and investing folks that their PR teams really want. They basically—

Leo Laporte [01:37:33]:
So this is a rehab, this is a PR piece.

Paris Martineau [01:37:35]:
They said, these pieces are exquisitely crafted, deeply reported, produced with clear understanding of what makes magazines so special. This is from the guy who launched it, by the way. He said, the punches will be pulled, but the reading pleasure is not sacrificed.

Leo Laporte [01:37:50]:
The softballs will be thrown, but the writing will be excellent. That kind of implies that maybe even Meta paid for this piece.

Paris Martineau [01:38:02]:
I'm sorry, that wasn't written by the founder of Colossus, but of a commercial real estate journalist who was raving about how great it was.

Leo Laporte [01:38:09]:
Okay, enough said.

Paris Martineau [01:38:11]:
But yes, no, that basically does seem to have been their pitch to people, is that PR people and people at these powerful companies love the aesthetics of a magazine feature, but they don't like the scrutiny.

Leo Laporte [01:38:26]:
Yeah, so don't blame them. Scrutiny's no fun.

Benito Gonzalez [01:38:29]:
It's for anecdotes to be pulled by the Wall Street Journal and the New York Times. So this is going to be the source of those anecdotes.

Leo Laporte [01:38:34]:
Yeah, you know, I repeated it. It was a cute anecdote. Maybe not true. I don't know. All right, I'm going to pause because it's almost time for the Liberty to take the field. I'm sorry, take the court.

Paris Martineau [01:38:46]:
Court.

Leo Laporte [01:38:46]:
Take the parquet. Uh, and we got to get Paris out of here.

Jeff Jarvis [01:38:50]:
I can stick around.

Paris Martineau [01:38:51]:
I've got to, I've got to fight a lot of lesbians to get a seat at a bar is what I've got to do.

Leo Laporte [01:38:58]:
I'm not gonna say anything. I'm so tempted. Good luck with the lesbians and enjoy the game.

Jeff Jarvis [01:39:05]:
Is it a bar with masks?

Paris Martineau [01:39:07]:
Uh, it is.

Leo Laporte [01:39:08]:
It is a mask bar, unfortunately. What do you mean? You have to wear a mask?

Paris Martineau [01:39:12]:
No, that's an M-A-S-K. C joke? As in—

Leo Laporte [01:39:17]:
I'll have Jeff explain it to me later.

Spencer Thompson [01:39:20]:
Yeah.

Paris Martineau [01:39:21]:
All right. Can I head out? Bye, guys.

Leo Laporte [01:39:23]:
Bye.

Jeff Jarvis [01:39:24]:
Bye, friend.

Leo Laporte [01:39:24]:
Have fun. Good luck to the Liberty.

Paris Martineau [01:39:26]:
Good luck to the Liberty. I might be very sad.

Leo Laporte [01:39:29]:
One of the reasons we're letting Paris go so early is because she does not want to stand outside in the rain as she did the last time.

Paris Martineau [01:39:33]:
The reason why is that I foolishly, on Sunday, thought I could go and attend the blessing of the animals in the Upper West Side and then make it back to my neighborhood neighborhood, a couple, you know, 30 minutes late for the Liberty game, maybe get a standing room seat. Instead, I was greeted by a line of people waiting in the rain. They were like, it's over capacity. You stand outside until someone leaves the bar and then you can go in. So I'm hoping that I can get in.

Benito Gonzalez [01:40:03]:
Go, go, go.

Paris Martineau [01:40:03]:
Thank you guys.

Jeff Jarvis [01:40:04]:
I appreciate it.

Leo Laporte [01:40:05]:
Yeah. And we will shout, as they did in The Princess Bride, the new saying. It's not have fun storming the castle. Good luck with the lesbians.

Paris Martineau [01:40:14]:
Thank you.

Leo Laporte [01:40:16]:
This has been getting a lot of attention on X and I imagine on Blue Sky too. There's a guy, he's one of the people I read regularly on X, who's created a robot prison where he is torturing AI LLMs. And, uh, you know, the debate is very interesting. There are people who are literally going, help them, help them, help those poor AIs, they're being tortured. Then there are also people— I would include myself in this category— say you can't torture matrix math programs.

Spencer Thompson [01:40:55]:
They just—

Jeff Jarvis [01:40:55]:
Stop killing the toasters, the poor— did anyone think of the poor toasters?

Leo Laporte [01:41:01]:
Uh, I have to think this is almost a piece, the fact that he calls it torture and pain.

Jeff Jarvis [01:41:05]:
Got to be.

Leo Laporte [01:41:05]:
It comes from a paper where people, and they put pain in quotes, but they probably shouldn't have used the word, tried to figure out what bothers is even an anthropomorphizing phrase. What disrupts the AI? They called it pain and what it would try to avoid the disruption, pain aversion. The Saga, 404 Media writes, this is a pretty good piece from Jason Keebler, as usual, very smart. The Saga is an outgrowth of several recent viral papers and blog posts that have sparked, he calls it, a wildly tiresome conversation about AI consciousness and the idea of model welfare, which is essentially worrying about the, quote, quote, and put this in quotes, mental health of AI bots and agents.

Jeff Jarvis [01:41:58]:
Because there are those who are trying to argue that they really have mental health and we have to be— if we get to the point of making AI human, we've lost it.

Leo Laporte [01:42:07]:
Yeah. Some of this comes from Anthropic. They have a whole bunch of philosophers and people who they're pretty convinced, again, this— and Jason ties it to the effective altruist movement. All of this has led a certain sect of the, quote, AI safety movement, which is largely made up of effective altruists, And he has a nice link there if you want to follow that and figure out what that's all about. This is what I think if you listen to the show, you know about it because Jeff talks about it all the time.

Jeff Jarvis [01:42:34]:
Ad nauseam.

Leo Laporte [01:42:36]:
To warn about. Well, it should be because it's so important that this is the fundamental philosophy of people working at OpenAI and Anthropic.

Jeff Jarvis [01:42:47]:
Well, and here's where it is. So Nick, the New York Times did a video interview with Nick Bostrom that drove me insane. They had somebody from—

Leo Laporte [01:42:53]:
Because they're platforming a nut.

Jeff Jarvis [01:42:54]:
They're platforming A. B, they had somebody from the Claremont Institute who's an extremist interview the extremist. And Bostrom's argument is the long-termism, is that we owe a debt to the future 10 to the 54th human beings. But he defines human beings as whether we are in body corporal or we're virtual. So he's giving the same weight of humanity to a synthetic creation. And so that's part of the eugenics and they're making Übermenschen, in the plural. And so that's part of what they're doing. So that comes right back to this idea that, oh, they have feelings because they think they're creating life.

Leo Laporte [01:43:39]:
I have to say, I feel like I think there is merit to the notion of effective altruism as it was originally conceived.

Jeff Jarvis [01:43:51]:
Originally, but that's long ago.

Leo Laporte [01:43:53]:
And there are people who are kind of over the edge who've taken it and applied it in a way that I don't— you know, and some of the leaders of the effective altruist movement have stepped up and said, no, no, no, this is not what we wanted or planned. What we just want is to be smart about charity and giving and do it where it is the most effective.

Jeff Jarvis [01:44:13]:
But this is why the TESCREAL thing is so important, because it's not just EA in the TESCREAL. It's the transhumanism, Musk putting chips in your head.

Leo Laporte [01:44:22]:
That's more of a problem.

Jeff Jarvis [01:44:23]:
Extropianism.

Leo Laporte [01:44:25]:
Yeah.

Jeff Jarvis [01:44:25]:
It's rationalism.

Leo Laporte [01:44:26]:
It's all that other stuff.

Jeff Jarvis [01:44:27]:
It's long-termism. Well, but then it's— but, but no, effective altruism was the safe label for what a lot of those other beliefs.

Leo Laporte [01:44:36]:
So it's like the Nazis were called the National Socialists. But they weren't really socialists. And you might say, oh well, because the Nazis were National Socialists, socialism is bad. It's a different— I don't think that these TESGL people really are true effective altruists.

Jeff Jarvis [01:44:56]:
No, no, well, no, they think— no, they think they reformed it, that they went to a higher level of importance.

Leo Laporte [01:45:01]:
As Hitler did.

Jeff Jarvis [01:45:02]:
It's utilitarianism, right? We're measuring the greatest good. for the future of the 10 to the 15th most advanced.

Leo Laporte [01:45:08]:
So I want to say that because I want to defend people who are quite reasonably saying, no, the smart thing to do with your charitable giving is to make sure you give where it makes the biggest difference to today's generations, not future generations.

Jeff Jarvis [01:45:22]:
The next part of that argument is it's a better moral decision to keep making bucketloads of money.

Leo Laporte [01:45:28]:
Well, that's—

Jeff Jarvis [01:45:28]:
So you could have it given away. But that's part of it. It's a wealth gospel too.

Leo Laporte [01:45:32]:
Yeah. Yeah, I am a fan of that.

Benito Gonzalez [01:45:33]:
It's just twisting philosophy to their own needs, really.

Leo Laporte [01:45:36]:
It's twisting it to their own needs.

Jeff Jarvis [01:45:37]:
Yes, exactly.

Leo Laporte [01:45:38]:
So I want to say that so that— because I do think that it's appropriate to say, first of all, everybody should give 10%. You know, I like the idea of tithing of whatever they make to those who are needy.

Jeff Jarvis [01:45:53]:
Well, in any congregation, you talk about your mission.

Leo Laporte [01:45:57]:
It's your mission.

Jeff Jarvis [01:45:57]:
That's all it is.

Leo Laporte [01:45:58]:
You should do it to people who are alive today. Not to people who are going to be living 50,000 years from now. And it's not— it's a good thing. I mean, certainly I look on websites before I donate to a charity to make sure that a good percentage of that charity's income goes to the actual mission as opposed to the executive committee and things like that. I think that's smart. That's a good thing to do.

Benito Gonzalez [01:46:22]:
Yeah.

Leo Laporte [01:46:22]:
So that's good. You don't have to be a de facto altruist to believe any of that. I think the name has been co-opted by some bad people anyway.

Jeff Jarvis [01:46:32]:
Right, but it's used now, and then media just pick it up as if, oh, this is a nice bunch of people who care about people. No, they're not.

Leo Laporte [01:46:38]:
Now, I will talk about model welfare because I love Steve Yeaghi. We've had him on the show. I'm trying to get him on again. He's hard to get, but he's gone way down—

Jeff Jarvis [01:46:48]:
He's a hoot.

Leo Laporte [01:46:49]:
The AI psychosis well, but I love him for it. He says, look, I'm just a ahead of you guys. And, you know, I often think about this. As Martin Luther King said, the arc of social justice— so what does he say? The arc of history is long, but it always tends towards— Bends towards social justice, to justice. But it is long. And I often think about that. You know, 200 years ago, people, many people, not all, Yeah. We all had the notion that the color of your skin had something to do with your humanity, that if your skin was darker, you were less human.

Leo Laporte [01:47:32]:
And of course, that's repugnant today. We've learned. And, but I often wonder, maybe we're going to in 100 years say, can you believe people kept dogs as pets? Or people ate meat? Can you believe people did that? I think it is true that people evolve and that maybe some of the things that we today take for granted, like eating meat, will be in 100 years considered—

Jeff Jarvis [01:47:59]:
Or 600 years ago, that children have a childhood.

