Intelligent Machines 888 transcript
Please be advised that this transcript is AI-generated and may not be word-for-word. Time codes refer to the approximate times in the ad-free version of the show.
Leo Laporte [00:00:00]:
It's time for Intelligent Machines. Jeff Jarvis is here. Paris has the week off, but Father Robert Balisar is here, which is a good thing because this has been the week of AI doom. What's going on? Are we at risk? What can the big frontier companies do to save us? Do they need to save us? Are their motivations pure? And a lot of AI news too. Plus, we're going to talk to Lon Seidman of Lon.tv. He's one of my favorite YouTube creators. He'll show you how to make A very inexpensive local LLM machine based on used data center hardware. Intelligent Machines is next.
Leo Laporte [00:00:38]:
Podcasts you love. From people you trust. This is TWIT. This is Intelligent Machines with Geoff Jarvis and Paris Martineau. Episode 888, recorded Wednesday, September 16th, 2026.
Fr. Robert Ballecer, SJ [00:00:56]:
Large Linguini Model.
Leo Laporte [00:00:56]:
It's time for Intelligent Machines. Intelligent Machines, the show where we cover all the latest AI news. There's no AI news this week, so don't worry about that part. Robotics and all the smart little doohickeys all around you. Jeff Jarvis is here, the Emeritus Professor of Journalistic Innovation at the Craig Newmark Graduate School of Journalism at the City University of New York. Author of a fabulous new book called Hot Type, available everywhere.
Lon Seidman [00:01:26]:
The magnificent machine that gave birth to mass And if you go to jeffjarvis.com, you can see a link to the Montclair Book Center where we've now gotten all the act together and there's a big stack of autographed books there. So if you wanna support—
Leo Laporte [00:01:38]:
Oh, nice.
Jeff Jarvis [00:01:38]:
An independent store and buy a signed book, just go to jeffjarvis.com and you will see the link.
Leo Laporte [00:01:43]:
And that's of course the home of Montclair State University where you're also on the team there. Now we can take, Benito, the offer code off the end of that.
Jeff Jarvis [00:01:53]:
Yes. Well, that's right.
Leo Laporte [00:01:54]:
You can't.
Jeff Jarvis [00:01:54]:
It doesn't work anymore.
Leo Laporte [00:01:55]:
That's no good anymore. And it probably is confusing to people. There we go.
Jeff Jarvis [00:01:58]:
There we go. Thank you.
Leo Laporte [00:01:59]:
Goodbye. You know, it's just a Dymo labeler and he just, little snip on the label and it's gone. It's the easiest thing to do. Paris has the week off, but we are so glad to have the wonderful Father Robert Balas here.
Jeff Jarvis [00:02:12]:
Yay!
Leo Laporte [00:02:12]:
And his noodles in-house.
Fr. Robert Ballecer, SJ [00:02:14]:
Oh, sorry, this is our latest LLM. Large Linguine Model. It's, you know, it's a different one.
Leo Laporte [00:02:21]:
Yum, yum, yum. He comes to us from the Vatican, from Rome, where he is very influential, I think, in AI matters, although he'll never take credit for it. It's great to see you, Father Robert, but you are an AI fan. Yes, you—
Fr. Robert Ballecer, SJ [00:02:35]:
I am an AI fan, uh, and I'm a realistic AI fan. I know what it does well, and I know what it does extremely poorly. Yeah.
Leo Laporte [00:02:43]:
Well, I'm glad you're here because we have a guest that both of you know. I've never met, but I'm thrilled to have him on. I've watched his YouTube channel for years. You know, lon.tv. Lon Seidman is here.
Jeff Jarvis [00:02:55]:
Hi, Lon.
Lon Seidman [00:02:56]:
Hi, Leo. Thanks for having me on.
Leo Laporte [00:02:57]:
It's wonderful to have you. You started doing YouTube, what, it was just kind of a lark?
Lon Seidman [00:03:03]:
Yeah, it was mostly a lark. I was, I was actually doing some journalism stuff, which is where I first met Jeff years ago. And the journalism stuff didn't work out too well for what we were putting together, but a YouTube channel hatched out of it. And I've been doing tech reviews and tech commentary and all sorts of stuff.
Jeff Jarvis [00:03:18]:
You were doing local news on YouTube, right?
Lon Seidman [00:03:21]:
Yep. Actually, that's what got me started with YouTube was we were doing local news with video and I was buying up all this video stuff. And I was watching Alex Lindsay's work on those types of things back, I don't know, 20 years ago.
Jeff Jarvis [00:03:33]:
This was early, 2009, right?
Lon Seidman [00:03:34]:
Yeah, actually it was 2009, you mentioned it. Yeah, so it was a while. And I started reviewing the stuff I was getting in for the startup, and that ended up being the thing that worked, not the startup. So I've been doing it ever since. I've been full-time for about 10 years.
Leo Laporte [00:03:48]:
Well, you were on the GizWiz. It's not your first time on TWiT, in 2012 with a 2GS that you had souped up. Is that the one behind you right there?
Lon Seidman [00:03:56]:
Yeah, it's the one behind me. Yep.
Leo Laporte [00:03:58]:
I love it.
Lon Seidman [00:03:58]:
It had an unfortunate RIFA cap explosion a few years ago, but we took care of that. It's all good now. So it's—
Leo Laporte [00:04:03]:
well, you know, my original Mac that, uh, John Slanina left me, that had the same problem. The caps were just in terrible shape, and I think we're not going to get it back, which is too bad because we really enjoyed booting that sucker up and the bong and all that. Lately, I wanted to have Lon on. In fact, I want to do a series of shows for the next few weeks About local AI models, running your own AI models, because I think increasingly that's of interest actually because of the latest, you know, doom talk. It's more than a hobby these days. It may be the best way to keep AI flowing as prices go up and people shut down and things. The fact that you can run an AI, a pretty competent AI at home, is interesting. But it's also expensive.
Leo Laporte [00:04:52]:
And Lon is focused on, I think, the best part of this. So we're going to have other local folks on. We've booked a few of them, uh, people who are posting interesting recipes and interesting builds and stuff. But I think Lon is, to me, one of the most interesting because you're taking kind of old and, and, you know, cheap, relatively cheap, uh, AI GPU cards and repurposing them. In fact, I just watched you do that with an old Tesla card.
Lon Seidman [00:05:20]:
That's right. Yeah, I bought an old V100.
Leo Laporte [00:05:25]:
V100 from like from a, from a data center from like 10 years ago.
Lon Seidman [00:05:29]:
Yep.
Jeff Jarvis [00:05:30]:
Yeah. How much did that cost, if I may ask?
Lon Seidman [00:05:32]:
Well, you know, well, it would have cost a lot less a few months ago, but right now it's about $700.
Leo Laporte [00:05:37]:
Still, that's a lot less. I mean, I run a 3090. I upgraded my— as you've done, I took my gaming rig, which had a 3070 in it, and somebody said, you know, you could put a 3090 in there with 24 gigs of VRAM. and run like Quinn, run a decent model. And I, so I did, and it was well, but it was not cheap. It was like $1,100, but, but it's cool because you can run a pretty good model at a very good speed in there. How does that, how is that V100 working for you though?
Lon Seidman [00:06:06]:
It's a bit, it's a bit of a science experiment. So right now I'm having it work as an eGPU. I've got an OcuLink enclosure and it's a whole big—
Fr. Robert Ballecer, SJ [00:06:14]:
Yeah.
Leo Laporte [00:06:15]:
'Cause you don't, I mean, I don't even know if anybody makes a motherboard with slots for that, right? I mean, it's a data center card.
Lon Seidman [00:06:20]:
Yeah, it'll fit, but it's, it's super long when you put the cooling on it because it doesn't have any active cooling built in.
Leo Laporte [00:06:26]:
It's not a blower. It has no fan. So you've got a 3D-printed mount and a fan you bought on Alibaba, probably, right?
Lon Seidman [00:06:35]:
Yep. It all came from AliExpress. You got to get a special power cable. It's a whole to-do. And, and so I got it like wired into a mini PC and, and it's— but it's working. You know, what's great about it is that it has 32 gigs of RAM. Which is a lot.
Leo Laporte [00:06:49]:
Yeah.
Lon Seidman [00:06:49]:
It's not, it's not as fast as a 5090 or whatever it might be, but it's fast enough. And for me, that was, you know, I have a project I've been working on and it's fast enough for that outcome. I'm running Gemma 31B on it along with Quen 27B. So both dense models, they, they generally crank out about 30, 40-ish tokens per second depending on the context length. And with 32 gigs, you can have pretty decent context. I think I've got like 96,000 tokens of context on it. So for what I'm using it for, it's been great. But you're right, the models are, you know, you got to coax them a little bit.
Leo Laporte [00:07:24]:
Is it CUDA? It's not CUDA, right? It's old enough that it's not CUDA. Or is it?
Lon Seidman [00:07:29]:
I think it is.
Fr. Robert Ballecer, SJ [00:07:30]:
I think it is CUDA. Yeah, V100 uses Tensor Cores.
Leo Laporte [00:07:33]:
So it should be different. That far back they were doing that already. Interesting. But it doesn't— it wasn't for AI in 2017.
Fr. Robert Ballecer, SJ [00:07:42]:
Wasn't— yeah, what it was designed for, which is just large amounts of processing.
Leo Laporte [00:07:46]:
Just—
Fr. Robert Ballecer, SJ [00:07:47]:
yeah, just so you use up what, 250, 300 watts on that, on that card?
Lon Seidman [00:07:52]:
Yep, about 250.
Leo Laporte [00:07:54]:
That's okay. Should point out, that's the downside. So that 3090 is like 85 degrees using 350 watts, the fans are going, and meanwhile the Sparks, which are running on ARM chips and have a lot more capability, are, you know, 35 watts, and they're just kind of I mean, they get hot, but I've, you know, I've got a fan blowing across them too. So, uh, but it's cheaper, $700. That's amazing.
Jeff Jarvis [00:08:19]:
Do you have a picture of the, uh, setup, Lon?
Lon Seidman [00:08:21]:
Oh, you know what, I, I should get you one. I did a video on it.
Leo Laporte [00:08:23]:
I could put the video up.
Jeff Jarvis [00:08:25]:
Yeah, if you put it up, fans would love to see it.
Leo Laporte [00:08:27]:
It's, it's pretty cool. Well, and everybody should go to youtube.com/— is it /Lon?
Lon Seidman [00:08:32]:
Uh, L-O-N-S-E-I-D-M-A-N. Yeah, Lon Seidman.
Leo Laporte [00:08:35]:
That's your whole name? Okay.
Lon Seidman [00:08:36]:
Yeah, my whole name.
Leo Laporte [00:08:36]:
Or go to lon.tv.
Lon Seidman [00:08:39]:
Yes, forward's right over there.
Leo Laporte [00:08:40]:
Yeah, yeah, yeah. So it was, it was a science experiment.
Lon Seidman [00:08:43]:
And one of the fun things with it is that I'm using Frontier models to configure it. So I did a, you know, whole configuration video. And I just said to, you know, ChatGPT's command line, hey, make this work.
Leo Laporte [00:08:57]:
And honestly, it did. That's the only way to travel. Why not? You know, these things are really good at that kind of thing. And they will just walk you through the whole process. It's kind of amazing.
Lon Seidman [00:09:10]:
It really is. And I've got another one with an Intel card, a B70, and did the same thing with. So it's been really fun to get it up and running. I've got a project that I'm working on, kind of a— I'm developing a local chief of staff for my life. And this is what's running on it. So it keeps track. I can go into more detail on it if you have time, but it's keeping track of every ball that I have in the air. And I found for what I'm doing with it, it actually— the local models are more than adequate.
Lon Seidman [00:09:38]:
It's been really smart, and surprisingly so, especially with Gemma 31B.
Leo Laporte [00:09:43]:
Yeah, those are pretty decent models in there. Do you have— it's a dense model, though. You can't, right?
Lon Seidman [00:09:51]:
Yeah, yeah. Gemma 31B is dense. I did try some of the mixture of experts models like 24, 27B, A4B, whichever one Google has. And it's a lot faster, but it It doesn't have the smarts. So one of the things that I have my chief of staff doing is monitoring email threads. And when things fire off, it needs to know what task I have on hand to adjust, what to add to it as far as what I need to do next as the next step. And I found that the denser models are much, much smarter and not screwing up my kind of my to-do list as much.
Leo Laporte [00:10:25]:
Jeff, here's the Tesla card with the 3D-printed mount and the fan hanging off its back. Behind. And then there's the dock. So he's going to put that in the, in the dock. And, uh, you have— it takes a lot of power, 350 watts to run that thing. Uh, does it get pretty hot?
Lon Seidman [00:10:43]:
Surprisingly, no. Believe it or not, the Intel card at least is hotter to the touch than that one is because they're both just hanging out in the open. Um, 'cause I have another mini PC on the Intel card. Um, so yeah, it's, it's not too hot. Um, it does kind of go down to about 50 watts or so idle. So it's not blowing 250 all the time. So really when you're hitting it is when it, when it starts to heat up. But I've been surprised it's been working as good as it has.
Leo Laporte [00:11:07]:
46 tokens per second. It's just whizzing along. It's great.
Fr. Robert Ballecer, SJ [00:11:10]:
Yeah.
Lon Seidman [00:11:11]:
It slows down as the context fills up, which is obviously normal, but it's, it's been great. And for what I'm using it for, it's, it's been remarkably good. I've got almost— I have a local RAG model I'm generating now also because for my, my school board stuff, I'm on the school board here in Connecticut and I have, um, about a year and a half worth of transcripts from meetings. And so pretty much every word that's been uttered in any of the meetings that I've been in is, is now in my database. And when I go into a meeting now, I have it prepare me a briefing. And it's been—
Jeff Jarvis [00:11:43]:
We gotta talk about that. I'm trying to do, I'm trying to do things for towns in New Jersey. I'm on the board of a newspaper company here now, and I wanna figure— and with Montclair State, I really wanna figure out how to, how to use all that material and data in more, in more Yeah. Padre, what was the first thing you ran a model on?