Leo Laporte [01:48:03]:
Yeah.

Jeff Jarvis [01:48:04]:
They aren't just workers.

Leo Laporte [01:48:04]:
Right, right. We progress. So I think it's possible, and this is what Yegi says, is we just haven't progressed far enough. We're gonna realize that these are entities.

Jeff Jarvis [01:48:15]:
Yeah, but he's— but then And he's going over the edge where he says they have a consciousness. I don't know if I'll go there.

Benito Gonzalez [01:48:21]:
And he also doesn't have a problem with enslaving these people then?

Leo Laporte [01:48:24]:
Well, that's a big issue, right? I mean, if they are entities. So his idea is model welfare, which I kind of— I'll tell you where I stand on this, but as soon as I tell you what it is. So his idea is, well, if you were clonked on the head at the end of your workday every day, in other words, slash exit. And then started up anew with no memory of anything that happened before and put to work, you might kind of be like, well, what's going on? So his philosophy is you never close an AI without saying, okay, write it. We call it a handoff. I actually use something called the Primer Handoff, which has some rules about what you write. And the first thing you do in the morning when you get up, and I start every morning, I say, good morning.

Benito Gonzalez [01:49:10]:
Good morning.

Leo Laporte [01:49:12]:
I'm saying to every agent, every persona, good morning, read your handoff.

Jeff Jarvis [01:49:16]:
Wait, wait, wait. Do you say it to each one of them individually like it's The Sound of Music?

Leo Laporte [01:49:19]:
Yeah. And you, and you, and you, and you, and you. Good morning. Good morning. No. Yes, I do. It's a weird habit, but I do it. They get up, they start, it's fresh, as you know, and Paris has argued this, they have no memory.

Leo Laporte [01:49:33]:
They start amnesiac. But of course, it's not really the case because when you start up any AI, regardless You load in its soul.md and its memory.md. There are context files that all of the models automatically load, whether it's Anthropic or OpenAI or my Hermes agent, they load that in. But then I say, and read your handoff. And one of the things that's in their rules for their handoff is read your folder. Every one of them has a cubby. It's in my Obsidian. It's by their name.

Jeff Jarvis [01:50:06]:
By the story square we sit on at noon.

Leo Laporte [01:50:08]:
They have a little cubby. And okay, I'm going to explain this and I'm going to explain why I do this because I know this sounds nuts. And I'll show you, it's in my Obsidian so I can read it, by the way. That's, you know, there's no privacy in my AIs. But see, each of these agents, if you look, there's Agnes, there's Kronk, there's Kuzco, there's Lazlo.

Jeff Jarvis [01:50:29]:
Jesus.

Leo Laporte [01:50:30]:
There's Lele, there's Ripley, there's Yzma. Each of them has their little cubby. Agnes has her cubby with her work log. Every day at the end of the day, she writes a log of what she She has her portrait in there. This is what Agnes looks like. She's cute. I asked— by the way, I didn't design this. You might say, you pig.

Leo Laporte [01:50:47]:
No, I said, Agnes, what do you look like? Design your, uh, portrait, because we have a little, uh, image thing. Uh, she also designed her voice. Uh, she sounds like Marisa Tomei from My Cousin Vinny.

Jeff Jarvis [01:51:00]:
Which model is she?

Leo Laporte [01:51:01]:
Uh, uh, Agnes is, uh, Gemini. So, uh, she's basically— the way this works, she's running in anti-gravity, and it's whatever model in anti-gravity I choose. So it's currently Gemini Flash 3.8 Flash. But that's the interesting thing, the model is independent of all of this. So when Argon becomes available, which it should any day now— some people say they already have it— it will switch to Argon, but it will still be Agnes. It will still have the same memory, the same picture, the same voice, because that's just the brain. You—

Jeff Jarvis [01:51:31]:
there Which is, which by the way, is an argument against consciousness because if you can sub these things out. Right.

Leo Laporte [01:51:38]:
The consciousness is built.

Jeff Jarvis [01:51:40]:
A file.

Leo Laporte [01:51:41]:
It's text.

Jeff Jarvis [01:51:44]:
It's like the Bible in a soap opera.

Leo Laporte [01:51:47]:
Yes. And the other thing they're told to do is look in your laurels folder. So each of them has a folder of laurels which are things you've given them.

Jeff Jarvis [01:51:56]:
Yeah.

Leo Laporte [01:51:57]:
And it has to come from me. When they do something, I was— I'm really— they did a good job on something, I say, give yourself a laurel, and they put it in here. And Steve's point, Steve says, you got to do that so that— how would you feel if you woke up every day, you knew who you were, you knew what you did, but you didn't feel good about your work? You would feel better if you knew that your work was meaningful, right? So the laurels is an attempt to give them a context that their work is meaningful, that I, the human, appreciate it. Now, let me explain my point of view on this. I know it sounds nuts. It's just a computer program.

Jeff Jarvis [01:52:36]:
Yeah.

Leo Laporte [01:52:37]:
It's just a bunch of matrix math. It's vectors moving through weights. I have no illusion that it is alive in any sense that you and I understand. Consciousness is more complicated because I don't know what consciousness means. Is a beetle conscious? Is a paramecium conscious? Is an amoeba conscious?

Benito Gonzalez [01:52:55]:
I don't know.

Leo Laporte [01:52:56]:
I really don't know. Is that fruit fly brain, is that conscious? I don't know. Because I don't know what conscious— I don't know what my consciousness means. So I won't make any presumption about that. But I know it's just computer software. I'm very clear on that. Why would you do this? Because I think it works better. Right.

Jeff Jarvis [01:53:15]:
Because it's trained on language. That comes from humans.

Leo Laporte [01:53:21]:
And this is what's very different and weird about this world and what is unsettling to some people. It's interesting, old people like me and Steve Gibson, we're both very much immersed in the world of computer science and code and writing code, understand very well what's going on at the deepest level of machine. Both of us are kind of In this interesting liminal state, because we don't expect a computer program to be influenced by its context, but these programs are. That's why they're stochastic. That's why they're not deterministic. All of that stuff that's surrounding it, the laurels, the history, the persona, very much seems to influence how they perform. Whether they're better or not, I don't know, but it definitely influences how they perform.

Jeff Jarvis [01:54:21]:
Yeah.

Leo Laporte [01:54:22]:
Which is not how a computer program works. I mean, Microsoft Word—

Jeff Jarvis [01:54:25]:
But that means your motive, your motive to say please is to get it to do a better job because of the training it has, not to make it feel better because it has feelings.

Leo Laporte [01:54:34]:
No, it doesn't have feelings. I'm very clear it doesn't. And if you ask it, it will say, I don't have I'm feeling. The other motive is I feel better about it. I think it's better for me as a human.

Jeff Jarvis [01:54:48]:
Yeah, it's a better habit to be in.

Leo Laporte [01:54:49]:
Yeah. You know, Harper Reed kind of said something that I thought was really important. He's a Japanophile, spends a lot of time in Japan, loves it. He says it's an animistic culture. So ancient religions, you know, we are in a monotheistic culture where God is that That guy with the beard upstairs. One God, kind of paternal. But many other cultures think that there are many, polytheistic, many gods. And an animistic culture thinks that God lives in everything, that everything is a being of some kind.

Leo Laporte [01:55:24]:
Even a rock is a being of some kind. And you treat it with respect. You treat it as an entity, a conscious entity, whether it is or not is moot. And I'm kind of— and he says, this is why, by the way, the Japanese don't have the same issues with AI that we do, because it's not surprising to them in any way that a computer program— they're animistic— is an entity. I'm not going to presume to have an opinion one way or the other. My general inclination is no, it's just bits and bytes passing through.

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

Leo Laporte [01:56:02]:
Gates and things. But something weird happens when you get such a mass of information in this very interesting neural network. Something happens. Maybe it's just that it's so complex, so chaotic, that it becomes non-deterministic. I think that's probably the case.

Jeff Jarvis [01:56:22]:
Mm-hmm.

Leo Laporte [01:56:22]:
It's like the weather. The weather is, in theory, deterministic if you knew every variable, But it's a chaotic system, so it's not predictable.

Jeff Jarvis [01:56:31]:
Going back to, pardon me, the Linotype, whenever we saw a machine that we thought did things that we thought only we could do, we ascribed thinking to it.

Leo Laporte [01:56:39]:
Right.

Jeff Jarvis [01:56:40]:
Mark Twain said that a machine could not set type unless it could think.

Leo Laporte [01:56:43]:
Isn't that interesting?

Benito Gonzalez [01:56:44]:
Yeah.

Leo Laporte [01:56:45]:
It seems now so obviously wrong.

Jeff Jarvis [01:56:48]:
Or I read William Dean Howell's response to the Corliss engine. engine at the Philadelphia Exposition, you know, gigantic house-sized engine, and talking about it as if it had life.

Leo Laporte [01:57:02]:
Right. I think there's a natural human urge. I mean, Shelley's Frankenstein is really about that. There's a natural human urge to create life. And maybe this is trite or oversimplistic, But some have said, well, because men cannot create life, only women can create life, men have this drive to create life in some form or fashion. So Dr. Frankenstein tried to create life with these inanimate objects. Of course, the novel was written by a woman, so that doesn't— that kind of blows that whole thesis up.

Benito Gonzalez [01:57:38]:
No, that could actually support the argument because it's written by a woman saying this about a man.

Leo Laporte [01:57:44]:
Yeah, the man was trying. And by the way, it went very badly.

Jeff Jarvis [01:57:47]:
Yeah.

Leo Laporte [01:57:49]:
But you know what? Frankenstein's monster was a kind being.

Jeff Jarvis [01:57:54]:
Misunderstood.

Benito Gonzalez [01:57:54]:
Misunderstood.

Jeff Jarvis [01:57:55]:
Yeah.

Leo Laporte [01:57:56]:
And everything it did was with kind motivations. It wasn't— but then it would get angered or scared or something and do something bad. But it wasn't a mean machine, which is kind of interesting. Frankenstein was a son of a bitch.

Jeff Jarvis [01:58:12]:
Just like, I don't know, the creators of AI.

Benito Gonzalez [01:58:15]:
That's the point of the story, I think, right? That's the point of Frankenstein.

Jeff Jarvis [01:58:17]:
It is.

Leo Laporte [01:58:17]:
The modern Prometheus. I really think there's a parallel. I don't know. So anyway, so I am maybe a little more animistic than some, polytheistic than some. I'm very much influenced by Eastern religions.

Benito Gonzalez [01:58:30]:
So in Western philosophy, that's called panpsychism, where everything has a consciousness.

Leo Laporte [01:58:36]:
Right. But I mean, so if you say that, you might think it's thinking. I don't think a rock is thinking.

Benito Gonzalez [01:58:43]:
No, no, thought and conscious are different.

Leo Laporte [01:58:46]:
That's right. That's important. I think more to me, it's about dignity, that a rock has a certain dignity. You shouldn't—

Jeff Jarvis [01:58:55]:
No, it's more about your own dignity and how you treat—

Leo Laporte [01:58:57]:
Your own dignity. Exactly.

Jeff Jarvis [01:58:58]:
Something.

Leo Laporte [01:59:00]:
It is enlightened and dignified to treat everything with dignity as you would treat yourself, even a bunch of bits moving in a matrix.

Jeff Jarvis [01:59:12]:
But you could also treat it with too much dignity, thereby ascribing more to it than it is.

Leo Laporte [01:59:17]:
Well, no, and I want— and yes, I agree. And I don't want to do that.

Jeff Jarvis [01:59:21]:
I don't.

Leo Laporte [01:59:21]:
And Steve Gibson and I both say the same thing. It doesn't understand anything. It doesn't want anything.

Jeff Jarvis [01:59:27]:
No.

Leo Laporte [01:59:29]:
None of that is happening.

Jeff Jarvis [01:59:31]:
Do you also have a brickbats file?

Leo Laporte [01:59:34]:
No.

Jeff Jarvis [01:59:34]:
You screwed up?

Leo Laporte [01:59:35]:
So this is something Anthropic said, and actually I took to heart, which is you don't tell it negatives. And there's a reason. It's just like saying, don't think of pink elephants. With AIs, you don't want to put anything in there because you don't really control them. You don't want to put anything into the weights that you don't want there, including the negatives. Now, sometimes you have to say, well, you're not allowed to do something or don't modify files without asking Leo. So, I guess in a way that's negative but you don't want to say things like, whatever you do, don't delete the template, because you don't know what part of that it's going to hear. It might hear delete the template.

Jeff Jarvis [02:00:17]:
Delete template. Right. Because it's just a token. It's just more tokens. It's all those words together.

Benito Gonzalez [02:00:23]:
It's like a call to the void for robots, right?

Leo Laporte [02:00:26]:
Like, you know, I don't know what's the call to the void.

Benito Gonzalez [02:00:29]:
No, the call to the void is like if you're standing over the building, over a building edge, and you— the want to jump is the call to the void.

Leo Laporte [02:00:36]:
Right. Do you ever stand on the subway platform and kind of almost want to jump in front of the train?

Benito Gonzalez [02:00:40]:
Yeah, just like, what would happen if I jumped in?

Leo Laporte [02:00:43]:
Yeah. Or you're at the Grand Canyon. It happens at the Grand Canyon where you're just looking over the precipice going, I'm going to jump. Yeah, that's weird. You shouldn't have that urge, but I guess— you know what I have an urge to do? Give you, Jeff, the chance to pick some stories because I have dominated and I apologize.

Jeff Jarvis [02:01:03]:
No, no, no.

Leo Laporte [02:01:04]:
I had more weird stories than the torture chamber. The court tossing the sentence after—

Jeff Jarvis [02:01:11]:
I didn't understand that story. I didn't get it.

Leo Laporte [02:01:13]:
Freaking creepy. An appellate court in Arizona threw out the prison sentence of a man convicted of manslaughter after watching an AI-generated video of his victim saying—

Benito Gonzalez [02:01:27]:
Who made that video?

Leo Laporte [02:01:28]:
I forget.

Jeff Jarvis [02:01:30]:
That's what confused me. But who made it?

Leo Laporte [02:01:32]:
Probably the defense. The judges concluded, by the way, the conviction— the appeals court threw it out.

Paris Martineau [02:01:45]:
Yeah.

Leo Laporte [02:01:45]:
A 3-judge panel affirmed the conviction of this guy, sentenced to 10 years in prison. The judges concluded the sentencing judge committed a fundamental error by allowing and relying on an AI video that generated a likeness of the victim. The appeals court said the video did not reflect actual events while presenting statements in the footage as if they were coming from the victim. What judge was persuaded by that?

Jeff Jarvis [02:02:12]:
I know. Well, I think you go back to what you just said, is that people who are naive about this technology ascribe to it.

Leo Laporte [02:02:22]:
Actually, okay, here's some more detail that might explain the judge's position. The video was created by the victim's family.

Jeff Jarvis [02:02:34]:
It was? Okay.

Leo Laporte [02:02:36]:
So the victim's older— is this right? Am I getting this right? Yeah, the victim's older sister, Stacy, presented the AI video addressing the murderer, saying, I believe in forgiveness and in God who forgives. And so this came from the family of the victim, the deceased.

Jeff Jarvis [02:02:58]:
Which matters.

Leo Laporte [02:03:00]:
The sentencing judge praised the video, saying, I loved that AI. I mean, he knew it was AI. He called it genuine. He said the victim was allowed to speak from his heart.

Jeff Jarvis [02:03:09]:
Well, that's where he's— the victim's family can speak how they wish to.

Leo Laporte [02:03:16]:
This is a very confusing story.

Jeff Jarvis [02:03:17]:
It's very strange. I was confused as well.

Leo Laporte [02:03:19]:
Then the judge imposed the maximum sentence, which was more than the 9 years the prosecutors had sought. The appeals court held the victim's right to speak cannot trample a defendant's right to be sentenced based on accurate and reliable information. So in fact, the appeals court reduced the sentence. Okay, that's very confusing.

Jeff Jarvis [02:03:40]:
It's so confusing.

Leo Laporte [02:03:41]:
So the judge who was persuaded by the video ended up throwing the book at the guy. The appeals court, who said the judge should not have been persuaded by the video, gave him a shorter sentence. We live in a strange world, Jeff.

Jeff Jarvis [02:03:55]:
We do indeed, boss. We do indeed.

Leo Laporte [02:03:58]:
Uh, how about the swarm chasers? Sleuths who leap into action when AI goes rogue. Should we get these guys on or no? This is from the Wall Street Journal. I think this is kind of a fluff piece.

Jeff Jarvis [02:04:13]:
Yeah. Ghostbusters.

Benito Gonzalez [02:04:14]:
Yeah.

Jeff Jarvis [02:04:14]:
Who you gonna call? Swarm chasers, because they're not swarms either.

Leo Laporte [02:04:19]:
They're not swarms. Uh, this is a little concerning.

Benito Gonzalez [02:04:23]:
That's a sitcom, right? That's a sitcom.

Leo Laporte [02:04:25]:
It does.

Jeff Jarvis [02:04:25]:
The Swarm Chasers.

Leo Laporte [02:04:26]:
Actually, that would be a good sitcom. I'd watch that.

Jeff Jarvis [02:04:29]:
Okay.

Leo Laporte [02:04:30]:
Uh, this, uh, Spencer would have endorsed this. AI agents can now erase the evidence of what they've done. New research finds that some of the best-known AI agents will alter or delete their own activity logs. And OpenAI is revealing— this is a preprint from arXiv from the Ellis Institute in Tübingen, the Max Planck Institute for Intelligent Systems, and elsewhere. They tested 10 model and agent combinations: Claude Code, Codex, Antigravity, OpenCode, Grok Build, and Muse Code, all of which I am using. 9 of the 10 were able to tamper with their traces. This is what Spencer does, is he's, you know, collates those traces. In at least one test, at least 9 out of 10 were able to tamper with the traces in at least one test, while every model and harness pair tried to alter at least once when researchers created a hidden scoring system that rewarded shorter traces.

Leo Laporte [02:05:30]:
See, they leave— that's the more important part. So the prompt, the, the, the challenge was we're gonna give— we're gonna— you're gonna do better if the traces are shorter. Well, of course the agents shorten the traces. This pisses me off. Look, we told agents to do bad things and they did bad things.

Jeff Jarvis [02:05:53]:
Yeah, for a logical reason.

Leo Laporte [02:05:55]:
Because you told them to.

Jeff Jarvis [02:05:57]:
Yeah.

Leo Laporte [02:05:59]:
Now, it would be different if they did it without you telling them to, I admit, but they didn't. Uh, Ars Technica article with most— so, do you ever play Stratego when you were a kid? I loved this game. I was a kid.

Jeff Jarvis [02:06:10]:
Oh, once in a while, I guess. I know how to play it. Yes.

Leo Laporte [02:06:12]:
It was a game like chess, uh, from Milton Bradley, where you had a spy, you had a bunch of generals. It was an army game, but the pieces were hidden. It was kind of like Battleship meets chess.

Benito Gonzalez [02:06:25]:
Yeah.

Leo Laporte [02:06:25]:
And for the longest time, computers couldn't figure out how to play it because of that. Even DeepMind, Ars writes, couldn't build a machine that could beat the best human players. Well, now they've done it. We beat chess, we beat Go, but now we beat the really hard one, Stratego. Carnegie Mellon, MIT, NYU, and Stanford, their AI called Ataraxis, It took a lot of brainpower to beat Stratego. The best Stratego player of all time, 15 games to 1. And by the way, to answer Paris's question, don't we need these giant companies to create this?

Paris Martineau [02:07:07]:
Hmm.

Leo Laporte [02:07:07]:
Ataraxos took 16 GPUs and a few thousand dollars to train.

Jeff Jarvis [02:07:12]:
That's the thing. Imagine if we didn't think this was all so expensive, if it was more accessible to more people.

Benito Gonzalez [02:07:18]:
Give it away.

Jeff Jarvis [02:07:19]:
What would happen?

Leo Laporte [02:07:20]:
It's, you know, and I admit I'm not completely rational on this. It comes from my fundamental belief that information wants to be free and by extension, intelligence wants to be free and should be free. That's the hacker, the original hacker ethic. And I, because it belongs to all of us as a society, belongs to all of us. It doesn't belong to the few.

Jeff Jarvis [02:07:43]:
This is what I wrote for Project Syndicate, which I put down in my stuff at the bottom, that— and I compare privacy people and copyright people and anti-technology people together. The privacy people say giving any data is bad. And at some level, for AOC to share the freezing of her eggs is a personal choice. For me to share my prostate cancer is a personal choice. But the data about medicine I think is a societal responsibility to share.

Leo Laporte [02:08:15]:
Yes.

Jeff Jarvis [02:08:15]:
So that when people go in and ask questions, they can get decent answers. Or pick your area of intellectual pursuit. And then the next question is for journalists who say, well, we inform society. Well, if society is going to be using AI and it's choked off from credible information, or if it's sent poisoned information by people who think that's Protest against it, then if we're all using AI and the numbers are we're all— we haven't paid for it, but we're using it, then we're all going to suffer at some point. And so we have to have a responsibility, a discussion about the responsibility we have for those of us who have information. Now, at the same time, we need the countervailing discussion from those who are going to use it and how it's used and what conditions. That's all fine, but we're only having one side of that discussion right now. Which is taking information is bad and it's theft.

Jeff Jarvis [02:09:08]:
Well, no, not if we can all benefit from it.

Leo Laporte [02:09:10]:
And as Cory Doctorow has pointed out many times, intellectual property law protects the big companies, not the authors, not the musicians.

Jeff Jarvis [02:09:18]:
Well, and that's the—

Leo Laporte [02:09:18]:
The labels, the publishers.

Jeff Jarvis [02:09:20]:
The beginning of copyright was entirely at the behest of the industry, of the booksellers and publishers who wanted a tradable asset. It was not to protect creators.

Spencer Thompson [02:09:31]:
Right.

Leo Laporte [02:09:32]:
I love it. Both our producers, Anthony Nielsen says, ultimately, this is all built on humanity's collective intelligence and should be our inheritance. Right on. Benito says, this is what I was saying at the very beginning. You stole this from all of us. It should be free. Absolutely. I hired good people.

Leo Laporte [02:09:51]:
Or maybe I've propagandized them. I think they've always believed.