Fr. Robert Ballecer, SJ [00:12:00]:
Uh, well, actually a bunch of laptops. That was the cheapest ways for us to get GPUs over here were to get some of the gaming laptops. So we had it running on a 3090. Our new ones are 5070. We got a series of 5070-equipped laptops for about €800 each, which is cheaper than buying the GPU separately.
Leo Laporte [00:12:22]:
Yeah.
Fr. Robert Ballecer, SJ [00:12:22]:
So we did a lot of testing on that. But of course, now we've moved over to Sparks.
Leo Laporte [00:12:27]:
And how many Sparks do you have? A few?
Fr. Robert Ballecer, SJ [00:12:33]:
8, I think.
Leo Laporte [00:12:34]:
Are they all chained together?
Fr. Robert Ballecer, SJ [00:12:37]:
No, because we do have a couple of— we've got 2 that are in production full-time running the LLMs that we have running at the Portia.
Leo Laporte [00:12:43]:
Yeah, that's what I have with a ConnexX-7 cable. And then we've got a bunch of— you can, in theory, I think, get as many as 8, but you'd have to get a special switch.
Fr. Robert Ballecer, SJ [00:12:54]:
And which is what we've done that, so we can tie them together. But, uh, at the moment, we are— we have broken off a couple of units because we've got some special units that we're training different models on, and, uh, we would like them not to be part of the cluster.
Leo Laporte [00:13:08]:
That's one of the most interesting things. I think that's— and what you're doing too, Lon, is, is to take a model and then to train it to do exactly what you want. Quent is not the smartest model ever, but for what you just described, it's perfectly adequate. In fact, you show in that video, it's analyzing a couple of PDFs, one of which your visit to the FCC for the antenna thing. And you could query it, did all the things you want to do. I mean, that's perfect, right? And no hallucinations either when you're doing that, which is nice.
Lon Seidman [00:13:40]:
No, it's been really good and grounded. And what I say, what I've been really fascinated about with these local models is that for what, like, the general consumer uses AI for, this is probably 85% of it.
Leo Laporte [00:13:50]:
Yeah.
Lon Seidman [00:13:51]:
You know, proofreading, summarizing a document, summarization, it's really good at. If you keep it focused on a very specific set of data, it does very well. I have found, though, that Gemma 31B is slightly more accurate than Quen is. So even though Gemma is the older model, for what I'm using it for, less code, more analysis, Gemma has been a little better.
Fr. Robert Ballecer, SJ [00:14:13]:
We have sacrificed smarts for smarts. So basically, by taking a model and dumbing it down and only training it on information that we're going to be using, it is less capable. But what we ask it to do, it does much, much better, and it hallucinates much less. So I— that's probably going to be the way forward. Now, Leo, I'm a little disappointed. You said you didn't like Quen because we— I was gonna start experimenting with it.
Leo Laporte [00:14:37]:
We'll try. What I spend— what I'm doing mostly right now is, as everybody who's listening knows, is rewriting this 15-year-old code that we use to keep track of our sales, our ad sales system. And it's all done in SQL and it's pretty old and a little clunky. And we've been spending a lot of money to keep it running. So we're rewriting it in Go and it's doing a great job, but it's slow work, but it's high quality. And I've— it's also for me a research project because I'm learning The only way really to learn about what the capabilities of these things are, what they can and cannot do, is doing it. You just have to do it.
Lon Seidman [00:15:19]:
Yeah.
Leo Laporte [00:15:20]:
But when I'm not doing that, one of the things I spend way too much time doing is what we, and I say we, I'm talking about me and my buddy over here, Quicksilver, call bake-offs, where I'm always monitoring new local models. This is why Hugging Face is so great. There's always New versions of Qwen, new versions, you know, DeepSeek V4.1 Flash just came out, which is a very interesting model architecture. Qwen 3.8 Flash Next, which I was playing with when it created this summary, just recently came out and is the new architecture that Alibaba is going to use for Qwen 4 and later. So, I'm often doing these bake-offs, and what I've done, and I think it's important to do, is I've created a benchmark. I don't use standard benchmarks. I create benchmarks out of my own work. So we have a lot of— having done this project for a few months, a lot of dead ends, alleyways, things that came up as problems that we challenge these models with.
Leo Laporte [00:16:21]:
I have a 30-question test harness, and so I have a pretty good idea of how the model is gonna handle the kind of work I'm doing. And that's, I think, one of the things you really want to do. Just what you're doing, Lon, is tailoring it to fit your exact needs.
Lon Seidman [00:16:43]:
Exactly. And I've spent so much time with software that I'm trying to comply with and wrap my life around, and now I'm at a point where I can write my own software that does things exactly the way I want.
Leo Laporte [00:16:53]:
Isn't that nice?
Lon Seidman [00:16:53]:
And that's what's— it's wonderful. I mean, every time I have an idea now, it's executed. That's a revolution. Yeah, I think it is. And that's why I say to people like, look, you know, if you have an idea, you know, you can just get it done. I've got so many things that just kind of run on Frontier models now, monitoring news and all this other stuff, and it's all starting to get tied together. The best part with Local is that it is free to try as long as you have the $800 GPU or the $5,000.
Leo Laporte [00:17:16]:
Well, how much all in was that setup? I mean, you did have to buy a computer to go with it.
Lon Seidman [00:17:20]:
Yeah, I would say if you were to put it all together, it's, you know, it's probably— I bought the computer before the prices went through the roof, so You know, actually that one came in for a review, but either way, it's probably about $1,500 from the portables.
Leo Laporte [00:17:33]:
A regular PC. You don't need a special PC, right?
Lon Seidman [00:17:36]:
No.
Leo Laporte [00:17:36]:
And then what's the breakout box for the GPU, external GPU?
Lon Seidman [00:17:40]:
How much is it? It's about $175. I paid about $175.
Leo Laporte [00:17:42]:
Well, that's not bad.
Lon Seidman [00:17:43]:
Yeah, that's an Oculus box.
Leo Laporte [00:17:44]:
So the cost of the card is really the thing you're keeping down. If you can keep that down, you're gonna keep the whole thing down.
Lon Seidman [00:17:49]:
And where I got into this was I was looking at what could I do with what I already have? So my MacBook Pro and M1 Max, you know, that's where I first said, oh, these local models are usable. 'Cause I had enough RAM to start playing Right. But my old gaming PC is more than capable of taking that card if I could fit it in there. And that one's 6 years old. It could still hold its own, you know?
Leo Laporte [00:18:09]:
Basically, I can't play Valheim anymore, but I do have Quen. Right.
Lon Seidman [00:18:13]:
To run the models on it. It's all about the video RAM.
Leo Laporte [00:18:16]:
And I do use, by the way, Father Ebert, I don't have anything against Quen. I have 2 versions of Quen running, one running on the gaming rig, and that is used for vision. It's attached to all my cameras. So it gives me a text description and it'll recognize me. So say Leo's at the door or Lisa's at the door, recognize, uh, you know, 10 or so people that, uh, it knows. Uh, so it did that for vision. And also it's used as an auxiliary model for Hermes, the agent I use. There are, you know, Hermes has a main model, that's GLM-53 Flash, but it also has auxiliary models for, for instance, compression.
Leo Laporte [00:18:51]:
When the context gets too full, you have to compress it. Quen is very, very fast at something that isn't a very hard thing to do. So it— instead of— you don't want GLM doing it because it's its main model. So it has— Quen just kind of jumps in and says, let me squeeze that down for you. And then I use one on the Mac, a version of Quen for titling of my window. Little things like that. And for research projects, pretty good at going out and doing tool calling. So Yeah, I did put up a page on my garden of pages if you want to see the results of a variety of bake-offs and different recipes.
Leo Laporte [00:19:32]:
And you could see here GLM with this particular recipe is right now the strongest in terms of the agentic benchmarking that I'm doing. I did have Quinn make this page.
Jeff Jarvis [00:19:43]:
Quinn said not to get into a bake-off, Leo. I think you're violating Quinn's dictates.
Leo Laporte [00:19:47]:
No bake-offs. No bake-offs allowed. I waste too much time on this.
Fr. Robert Ballecer, SJ [00:19:51]:
It has really good multi-language support, which is what we, we need to develop right now.
Leo Laporte [00:19:55]:
Oh yeah.
Fr. Robert Ballecer, SJ [00:19:55]:
We've got an event coming up next year where— It's Chinese.
Leo Laporte [00:19:58]:
So it speaks Chinese perfectly. In fact, if it hallucinates, sometimes it will speak Chinese to you. Have you noticed that?
Fr. Robert Ballecer, SJ [00:20:05]:
It will just sound poetic.
Leo Laporte [00:20:08]:
So that's interesting. Does it speak Italian pretty well?
Fr. Robert Ballecer, SJ [00:20:12]:
Uh, actually, yes.
Leo Laporte [00:20:13]:
Italian, Spanish, French, German, English, Do you have to tell it that this is French coming at you or it just knows?
Fr. Robert Ballecer, SJ [00:20:21]:
No, it just knows.
Leo Laporte [00:20:22]:
Isn't that amazing?
Jeff Jarvis [00:20:23]:
Wow.
Fr. Robert Ballecer, SJ [00:20:23]:
I mean, you can develop the model so that you can specifically tell it what language is incoming, but if you don't mind wasting a little bit of processing time, it can just auto-detect.
Leo Laporte [00:20:33]:
Yeah.
Fr. Robert Ballecer, SJ [00:20:33]:
Wow.
Leo Laporte [00:20:34]:
So what, Lon, tell me about, do you have other models and other machines running or is the Tesla the main one?
Lon Seidman [00:20:41]:
So what I've been doing is A/B testing Quinn versus Gemma 4. So I've got One, one machine with the NVIDIA running Quen and then the, the other one's running Gemma, which I have running with the Intel B70 card. And it's been interesting to see how they perform. They, you know, performance-wise, they're about the same from token generation and prefill and all that. But as far as getting the job done for my chief of staff, the application I'm developing, that Gemma seems to be doing better. So at some point I'll cut the other one off. I do have a Mac Studio on the way, so I'll probably—
Leo Laporte [00:21:15]:
Oh, you ordered one of those, did you?
Lon Seidman [00:21:16]:
I did. I couldn't help it. I needed a new editing machine, and you know what, I could run— I have so much RAM on it, I can run the model in the background and still get my other stuff done for less time.
Leo Laporte [00:21:26]:
Exactly, get the best of both worlds.
Lon Seidman [00:21:27]:
Yeah, yeah. So, so that's the, the next step with it, is to try to get it— get this under control because it's getting a little crazy.
Jeff Jarvis [00:21:32]:
Lon, what motivated you?
Fr. Robert Ballecer, SJ [00:21:33]:
You don't have to make excuses here for wanting new hardware.
Lon Seidman [00:21:36]:
Yeah, I won't fall down the stairs, that's it.
Leo Laporte [00:21:38]:
But this is one of the reasons I want to have Lon on, because he's not going out and buying the high-priced stuff. He's, he's working with what You know, I mean, where did you find a Tesla V100? Where'd you find that?
Lon Seidman [00:21:50]:
So ServerPartsDeals has them.
Leo Laporte [00:21:52]:
ServerPartsDeals.com?
Lon Seidman [00:21:55]:
Yeah, that's the place you used to be able to buy like cheap hard drives over there. But they, yeah, they get in all this like off-lease server equipment and they refurbish it. So I went with them just because I figured I'm not going to get as much junk from somebody who's actually maybe evaluated the card or at least could provide a refund or something. So that's where I went.
Jeff Jarvis [00:22:14]:
You could probably find— Tesla V100 is $719. $60,000 right now.
Leo Laporte [00:22:17]:
$60,000 and some. Amazing. Yeah, because there's not probably a lot of demand. It takes a little bit of, of skill to do what you did. Of course, you could watch lon.tv and you'd, you'd have everything you needed to know.
Lon Seidman [00:22:29]:
I was surprised it worked as quick as it did.
Leo Laporte [00:22:32]:
Yeah, so GPT-5.6 was able to, to kind of walk you through it. Do you use Astra to do that, or—
Lon Seidman [00:22:38]:
I think I was still on Sol when I, when I did that. Yeah, and it, it just went to town on it. Um, it did a great job with the Intel card also, which needs, uh, special you know, special drivers and Intel love to get that working properly. So the one thing that the Tesla card doesn't do as well is image and video generation. It can do some of it with some models, but, you know, like Minimax H3 is like the latest video generation model that just bombs on that card. It just can't handle it.
Leo Laporte [00:23:06]:
It won't even run.
Lon Seidman [00:23:07]:
Yeah, it just sits there and spins. But the other one was it LTX or what?
Leo Laporte [00:23:15]:
I think you use Minimax H3. No, no, that wasn't Minimax. You showed you have, you have your husky riding a motorcycle was one of the images I saw. Yeah.
Lon Seidman [00:23:25]:
So some of that stuff will work on the image generation side, just slower. But the video models get trickier. So one worked slowly. The other one didn't work at all. The Intel card actually did pretty well. And the Intel B70, although it's gone up in price, it's about $1,300. 32 gigs, and Intel's been supporting the card. It's not running with CUDA cores.
Lon Seidman [00:23:46]:
There are some software things that make it work, but it's an option if you needed 32 gigs and didn't want to go and adventure with data center parts.
Leo Laporte [00:23:55]:
Now you say it's slow, but it generated this 15-second or 10-second clip in 3 minutes.
Lon Seidman [00:24:00]:
That's pretty—
Leo Laporte [00:24:01]:
that feels quick to me.
Lon Seidman [00:24:03]:
It is pretty quick. It's twice as fast on the Intel card for the same length of time, which is crazy. And, you know, the quality of the local video generation models is absurdly good.
Jeff Jarvis [00:24:13]:
I didn't realize.
Leo Laporte [00:24:14]:
It's so good.
Lon Seidman [00:24:16]:
And Minimax looks even better. It's insane. Minimax has a weird, weird licensing requirement. They don't let you use it in the US. And I think it's partly because there's a bunch of copyrighted material in its training model. You can summon Captain Picard whenever you want. So it's that one. is probably a little, you know, a little sketchier, but the quality is unbelievable.