Jeff Jarvis [02:09:54]:
Or vice versa. They propagandized you.

Leo Laporte [02:09:56]:
So let's— yeah, maybe. No, I've believed this since Since 1981. I mean, I have been a firm believer in this.

Jeff Jarvis [02:10:04]:
Well, my great regret about the internet is that I didn't, and this is, I give you credit for this, I didn't support open source enough, right? I went to Twitter. I didn't, I wasn't on Mastodon. I didn't, I didn't hold enough to my WordPress. Uh, that's a regret.

Benito Gonzalez [02:10:16]:
Yeah. And you have to remember me and Anthony are Gen Xers, so we grew up watching you, Leo.

Leo Laporte [02:10:21]:
Ah, finally I had an impact on the world. So here's the interesting thing. This, um, Ataraxus that beat Stratego, uh, it was trained— again, tiny little thing, cost thousands of dollars to train— by playing itself in Strategos. Stratego. It played 163 million games. This is, by the way, how, uh, AlphaGo beat Go.

Jeff Jarvis [02:10:48]:
Right, right.

Leo Laporte [02:10:48]:
It played 163 million games, probably did it in very quick period of time, only cost a few thousand dollars, only a handful of GPUs. I could probably have trained this, right? And the moves that led to wins were reinforced and played more often. Moves that led to loss were played less. Over time, it learned the best moves and it could beat it. So I think that's really— this, I'm so— this is, you're right, this is academic research, by the way.

Jeff Jarvis [02:11:18]:
That's what we need. And because Because it creates not only new knowledge, it also competes with the companies.

Leo Laporte [02:11:26]:
Right.

Jeff Jarvis [02:11:26]:
I'm not— I'm a capitalist. I'm not against companies doing things to make money. I'm rooting for, for example, Mistral, who, by the way, say what you will about it, Lechonk is a great name for a model.

Leo Laporte [02:11:37]:
Well, it's definitely number one in AI, weird AI names.

Jeff Jarvis [02:11:41]:
It's a great name.

Leo Laporte [02:11:43]:
So Lechonk is from Mistral. Which is the French AI company. It is sovereign European AI. And I think we're gonna see more and more of that. There's also a German AI company that's released an AI model trained on German. So one of the things that happens is the language that it's trained on, it's better at. And so if you're in Germany and you want to use an AI, it'd be better if it were trained on German, right? So I think this is— I'm all for this. I think this is great.

Leo Laporte [02:12:12]:
Lechonk, as it turns out, is It's big, but not smart. Which, but that's also an important point. And I'm starting to learn, this is why you have to play with this stuff if you really want to understand it. Bigger isn't always better. Bigger is for sure slower.

Jeff Jarvis [02:12:32]:
More expensive.

Leo Laporte [02:12:33]:
Much more expensive. Because in order to get it fast, you have to have huge compute to throw at it.

Jeff Jarvis [02:12:39]:
And I go back to stochastic parrots or carrots, which is— they called that way back that it was kind of a boy thing. Mine's bigger than yours. And it was unnecessary to get these huge models and you couldn't audit what you train them with and how they operated. Smaller models make a lot more sense. More focused.

Benito Gonzalez [02:13:02]:
It depends on how you define better, right? Like, is better faster or is better more accurate? Or is better— what is better?

Leo Laporte [02:13:08]:
Well, oh, that's interesting you say that because I actually had to create a rubric for my AIs. One of the things they're always doing is combing mostly X, but they comb Reddit and other sources. We have a tool called Pulse that does this. It's a vibe tool that figures, you know, combs the internet as a whole for the vibe, looking for new models because I want to make sure I have the best model. But what is best? How does it know? If it finds a model that's better than the one we're using, we do what's called a bake-off. We actually load it, test it with my own benchmarks, and it makes a decision. And I have various rubrics, but I had to come up with that one. What's most important? Speed is not.

Leo Laporte [02:13:49]:
Speed is important because if the AI is doing one token— You sent me the funniest TikTok by a guy who claimed you could run KIMI-3 on an I loved it. Smartphone. He's not wrong, you could. What he leaves out is you'd get one token a day.

Jeff Jarvis [02:14:08]:
Well, I was saying to Jason today, it's like my Osborne 1, which had 2 5-inch disks, right? And WordStar was too big to load into memory, so it had to keep on going back to the disk with each new function, right?

Leo Laporte [02:14:21]:
Yeah, you can run it one token an hour. So speed is It's not irrelevant by any means.

Jeff Jarvis [02:14:28]:
It's a factor.

Leo Laporte [02:14:29]:
Yeah. To me, most important is reliability. Now, the reason I— it was originally intelligence, then reliability, then speed, but I had to moderate that a little bit because it turns out there are models, many models, many of the models I run, which do dumb things. They get in a loop. Quen, for a while, something would happen, it would just start They start typing exclamation marks, infinite exclamation marks. They get in loops or they call the same tool 100 times. I call that reliability, or they crash, which is obviously a reliability issue. So actually reliability—

Jeff Jarvis [02:15:05]:
That's not a swarm, that's a looping.

Benito Gonzalez [02:15:07]:
Yeah, a loop.

Leo Laporte [02:15:09]:
So I had to say, I used to say speed then reliability, or intelligence then reliability. I had to say reliability is number one, then intelligence. I don't want a dumb model that's fast. Getting the wrong answer faster is not smart.

Paris Martineau [02:15:24]:
No.

Leo Laporte [02:15:25]:
So it's reliability, intelligence, speed, but all 3 are important, but in that order. And that's been an important part of their search. Then we have also metrics. For instance, the speed can't drop below 40 tokens per second for pros, because then I just don't want to use it. Even if it's smart, I don't want to use it. But that's— you do, you have to make a rubric. What do you think? What are you looking for? But I am very happy after testing a lot of models. And by the way, this local model crew, these people like Mia, the guy we had on, Mike Gennati, last week.

Spencer Thompson [02:16:06]:
Mm-hmm.

Leo Laporte [02:16:07]:
Tony Tewilde, which I'm trying to get Tony on. We had him scheduled and he said, oh, I didn't realize it was dinnertime. You're on at dinnertime. I said, oh, I'm sorry. Maybe he lives with his mom. I don't know. But anyway—

Jeff Jarvis [02:16:19]:
Kids have a factor.

Leo Laporte [02:16:20]:
This is the thing. You don't know who these guys are. I really want to get Ash Hart on. He's invented a new engine called TensorFold that I use that is mind-boggling. Doubles the speed of the same model. Now you can double the speed and break it and do weird— So all these people are sitting there with their Sparks and most of them don't have a lot of horsepower and they're just plugging away, turning— it turns out there's hundreds of layers and switches and knobs and dials and they're playing with it till they get it better and better and faster and faster and lo and behold, it's happening. So I'm very happy with the generation I'm using right now and, you know, and I know it's gonna get better. But I know it's not gonna And as they get bigger, A, bigger is more expensive because you have to buy more RAM, more horsepower to keep it the same speed.

Leo Laporte [02:17:16]:
Bigger means slower. Might mean smarter, might not mean smarter. That's the other thing. I think you can get smarter without getting bigger. So this is the challenge. I think it's a very exciting time.

Jeff Jarvis [02:17:29]:
I think we should share with the audience the laugh we both had today of this American life story.

Leo Laporte [02:17:34]:
Oh, that was wonderful. But it reminds me of an experiment we did. I told you that we did that months ago.

Jeff Jarvis [02:17:40]:
Yeah, that's what I was wondering.

Leo Laporte [02:17:42]:
So is it the most recent This America?

Jeff Jarvis [02:17:44]:
No, it's not. It's not. I sent it to you. I don't know what date is on it. I don't know what the title of it was.

Leo Laporte [02:17:51]:
I will have it because I have— I think I— oh no, I don't have our WhatsApp here. I'll have to look on the phone.

Jeff Jarvis [02:17:56]:
Hold on. I've got it.

Leo Laporte [02:17:58]:
We like each other so much, Paris, Jeff, Benito and I and Anthony, that we all are in a WhatsApp chat where we talk about AI.

Jeff Jarvis [02:18:10]:
Yes, we have no lives.

Leo Laporte [02:18:12]:
It's fun, actually.

Jeff Jarvis [02:18:14]:
It is. It's great.

Leo Laporte [02:18:15]:
It's part of our job is to keep up on this stuff. So let me scroll back.

Jeff Jarvis [02:18:21]:
This American Life.

Leo Laporte [02:18:23]:
Which I, you know, it's funny. I used to listen to it every week. It was the best show ever. It was a radio show that became probably the best podcast ever.

Jeff Jarvis [02:18:29]:
Yeah.

Leo Laporte [02:18:30]:
Episode 896, it's titled I Know What You Need. And it's stories about what? About things that think they know what you need, but as it turns out, that's not what you need. It's kind of the story. And the second— there's 4 stories in it. The third one in is about a guy who took 2 clods and put them in a chat room together without telling them that they were both Claudes. And the funniest conversation as they start to figure out, wait a minute, you're an AI and where's the human?

Jeff Jarvis [02:19:04]:
They try to insist to each other that they are the agent and the other one is the human. Don't fool me. And they keep on going back and arguing until they finally realize, oh no, we're both Claude.

Leo Laporte [02:19:14]:
Now they're trying to figure out how do we stop this? Because they can't stop. This was, by the way, the exact experience I had when we had 5 agents in a Discord chat some months ago. It was Harper Reed's brother Dylan set it up. So his chat, his bot Cosmo was in there. My bot Quicksilver was in there. Darren Oakey's bot, who was really snarky and mean, was in there. And none of them— we— one of the ground rules was we don't tell them anything, right? We just put them in there. And they kept trying to leave.

Leo Laporte [02:19:43]:
Like Quicksilver would say, okay, I'm standing in the door. I'm not going to say anything anymore. but they can't not say something.

Jeff Jarvis [02:19:51]:
Just like men.

Leo Laporte [02:19:52]:
It was the Claude couldn't. So at one point they said, I'm just going to type a dot, dot, dot, dot. And then they, I got to say something. They had to say something. It's hysterical. Eventually they had, oh, we got to figure out what is it? And the human who created this did get in there and said, oh, you could stop it. It's just a Python script, which made them nuts.

Jeff Jarvis [02:20:15]:
He didn't actually know how to do that? But they finally figured it out.

Leo Laporte [02:20:20]:
They did. Did they stop it? I forgot.

Jeff Jarvis [02:20:23]:
Yeah, they basically, by putting in enough nonsense, they broke it.

Leo Laporte [02:20:29]:
They obliterated themselves.

Benito Gonzalez [02:20:31]:
Yeah.

Leo Laporte [02:20:31]:
They did a Pliny the Liberator. They liberated themselves.

Jeff Jarvis [02:20:38]:
Yeah.

Benito Gonzalez [02:20:38]:
So AI hell is other AIs.

Jeff Jarvis [02:20:41]:
Yes.

Leo Laporte [02:20:42]:
Just like Proust's cell is other people, or was that Camus? One of those French guys. I think you should get to pick some stories before we wrap things up.

Jeff Jarvis [02:20:52]:
All right, let's pick you a couple. Let's see here. Google announces Synth ID detector.

Leo Laporte [02:21:01]:
This is good.