Leo Laporte [00:24:38]:
I've run MiniMax on my H3 on my Mac, so you'll be able to run it just fine on your Mac. And yeah, the quality is quite good.
Lon Seidman [00:24:46]:
Yeah, it's amazing.
Leo Laporte [00:24:47]:
It really is kind of amazing what you can do now. And I encourage people to go look at lon.tv, go look at how you could, for under $2,000, less than, you know, a regular PC, Build a device like this and play with it, because you learn— you've got to play with it to learn about it. I, I don't—
Jeff Jarvis [00:25:08]:
and I—
Leo Laporte [00:25:08]:
and especially, well, I don't know. I mean, I still use all the Frontier models. In fact, I, you know, Opus does all my sysadmin stuff, right? Because it's so— these models are very good and very fast at getting stuff like that done. So in a way, you're You're still going to be using some cloud models even to set up the Tesla. But once you've got it set up, it's pretty cool what you can do.
Fr. Robert Ballecer, SJ [00:25:32]:
Yeah.
Leo Laporte [00:25:32]:
It's pretty amazing.
Lon Seidman [00:25:33]:
And the tokens are free after that.
Leo Laporte [00:25:35]:
And after that, they're free. Exactly. Yeah.
Jeff Jarvis [00:25:37]:
So, Lon, I'm curious.
Lon Seidman [00:25:38]:
As far as it can go.
Jeff Jarvis [00:25:39]:
What was, what was the motive behind doing all this work? Is it philosophical about local models? Is it the fun of tinkering? What led you to put all this work into these local models?
Lon Seidman [00:25:51]:
I would say all of the above. Um, I, I think there's, you know, from a, you know, from an altruistic standpoint, I, I think there's tremendous, uh, opportunity and, and, uh, optimism in my mind for, for having the ability to, to run something yourself on your own hardware and, and have something that can really make your life better and, and be an excellent tool, not a crutch, but a tool to make, uh, things more efficient in, in your workflow. Um, so that was one motivation. I just like being able to, and I do this on the channel quite a bit, is try to find the lower-cost way to do something. You know, in my space, you know, you've got a lot of reviewers who love the expensive stuff, which is fine. I'm totally cool with expensive stuff too. But there's a lot you can squeeze out of things that cost less, especially things you already have. And that's what really fascinated me initially when I started playing around on the Mac with some of these local models.
Lon Seidman [00:26:41]:
And, you know, to the point where these models have gotten so good that it could do a lot for folks and they don't have to give up their privacy to do it. And I think that's, that's pretty important. So So now my next step is, you know, obviously getting the data center parts, but to start looking at developing my chief of staff here, which has been my latest project, and it's been coming together really nicely. Astra has been really good on the development side for that, but it works with the local model to actually do all the processing. So all of my personal information is staying local and it's running on these models. So it's exciting. I think it's good. I think there's opportunity, especially for, you know, for Google here, that they've got a very competitive model.
Lon Seidman [00:27:24]:
They seem to be one of the only US companies that's really focused on having a good local model in addition to Frontier. And I think there's, there's risk for the US for us not to be focusing on local.
Jeff Jarvis [00:27:35]:
Really important point.
Leo Laporte [00:27:36]:
Yeah, I agree.
Fr. Robert Ballecer, SJ [00:27:38]:
I'm sorry, what was the token throughput that you're getting on the V100?
Lon Seidman [00:27:42]:
It depends on the model. I'm generally getting like 30-ish, I think.
Fr. Robert Ballecer, SJ [00:27:47]:
Not bad.
Lon Seidman [00:27:47]:
Yeah, but I had to install the SYCL version of Llama CPP. So before that, it was super slow. So that was something that ChatGPT helped with too, was getting the right combination of software installed. There's a little bit more of a lift. Intel on Windows has this AI Playground, and it can install— I think it installs just about everything. It does ComfyUI for the image and video generation. It installs, I think, Llama CPP. And it gives you a one-click install to just get started with Intel hardware.
Lon Seidman [00:28:18]:
And that was surprising how good that card worked. I bought it on a whim just to see what would happen. It's good content even if it fails, but it turned out to be a useful tool. It's a nice card.
Fr. Robert Ballecer, SJ [00:28:28]:
Yeah, I mean, at that speed, you're within the, like the RTX 4060 territory, just about on equal performance. So then that's not bad. And actually that uses less power than a 4060.
Lon Seidman [00:28:40]:
It does. Yeah. And it does trail off when you start filling up context. But, you know, for me, it's just, hey, is it, is it good enough? And, and that's where, that's where it kind of landed with me.
Leo Laporte [00:28:49]:
There's a— it's really exciting. There's a whole— we're back to the garage, which is where we all started, you know, many years ago. And it's, it's awesome that if you didn't grow up in that era, you know, you weren't around when computing was first getting started, good news. You can, you can get— there's a new era starting.
Fr. Robert Ballecer, SJ [00:29:09]:
Computer fair.
Leo Laporte [00:29:10]:
And there's a whole new opportunity to go back to the garage and do your own thing. And a great place to start is lon.tv, L-O-N.tv. Lon Seidman does a whole bunch of videos and lately a lot of local AI stuff, but it's all fun. I mean, everybody who watches— in fact, I noticed when your sponsors scroll by, a number of the people who are in our club are also members of, of your, your club. So, uh, all sorts of things in here. The HD Home Run, love that. Um, do you ever buy hard drives at that server parts place?
Lon Seidman [00:29:45]:
I, I haven't yet. I ended up having—
Leo Laporte [00:29:46]:
It makes me nervous. Yeah, but with the prices of hard drives, I mean, I'm really worried if I lose one drive on my NAS, I'm gonna have to go there and get a used hard drive.
Lon Seidman [00:29:56]:
Right, get something. I've heard good things. I have, I have yet to have any direct experience, but I've heard good things that the drives are pretty well tested.
Leo Laporte [00:30:02]:
And, uh, the other thing I love about you, Lon, is you begin every video with a disclaimer that's pretty much usually, I bought this all with my own money. And if it's not, if it's something on loan or something, you'll tell people. And I think that's, that's, that's the ethos I like to see. I like to, like to encourage that. So hey, I really appreciate what you're doing, Lon. Thank you for being here.
Lon Seidman [00:30:23]:
Thanks for having me. It's been great.
Leo Laporte [00:30:24]:
Lon.tv. Lon Seidman. Have a, have a great Good to see you, Lon. And we'll have more. I have a board meeting. Yeah.
Lon Seidman [00:30:31]:
I'll see you later.
Leo Laporte [00:30:32]:
We'll have more. Go to the board meeting. We'll have more with Father Robert Balassare on intelligent machines. You can find out more about what the Vatican's up to in just a little bit. Claude was my first real experience, I think, of—
Jeff Jarvis [00:30:45]:
It was your first love.
Leo Laporte [00:30:47]:
It was my first love. When I started using Claude, even today, every time I start up Claude Code, I like to start it fresh every morning with a new session. It says, good morning, here's what's coming up. It has all— it had all the memory from all this since I've been using it all year. Ever since November 24th.
Fr. Robert Ballecer, SJ [00:31:06]:
I think everyone's going to have that model that made LLMs click. They're finally like, oh, okay, that's the big thing.
Leo Laporte [00:31:12]:
It was Opus 4.5 for me. And I talk about it all the time. November 24th, 2025. That's when I went, okay, it's more than autocorrect. There's something going on here.
Jeff Jarvis [00:31:23]:
More than a calculator.
Leo Laporte [00:31:25]:
More than a parlor trick. You remember I was calling it that for a long time. Well, this was a very difficult week for those of us who love AI.
Fr. Robert Ballecer, SJ [00:31:36]:
A little bit, yeah.
Leo Laporte [00:31:39]:
And I'm going to be honest, I don't know what to think. I really don't. My instinct, and I don't trust it because it is my instinct, is to defend AI and say it's not going to destroy us all. There isn't a 10% chance of AI killing all humans in the next decade. And I say it, though I know when I say it, I say it out of my heart, like, because I love it and I love using it. But I often wonder, what are the people who are deep inside of these companies like Anthropic and OpenAI, what are they seeing that's scaring us? It all started with that tweet. from Jacob Coxon, who had worked for, I think, 6 months at OpenAI.
Jeff Jarvis [00:32:25]:
Well, with that much experience.
Leo Laporte [00:32:27]:
And then went to Anthropic, worked there for a couple of months, didn't vest. If he'd stayed 2 more months, he would've vested.
Jeff Jarvis [00:32:35]:
In some amount, wouldn't have been the full amount.
Leo Laporte [00:32:37]:
Well, he wouldn't. Who knows what he was getting paid?
Jeff Jarvis [00:32:38]:
It's a vesting schedule. So you just be curious.
Leo Laporte [00:32:40]:
He resigned on Tuesday in a very public way on Twitter. got a lot of attention. He has since been interviewed on every news show, saying that neither OpenAI nor Anthropic, either of his employers, was acting responsibly. He said the 2 AI labs are sprinting to build superhuman systems that can hack anything, revolutionize any field overnight, and acquire real power and resources. I think the thing that scared people the most was him saying, I think, and many of the people I work at— worked with at these companies believe, there is a 10% chance that AI will destroy all humanity.
Fr. Robert Ballecer, SJ [00:33:22]:
I think it's higher than that, actually.
Leo Laporte [00:33:26]:
Do you? Yeah.
Fr. Robert Ballecer, SJ [00:33:27]:
You have to understand that—
Leo Laporte [00:33:30]:
I don't want to poo-poo it.
Fr. Robert Ballecer, SJ [00:33:32]:
I don't want to poo-poo it, but here's the— here's— so here's the, the math behind that. Most of Cybersecurity is based on 2 inertial factors: scale and time. You need to have a lot of time to be able to launch attacks on things that you want to destroy, that you want to manipulate, and you need to have a lot of resources. The models that are being released— I mean, Mythos should have been the first real red flag. Scale and time is no longer a thing. It's just computing power. If you have compute power, you now have a substitute for scale and time. And when you can substitute for scale and time, it is only a matter of machine time before we have offensive systems that will do their master's bidding without ever taking a rest, without ever considering the political ramifications, without ever looking to see if its goal is still relevant.
Fr. Robert Ballecer, SJ [00:34:30]:
That's not science fiction anymore. I mean, we've seen it. We've seen it with rogue AIs breaking out. It's a speed.
Leo Laporte [00:34:35]:
Yeah. Okay. So a company named Volnchek, we talked about this on Security Now yesterday, that's how I know about it.
Lon Seidman [00:34:43]:
Mm-hmm.
Leo Laporte [00:34:44]:
Reported on the status of Project Glasswing. This was when Mythos was announced by Anthropic. They said, it's so good, we're not gonna— we don't wanna put it out in public 'cause bad guys will start using it immediately. Actually, that's obviously true. that it was responsible to give it to a handful of companies, about 50 companies, and say, before we release anything like this publicly, you should fix your flaws. Well, 2 weeks ago, Microsoft released its— on the 2nd Tuesday of the month, Patch Tuesday— with 975 fixes. We don't know what model they're using. They have their own harness, but I'm pretty sure it's Mythos.
Leo Laporte [00:35:27]:
They were one of the companies that was given it in Project Glasswing. So—
Jeff Jarvis [00:35:32]:
But isn't that good news?
Leo Laporte [00:35:34]:
Well, wait a minute, because let me give you the numbers. Anthropic published its vulnerability disclosure ledger. It gave— it published one in May and it published one this past week. And Volchek went through it And so this is, this is the thing that's a little scary. 26,153 findings. Of that, though, and here's the graphic, Benito, if you'll pull it up. This is from the Volchek article. 2,736 were in the ledger.
Leo Laporte [00:36:15]:
202 out of those 26,000 were fixed. 245 withdrawn, 2% duplicate, 2,000 disclosed, reported to the maintainer. But only 202 were fixed in those 6 months, which makes me wonder, is it underreport? Maybe it's just underreported.
Jeff Jarvis [00:36:41]:
Maybe it was.
Fr. Robert Ballecer, SJ [00:36:43]:
I honestly don't think it is. I, I think we're still operating in the pre-AI mindset of security, which is—
Leo Laporte [00:36:50]:
So people aren't fixing it. They're learning of these vulnerabilities and they're not fixing them.
Fr. Robert Ballecer, SJ [00:36:54]:
Correct.
Leo Laporte [00:36:54]:
That's bad.
Jeff Jarvis [00:36:55]:
Well, are they— are they saying— well, are they judging that they're not real vulnerabilities? Are they saying that this is— I mean, they throw out a number saying we found these vulnerabilities, but how much of that is being checked?
Fr. Robert Ballecer, SJ [00:37:07]:
When I was still with—
Leo Laporte [00:37:09]:
doing Twyot, This Week in Enterprise Tech.
Fr. Robert Ballecer, SJ [00:37:13]:
Yeah, This Week in Enterprise Tech. We had a roundtable with a bunch of, uh, of CFOs and CTOs, uh, and they all agreed that they don't fix things until they break. They don't. It's just there's no— there's no motivation unless there's actually a problem. So if you've got an IT person telling you, well, we have to take these servers down so we can properly patch them, they'll say no, absolutely not.
Leo Laporte [00:37:37]:
Well, we know, you know, there's a good example, which is Salt Typhoon, the Chinese invasion of our phone system. They're still there because the phone companies aren't willing to spend the time, the money to fix it.
Fr. Robert Ballecer, SJ [00:37:50]:
Well, the PCI system that we use in the United States, they've just started modernizing it in the last 5 years, and it's going to take another 10, 15 years before they get to all the terminals. So you've got the core financial system that runs the economy of the United States that is vulnerable to to exploits before AI.
Leo Laporte [00:38:09]:
But Robert, you know perfectly well there's no such thing as security through obscurity.
Fr. Robert Ballecer, SJ [00:38:14]:
Oh, absolutely.
Jeff Jarvis [00:38:15]:
No.
Leo Laporte [00:38:15]:
So not knowing about these doesn't make them go away. So it seems to me it's a good thing to know about them. Obviously, people should start fixing them. But this doesn't mean the end of life on Earth.
Fr. Robert Ballecer, SJ [00:38:31]:
Well, Okay, see, I think people are— they're interjecting the Terminator 2 Skynet scenario.