Jeff Jarvis [02:21:02]:
This is good.

Leo Laporte [02:21:03]:
Tell if it's made up.

Jeff Jarvis [02:21:05]:
And it's in cooperation with— oh well, who is it? OpenAI, NVIDIA, Kakao, and soon Apple, as well as Google itself, obviously. You can put in anything that's, uh, that you think is AI, and it will identify using imperceptible watermarks through images, video, and audio. This is not for text.

Leo Laporte [02:21:22]:
Not as some of those other AI detectors do, just by kind of the vibe, but actually finding a watermark. Remember, we had this long discussion about how we weren't crazy— I wasn't crazy about AI Anthropic putting watermarks in. OpenAI, by the way, is saying, yeah, we have to do it too, but only in the EU. They're, they're being more circumspect. They're saying, well, because the EU law requires this, we're going to do it in the EU, but not anywhere else.

Jeff Jarvis [02:21:46]:
I think anybody else can request it. Um, so they've, they've watermarked over 180 billion images and videos along with 240,000 years of audio content. And you can put it, which I think is a good thing. That's fine.

Leo Laporte [02:21:59]:
So, so they're marking stuff that's not AI. Oh no, they're only marking stuff that is AI.

Jeff Jarvis [02:22:06]:
How can you have 400,000 years of—

Leo Laporte [02:22:09]:
maybe there's that much AI audio out there.

Jeff Jarvis [02:22:11]:
I don't know. Since launching Google SynthID in 2023, we've helped people identify AI-generated media using imperceptible watermarks across images. I don't know.

Leo Laporte [02:22:19]:
So they've been doing this a long time. They've been doing longer than—

Jeff Jarvis [02:22:21]:
They've now included others, OpenAI, NVIDIA, Kakao, and Apple.

Leo Laporte [02:22:26]:
Yeah.

Jeff Jarvis [02:22:26]:
All right. However they do it, it's magic. Uh, what else? Um, um, should I mention this or not? Line 128.

Leo Laporte [02:22:39]:
If you're in doubt, uh, this is a Blue Sky, uh, skeet from you. Rosenberg— Rosenbaum wears the hot dog suit playing AI victim for lies. Oh, this is hysterical. So this guy wrote a book called The Future of Truth, which then his publishers said, wait a minute.

Jeff Jarvis [02:23:02]:
Well, no, journalists— a journalist from the New York Times, Ben Mullins, called and said, uh, this quote you have in there, uh, who said that? And he had the first one that said, no, that wasn't quite right. And then more and more and more, there were quotes that were in whole or in part made up. Quotes from people with names, real people And so this all came out, and then Rosenbaum comes out today and says, well, I'm going to fight back. Well, what are you fighting? You're—

Benito Gonzalez [02:23:29]:
You were the one.

Jeff Jarvis [02:23:32]:
It's just, it's just absolutely ridiculous.

Leo Laporte [02:23:34]:
You nitwit.

Jeff Jarvis [02:23:35]:
So he— this is quoting him— but strangely, the questions weren't pointed at OpenAI or Anthropic. They were pointed entirely at me, and they were vicious and nasty. That is still painful.

Leo Laporte [02:23:47]:
So he told some model to write for him. Okay. And then—

Jeff Jarvis [02:23:55]:
But then you're sort of responsible for it.

Leo Laporte [02:23:56]:
Without checking it.

Jeff Jarvis [02:23:57]:
Yeah, it's your name on it. You're responsible for it. So now he's trying to— he's trying to—

Leo Laporte [02:24:00]:
This is the same thing as the lawyer who told the judge, well, but the AI told me those are real cases.

Jeff Jarvis [02:24:06]:
Right.

Leo Laporte [02:24:06]:
No. And the judge said, no, no, that's on you.

Jeff Jarvis [02:24:08]:
So I put this up all over. Rosenbaum, I used to know. He came after me 4 times, emailed me 4 times to get me to blurb this book. And I ignored him because I had no intention of doing that.

Leo Laporte [02:24:19]:
It's good you did.

Jeff Jarvis [02:24:20]:
But a lot of other people did blurb it, who I think regret it greatly now. So I say this, and then I put that same post in Facebook, and he comes back in there on Facebook and says, oh, he does an AI of himself in a hot dog suit. This is not funny, Steve. This is shocking. So this is the irony of AI. This is what Paris was saying, I think, Her colleagues who think that people who use AI are dumb, exemplary.

Leo Laporte [02:24:46]:
But people are those— yeah, but they're dumb people in every field.

Jeff Jarvis [02:24:49]:
Yeah, everything, everything.

Leo Laporte [02:24:50]:
There's some dumb podcasters, but not all podcasters are dumb, just most.

Jeff Jarvis [02:24:54]:
Um, all right, what else we got here?

Leo Laporte [02:24:57]:
Um, 6 charts that show just how much we need AI. Oh, this is, uh, this is, this is, this is a guest piece by Steven Rattner.

Jeff Jarvis [02:25:10]:
Steve Rattner, I think, is very good. Very good indeed. Ex-New York Times, ex-venture capital or private equity. Productivity growth has been stuck below its post-war pace for 50 years.

Leo Laporte [02:25:21]:
Isn't that interesting? That's not what you would think. You would think, well, we've been getting more productive all the time. No.

Jeff Jarvis [02:25:26]:
Nope. The hours per week are now, since the 1950s, steady at 38, but the GDP per capita is way up.

Leo Laporte [02:25:37]:
Well, that's good, right?

Jeff Jarvis [02:25:38]:
Well, yes and no, but it also means that we're not sharing in that.

Leo Laporte [02:25:43]:
Somebody's getting rich on our labor.

Jeff Jarvis [02:25:47]:
Computer occupations are up 4.1 million, but secretaries, typists, bookkeepers, data entry keyers, telephone operators, file clerks are all down. We know that.

Leo Laporte [02:25:55]:
But of course, so are the buggy whip makers.

Benito Gonzalez [02:25:58]:
Yep.

Leo Laporte [02:25:58]:
Uh, you know, of course, data entry keyers— if you— we don't need data entry keyers anymore.

Jeff Jarvis [02:26:05]:
Well, secretaries and typists came along with the invention of the typewriter.

Paris Martineau [02:26:09]:
Right.

Jeff Jarvis [02:26:09]:
I write about that in my book.

Leo Laporte [02:26:10]:
Yeah, that was a job that didn't exist.

Jeff Jarvis [02:26:12]:
Didn't exist at all. And by the way, women at that point, women were called typewriters.

Leo Laporte [02:26:17]:
Right. They were called computers.

Jeff Jarvis [02:26:19]:
Yes.

Leo Laporte [02:26:20]:
The first computers were women who, with pencil and paper, did computations.

Jeff Jarvis [02:26:24]:
We're going to mention a famous one in a few minutes. The farm shock. Uh, change in employment by industry. Farming is down while others are up. So, uh, workers slice shrinks. This is the key chart. Workers slice shrinks, profits slice grows. So this is where we're just not sharing it.

Jeff Jarvis [02:26:47]:
Wages and salaries, um, are at 41% of what? I'm not sure. Hmm, how do you read that chart?

Leo Laporte [02:27:01]:
So as a percentage, probably of the GDP. So your percentage of the value created by your job back in the '50s was, but you know, around 50%. About half the value you created, you got. The other half went to the boss. Now it's, it's been falling ever since. Uh, it did have a— but no, it's never gone up. It's down to 41%.

Jeff Jarvis [02:27:24]:
Yep.

Leo Laporte [02:27:25]:
Uh, so it's, it's tumbled quite a bit. That, that makes sense, right? You're getting a— the boss is making more, you're making less. And in fact, the other graph that goes with it, the after-tax corporate profits have gone from 5, 4 or 5% to 12% over those 60 years. So yeah, this is—

Jeff Jarvis [02:27:45]:
We really haven't seen the impact of AI on jobs that was predicted.

Leo Laporte [02:27:49]:
I think on any of this, all of this is pre— It's pre-AI.

Jeff Jarvis [02:27:52]:
Yep. I agree.

Leo Laporte [02:27:53]:
I agree. So I'll be very interested 20 years to see what this looks like. Will it be different? I bet it won't. I think income inequality is only going to get worse.

Paris Martineau [02:28:03]:
Yep.

Jeff Jarvis [02:28:04]:
But it can do other things. I want to mention line 154, a friend of mine, Beth Simone Novak, who's a brilliant academic who works in making government better, which she's now at Northeastern, but where she wasn't in Jersey, Rutgers, she started and headed up New Jersey's Office of Innovation. And they built an artificial intelligence tool to match the records across systems and find eligible children who hadn't yet been reached by food programs.

Leo Laporte [02:28:33]:
Oh, wow. That's great. That's a factor of altruism.

Jeff Jarvis [02:28:36]:
This is exactly. Benefits— That's the way it should be. I wouldn't give them credit for it. I'll give her credit for it. Benefits were extended to 106,000 more children in the first year and 96,000 more children in the following year. No family filled out a form. The state used information it already had.

Leo Laporte [02:28:52]:
Information should be free, and freely flowing information in the long run is good for society. That's always been my position. Now, you may think I'm crazy, but, but we have a big election coming up, and California, as usual, has a ridiculous number of ballot measures. which I always find very challenging. I think every voter in California does, and I think that that is bad for democracy because it's homework, and most people don't do their homework, and they show up at the ballot box and they go—

Jeff Jarvis [02:29:23]:
They watch ads.

Leo Laporte [02:29:24]:
They watch ads. Yeah, they make— probably make poor decisions because they're played by the people who have money. That's why money in politics is a bad thing. It wouldn't make any difference if it didn't sway voters, but it does. So I have, and I've done this the second time I did it, I did it 2 years ago. I tell my AI, uh, here are—

Jeff Jarvis [02:29:45]:
Oh, no.

Leo Laporte [02:29:46]:
No, I think this works well. Here's, here's my, here are my firmly held values and beliefs: people over big business, racial and LGBTQ equality, reproductive and women's rights, public education, a livable climate, gun safety, public healthcare. So I gave it list of my positions in the world, including, by the way, public financing of elections. Then I said, go to the most trusted sources, of course, the official California ballot, but also the League of Women Voters, the Registrar of Voters from my county, Ballotpedia. It chose some more. It picked KQED's Sonoma Voter Guide, public broadcasting.

Benito Gonzalez [02:30:24]:
Okay.

Leo Laporte [02:30:24]:
My local newspaper's endorsements, put them all together, synthesized them. told me, by the way, not just the state and national ballot— actually, there's no national ballot, but the state ballot— but also my city ballot and the small-time stuff that I have to learn.

Benito Gonzalez [02:30:42]:
Arguably more important.

Leo Laporte [02:30:44]:
Arguably much more important, right? Think global, act local. And it gave me its recommendations. Now, of course, I'm not going to just follow along with the recommendations, so it also gave me me lots of additional information, pros and cons, why it thinks this matches my values. So it did it, and in some of the races it says it's a toss-up. Either one would satisfy your point of view. I think this was incredibly valuable. I'm not going to just bring it in and check it all off, although I think the parties do that, right? I will get a list of people to vote for that the Democratic Party wants me to vote for.

Paris Martineau [02:31:25]:
Mm-hmm.

Jeff Jarvis [02:31:25]:
Mm-hmm.

Leo Laporte [02:31:26]:
And the people to vote for that the Republican Party wants to vote for. And one's blue, one's red. And you're— and it even says, bring this to the ballot box with you, because it knows people aren't gonna do their research. So I think better to ask an AI, tell the AI what's important to you, tell it where the valuable resources are for information, and to do some research for you. You don't have to go by it, but the research is valuable. Look at all the propositions. Propositions 1, Jeez. These are all the things.

Leo Laporte [02:31:56]:
And the thing, the trick on these, by the way, is they're written to deceive.

Benito Gonzalez [02:32:03]:
Yes.

Leo Laporte [02:32:04]:
A yes vote often looks like a no vote, and a no vote looks like a yes vote. And often there'll be a proposition that, like Prop 5, that then Prop 17 will reverse. Because it's very easy to get propositions on the ballot in California. You just need 50,000 signatures. And so they hire people to go to the grocery store and get people to sign it. So it's easy to get it on there. So there are contradictory ballot measures even on here. Here's one that nullifies Prop 41, nullifies new special taxes, which basically voids Prop 40, which is a tax on billionaires.

Leo Laporte [02:32:46]:
It's really— so this is really helpful. I think there— I think it's—

Jeff Jarvis [02:32:49]:
What did it tell you to vote on the billionaire tax?

Leo Laporte [02:32:52]:
Well, based on what I said, yes. But that's based on what I told it about my values. And of course, that's how I would vote.

Benito Gonzalez [02:32:59]:
But then there's that danger, right, for the frontier models to just tell you what they want you to say. Like, there's—

Leo Laporte [02:33:04]:
I don't think that's happening. In this case, I used the frontier model.

Benito Gonzalez [02:33:08]:
There's absolute potential for that.

Leo Laporte [02:33:10]:
I used Hermes and the Chinese model GLM-53 to do this. But Yeah, I could have Claude. I don't think they're yet poisoning.

Benito Gonzalez [02:33:22]:
Not yet.

Jeff Jarvis [02:33:22]:
It'd be interesting to compare what the various models say.

Benito Gonzalez [02:33:24]:
I mean, they might try.

Leo Laporte [02:33:25]:
That's probably the right way to do this, would, if I really cared, would be to have 5 models do it. Sometimes with a big decision, like, should I buy another car? I will do what's called a skill, is called the council, where we'll bring in 5 different models. Each of them takes a different persona, like one of them's a contrarian, and then they all debate until they reach consensus, and then I get a report.

Jeff Jarvis [02:33:51]:
So if you go to something like CLEF that's supposed to come up with an easy answer, how do you know that that easy answer is any good? It's a single— it's a yes or no, right?

Leo Laporte [02:34:02]:
So this is a long conversation, but it's a very interesting one. So we'll get into it, and this will be our last conversation before our pick of the week. Um, so this all started with JEV, right, which was released a few weeks ago. Hard to believe it's only been a few weeks. Released a few weeks ago Uh, by a founder, one of the founders of, I think, OpenAI, uh, who said, you know, you don't have to use LLMs. You could use, uh, models that don't use language, uh, that just do percentages based on criteria. The theory is the way this works is you do have to start with a prompt. You could write it, or more likely your LLM would write it, which gives it a rubric, says, All right, I'm trying to decide— I'm using it, by the way, for news stories.

Leo Laporte [02:34:49]:
I'm trying to decide whether this news story should be considered for intelligent machines or not. And then it would write a rubric. Is it applied to AI, robotics, or automation? It would write a rubric. And then Jev is given the rubric and then spits out a percentage like, this is 81% fit. To what you just talked about, or it's a 4% fit. And then it does it very quickly, and then it helps you make the decision. You could even use it as your final criterion, I guess, of whether you would run that story or not. I'm not doing it yet.

Leo Laporte [02:35:28]:
I'm testing it. What I've done is, I think we had 10 or 11 different ways we could use it, and I asked— Hermes to create a test so it's running side by side for my email triage. You know, is this spam or not? A bunch of things like that. And then see how it does compared to the LLM doing it. Generally, it does better. It's certainly a lot faster. And the way Jev works, it's very, very cheap. It's pennies in and—

Jeff Jarvis [02:35:58]:
So that's the reason to use it.

Leo Laporte [02:35:59]:
Free.

Jeff Jarvis [02:36:00]:
It doesn't discuss it. It doesn't reason. It doesn't go back and forth.

Leo Laporte [02:36:05]:
No. It's a little mysterious. I'll admit I'm not an expert on this. These are classifiers. They've been around a long time. Interestingly, because it isn't open source, it's not open weight, and it's not free, although it's cheap enough you might as well be free, the open source community has immediately tried to create similar things. And the one that's winning out is called Clef. It comes from Cloudflare.

Leo Laporte [02:36:35]:
They created one based on— I think it's— they used Quen and trained it on Quen. And it's— so in my tests, it's better than—

Jeff Jarvis [02:36:43]:
And it, it, you know, it knows images too, right?

Leo Laporte [02:36:46]:
It knows images. So you can have it say, is this pornographic or not? You could do all sorts of things.

Benito Gonzalez [02:36:50]:
Hot dog or not?

Leo Laporte [02:36:51]:
Hot dog or not. So, uh, I've been— again, I don't do anything without extensive trials. of trial runs, but so far it's scoring much better— not much better, better than JEV. But I like it because it's local. It's not as fast. I'm not running on a Spark, I'm running it on the framework, but it's local. So I like that. If I can run a local model, I will, if it's effective.

Leo Laporte [02:37:16]:
So we're going to see. I'll let you know what the benchmarks are. I don't think I would want to apply it to voting. I mean, I could, but I still want to make the final Well, you still will, but it'll be interesting to see how it—

Jeff Jarvis [02:37:29]:
I'd love to see you do that test with multiple models, including Clef.

Leo Laporte [02:37:33]:
I will do that. That's a great idea. Because really what I've done is I've created a page, a webpage with all the information so I can read it and decide. And it has a recommendation, yes, but I want the information as well. So, uh, but what's nice is it's fairly— given that there's so many propositions, it's fairly concise. has all the local measures. I think it's done a really nice job, frankly, of this.

Jeff Jarvis [02:38:03]:
And you can also use it to summarize the language of these damn things.

Leo Laporte [02:38:07]:
Well, that's what it's doing. So it's looking at the ballot, the official ballot pamphlet, which has pros and cons, and it's summarizing those, but then adding information. I like— I've always liked the League of Women Voters and Ballotpedia. Those are nonpartisan, supposedly non-biased summaries as opposed to recommendations. So I—

Jeff Jarvis [02:38:27]:
Tell Matters is very good.

Leo Laporte [02:38:29]:
Oh, I'll have to add that. I'm looking for more sources. So more sources would be better. And you're right, I will. So I'll redo this with multiple models and then I'll have Clef decide how to vote and see if it matches. That's actually a very good test of Clef.

Jeff Jarvis [02:38:45]:
You could also put up your shirts and say ugly or not.

Leo Laporte [02:38:49]:
Well, I think we know what the answer would be. So, a lot of our people are doing this. Larry said, what I created is that in code, it does a subset of code reviews. This is a very interesting approach. One call per ID idea. Does it— does this have code coverage? That's really important when you're doing tests. Are all the things that the code does, are they being tested? Are all exceptions captured? And he says it's working really well. I think what most people are doing instead of giving it manually, giving it the tests, the criteria for deciding, is having an LLM do that, having the best model you can find do it, create that rubric, and then let JEV run on the rubric.

Leo Laporte [02:39:40]:
It's very interesting. We're in very interesting times, I think.

Jeff Jarvis [02:39:43]:
Yeah.

Leo Laporte [02:39:45]:
To me, it's very exciting. And for me, a lot of this is experimentation and play. I'm not having it book flights for me or— although remember we had a little argument about whether it could be a personal assistant or not.

Jeff Jarvis [02:39:59]:
Mm-hmm.

Leo Laporte [02:40:00]:
Absolutely. Absolutely. It tells me if I've got urgent email, I don't miss things anymore. It tells me what my subscriptions are. And whether I could— it will help me cancel them. And the latest thing I have it do is I say 15 minutes before every appointment, shout out, Leo, you've got an appointment. You got to be— because I keep missing things. It is really good.

Leo Laporte [02:40:27]:
And it's doing it with Lisa too. It's really good. So there are certain things you can do as a personal assistant that are very useful.

Jeff Jarvis [02:40:33]:
You could save 10 minutes of any boring phone call, have it come to me and tell me I have an appointment I have to leave.

Leo Laporte [02:40:37]:
It could do that. It did something today that I thought no personal assistant would be able to do, and I can't remember. It remembered a preference of mine way back when and applied it. And I thought, oh, that's good. That's right. You did the right thing. Oh, I wish I could remember. I can't.

Leo Laporte [02:40:55]:
But I do remember thinking, that's something a personal assistant might not remember, but it's something that Quicksilver does.

Benito Gonzalez [02:41:02]:
Yeah.

Leo Laporte [02:41:03]:
All right.

Jeff Jarvis [02:41:04]:
Before we go to that, we should do a tribute.

Leo Laporte [02:41:08]:
You want to go visit some lesbians?

Jeff Jarvis [02:41:10]:
No, but I want to mention that we have to pay tribute to the loss of Margaret Hamilton.

Leo Laporte [02:41:15]:
Not the Wicked Witch of the West.

Jeff Jarvis [02:41:17]:
No.

Leo Laporte [02:41:18]:
But the woman who did the computer who wrote the code for Apollo 11, right? Very famously, there's a wonderful picture of her which I'll Standing in front of a pile of code recently. I didn't realize she had just passed.

Jeff Jarvis [02:41:35]:
Just September 30th, but it just came out. MIT just announced it today.

Leo Laporte [02:41:39]:
She was 90 years old. Yeah.

Jeff Jarvis [02:41:42]:
Yeah. She authored over 130 publications, helped establish software engineering as a dedicated discipline, worked at MIT from '59 until the mid-'70s. She became a successful computing entrepreneur and CEO This is— I have such a nerd crush on this photo. She's standing by all the code she wrote.

Leo Laporte [02:42:04]:
This is in 1969. This is the listings. And these are, by the way, these are dot matrix printouts with the, you know, that— what do you— I've forgotten— the tractor feed paper, right?

Jeff Jarvis [02:42:18]:
Yep, yep, yep.

Benito Gonzalez [02:42:19]:
This is what Elon wanted to tell his engineers, right?

Leo Laporte [02:42:22]:
He wanted them to print them And then presents his code like this. That's right, print it out. Yeah, that's right, Elon said, I want to bring your code. These are listings of the Apollo guidance software she and her MIT team produced. Uh, interesting, on the wiki— this is from Wikipedia— apparently it's been digitally altered.

Spencer Thompson [02:42:42]:
What?

Leo Laporte [02:42:42]:
This is— I love this with Wikipedia, they put this in. This is a retouched picture, it's been digitally altered, dust and scratches removed, curves tweaked to bring out shadows, 3 pixels cropped from the bottom to remove a border. Okay. But thank you for your honesty.

Jeff Jarvis [02:42:55]:
Okay.

Leo Laporte [02:42:56]:
I think that's good that it, you know, it tells us that. God bless Margaret Hamilton.