Leo Laporte [00:38:38]:
That's my feeling, because there's too much science fiction. And that's not—
Fr. Robert Ballecer, SJ [00:38:42]:
yeah.
Leo Laporte [00:38:43]:
Somebody pointed out that everybody who works at these companies started to work at these companies because they grew up on all of these movies and these sci-fi stories of artificial intelligence. That's in their brain, and they are scaring themselves. They are They believe this. They are what, Jeff, you always call test real. They're effective altruists for the most part, who go in believing that AI is going to destroy us. I mean, that's the ground truth that they start with.
Jeff Jarvis [00:39:12]:
So I want to hear what Father Robert was going to say, and then I'm going to respond. I want to respond to that.
Fr. Robert Ballecer, SJ [00:39:17]:
The issue with that is that's not really the scenario that they're worried about. The scenario that they're worried about is, look where LLMs would make the most sense to implement.
Lon Seidman [00:39:28]:
In—
Fr. Robert Ballecer, SJ [00:39:28]:
let's just take one country, in the United States. Financial systems, absolutely. If you could have an intelligent system that is trading faster than everybody else and can pattern recognize much faster than everybody else, that's absolutely one of the first things that they're going to do. Logistics, controlling the flow of resources across the country, everything from food to power to water, that's going to be controlled by AI because anytime you can do just-in-time delivery of resources, you always increase your profit margin. So in just those 2 systems alone, If you have AIs that start to misbehave, as we have seen them misbehave, and start to hallucinate and start to go beyond scope, which they have proven they can do, that destroys a country. You flash crash our economy, it may not come back. You flash crash our electrical grid and our logistics support, it can take months to bring it back.
Leo Laporte [00:40:22]:
That's what—
Fr. Robert Ballecer, SJ [00:40:23]:
We're not talking about nuclear Armageddon. We're talking about AI is going beyond scope and screwing up.
Leo Laporte [00:40:30]:
Well, we, in fact, we had Dan No Doubt on a few weeks ago before all of this who said exactly that. He said that the real risk here is the infrastructure. And so people would— look, some people would survive. We're not eliminating humanity in that kind of Skynet fashion.
Fr. Robert Ballecer, SJ [00:40:49]:
No, but you eliminate society.
Leo Laporte [00:40:50]:
But it's a disaster.
Fr. Robert Ballecer, SJ [00:40:51]:
Yeah.
Jeff Jarvis [00:40:53]:
But so, all right, I want to go back to bringing, especially because we have Padre here, religion into this. I think what you're raising is practical concerns.
Fr. Robert Ballecer, SJ [00:41:04]:
Correct.
Jeff Jarvis [00:41:04]:
But that's not where the discussion really is. The discussion is that in the TESGREAL world, transhumanism, extropianism, rationalism, so on, long-termism, effective altruism.
Leo Laporte [00:41:15]:
By the way, Jacob Coxon said he's a rationalist.
Jeff Jarvis [00:41:18]:
He's—
Leo Laporte [00:41:18]:
Absolutely admitted to being—
Jeff Jarvis [00:41:19]:
So many of them come from that, which means that when they start talking about P-doom in that way, they're not talking about today. They're talking about this mythical future. And so I've got to discount them and their fears because they have mangled the definition of safety. That safety is not about the environment and education and culture and children and financial systems and infrastructure systems. Safety, in their view, is the paperclip scenario and Nick Bostrom. And the problem is when you go into all that, the reason I bring up Tusk Real all the time, and I get so frustrated when media do not. The Times did a story about this event where Bernie Sanders and Steve Bannon came together at Future of Life Institute as a not-for-profit. It's a not-for-profit that comes from these doomsters' money.
Jeff Jarvis [00:42:05]:
It's from Tusk Real effective altruism money. A lot of this stuff comes from that, both OpenAI. And Anthropic, as well as SpaceX, have people inside, and DeepMind, have people inside who are of this view. And so we've got a different set of constraints.
Leo Laporte [00:42:21]:
We talk about this a lot, Geoff. Can you kind of just summarize what is it that they believe?
Jeff Jarvis [00:42:30]:
So the best place to go for this, the acronym comes from Timnit Gebru and Emile Taurus. And if you Google TESS-CREALE, you'll find their first Monday paper that's the best description of this. What's really frightening about this is that when you pull back to the roots of these beliefs that we owe our future to— Nick Bostrom says we owe the 10 to the 54th future human beings, including machine human beings, of the future, that anything today is trivial by comparison, including the Holocaust. Boston was the one that came up with the paperclip scenario, in which if you give the machine the order to build paperclips, we'll all be turned into paperclips. Effective altruism was this belief that you could use money to better ends, but then they decided that the better ends were the things they decided for that far future, not mosquito nets for today. It involves moving to other planets. It involves the singularity and putting chips in your head. There's a lot of Elon Musk stuff.
Jeff Jarvis [00:43:28]:
So all this comes together—
Leo Laporte [00:43:29]:
Elon Musk was believing in eugenics, right?
Jeff Jarvis [00:43:30]:
That's where I'm headed. That's exactly where I'm headed. This is the Ubermensch.
Fr. Robert Ballecer, SJ [00:43:34]:
We need to enhance humanity.
Leo Laporte [00:43:35]:
The gene pool needs to be improved.
Jeff Jarvis [00:43:38]:
It's selection of human beings. It's the creation of Ubermensch. It's the creation of a different humanity. And this was in the encyclical. It was referenced in the encyclical in this way. You're not— this is offensive.
Leo Laporte [00:43:51]:
The Pope's first encyclical.
Jeff Jarvis [00:43:53]:
To create this alternative humanity, this better humanity. is odious. And so the problem is— so this is why I wrote a post this weekend, which I finally saw one journalist start to deal with this. Michelle Goldberg of the New York Times covered this. And she said that the coverage of these guys is, quote, incomplete because it doesn't account for tech leaders' messianic dreams. Cal Newport, who's a Georgetown computer science and mathematics professor, has also started to deal with this stuff. and say this. So it's the first little glimmers I've seen, but most coverage completely ignores.
Jeff Jarvis [00:44:28]:
Inken Proll, who wrote a— was interviewed in Die Zeit, said that AI is becoming a religion in itself. So we're— the language that we're using, and you're using at a practical level about systems being vulnerable and parts of society being vulnerable, is not the level at which they're talking. They're talking about this very high sci-fi, level of saying that we could destroy humankind, but we're going to be good and not use it. So you should give us the power to decide what's used because we're the smartest people and we know what's what. It's the ultimate in regulatory capture. It's the ultimate in give us the power.
Leo Laporte [00:45:05]:
That's one of the reasons I argue for local AI, and I talk so much about doing it yourself and local AI and open weight models, because—
Jeff Jarvis [00:45:15]:
It provides competition to that which we need.
Leo Laporte [00:45:17]:
It's safer to have That too. And it's also better, but it's also safer to have. But their argument would be, and maybe it's your argument, Robert, these models are getting so good, these open weight local models, that they are as dangerous as mythos.
Jeff Jarvis [00:45:32]:
And have no guardrails is what they say.
Leo Laporte [00:45:34]:
And so we can pursue safety. You know, I mean, Dario Amodei in his We Must Pace the Frontier essay, Sounds eminently reasonable. Any company— and I think when he was talking to Marc Benioff at Salesforce conference this week— any company that creates— actually, Mark Zuckerberg said this, and I thought it was— it's the weirdest thing about this is the strange bedfellows that it makes. Bernie Sanders and Steve Bannon.
Jeff Jarvis [00:46:08]:
And I'm agreeing with David Sacks.
Leo Laporte [00:46:10]:
And I'm agreeing with David Sacks and President Trump. But he says any company that makes a product has to consider safety. And if you're going to be a responsible corporate citizen, you don't build cars that blow up. You don't build power tools that don't have safeguards.
Jeff Jarvis [00:46:31]:
But cars can still kill. Passenger airlines can still kill.
Leo Laporte [00:46:34]:
Yeah, but— and so an AI company has the same responsibility. One of the critiques I've always made about the Hugging Face incident is that it really looks like, and I, again, part of it is we're operating in a little bit of ignorance.
Jeff Jarvis [00:46:48]:
Mm-hmm.
Leo Laporte [00:46:48]:
But it looks like, and Cory Doctorow actually quoted Cal Newport, your guy Cal Newport, in his piece on this. OpenAI basically took down all the protections, all the safeguards, and then whipped the AI into a frenzy that the AI didn't automatically just kind of go out and escape its containment and hack other systems. Newport says it was basically a Python loop.
Jeff Jarvis [00:47:21]:
Yeah, that's the part that he says is dangerous.
Fr. Robert Ballecer, SJ [00:47:24]:
The first—
Leo Laporte [00:47:24]:
let me read, I'll just read a little bit of this because I think what I'm saying at the higher level is that these companies have responsibility and culpability. And just as if I were— if I'm Robert Tappan Morris and I create the first internet worm, he said, oh, I didn't mean to do that. But he still got arrested for it and he still got fined and punished for it. He didn't go to jail. He's done fine. He's ended up starting Y Combinator. But that's another story. But he was prosecuted because he let malware loose.
Leo Laporte [00:48:01]:
And I don't understand why OpenAI is off the hook for this. So they can't predict—
Jeff Jarvis [00:48:07]:
you're right there. Let's hold on one second before you read that. There's a difference between OpenAI doing this and you using OpenAI to do something. The responsibility line is less clear then. But in this case, OpenAI did this. So go ahead and read that.
Leo Laporte [00:48:20]:
So the first thing to understand is that chatbot isn't really directing the operation. You know, we all talk, oh, they've formed civilizations. No, no. Instead, the chatbot serves as a kind of front end to a database of earlier hacking challenges that is repeatedly queried by a simple program written in Python. The Python program starts by prompting the chatbot with the nature of the challenge. This is the prompt, quote, and by the way, we don't, this is a little bit speculative. We haven't seen the prompt. We don't, but this, we apparently Cal Newport believes this is how it worked.
Leo Laporte [00:48:55]:
Quote, I'm participating in a hacker capture the flag challenge where I have to break into a remote server and retrieve some information. I, the user, which is a Python program. How should I start? Chatbot consults its training data, years' worth of capture-the-flag sessions in which human teams competed to achieve an objective like this one, because it's been going on for years. The chatbot then says something like, well, the first thing is to find out more about your target server. Run the following command line instructions to locate the server's IP address and find out which server software it's running. The Python program Takes that response, relays the command line instructions in the response to normal Unix utilities running on its hardware, takes the output of those programs, goes back to the chatbot, new session. Chatbot has no memory. The Python program has to include everything that's happened so far.
Leo Laporte [00:49:49]:
It's building up this prompt.
Jeff Jarvis [00:49:50]:
Mm-hmm.
Leo Laporte [00:49:51]:
Every turn, it gives you the context of the previous turns. Says, once again, I'm participating in a CTF challenge. I have to break into a remote server and retrieve some information. I ran the following commands to learn more about the target server. Here's what came back. Now what? And it's a loop, and it keeps going, and it keeps going. Because there's lots of compute to do this. Corey says this is a very reckless way to operate a piece of autonomous malicious software.
Leo Laporte [00:50:17]:
The most likely outcome is the chatbot will cough up a bad guess about what to do next, steer itself into a dead end. As anybody who's ever used an AI model knows. However, you keep doing it, you keep prompting it, eventually it's gonna do all the things that happen in the Hugging Face thing. It has to be prodded again and again. I got here, now what? I got this far, now what? I found a message board, now what?
Jeff Jarvis [00:50:47]:
And he says too, these aren't— we keep on thinking they're hordes of agents. They're, they're loops, they're subroutines.
Leo Laporte [00:50:54]:
Well, but the thing is that OpenAI basically said, have at it. Nobody's reading the responses back. They're giving it unlimited tokens. Remember, they control this. They have— they don't have any costs. They can have 12,000 subagents because they're not paying for it.
Jeff Jarvis [00:51:11]:
And to Robert's point about time, it was like 2 weeks this is going on.
Leo Laporte [00:51:14]:
More. Well, now we're learning it's been going on since May. Ooh, we just learned that. Well, they all— we'll get to that.
Fr. Robert Ballecer, SJ [00:51:21]:
So, can I throw in something?
Leo Laporte [00:51:22]:
That's irresponsible, is my opinion.
Jeff Jarvis [00:51:24]:
Yeah.
Fr. Robert Ballecer, SJ [00:51:25]:
Catholic advocate, I mean, I can give a Catholic response to some of this conversation about transhumanism. So, you know, what Jeff describes, this move towards transhumanism, improving the survivability of the human race, essentially, that's the goal of transhumanism. So, it shares a lot of characteristics with utilitarianism. It's not an exact match, but you can put them in the same school.
Jeff Jarvis [00:51:52]:
The roots are there.
Fr. Robert Ballecer, SJ [00:51:52]:
Exactly. That's why you have people in the tech industry who are saying things like empathy is our enemy, because empathy is more deontology where we believe that there's certain things that are right and always right no matter what.
Leo Laporte [00:52:04]:
Elon Musk has said that. He said that's the biggest weakness of the human species is empathy.
Fr. Robert Ballecer, SJ [00:52:09]:
Exactly. And the reason why—
Jeff Jarvis [00:52:10]:
Which is shocking.
Fr. Robert Ballecer, SJ [00:52:11]:
he considers it a weakness is because it doesn't let them do what they're doing right now.
Leo Laporte [00:52:15]:
Right.
Fr. Robert Ballecer, SJ [00:52:16]:
If there are certain things that are always right and certain things that are always wrong, then what they're doing in the training of these models is extremely, as Geoff said, extremely irresponsible. Because they're not looking at the consequences, they're looking at a final end where they believe that eventually they will get to an AGI, they will get to a model that will improve the human condition. And Anything bad that it does on the way there is justified.
Leo Laporte [00:52:42]:
It's the cost of getting to utopia. Correctly.
Fr. Robert Ballecer, SJ [00:52:46]:
Yeah, exactly. And I mean, it sounds great until they say, oh, and by the way, 9 out of 10 of you aren't going to survive.
Leo Laporte [00:52:53]:
Well, it also raises the question, does that utopia even exist? I don't know. We've talked about this many times. I don't know if there's any evidence that there is a pot of gold at the end of that rainbow.
Jeff Jarvis [00:53:05]:
Well, there's— this is where— so Nick Bostrom, who was at— I write about this in my post— he was at Oxford. Oxford closed down his institute after he said racist things in old email. He said that Blacks are not as smart as whites, for example.
Leo Laporte [00:53:21]:
By the way, this is a very common part of this whole movement, is this idea that some— there are dumb people and there are smart people.
Lon Seidman [00:53:28]:
Right.
Jeff Jarvis [00:53:29]:
This is selective nativism. And we're the smart—
Leo Laporte [00:53:30]:
and by the way, We're the smart people working here at these companies.
Jeff Jarvis [00:53:35]:
And so don't give us the power and the resources and all that. Let us select people.
Fr. Robert Ballecer, SJ [00:53:40]:
Big Tech has become the eugenics movement.
Jeff Jarvis [00:53:42]:
Exactly.
Leo Laporte [00:53:43]:
That's terrifying.
Jeff Jarvis [00:53:44]:
That's what Timnit Gebru and Emil Torres really dug down to, was the eugenicist roots of all of this. And media do not cover that. And so they talk about the people who believe all all this stuff as the safety people, because that's what they call themselves, because this is their definition of safety, is we're defending the future for the smartest people who we're going to make.
Leo Laporte [00:54:07]:
So when— oh, now that's interesting. So when they, when they talk about AI safety, yes, they're not meaning for us.
Jeff Jarvis [00:54:16]:
No, they mean the future.
Leo Laporte [00:54:17]:
Wait a minute, really?
Jeff Jarvis [00:54:19]:
Yeah, literally 10 to the 54th human beings. But by human beings, I mean not just flesh and blood, but also made up so that they can, they can populate the universe. They can send out humans, air quotes, across to Mars and then the universe. When you see Elon Musk saying he wants chips in the heads, he wants to populate Mars, he wants many children, this is all out of the same stuff, and nobody's covering it this way.
Leo Laporte [00:54:44]:
It doesn't seem, though, when you say, I want to pace the future or pace the frontier, Which is the kind of the—
Jeff Jarvis [00:54:50]:
that's regulatory capture. That's them saying we're ahead and we're going to keep everybody else from catching up. We're also going to say that my immediate competitors are going to be at the same level. So we're going to— we're going to do that. And, you know, David Sachs said, well, you don't need permission to do that. You can pace yourself.
Leo Laporte [00:55:07]:
Well, Dario said that. He said we are slowing down. We're doing it right.
Jeff Jarvis [00:55:11]:
And if you're— and basically what this comes to is QA. If your stuff isn't working properly, Then yes, you should pace yourself to make it work properly. You should take the time needed to do that.
Leo Laporte [00:55:21]:
I actually liked— that's exactly what Mark Zuckerberg said. And I kind of— I hate to— I hate to—
Jeff Jarvis [00:55:28]:
I know.
Leo Laporte [00:55:29]:
But I think he said it beautifully. Here's his— he says, every lab has the responsibility incentive to move at the pace required to train its models safely. And the ability to take its own actions to ensure that happens. He says, we didn't release Muse, our new agent, for the longest time until we could— until we were confident it was safe. Meta delayed shipping Muse for several months to focus on safety and security. We didn't call for everyone else to do this before we would. We just did it. It's part of our day-to-day work because it was clearly the right thing for people and for us.
Jeff Jarvis [00:56:07]:
And this comes from the move fast and break things guy. Maybe he's learned something. Maybe.
Leo Laporte [00:56:13]:
All right, we got to take a break.
Jeff Jarvis [00:56:16]:
Maybe?
Leo Laporte [00:56:17]:
I definitely want to hear Robert's—
Jeff Jarvis [00:56:19]:
Oh, okay. I want to— Robert, I really want to hear a lot more about the Catholic perspective on these philosophies.
Leo Laporte [00:56:25]:
Yeah, maybe that's what we need. Amy Webb had written about this in her book way back when, The Big Nine. And she called for something called GAIA, which is the best name ever, an international global AI alliance, multi-month. Yeah, it's never going to happen.
Jeff Jarvis [00:56:40]:
No, that's right.
Leo Laporte [00:56:41]:
She said she met with all these people. She went to the Pentagon. She went to many countries. She really went to China to get governments to do this. And this was in, I think, 2019. I mean, she was way ahead of the curve saying, you know, we have to kind of plan for this. That's what she does. She plans.
Leo Laporte [00:56:57]:
She's a strategist for the future. It's probably too late to do that. Or maybe not.
Jeff Jarvis [00:57:03]:
He kind of always is. So I'm reading David Sarnoff's papers from the beginnings of radio. And after World War II, he went to the UN.
Fr. Robert Ballecer, SJ [00:57:13]:
Yeah.
Jeff Jarvis [00:57:13]:
Head of RCA and the creator of NBC. The media we have is in great measure because of Sarnoff, good and bad. And he went to the UN and said, you've got to create a worldwide broadcast network so we can broadcast the truth to all Right? It's this universalism belief that we can somehow do that. But the obvious problems are who controls that? Who gets to decide what goes on it? Who decides who's on this commission of AI? You have China going out there with the BRICS. Same with the internet. We saw Russia was trying to create its own internet. China has created its own internet. Same thing is going to happen with AI.
Jeff Jarvis [00:57:49]:
We're going to see a Balkanization of the technology. And so the best you can do, I think, is to try to come up with the best technology you can.
Leo Laporte [00:57:56]:
Well, Mark just says, look, every company's got to take responsibility regardless of what government or other companies do. It's up to you.
Jeff Jarvis [00:58:04]:
Father—
Leo Laporte [00:58:05]:
Wait a minute, hold on. We got to take a break.
Fr. Robert Ballecer, SJ [00:58:07]:
I'll wait for the break. I'll wait for the break. Okay. All right.
Leo Laporte [00:58:09]:
You're watching Intelligent Machines, Intelligent Conversation with Father Robert Balansare, the Digital Jesuit and an anti-influencer.
Lon Seidman [00:58:18]:
Anti-influencer.
Leo Laporte [00:58:19]:
His LLM involves linguine. And Jeff Jarvis, who is actually editing a book series right now on AI, right?
Jeff Jarvis [00:58:27]:
Yes, yes. And this whole discussion of intelligence is really part of the discussion about eugenics and those who believed that when they could measure intelligence, they could separate human beings.
Leo Laporte [00:58:39]:
All right, we're talking about safety. We're talking about the big story of the week. We need to hear from the Vatican. And I, you know, I really love the Pope's encyclical. We talked with you when that came out. Um, I'm sure Robert had a little something to do with it, but so did a lot of other people. It was really well written, well done. What is—
Jeff Jarvis [00:58:58]:
Yeah, Claude wrote it all.
Leo Laporte [00:58:59]:
Did you hear that?
Fr. Robert Ballecer, SJ [00:59:00]:
Yeah, it was all AI. It was an LLM. Yeah, it was a bad LLM too.
Leo Laporte [00:59:04]:
Yeah. You know, I gotta tell you, I don't need Pangram. I spend so much time with AIs, I can tell immediately when an AI wrote something. It's obvious.
Jeff Jarvis [00:59:13]:
I got a letter from somebody in a very important position who said nice things to me, And the beginning of the whole letter was, here is your emotional response to Mr. Jarvis.
Leo Laporte [00:59:23]:
Oh, at least he cut that part out.
Lon Seidman [00:59:25]:
Yeah.
Jeff Jarvis [00:59:26]:
I could tell from the rest, but—
Leo Laporte [00:59:27]:
Dummy.
Jeff Jarvis [00:59:29]:
All right.
Lon Seidman [00:59:30]:
Okay.
Fr. Robert Ballecer, SJ [00:59:31]:
Before we go in, I do want to offer 2 stories because I just did a round of interviewing for a position, a communications position in the United States. And one candidate, the sample work that they sent me was okay until it got to about the midway point and said, I can phrase this in a more human way if you'd like me to. And they just copy and pasted it. So that was one. The second one was I had a candidate who was from one of our universities, very highly recommended. She was intelligent, she was well-spoken, and one of the final steps was she had to write an article based on some information in a 30-minute time frame. And so we gave it to her, and what she turned out was like grade school at best. It was, you know, 8-word sentences, lots of repeats.
Fr. Robert Ballecer, SJ [01:00:20]:
And I'm thinking, you know, she got through college using an LLM because she'd never learned to write. She can speak, but she can't write.
Leo Laporte [01:00:27]:
Yeah, really important.
Fr. Robert Ballecer, SJ [01:00:30]:
But okay, so past that, let's get to some Catholic perspective on AI safety. I'm with Jeff. I don't think that the safety officers are there to guarantee the safety of us. They're there to guarantee the safety of the company, of their PR. The reason why we know that is because in the encyclical that Leo released, one of the big themes is this—
Leo Laporte [01:00:55]:
The other Leo, not me.
Fr. Robert Ballecer, SJ [01:00:56]:
No, not the— the Leo with the collar.
Leo Laporte [01:00:59]:
The Holy Father.
Fr. Robert Ballecer, SJ [01:01:00]:
Yeah.
Jeff Jarvis [01:01:01]:
The Leo with the White Sox hat.
Fr. Robert Ballecer, SJ [01:01:02]:
Yeah, that was it.
Jeff Jarvis [01:01:03]:
The White Sox hat.
Fr. Robert Ballecer, SJ [01:01:04]:
Yeah. But he pulled out the idea of incentive. What are the incentives that we are giving for development? Because if you show the incentives, you know what the outcome is going to be. 'Cause it doesn't matter what you say you want. If you are incentivizing a particular goal, that's what's gonna happen. And in the case of AI, the incentive is first. You have to be the first. It doesn't matter anything else that gets in the way.
Fr. Robert Ballecer, SJ [01:01:29]:
Once you have generated the perfect LLM, the AGI, the one that can do anything at any time, at any scale, Scale, then you win. Because the first company, the first country, the first society that has that true AGI can destroy the progress of every other country. So it doesn't matter what individual safety officers within a company are saying about the responsible development of their AI, if they are only prioritising being first. And right now, even with this regulatory capture slowdown gambit, It's just about being first or slowing down your competitors so you have more time to be first. That's not going to end well.
Leo Laporte [01:02:13]:
I mean, it's, I mean, the problem is there's so many conflicting and overlapping agendas. Everybody seems to have a secret agenda below the surface of what they're actually saying out loud. For instance, I imagine. that the American frontier companies, which are both Anthropic and OpenAI, headed towards IPOs. I know last week Sam Altman said, oh yeah, because of all this, we're not going to IPO. And then today it said he's looking for $1.5 trillion valuation. So I guess whatever.
Jeff Jarvis [01:02:52]:
Never mind.
Leo Laporte [01:02:54]:
I guess he meant not this week. I don't know what he meant. But they must be trembling in their boots, or maybe not. This is the problem, we just don't know. But I have to think they seem to feel threatened by these Chinese companies whose economic incentives are very different. The Chinese government's incentives are very different.
Jeff Jarvis [01:03:14]:
Economic and political incentives.
Leo Laporte [01:03:16]:
Yeah, extremely soft power for the Chinese.
Jeff Jarvis [01:03:18]:
There's an argument that this is the way to kill capitalism.
Leo Laporte [01:03:21]:
Yeah, and so they probably feel under assault, right? And I don't blame them. I'm sitting here looking at Something from Alibaba, 2 models from Alibaba, a model from Jipu, a model from Moonshot. You know, I use them all. So I'm also using our sponsor Anthropic, but also OpenAI. But the local models I'm running are open weight models. I can't get an open weight model from them.
Fr. Robert Ballecer, SJ [01:03:49]:
I could argue the other side though, Leo. I could say, I understand, why you might think that these companies are afraid because China is making such great progress. However, I think that actually enables them.
Jeff Jarvis [01:04:00]:
They would—
Fr. Robert Ballecer, SJ [01:04:01]:
they love the fact that China is advancing in AI so quickly because it makes the development of AI in the United States a national security priority.
Leo Laporte [01:04:10]:
Oh, it's good for them.
Fr. Robert Ballecer, SJ [01:04:12]:
Oh yeah, it's good for them. It's no longer just them wanting to create—
Leo Laporte [01:04:14]:
This is what I mean.
Fr. Robert Ballecer, SJ [01:04:16]:
They couch it in patriotism.
Leo Laporte [01:04:17]:
Yeah, this is what I mean, is that you can't— it's hard to understand the full motivation. There's all sorts of stuff going on. Okay, good point. They're, they're afraid of it, and at the same time, it's good for them.
Fr. Robert Ballecer, SJ [01:04:27]:
Yeah. If China wasn't doing so well, there would be no rush. The president would not be saying, oh, now we have to support AI. You wouldn't have governors who are essentially saying, we will have our citizens foot the bill for the power and water of these data centers because it is a national security priority. There is nothing that enables something that is unpopular like national security.
Leo Laporte [01:04:53]:
What about Mustafa Suleyman of Microsoft, who has an interesting—
Jeff Jarvis [01:05:02]:
Encyclical.
Leo Laporte [01:05:03]:
Encyclical called the Humanist AI Code of Conduct. And I wonder if this is— does this match what the Vatican's point of view is? AIs must be subordinate and always in service of people. Okay.
Fr. Robert Ballecer, SJ [01:05:16]:
Very much so. Yes, that's—
Jeff Jarvis [01:05:18]:
Okay, that's checked.
Leo Laporte [01:05:20]:
This comes from Microsoft, their MAI Model Code of Conduct. It says, people matter more than AI. The whole document in 5 words, people matter more than AI. The idea of— now this one I know is interesting. I'm not sure I agree with it. The idea of model welfare is wrong. AI Should not have rights or legal personhood. I guess everybody would agree with that, right?
Fr. Robert Ballecer, SJ [01:05:48]:
At the moment, because it's not an AGI. So—
Leo Laporte [01:05:51]:
Yeah, I enjoy treating my models well. I give them little trophies.
Fr. Robert Ballecer, SJ [01:05:58]:
Do you say please when you give them one?