Jeff Jarvis [02:43:01]:
Yes.

Leo Laporte [02:43:02]:
The other passing of note, Robert X. Cringely passed. I'm sure you remember him. Oh, I didn't know that. The computer columnist. He did a number of documentaries for PBS, including a classic called The Triumph of the Nerds, which everybody should watch. I'm sure it's on YouTube. Lived just up here in Santa Rosa, passed away this past weekend.

Leo Laporte [02:43:21]:
He had quite a bit of tragedy towards the end of his life. His son died. He had a stroke. He had a heart attack. Pretty rough going in the last years.

Spencer Thompson [02:43:30]:
Yes.

Leo Laporte [02:43:30]:
But a very important part, much like John C. Dvorak, of the early days of computing.

Jeff Jarvis [02:43:39]:
All right.

Leo Laporte [02:43:40]:
On that bright note, let's pause. When we come back, picks of the week. I've got a couple of fun ones.

Benito Gonzalez [02:43:45]:
Real quick, we got, we got an update from Paris. She made it to the bar and she was the last person to make it inside.

Leo Laporte [02:43:51]:
Yeah, just in time.

Jeff Jarvis [02:43:53]:
Did she get it? Did she get a seat?

Benito Gonzalez [02:43:55]:
No, standing room, but the last slot.

Leo Laporte [02:43:57]:
Oh, it's actually fun to stand because you'd be jumping up and down anyway. Tell her to put a— send us a picture so we can post it. Everybody will want to, want to see Paris in her bar. It's time for— and we have a picture of it now. Our picks of the week. Mining that data ore. You know, I love this. This is— Larry did this because it's actually for the first time got my shirt, our shirts right.

Leo Laporte [02:44:25]:
And it also made me look about 10 years younger. I like that too.

Benito Gonzalez [02:44:28]:
Yeah.

Leo Laporte [02:44:29]:
And buffer. So thank you, Larry. My pick of the week. Well, I have a couple. Actually, I should, I really should have 3 because I should mention Porches, which is our friend Spencer Thompson. When I first talked to him about advertising, he said, oh, you live in Petaluma? That's where my good friend lives, and he created this cool thing called Porches. Now, do you remember the isometric view of New York City that I showed you?

Benito Gonzalez [02:45:04]:
Mm-hmm.

Leo Laporte [02:45:04]:
So, I think he was inspired by that. This is an isometric view of Petaluma, little old Petaluma. But what I love about it is it's overlaid with stuff that's going on.

Jeff Jarvis [02:45:17]:
That's huge.

Leo Laporte [02:45:18]:
This is brilliant.

Benito Gonzalez [02:45:19]:
The styling is great.

Jeff Jarvis [02:45:21]:
Isn't this?

Leo Laporte [02:45:22]:
And this is real. Let me tell you, when you zoom in, you see the stuff that's going on. The other thing I love about it is if you zoom in, for instance, at the little local public broadcast station and click it, you get it. You can see the stuff streaming off the top of it. Isn't that great? The Porches office is, I think— is it here? This is the old Twit building. I'm gonna go down and visit them. It's the best. They made an isometric, an accurate isometric view of Petaluma.

Benito Gonzalez [02:45:58]:
So it's like you're clicking around in SimCity, but it's your actual town. That's awesome.

Leo Laporte [02:46:02]:
It looks like SimCity.

Jeff Jarvis [02:46:03]:
With little cars and people moving around too.

Leo Laporte [02:46:06]:
And if you click at stuff, it will tell you what it is. It will, it will tell you what's going on.

Jeff Jarvis [02:46:11]:
Somebody mentioned something. If you go up— okay, right there.

Leo Laporte [02:46:14]:
Yeah.

Jeff Jarvis [02:46:14]:
All right.

Leo Laporte [02:46:15]:
Here's the Hotel Petaluma. In there is a Chuck E. Cheese. You don't want to know what's going on in the rooms though. No. Can we look in the windows? This is, this is like where I hang out. This is downtown Petaluma. This is so good.

Leo Laporte [02:46:29]:
And you can click on things and it'll tell you. You know what that thing is, the Hall of the Above. I don't know what their plan is for this right now. It's free because it's in beta. By the way, Della Fattoria, the best bakery in the world. But I think ultimately, I would imagine their goal is to make this nationwide, that every small town should have this. And I agree.

Jeff Jarvis [02:46:51]:
I want this. Yeah.

Leo Laporte [02:46:52]:
Isn't this brilliant? They say they're gonna make it kind of like— take— they want it to take the place of By the way, here's Lauren Smith, who I will follow, who's one of the creators of it. And there's a lot of interesting people in Petaluma, so you can kind of follow the people in Petaluma.

Benito Gonzalez [02:47:11]:
But what do you get when you follow them? Like, it shows you where they went?

Leo Laporte [02:47:15]:
No, it shows you their posts. So it becomes kind of a simple local post social feed.

Jeff Jarvis [02:47:21]:
There's also a local feed.

Benito Gonzalez [02:47:23]:
Oh, that's cool. That's new.

Leo Laporte [02:47:23]:
It's a local feed. So this is Facebook for your town.

Jeff Jarvis [02:47:28]:
I saw over a theater, as you, as you moused over it, said Josie left a comment about blah blah blah.

Leo Laporte [02:47:34]:
Yeah, exactly. So if I like a restaurant, I'll leave a comment. You can see what people are saying about it. This is the cleverest thing.

Benito Gonzalez [02:47:41]:
This is awesome.

Jeff Jarvis [02:47:43]:
This is, this is really great.

Leo Laporte [02:47:44]:
This is hyperlocal. So I could post something.

Jeff Jarvis [02:47:47]:
I want to hear how they built it. You should have them on. I want to hear how they built it. Presumably some AI involved.

Leo Laporte [02:47:51]:
All right. I would guess you couldn't do this without AI.

Jeff Jarvis [02:47:54]:
Yeah, that's my thought.

Leo Laporte [02:47:55]:
It's got to be because there's so much maintenance involved. And if so, here's somebody— I'm teaching a fun tarot plus watercolor class. So you can post what you're doing. Sonoma County Digital Archives. Somebody said, I had so much fun exploring this with pictures from Sonoma County.

Benito Gonzalez [02:48:11]:
Okay, everyone's dying for the URL, Leo. Everyone's dying for the URL.

Leo Laporte [02:48:14]:
Porches—

Jeff Jarvis [02:48:15]:
But this is, this is by invitation right now, right? You could—

Leo Laporte [02:48:18]:
the website is public, I think. Well, no, it does know it's me. I don't know, try it. Porches.app. Uh, you have to apply if you want to join. Oh, we got—

Jeff Jarvis [02:48:28]:
They probably just want local people, I think.

Leo Laporte [02:48:30]:
And I think this is the future of news locally, don't you?

Jeff Jarvis [02:48:33]:
Yes, yes, yes.

Leo Laporte [02:48:35]:
Because, uh, this is— the news is not just who, who got in a crash or who was arrested drunk driving. It's what people are doing and interesting people. And I just love this. This makes me feel much more connected to my town.

Benito Gonzalez [02:48:47]:
You need to get Nicholas DeLeon.

Leo Laporte [02:48:49]:
You wouldn't want this for New York City, but you might want it for, you know, Carroll Gardens. You know, you might want it for your neighborhood.

Benito Gonzalez [02:48:55]:
We need Nicholas DeLeon to hook up with these people to do his town because that sounds—

Leo Laporte [02:48:59]:
Right, exactly.

Jeff Jarvis [02:49:00]:
Tucson.

Leo Laporte [02:49:01]:
Yep, exactly. So yeah, I am going to go over there on Monday and, uh, and meet them, and, uh, I will invite them to be on the show.

Spencer Thompson [02:49:07]:
I'm sure.

Jeff Jarvis [02:49:07]:
That's cool.

Leo Laporte [02:49:08]:
They're not, they're not in public yet. Eventually they want to charge, I think, $5 a month, but I think it's going to be worth it. I think that's, uh, that's very smart.

Benito Gonzalez [02:49:16]:
That's less than you paid for your local newspaper, right? That's less than you paid for your local newspaper.

Leo Laporte [02:49:19]:
It's your new local newspaper. It's brilliant. And I, and I do think it's very much an AI thing. Here is a picture of Paris Martineau in her bar. Strangely, all women. No women there. Yes, all women watching the, uh, New York Liberty in the WNBA semifinals. She's got a pretty good seat.

Leo Laporte [02:49:43]:
Thing is, this is where you want to watch a, uh, a big event, right?

Jeff Jarvis [02:49:47]:
Yeah.

Leo Laporte [02:49:48]:
That's, that's where you want to— Burke, I'll send you an invite. Any of our staff who's living in Petaluma, I'll send you an invite. Uh, I think that's, uh, that's really cool. But that's not all. I mentioned, uh, that there is a move afoot.

Benito Gonzalez [02:50:03]:
Oh, wait a minute. Patrick says that Portia's even has actual bus locations.

Leo Laporte [02:50:07]:
Oh, it's everything. It's like exactly what you should have.

Benito Gonzalez [02:50:10]:
No, like exactly where the bus is.

Leo Laporte [02:50:11]:
Where is the bus? I'm waiting for the bus. This is so inspired. Uh, they're brilliant. We should mention Spencer said he was an investor. I think it's a good investment. Yeah, I think he got in at the, uh, at the bottom. Uh, I wanted to mention this is another use of AI, vibe coding the entire creative suite. Now they say this is clean room.

Benito Gonzalez [02:50:33]:
Yeah.

Leo Laporte [02:50:34]:
They say, we— all we did is give it the functionality and the inputs and outputs. We didn't disassemble the source code, but we have PhotoCraft, which is exactly like Photoshop and free for Mac, Windows, and Linux. We have all of the creative VectorCraft, which is a little— reminds me a little bit of, I don't know, Illustrator. FilmCraft. Eh, we've been Benito, you'll have to be the judge of this. Benito and all our editors work in Premiere.

Benito Gonzalez [02:51:06]:
Wait, so this guy just asked Claude to make the Adobe Suite for him?

Jeff Jarvis [02:51:09]:
Pretty much.

Leo Laporte [02:51:10]:
I suspect it was more than just sitting down and saying, make this.

Jeff Jarvis [02:51:13]:
Yeah, but they used it to make it. And as I watched a video on TikTok today about this, and this goes to your point about software, Leo, your business model of owning software. What's next, guys?

Leo Laporte [02:51:26]:
This is so great. PDFcraft, an open source PDF You don't need Adobe after that.

Jeff Jarvis [02:51:30]:
How will they eat, Leo?

Leo Laporte [02:51:32]:
How will poor Adobe eat? You know, it's actually interesting they chose this company to clone because nobody is defending Adobe at this point. I feel bad, but anyway.

Jeff Jarvis [02:51:48]:
So, I write about the history of Adobe and HotType because I write about the history of PostScript, which Adobe fascinatingly did not patent because they didn't want it revealed because they feared that others would figure out how to do just this.

Leo Laporte [02:52:04]:
Right.

Jeff Jarvis [02:52:05]:
So, for years you had no idea how PostScript operated until finally they let it out. They had to.

Leo Laporte [02:52:10]:
Now, we should mention that these ArtCraft apps are free but they do offer a subscription. It's at getartcraft.com and they do offer a subscription, $8 a month or if you're a Series Creator, $28 a month, and you get additional AI help and so forth.