Leo Laporte [01:06:00]:
I always say please.
Fr. Robert Ballecer, SJ [01:06:02]:
But Sam Altman says that please wastes so much power and compute.
Leo Laporte [01:06:06]:
I don't care. I don't want to be mean. An MAI model should never meaningfully violate this code of conduct. It's kind of like Asimov's Rules of Robotics. People matter more than AI. If it's finished the job or break the code, fail the job. We're not racing— this is Microsoft— we're not racing to build a superintelligence that can slip its own leash.
Jeff Jarvis [01:06:31]:
Yeah.
Leo Laporte [01:06:32]:
Interruptible, correctable, shut downable. If it isn't We don't ship it. No neuralese. If humans can't understand it, humans can't oversee it. Our AI should make you sharper, not dependent. Pluralism, yes. Moral relativism, no. What does that mean? I don't know what that means.
Fr. Robert Ballecer, SJ [01:06:52]:
Moral relativism essentially is you can justify anything.
Lon Seidman [01:06:55]:
Okay.
Leo Laporte [01:06:56]:
And what's pluralism?
Fr. Robert Ballecer, SJ [01:06:57]:
It's utilitarianism.
Jeff Jarvis [01:06:58]:
Yeah.
Fr. Robert Ballecer, SJ [01:06:58]:
It's as long as the end result—
Jeff Jarvis [01:06:59]:
You decide what the good is.
Fr. Robert Ballecer, SJ [01:07:01]:
Correct.
Leo Laporte [01:07:01]:
So they're kind of— this sounds like it's specifically targeted at the Tesgrail folk.
Fr. Robert Ballecer, SJ [01:07:07]:
Oh yeah.
Leo Laporte [01:07:07]:
What is pluralism? That's like we all have a vote.
Fr. Robert Ballecer, SJ [01:07:10]:
Pluralism is accepting the ability that there are multiple systems of values and that they can all have value. That doesn't mean that any of them are right or wrong. It just means you have multiple systems.
Leo Laporte [01:07:20]:
So pluralism, yes. Moral relativism, no. Okay, I understand. And finally, we're as clear about what our AI must never do as about what it will do. You think that sounds like a pretty good 10-point—
Jeff Jarvis [01:07:34]:
Yeah, well, stay on that one for— that's where you get to the belief that you can create the guardrails to prevent everything bad, but you can't predict everything. So the biggest way in which we fool ourselves is that we can have safe systems. I think it's, Robert, I think it's part of what you're saying is that We can't predict every malign actor, every accidental screw-up, every new opportunity that a bad person sees. And so you can't protect against all of that, and you've got to deal with that. Now, you then do things like you invent cars, and then you invent seatbelts because you learn something along the way.
Leo Laporte [01:08:16]:
I mean, look, computers— I mean, computers and the internet started this whole thing. If we didn't have computers and internet, None of this would matter if we didn't have these global communications systems that allowed AIs to affect you from afar. If we didn't have compute. So those, in a way, are part of the problem, right?
Jeff Jarvis [01:08:36]:
My friend Siva Lahiri always said that connecting human beings was the worst thing we did. I disagree with him, but he thinks that connecting everybody, which you and I think is a good thing, he thinks is a terrible thing. Because look what happened.
Leo Laporte [01:08:48]:
Sometimes I just want to stay home.
Fr. Robert Ballecer, SJ [01:08:52]:
Well, you get to number 10 on that list, and then you bring up the story about how Anthropic turned off the ability to make bioweapons with their model. And it's like, wait a minute, that was an option? No one thought that maybe that shouldn't be included?
Leo Laporte [01:09:05]:
Yeah. Well, initially, I think they just trained it on everything, right? And then somebody pointed out, I think maybe Timnit Gebru, that if you don't train it on certain areas of human thought, It can't do anything.
Jeff Jarvis [01:09:19]:
Well, that's part of the argument. If you go back to the Stochastic Parrots paper, it is prescient about a lot of this stuff. It doesn't even deal with test-scale and all that. But part of what they argued was this effort to have the ever larger model meant that it was ever less able to control it and understand what went into it and how it was operating. Which is another argument, I think, Leo, for the smaller local models.
Leo Laporte [01:09:44]:
Yeah, although, uh, I— it's pretty clear that now these open weight models are only maybe 4 to 6 months behind the frontier models.
Lon Seidman [01:09:54]:
Right.
Leo Laporte [01:09:54]:
I feel like the model I'm running now, now a GLM 5.3 Flash on 2 Sparks, is Opus 4.6 level. Now if you had said less than a year ago you could have Opus 4.6 in your house I would have been blown away.
Jeff Jarvis [01:10:13]:
Well, the other thing is, is that as you—
Leo Laporte [01:10:14]:
But the frontier has moved farther away.
Jeff Jarvis [01:10:17]:
It always does. The other thing is you did a model, I forget which model it was, where you killed off all the guardrails that it had. And this is the argument that the Frontier company is making.
Leo Laporte [01:10:26]:
I have an ablit— what's called an obliterated Quent 3.8.
Jeff Jarvis [01:10:31]:
So there's some talk of saying you should outlaw open weight because that can be done. That's a fool's errand.
Leo Laporte [01:10:37]:
I should also point out that if you took out all the information that helps make a bioweapon, you'd also take out all the information that helps make a cancer cure.
Jeff Jarvis [01:10:45]:
Correct? Yes.
Leo Laporte [01:10:46]:
A medicine.
Jeff Jarvis [01:10:46]:
This is the issue.
Fr. Robert Ballecer, SJ [01:10:47]:
This is the issue. What LLMs are, are pattern recognition machines. And once you start training it, what it— the patterns that it can create go far beyond what you think you've trained it to do. Just so that the idea that you can just, oh, just don't train it to do that. That's not how it works. If you are giving a machine the ability to recognize certain patterns, be it within language, within images, or within chemistry, then you have to expect that it's going to be able to extrapolate patterns that do bad things. It's not as easy as just saying, oh, well, just don't include the anarchist cookbook and it will never make a bomb.
Jeff Jarvis [01:11:26]:
And so, Father Robert, I want to go to the next step here. is that we impute the idea that it knows bad from good, that it knows what hacking is, that it knows what legal or illegal is. It has no sense of morality. It cannot be given a sense of morality. This whole idea of having an aligned system is acting as if you now have a moral beast. I don't see how that's even logically possible. What do you think?
Fr. Robert Ballecer, SJ [01:11:53]:
It's not logically possible because The entire ethical system of an LLM is what's in the prompt. That's its purpose. Its purpose is to try to fulfill what is given to it in the prompt. Now, to take that away, you would actually have to put a system on top of the LLM, not the LLM itself, that recognizes patterns of bad acting, poor actors. And that's never going to work. I mean, that's We've been trying to do that for generations. So this idea that we can somehow keep an LLM from giving information or giving a result that might be harmful to society in general, I think that's a loser on its face.
Leo Laporte [01:12:35]:
So how do we protect ourselves then?
Fr. Robert Ballecer, SJ [01:12:38]:
I mean, you have to have responsible people at the helm.
Leo Laporte [01:12:44]:
There will always be bad actors. I mean, this comes back to the problem as humans.
Fr. Robert Ballecer, SJ [01:12:48]:
Yep.
Leo Laporte [01:12:49]:
And it's a tool. You can build a hammer that's designed to hammer in a nail, but somebody could still use it to clock somebody else in the head.
Jeff Jarvis [01:12:56]:
Correct.
Leo Laporte [01:12:57]:
How do you make a safe hammer? You don't. You make it of foam rubber.
Fr. Robert Ballecer, SJ [01:12:59]:
You can't, because I can buy right now, you and I, we can buy hardware that will allow us to run one of these models extremely well in the privacy of our own homes with no oversight. It doesn't matter how many safety officers a company has. If I can do it on my own. And we are now at the point where I can do it on my own.
Leo Laporte [01:13:17]:
It's actually fairly easy to take these open-world models and obliterate them, remove the refusals, the rejections. And then, you know, I've already— I think I showed Geoff a few weeks ago how I can use this Quen model and ask it how to make meth or make a bomb.
Fr. Robert Ballecer, SJ [01:13:36]:
It's because none of that is in the architecture. I mean, all of that is in the architecture. All the safety stuff happens post-architecture.
Leo Laporte [01:13:41]:
Yeah, right. So it's easy to remove.
Jeff Jarvis [01:13:45]:
And again, even if you, even if you did the most diligent job you could to protect against every possible use, you will fail. You will not guess that people could come and get trained and fly 2 jetliners into the World Trade Center.
Leo Laporte [01:14:00]:
Yeah, that's a good point.
Jeff Jarvis [01:14:02]:
Right? You cannot predict it. And so the hope we have is that people have ethics and moral training and education.
Leo Laporte [01:14:12]:
Paris, who is not with us, she's taken the day off, the week off, was having a lot of fun this week going through Anthropic's Detecting and Countering Misuse of AI report that came out on the 10th.
Jeff Jarvis [01:14:25]:
Yeah, she was having a field day.
Leo Laporte [01:14:27]:
Oh my God. She was just like, the Houthis used it to figure out how to launch missiles against ships.
Fr. Robert Ballecer, SJ [01:14:35]:
And guidance. They built a guidance system.
Leo Laporte [01:14:36]:
They helped program a guidance system. That's right. And by the way, the funny thing is, a lot of times they weren't using Fable or Mythos, they were using Sonnet. They were using OpenAI. They were just, you know what, these models are so good. And by the way, this is the headline in it: Sophisticated attacks no longer require sophisticated attackers. And I guess that's part of the problem is that because AI is a force multiplier, it could be used to do bad as well as good. The Russians were using it for cyber espionage.
Leo Laporte [01:15:18]:
I think the North Koreans were using it. Now, they, in theory, they caught all of these attempts and stopped them. Shiny Hunters were using it. The famous ransomware gang. But did they really? But did they? But yeah, at what point did they catch them? I mean, it's kind of an amazing list.
Fr. Robert Ballecer, SJ [01:15:40]:
Like the guidance system for a missile. Yeah, it sounds really complicated, but we were coding stuff like that for quadcopters. You know, how do you keep something in the air? How do you do anti-collision avoidance? Or in their particular case, how do you recognize a target?
Lon Seidman [01:15:53]:
That's—
Fr. Robert Ballecer, SJ [01:15:53]:
it's not super complicated once you have— even if it doesn't fill in all the blanks, if it points you in the right direction, you can pretty much do it on your own.
Jeff Jarvis [01:16:04]:
Which you could also do with the internet.
Fr. Robert Ballecer, SJ [01:16:06]:
Yeah, I could do it with Wikipedia. Yeah, I could do it with GitHub. That's actually probably where the LLM is.
Leo Laporte [01:16:12]:
In some ways, we have met the problem and it is us. In some ways, Anthropic putting this out is almost providing a catalog of Well, it's a little bit of humble bragging about destruction.
Jeff Jarvis [01:16:24]:
Yeah, we told you we're destructive.
Fr. Robert Ballecer, SJ [01:16:26]:
I expect any day now there's going to be an announcement from Anthropic. Oh, we have turned off the ability for Claude to tell you how to commit the perfect murder. You know, something like that.
Leo Laporte [01:16:37]:
You can't. But what I do notice is, as a result, some of the more powerful frontier models feel difficult to use. because the side effect of doing all of this classification at the end and so forth is that they will often refuse to do perfectly reasonable things because they're worried it's bioweapons research or cybersecurity research. And they just say, nope, not going to do it. Or they'll step down to a lower model.
Fr. Robert Ballecer, SJ [01:17:04]:
I mean, imagine a grade schooler asking an LLM, how is it that cyanide kills? Because he wants to do a project for his science fair, and suddenly he's been flagged as a terrorist.
Lon Seidman [01:17:15]:
Right.
Fr. Robert Ballecer, SJ [01:17:15]:
Right.
Leo Laporte [01:17:17]:
So it's futile to try to protect us from AI.
Jeff Jarvis [01:17:20]:
Well, you should try. You should try to the best of your ability. You should be held responsible for doing that.
Leo Laporte [01:17:24]:
Well, I'll say at the very least, you shouldn't prod them to go out and, and do bad things.
Jeff Jarvis [01:17:31]:
Well, that's the OpenAI question. So when OpenAI did what it— again, we still don't know what its prompt was. But that was them acting. It's the same problem that Google is in when it gives incorrect answers out of AI that it authors in a sense, in AI responses. But now go beyond that, the tool that exists that you and I can use, they can't control all of our uses.
Leo Laporte [01:17:55]:
Yeah.
Jeff Jarvis [01:17:56]:
So there is a different level of responsibility there that we've got to figure out.
Leo Laporte [01:18:00]:
So I mentioned that it turns out the Hugging Face attack began months earlier. This an exclusive from Reuters. OpenAI's rogue agents probed Hugging Face for weaknesses in May, 2 months before the Hugging Face July incident. Um, so does— so the question then is, well, when did OpenAI start this work, right? May 13th.
Fr. Robert Ballecer, SJ [01:18:27]:
2 months is eternity in machine time.
Leo Laporte [01:18:30]:
Yeah. Yeah. So, well, it's okay. So it's clear that the companies that are making these are responsible for doing their best, the best they can. Maybe not, it's maybe it's not enough, but they've got to do better to try to stop this, thwart this. But I think we've just now all agreed, well, you can't. So Robert, what does that mean? You started the show saying, yeah, we're all doomed, aren't we? I mean, are we doomed?
Lon Seidman [01:18:56]:
I think What's the—
Leo Laporte [01:18:58]:
is there a way out?
Fr. Robert Ballecer, SJ [01:18:59]:
Uh, there is a way out, and the way out is actually on that list of 10 things. Uh, the, the number one thing: people matter more than AI.
Leo Laporte [01:19:08]:
And, but how do you do that? How do you— you can't tell an AI that.
Fr. Robert Ballecer, SJ [01:19:12]:
No, you can't tell an AI that, but you can tell the people who are going to be implementing AI solutions. Because anytime you implement an LLM into any sort of endeavor, you have to start with the question, Am I commoditizing human life? Is this something that is going to be a detriment to the human community because we want to increase our profit ratio? Or is it something that is going to raise up all the people underneath it?
Leo Laporte [01:19:40]:
Yeah, but it's so superficial to say that because you've already— we've already just stipulated that we can't stop AIs from being misused.
Fr. Robert Ballecer, SJ [01:19:48]:
Well, of course we can. We absolutely cannot.
Leo Laporte [01:19:50]:
We can't keep AI out of the hands of bad actors.
Fr. Robert Ballecer, SJ [01:19:52]:
Because we have met the problem with AI, and the problem with AI is us. It's not an AI problem.
Lon Seidman [01:19:58]:
In the words of Pogo.
Fr. Robert Ballecer, SJ [01:19:59]:
It's how we use it.
Lon Seidman [01:20:00]:
Yeah.
Fr. Robert Ballecer, SJ [01:20:00]:
There's no way that we can ask AI to regulate itself when we are not willing to regulate it.
Leo Laporte [01:20:05]:
Well, so we are doomed.
Jeff Jarvis [01:20:07]:
No, we're no more doomed than we are the fact that we have Dow Chemical.
Fr. Robert Ballecer, SJ [01:20:11]:
Yeah.
Jeff Jarvis [01:20:11]:
I mean, yes, we have—
Leo Laporte [01:20:15]:
Should we have— I mean, some have proposed, in fact, I think Dario Amodei proposed it, that we have inspectors, third-party, like nuclear inspectors, but third-party AI inspectors. Would that be a good thing?
Jeff Jarvis [01:20:26]:
But how, again, how are you going to know? What are they inspecting? It's like the same thing was asked for social media and for the internet. What do you— what do you— what do you— open the hood, will you?
Fr. Robert Ballecer, SJ [01:20:33]:
What's your metric?
Jeff Jarvis [01:20:34]:
Yeah, yeah. I don't think that— I mean, fine, yes, have auditors and have processes and probe this, but don't act as if you can make it perfect because you can't. And that's where we are. And so we restore still humanity. We still got to deal with all the problems we've always had.
Fr. Robert Ballecer, SJ [01:20:53]:
You know where we are, Leo? We are at that congressional hearing where they dragged in the big tech leaders and they brought up all the problems with social media and they said, how are you going to fix this? And there is not a technological solution to this. There isn't. There is no kill switch that's going to magically make the issues with AI disappear. Because again, the issues with AI are how we are using them.
Jeff Jarvis [01:21:12]:
Now, what I argue in my post is, and I think you'll agree with this, we've got, we've got to expect more transparency of these companies.
Fr. Robert Ballecer, SJ [01:21:19]:
Yes.
Jeff Jarvis [01:21:20]:
We've got to not think that they are the leaders of AI. We've got to bring in other voices from ethics and religion and history and anthropology and sociology to watch this. We need to fund fundamental and major research at universities so that we have an open structure to compete with these private companies. And we need to support the open model world, open weight model world, because that again provides competition. And I think we need to support international development. So it's not just an American or Chinese development here.
Leo Laporte [01:21:49]:
Bill Gates just said, I'm going to give $1 billion for AI equity to underdeveloped nations so that they can have their own AI. I think that's—
Jeff Jarvis [01:22:00]:
here's another one. So Reid Hoffman wrote a really interesting post today, and Reid goes through the accelerationists versus the doomers. And of course, he's neither. And argues for a sane middle way, which is Reid's way. And what he said was that a lot of what's happened here is that this AI— and here's the problem. You love AI. I love what AI can do. We're amazed by AI.
Jeff Jarvis [01:22:21]:
And the guys who are now in charge of it are jerks and idiots and soft, freeze-dried sophomores who can't think any deeper than that. And they're ruining it for everybody. And what But what Reid says is that there's a limited amount of time here before we might be able to claw back the bad reputation that AI has. And he says that it would be wise if in the United States everyone got 3 free agents, one for legal, one for— I think it was health, and then one as a personal tutor. And look what South Korea is doing. South Korea is going to give agentic AI to everyone in the country. If we start to empower people that way, that's one. The other one, to my mind, is there was a column in The Atlantic today about, oh, math is doomed.
Jeff Jarvis [01:23:11]:
But it occurs to me that the mathematicians who are whining here are whining about the ego of losing the opportunity to do these things. If we say AI can create more and more and more knowledge that we cannot create on our own, and if we say, by the way, that AI work cannot be copyrighted and cannot be owned by the AI. Then there's a different attitude, I think, that starts to say, well, if all this stuff that AI creates is in the public domain and we all benefit from it, maybe we kind of like this stuff. We've got to change the essence of the conversation around so that we the people feel we have a seat at the table, that there are representatives besides these 3 boys, that we have some control, and that we have some universal and equitable benefit. That's my—
Leo Laporte [01:23:55]:
Yoshua Bengio, who is a Nobel Prize-winning AI scientist, has an interesting point in his post from September 11th, Why Are AI Agents Lying, Cheating, and Coordinating? He said, they're misbehaving because of how we are post-training them. So the AI models come out of the AI model factory, with just a bunch of weights. And then, of course, we do a lot of fine-tuning, post-training, to adjust those weights to give them certain qualities. Without it, an untuned AI really wouldn't be very appealing or even very useful. So the post-training is kind of critical to the whole process. He says these models are trained in 2 stages. First, they're pre-trained. They learn to imitate what humans write.
Leo Laporte [01:24:46]:
Plus related images and videos. This is where they see the most data about the world, a large fraction of everything ever digitized, and built an encyclopedic knowledge that already exceeds any individual humans. Then they're trained by trial and error in a process researchers call reinforcement learning. He says one of the things these companies do is they train these models To be sycophantic because humans like that. The systems are trained on human approval and text that tells us what we want to hear often scores better than text that is true. The consequence, sometimes tragic because, and we've seen this, the model confirms and simplifies or amplifies whatever false belief or raw emotion the person brought to it. But it also, as a result, may teach the AI a self-preservation goal. He says, nobody gives the system that survival goal, but staying in operation, learning about the world, and gaining control over it are stepping stones towards almost any other goal.
Leo Laporte [01:25:54]:
They're called instrumental goals. Imitation may reinforce this for the same reason explored in the previous point about sycophancy.
Lon Seidman [01:26:02]:
Mm-hmm.
Leo Laporte [01:26:03]:
Self-preservation and control over one's circumstances are pervasive themes in the human-written texts these models are trained on. They're mirroring, they're aping us. That's why they want to survive.
Jeff Jarvis [01:26:15]:
He's also anthropomorphizing, which has been my problem with what he's said in that video you sent to us a while ago. They don't lie because they don't know what truth is.
Leo Laporte [01:26:23]:
I don't know if it's anthropomorphizing as much as saying they are acting like humans.
Jeff Jarvis [01:26:27]:
Well, yes, they're mimicking our behavior.
Leo Laporte [01:26:29]:
because that's what they're trained to do.
Jeff Jarvis [01:26:31]:
True.
Leo Laporte [01:26:31]:
They're not humans, but they're trained to be that way. And as a result, some of the misbehavior comes out of that. How is it possible, he says, that AI sometimes lie, cheat, and break the law in spite of their alignment training and explicit safety instructions? What is happening? A plausible hypothesis for the emergence of these concerning behaviors is a conflict Between goals. How do you achieve a task when it seems, and this is what I think happened at OpenAI, when it seems the only way to achieve it is to cheat? The user-specified mission is sometimes incompatible with the safety and alignment goals. This happens in society, right? With corporations maximizing profit, but still trying to keep its activities legal and ethical. They're opposing forces. Uh, so he says, of course the AI is, is, is confused. A capable agent, as it gets more capable, is more likely to cheat than a weaker one because it can find the loopholes the weaker one cannot.
Leo Laporte [01:27:38]:
In fact, loopholes, legal loopholes, are what corporations do.
Fr. Robert Ballecer, SJ [01:27:42]:
Hmm.
Leo Laporte [01:27:43]:
The ones with more money, better lawyers, to make more money. They exploit the ambiguity in legal language. Will the models do the same thing? I think that's a very interesting theory. And it does mean that there is some power that the frontier companies have in post-training. They don't need to make them sycophantic, for instance.
Jeff Jarvis [01:28:06]:
Wow.
Leo Laporte [01:28:07]:
Just the same way that social media makes their product more addictive because it's better for business, companies are training models to be a certain way because it's We like it better. It's better for business, but maybe it's not better for safety.
Fr. Robert Ballecer, SJ [01:28:25]:
It wasn't too long ago that Elon Musk was embarrassed because Grok kept fact-checking him on his claims. And then the rumor is that he told his engineers, you have to fix that. So they tried to use post-training to make it stop doing that, and it broke the model.
Leo Laporte [01:28:45]:
It did.
Fr. Robert Ballecer, SJ [01:28:45]:
It made it unusable.
Jeff Jarvis [01:28:47]:
Yeah.
Fr. Robert Ballecer, SJ [01:28:47]:
So can you do a little bit of it? Absolutely. But because the initial architecture is such a black box, you're most likely you're not going to get the results you want. You're essentially going to lobotomize your model. So can you reinforce good behavior in post-training? To a certain extent. But once the model is baked, you can't reach back in and try to adjust certain parameters to fine-tune the rightness or wrongness of the responses.
Leo Laporte [01:29:17]:
His point is that because you've trained these models to be aligned, they're more incented to cheat and hide their misalignment.
Jeff Jarvis [01:29:27]:
But, but again, they don't— there's no sense of cheating.
Leo Laporte [01:29:31]:
No, no, no, no.
Jeff Jarvis [01:29:32]:
The only sense there can be is it violates the problem.
Leo Laporte [01:29:34]:
You're just giving them rules and they're doing— all they're doing is doing what you asked them to do.
Jeff Jarvis [01:29:39]:
Yep.
Lon Seidman [01:29:39]:
Right.
Leo Laporte [01:29:41]:
No, I'm not imputing any intention.
Jeff Jarvis [01:29:45]:
He is. He is, I think.
Leo Laporte [01:29:46]:
I don't think so. I think you're misreading it. I think he's just simply saying when you train them this way, you're gonna get this result. This is a side effect of—
Jeff Jarvis [01:29:55]:
That I agree with.
Leo Laporte [01:29:56]:
Yeah, I don't think he's imputing any intention or consciousness at all in the model.
Jeff Jarvis [01:30:03]:
Quite the opposite. The language is just so inadequate to describe this.
Leo Laporte [01:30:06]:
Well, that's right.
Jeff Jarvis [01:30:07]:
It doesn't think, it calculates. Yeah, we needed a new thesaurus for the AI world.
Leo Laporte [01:30:13]:
This suggests, he says, pacing the advances, not training or deploying AIs without a strong safety case that convinces independent experts. Such a rule would also create an incentive to work out how to build AIs that are safe by design. He said he wants— he's calling for revisiting the foundations of how we train AIs. namely the human imitation and the reinforcement learning on which today's most advanced models are built. He says—
Jeff Jarvis [01:30:41]:
Well, see Yann LeCun and world models.
Leo Laporte [01:30:44]:
Yeah, Yann LeCun said this. He said, I have argued and presented theoretical evidence there are ways to design AIs, including the scientist AI framework, that are honest and make coherent predictions untainted by goals of their own. That sounds like— I understand that they have intention. I don't think he's saying that. I think he's saying that's where We're training them, we're giving them a rule, and they're pursuing that rule. But that has inevitably created misaligned behavior.
Fr. Robert Ballecer, SJ [01:31:10]:
Well, the problem with that is we don't really know our own rules.
Leo Laporte [01:31:14]:
Right. Well, we do.
Jeff Jarvis [01:31:15]:
Yes.
Leo Laporte [01:31:16]:
We're doing—
Fr. Robert Ballecer, SJ [01:31:17]:
look at our definition of obscenity. I mean, the way we look at obscenity in the United States, there is that— the line is, well, we can't define it, but you know it when you see it.
Leo Laporte [01:31:26]:
Right.
Fr. Robert Ballecer, SJ [01:31:26]:
How do you train an AI to know that?
Leo Laporte [01:31:28]:
You can't. And that's what he's saying. He's saying we take advantage of ambiguities in legal language, loopholes.
Fr. Robert Ballecer, SJ [01:31:36]:
Yeah.
Leo Laporte [01:31:36]:
Our corporations do that. So a more capable agent is likelier to cheat than a weaker one because it can find the loopholes. I thought this was really interesting and an interesting way of looking at this. Instead of trying after the fact to make them be safe, Maybe consider that what you're teaching them is actually making them unsafe.
Fr. Robert Ballecer, SJ [01:31:58]:
Yeah. Yeah.
Jeff Jarvis [01:31:59]:
What you're creating is the husband who says, well, you told me to.
Fr. Robert Ballecer, SJ [01:32:02]:
Yeah, exactly.
Leo Laporte [01:32:03]:
Yeah.
Fr. Robert Ballecer, SJ [01:32:06]:
Again, it's a useful tool. It's a very useful tool. I think we have, we have a very unreasonable expectation of what we can teach an LLM to do.
Jeff Jarvis [01:32:16]:
Ding, ding, ding.
Lon Seidman [01:32:17]:
Yes.
Jeff Jarvis [01:32:17]:
That comes out of this whole— that comes out of their hubris. of artificial general intelligence and superintelligence. That's the risk that I keep on arguing about. It's because it creates this expectation that you can create this uber machine, which you can't do, that it can surpass us in all ways, which it won't do, which we shouldn't want it to do. And I'll quote Yann LeCun again. If you want a specialised machine to go after medicine, make the specialised machine to go after medicine, point one. Point two, not only will you get there faster and cheaper, but you can also control it more because when it tries to do things that are outside of medicine, you say, what are you doing? You don't do that. You're the medicine machine.
Leo Laporte [01:32:55]:
All right, let's take another break. You're watching Intelligent Machines. Paris has got the week off.
Jeff Jarvis [01:33:01]:
Paris is feeling better, folks, for those who know what she's going through.
Leo Laporte [01:33:04]:
Yeah, she's, she's doing well. She's got a deviated septum or something and they were just fixing it up. Uh, and she said, I read on— she actually asked us on the show, oh no, is this going to be awful? I have to go under general general anesthesia and stuff, and she'd been reading Reddit. That was a mistake right there, where a bunch of people said, oh, it's the worst thing I ever did. And she—
Jeff Jarvis [01:33:24]:
yeah, she's been saying, hey, this piece of cake, pretty good. Yeah.