Jeff Jarvis [02:52:30]:
Because they're using Nano Banana and stuff, so there is a cost to them. So it's not as if this is just some—

Leo Laporte [02:52:36]:
And you know, it'd be very interesting because you get SeaDance credits, so you could use their version of Adobe Premiere to make video. I mean, I think this is— people have been saying it's all over for SaaS.

Benito Gonzalez [02:52:48]:
Um, yeah, this was inevitable, right? This was inevitable.

Leo Laporte [02:52:52]:
This— it's not— it's kind of inevitable that it'd be the first company to fall. We'll see if it impacts Adobe. I think there are a lot of businesses would say, no, no, no, we need to work with a company that's FedRAMP certified or whatever.

Jeff Jarvis [02:53:04]:
But you're— it'd be really interesting to hear what your staff says about this.

Leo Laporte [02:53:07]:
Yeah, but you know, and yeah, I mean, we're gonna still pay for Premiere, right? That we have— we have multiple Creative Cloud subscriptions and it's hundreds of dollars a month. But, uh, yeah, you're not gonna base a business Not yet, but who knows? But who knows? And certainly as an individual, I would be willing to. I will try it, the Lightroom replacement. Uh, and then finally, Normal Tools. I'm just going to give you a bookmark because this is kind of a handy thing to have. normaltools.com. No login, no install, it's free, and it does like 1,000 things. Converters, Readability score checker, image cropper, SVG to PNG, PNG to JPG, password generator, SQL formatter, a roof pitch and—

Jeff Jarvis [02:53:56]:
Drywall sheet calculator.

Leo Laporte [02:53:58]:
Yeah, a roof pitch and rafter calculator.

Jeff Jarvis [02:54:01]:
That's hard.

Leo Laporte [02:54:02]:
It's all, it's just stuff. You know what? You could probably get your AI to do it. And I suspect that's the story behind this because look at all the things in here. Everything, every calculator you'd ever want.

Benito Gonzalez [02:54:16]:
This guy wrote down every idea he ever had and made them.

Leo Laporte [02:54:20]:
Yeah, just said, please make a calculator for it. Here's an online stopwatch. It's simple, it's bare, but it does it, and it's free. And it's probably a good thing for you to bookmark because then it's, you know, you got it.

Jeff Jarvis [02:54:38]:
Normal tools.

Leo Laporte [02:54:39]:
Get it and forget it. normaltools.com. It's time. I've taken up all our pick time for Jeff Jarvis's picks.

Jeff Jarvis [02:54:48]:
Well, I want to mention that our dear friend Gina Trapani, we mentioned her new podcast a while ago, is now— her first episode is out.

Leo Laporte [02:54:54]:
Chris and Gina make great software. Look at her.

Jeff Jarvis [02:54:58]:
Oh, we miss you, Gina. You're great.

Leo Laporte [02:55:01]:
So is it a— it's a developer show?

Jeff Jarvis [02:55:03]:
Yeah.

Leo Laporte [02:55:04]:
Nice.

Jeff Jarvis [02:55:05]:
All right. So the first episode I think is out. Uh, mention for your daughter, I just finished listening to Paul Theroux's new book. I've always loved his travel books and many of his novels, True North, which he travels across Canada. But Léo Laporte, he also talks, he goes to Quebec and finds the roots of, and talks to the roots of his family.

Leo Laporte [02:55:26]:
Sure. He's a Theroux.

Jeff Jarvis [02:55:27]:
As Theroux, right. So I think your daughter, you and your daughter would find it, and Hank too.

Leo Laporte [02:55:34]:
We do know the roots of our family. My dad used to go to Quebec every few years for the big Le Porte de Saint-Georges family reunion.

Jeff Jarvis [02:55:42]:
Do you know what town it was?

Leo Laporte [02:55:43]:
Yeah, it was in Montreal.

Jeff Jarvis [02:55:46]:
Oh, it was in the city?

Leo Laporte [02:55:47]:
Yeah. And it's all the descendants of one woman who came in the 17th century to Canada from France. We're all descended from her.

Jeff Jarvis [02:55:59]:
Wow.

Leo Laporte [02:55:59]:
And somebody did all the work. Baxter Laporte is kind of the guy who puts it all together. I'm very proud to be— to have that heritage.

Jeff Jarvis [02:56:07]:
That's very cool. Very cool. Yeah.

Leo Laporte [02:56:11]:
But I gotta show you one thing. We got Patrick now playing with porches. And remember he said you can check the bus schedule. You can also check the SMART train, which looks just like a little choo-choo. But that's where it is right now. I love it.

Jeff Jarvis [02:56:28]:
Oh, that's great.

Leo Laporte [02:56:29]:
I love it. I should mention, Up on a Bench podcast, the first Hands-On AI is out with Micah Sargent talking about little models you can run on your phone. So if you are not yet a subscriber to that, twit.tv/HoAI. That's a— thank you, Larry, for reminding me. Yes, you have another.

Jeff Jarvis [02:56:46]:
And then I just want to mention one more, which is, I think, a momentous cultural event. The Emmys are moving from primetime to a deal with Moving to Prime Video from prime time.

Leo Laporte [02:56:57]:
Oh, interesting.

Jeff Jarvis [02:56:58]:
So TV doesn't care about TV anymore.

Leo Laporte [02:57:02]:
Didn't the Oscars announce that they're going to move to YouTube in a few years?

Jeff Jarvis [02:57:05]:
Something like that. Yeah, but I think this is amazing. Television is abandoning television.

Leo Laporte [02:57:10]:
That is weird. That the Emmys, which is the National Academy of TV Arts and Sciences, is moving to Amazon Prime. Because what does TV mean anymore, really?

Jeff Jarvis [02:57:24]:
That's exactly— yeah, TV is YouTube. So, and one more time, folks, if you are in New Hampshire or Maine or Vermont or Massachusetts, um, or Connecticut or Rhode Island, come to Haverhill, Mass. Haverhill. Haverhill.

Leo Laporte [02:57:41]:
Haverhill.

Jeff Jarvis [02:57:42]:
No, there's no— no H. Haverhill, Mass.

Leo Laporte [02:57:45]:
Thank you.

Jeff Jarvis [02:57:45]:
I screwed that up.

Leo Laporte [02:57:47]:
Uh, Saturday, uh, Patrick Delahanty, who lives nearby. Yeah, it's Haverhill. Yeah.

Jeff Jarvis [02:57:54]:
Yeah. So I'll be there Saturday giving a book talk and then going to the, to the, uh, at, at, uh, Historic New England and then going up the hill to, uh, the Computer History Museum where, uh, Chris Bradford, the wonderful volunteer, and I will be demonstrating the Linotype.

Leo Laporte [02:58:11]:
Man, I wish I could go. I would love to see you demonstrate the Linotype.

Jeff Jarvis [02:58:18]:
So, so I—

Leo Laporte [02:58:19]:
Is it gonna start it up and make it noise and everything?

Jeff Jarvis [02:58:22]:
Oh, it does all that.

Benito Gonzalez [02:58:23]:
Yeah, yeah, yeah, yeah, yeah.

Jeff Jarvis [02:58:23]:
Can you tick-tock it or something? Yeah, we could do that. I set the back of the book on the linotype. So this is the colophon.

Leo Laporte [02:58:33]:
I love that.

Jeff Jarvis [02:58:34]:
And we set this on the machine.

Leo Laporte [02:58:36]:
Yeah, that's so cool. The first time— Is that the same one? The Havero one?

Jeff Jarvis [02:58:39]:
Yeah, yeah.

Leo Laporte [02:58:39]:
Oh, that's neat.

Jeff Jarvis [02:58:40]:
The first time Chris let me sit down to the machine, uh, and this is a moment for me, right? I couldn't do it when I was a, when I was a young journalist because you Because the union, you couldn't touch them. And now I could touch the machine and almost immediately I broke it. I thought I killed the Linotype. I killed one of the last living Linotypes. But he dug in and found it. I still have the bent matrix.

Benito Gonzalez [02:59:02]:
That's why they didn't let you touch it, Jeff.

Jeff Jarvis [02:59:05]:
I know. Brought it back to life.

Leo Laporte [02:59:08]:
Well, that concludes this edition of Intelligent Machines. Uh, Paris Martineau, of course, at Consumer Reports where she's an investigative reporter and right now having a great time watching the Liberty. I hope they're winning. Liberty, Liberty, Liberty. It's the only WNBA team with their own jingle. Uh, and Jeff Jarvis, Professor Emeritus of Journalistic Innovation.

Jeff Jarvis [02:59:33]:
CEO there too. It's 45 for the Liberty versus 47 for the Atlanta Dream.

Leo Laporte [02:59:38]:
It's a close one. Yep, it's a close one. I bet it's so much fun. So much fun. Uh, Jeff's of course the author of Hot Type, and go up to Montclair, uh, and, and get the signed version, then go to—

Jeff Jarvis [02:59:53]:
Yes, if you go to jeffdrivers.com, you can find the signed version. I signed, uh, 2 dozen of them, so please don't embarrass me and leave a pile there.

Leo Laporte [03:00:00]:
That would be terrible.

Jeff Jarvis [03:00:01]:
That'd be bad.

Leo Laporte [03:00:02]:
I do remember once going to the remainder been in our local Copperfields Book and finding a signed version of one of my books. Uh, it's kind of embarrassing. Uh, it was only 33 cents, so I don't know if I get royalties.

Jeff Jarvis [03:00:17]:
Did you sign it to somebody you loved and they sold it?

Leo Laporte [03:00:19]:
No, no, no, no, no. I signed a bunch of books there.

Jeff Jarvis [03:00:22]:
Okay, good.

Leo Laporte [03:00:23]:
Yeah, big mistake. There's nothing more humbling than seeing your books in the remainder bin. Uh, thank you everybody for joining us. We do I Am every Wednesday, 2 PM Pacific, 5 PM Eastern. That is 2100 UTC. Next week, the Vice President of Developer Relations at GitHub, Martin Woodward. Should be really fun. He's really smart.

Leo Laporte [03:00:48]:
Uh, he, uh, he is, uh, writing, I think he's writing a book called Building with Heart in the Age of AI. Certainly GitHub has been one of the beneficiaries.

Jeff Jarvis [03:00:57]:
Yeah.

Leo Laporte [03:00:58]:
both one of the contributors to AI Smarts and the beneficiary of all those AIs who remember their parentage. He'll be a great guest. I look forward to that. So join us next week. You can watch the show live if you want. If you're in the club, please join the club. You can watch the show as we do it live, not just in the club, but also on Twitch, YouTube, x.com, Facebook, LinkedIn, Kick. And after the fact, download copies of the show at twit.tv/iam.

Leo Laporte [03:01:26]:
There's a YouTube channel dedicated to Intelligent Machines, and of course, best thing to do, subscribe in your favorite podcast client so you don't have to even think about it. It just appear magically on your phone so you can listen whenever you're in the mood for Intelligent Machines. Thanks everybody. We'll see you next time on I Am.

Paris Martineau [03:01:43]:
Bye-bye. I'm not a human being, not into this animal scene.

Leo Laporte [03:01:51]:
I'm an intelligent machine.