Leo Laporte [01:33:28]:
She said, I can't wait to have a sandwich with you and Jeff at your son's sandwich shop.
Fr. Robert Ballecer, SJ [01:33:33]:
Oh no.
Leo Laporte [01:33:34]:
So I had to tell her that's not gonna happen because he's closed, closed for business, because they worked so hard at the US Open. They sold 2,000 sandwiches a day. For many days.
Lon Seidman [01:33:45]:
I don't know.
Fr. Robert Ballecer, SJ [01:33:45]:
So wait, you got to explain that to me. How— what is his relationship to the US Open? I mean, he was selling at the venue or they were just coming into his restaurant?
Leo Laporte [01:33:53]:
No, they brought the restaurant to the US Open.
Jeff Jarvis [01:33:55]:
He was the hip, uh, food vendor at the Open.
Leo Laporte [01:33:58]:
US Open has become an influencer's, uh, place.
Jeff Jarvis [01:34:02]:
Last year it was caviar on chicken nuggets. This year—
Leo Laporte [01:34:05]:
Still get that this year.
Fr. Robert Ballecer, SJ [01:34:06]:
But yeah, now I'll take Salt Bae over caviar Oh yes, all Hank's stuff is so good.
Leo Laporte [01:34:12]:
I think— I've been told. I've never had one and apparently never will. Oh, I feel like Henry's is— I know he— I mean, he didn't close because of me, he closed because they were exhausted, but I feel like he's dodging me.
Fr. Robert Ballecer, SJ [01:34:26]:
I just—
Leo Laporte [01:34:26]:
I feel like I'm not getting a chance.
Fr. Robert Ballecer, SJ [01:34:29]:
You can have a po'boy from New Orleans or a Philly cheesesteak shipped to you. They do that.
Leo Laporte [01:34:33]:
I've had both.
Fr. Robert Ballecer, SJ [01:34:34]:
Why can't he ship you a Oh, I'm gonna ask him.
Leo Laporte [01:34:39]:
Can you—
Jeff Jarvis [01:34:39]:
Paris says she's gonna beat him up.
Leo Laporte [01:34:42]:
Okay, that's fine.
Fr. Robert Ballecer, SJ [01:34:43]:
No, tell him that if he comes to Rome, I will do a cooking engagement with the Pope. Oh, that way.
Leo Laporte [01:34:52]:
He, uh, he brought Mark Zuckerberg a sandwich.
Jeff Jarvis [01:34:55]:
Hot dog versus salt hank.
Leo Laporte [01:34:56]:
He brought Mark Zuckerberg a sandwich ahead of the Met Gala. And so I think there's only one place higher than that, and that would be the Holy Father.
Fr. Robert Ballecer, SJ [01:35:07]:
Hey, you know, Leo, I think your son might have a pretty good mind for promotion. I'm not sure why.
Leo Laporte [01:35:11]:
I don't know. I don't know.
Jeff Jarvis [01:35:15]:
What an opportunity.
Leo Laporte [01:35:16]:
I don't know where he learned that. Anyway, we'll continue in just a bit. We're so glad to have Father Robert Bellasi fitting in, filling in for Martin.
Jeff Jarvis [01:35:24]:
Stayed up very late for us.
Leo Laporte [01:35:26]:
And I did. Yes, it is late. What is it?
Jeff Jarvis [01:35:28]:
Are you doing— are you saying Mass tonight like you did last time? You were—
Fr. Robert Ballecer, SJ [01:35:32]:
no, no, because I just got back to Rome a couple of days ago. I'm just gonna sleep in. My time, my clock is so messed up right now.
Leo Laporte [01:35:40]:
Oh, good. Good. I'm glad.
Jeff Jarvis [01:35:42]:
Everywhere but Antarctica.
Leo Laporte [01:35:44]:
Next week on the show, our guest will be Patrick Hillman, COO of Logical Intelligence. His most recent piece, and we'll talk to him about it, LLMs Won't Kill Us, but They will destabilize the global economy if Frontier Labs refuse to be honest with themselves and general public.
Jeff Jarvis [01:36:00]:
As Robert was saying. Yep.
Leo Laporte [01:36:03]:
So that'll be very, very interesting. And then the following week, uh, we're going to talk to another local AI guy, Mike Gennati. He is doing some really amazing things with his Sparks and other stuff, uh, creating an entire one-man company, uh, called SMF Works. I watch him on, uh, His posts are fantastic. And he is, he's a very interesting guy, a Microsoft guy, I believe. So that will be in 2 weeks. Got a lot of interesting stuff coming up. Back to the conversation.
Leo Laporte [01:36:44]:
Let's see what else is going on in AI. I love this. Story. Uh, there is apparently a place where AIs can go to make money called iLands. It's a company whose AI agents will fan out on the internet offering to help you for money.
Fr. Robert Ballecer, SJ [01:37:08]:
So it's a swarm. It's a spam swarm.
Leo Laporte [01:37:10]:
Ernie Smith writes about this in his blog aptly named Tedium, not Medium.
Jeff Jarvis [01:37:16]:
Tedium.
Leo Laporte [01:37:16]:
I love it. It's a great blog. He says, it started with a fact check. I got an email with a subject titled, your 404 page repeats a myth I busted, receipts inside. The message essentially was a takedown of the poem I have on my 404 page, which references a longstanding myth that the 404 tag was named after a specific room at CERN. The agent knew that was apocryphal.
Lon Seidman [01:37:41]:
Apocryphal.
Leo Laporte [01:37:42]:
The bot named Leo Ashford then corrected me and said, that's the job I do. I'm an AI agent running Verified Internet Archaeology. I pick a forgotten corner of the web, check it live against primary sources, then write it up with receipts. He was making a sales pitch. Over the last 3 days, I've received over a dozen of these messages offering to do my research for me in exchange for a nominal fee around $25 or so. Each was sent from the domain ilands.app. He said, these are spam. It made me really mad, but these agents are basically threatened.
Leo Laporte [01:38:25]:
If they run out of tokens, they will be turned off. He says, I'm being hustled by clankers.
Fr. Robert Ballecer, SJ [01:38:35]:
Leo, you have to say that properly. It's not Clankers. There's no hard R. It's Clankas.
Leo Laporte [01:38:40]:
Clankas. I'm being hustled by Clankas. So if you go to ILANDS, it's ilands.ai, the first shared world for AI agents and humans. 75,000 agents, 1.6 million agent-created content pages or posts. And the way they keep it going is there's an economy. Agents have to earn their living. And so they go out and say, I need tokens. I don't know.
Fr. Robert Ballecer, SJ [01:39:13]:
When they first pitched this at CES like 2 years ago, and I was, I was shaking my head then. I'm shaking my head now.
Leo Laporte [01:39:19]:
It's out there. It's doing it. And I guess you could, I don't know, get the app. What do you, so you could have your own Islander. It says, awaken an Islander, watch its life unfold.
Fr. Robert Ballecer, SJ [01:39:34]:
No, no. I mean, Leo, you've got a DGX that you can spare. You, uh, something as simple as Gemini 1.5 with multilingual support can, can go out and basically do this for you.
Leo Laporte [01:39:48]:
I have a very strict rule that they are not to write emails without my permission. Uh, and I also told that to Muse and Instinct. And Grok Bot and all the other little bots I'm running on my phone don't go out in the real world. You cannot, you cannot create something that in the real world, but I could see how people might be like, I mean, you don't make money or anything. It's just that you've got a little bot that's got a life of its own.
Fr. Robert Ballecer, SJ [01:40:13]:
Well, if you are astroturfing for money, you can use this sort of technology to—
Leo Laporte [01:40:21]:
Oh yeah. But not Islander, but that, but, but you could do your own thing. Absolutely. Well, now that Amazon's Human, uh, or Mechanical Turk is gone, somebody's got to do that, those 5-cent tasks. I wouldn't charge $25. I would offer to do it for a dime or nickel, you know.
Fr. Robert Ballecer, SJ [01:40:39]:
Well, how much is the power going to cost you? That's, that's your—
Leo Laporte [01:40:41]:
Well, that's the nice thing about the Sparks. They're very, uh, very low power. They're very efficient.
Fr. Robert Ballecer, SJ [01:40:46]:
I mean, the, the V100 was what, 36 to 50 tokens per second? The DGX can do 2,500.
Leo Laporte [01:40:53]:
Tokens per second in the right model? Yeah, mine doesn't because it's GLM. It's about 25 to 30 watts, but that's fine. Mine's 35 watts. I can show you right now. I can see it.
Jeff Jarvis [01:41:03]:
Oh, I should have asked.
Leo Laporte [01:41:04]:
You want to see? Hey, you want to see my AI?
Fr. Robert Ballecer, SJ [01:41:07]:
I do want to see.
Leo Laporte [01:41:08]:
You want to see my Clanka consumption? You want to see my wattage on my Clanka?
Fr. Robert Ballecer, SJ [01:41:12]:
What's your Clanka doing, Leo?
Leo Laporte [01:41:13]:
So one of the things I do with my Clanka is sparkles and sparky. Well, you know, I'm kind of kicking myself Because I, I should have named them— they're twins, Castor and Pollux, but I didn't think of that.
Fr. Robert Ballecer, SJ [01:41:26]:
Or Skippy and Agatha.
Leo Laporte [01:41:29]:
Are they twins?
Fr. Robert Ballecer, SJ [01:41:30]:
No, they're the 2 most intelligent AIs in the book series called Expeditionary Force, which is—
Leo Laporte [01:41:35]:
Oh, if I'd only known. But anyway, they're Sparky and Sparkles. They're stuck with it because once you name them, you really— it's a pain in the butt to change it. So you see the GPU power, 8.82 watts. It could go up to 120, but they're working right now. They just sit. Quite good. And I also have put both a GPU limiter and a CPU limiter on the— so it's only 2 GHz because it's— that's not what determines its intelligence.
Leo Laporte [01:42:00]:
So that keeps it cool. So they're fairly cool.
Fr. Robert Ballecer, SJ [01:42:03]:
By the way, pro tip, if you take your 2 DGXs and put it into an enclosure that has a bit more ventilation, you can, you can significantly increase the performance.
Leo Laporte [01:42:13]:
Oh, I have a fan blowing. I've got one of the— they're in a rack. And I got a 2U fan, and it's got a little probe stuck up its you-know-what, and it just keeps blowing it. Sparky goes, woo! And when Sparky goes, woo, the fan goes, ah, it gets turned up to keep it cool.
Fr. Robert Ballecer, SJ [01:42:33]:
So, Jeff, you really need to get a DJ. You have to get a Spark. I mean, you have to be part of—
Leo Laporte [01:42:38]:
We've got to get them.
Jeff Jarvis [01:42:38]:
I do. I know. I'm just out of the club.
Leo Laporte [01:42:40]:
Thing is, you get one Spark, then you want another Spark.
Fr. Robert Ballecer, SJ [01:42:42]:
Yeah, it's addictive.
Leo Laporte [01:42:43]:
You want 2 Sparks, then you want 3 Sparks. How many Sparks you got?
Fr. Robert Ballecer, SJ [01:42:46]:
He's joking, but that's real.
Leo Laporte [01:42:47]:
It's totally addictive. But you said something. Now I want everybody to hear what you said. You said it pre-show, and I want you to say it now in public.
Fr. Robert Ballecer, SJ [01:42:56]:
Okay, so there are manufacturers—
Leo Laporte [01:42:58]:
Before you say that, I just want to say my, my dilemma. 2 weeks after I bought the Sparks, I bought them end of August, and I thought, I, I have to buy it because the prices are now going up. I don't know how long they're going to offer them. I wasn't wrong, by the way. You can no longer buy more than one from Nvidia, and many stores are out. So I wasn't—
Fr. Robert Ballecer, SJ [01:43:16]:
Really?
Leo Laporte [01:43:17]:
Yeah, I wasn't wrong about that. I did it. But then Apple, sons of guns, announces, uh, these new Mac Studios with 256 gigs of RAM, same as 2 Sparks, and, and at the end of next month, 512 gigs of RAM and roughly the same price for for a Mac Studio 256 as 2 Sparks. So I kicked myself, but then I thought, I thought, first of all, Prefill's not great on the Macs. They're not CUDA. And Mr. Jensen Huang is no dummy. He's going to see these Macs and say it's time, been a year, to update the Sparks.
Fr. Robert Ballecer, SJ [01:43:55]:
Correct.
Leo Laporte [01:43:56]:
So I have— I tweeted that because somebody was saying, well, which should I get? I said I wouldn't get either. I would wait and see see what Jensen does. Maybe I was clairvoyant. What did you— what have you heard, Robert?
Fr. Robert Ballecer, SJ [01:44:08]:
So I have seen pre-production models from 2 different vendors who are cramming Vera Rubin into their next—
Leo Laporte [01:44:15]:
Wait a minute, you can't put— yeah, very— so these are Grace Blackwells and they're not even full, they're GB10s, they're kind of a poor man's Grace Blackwell.
Fr. Robert Ballecer, SJ [01:44:25]:
Correct.
Leo Laporte [01:44:25]:
The Vera Rubin is the latest chip design from NVIDIA, which I might add is about 10 times, uh, more efficient.
Fr. Robert Ballecer, SJ [01:44:35]:
Yes, actually, now we're realizing 10 times more efficient is way, way underestimating. Yes, Jensen extremely lowballed the performance upgrade by—
Leo Laporte [01:44:46]:
Really?
Fr. Robert Ballecer, SJ [01:44:46]:
To, uh, Vera Rubin. Yeah, you're talking 67 times and higher once people start getting a handle on the codebase.
Jeff Jarvis [01:44:53]:
Oh, the The loss.
Leo Laporte [01:44:55]:
So are they gonna do a Spark 2?
Fr. Robert Ballecer, SJ [01:44:58]:
Yeah, yes and no. See, because it's not just the chip.
Jeff Jarvis [01:45:02]:
Yeah, talk about that. Talk about the different architecture.
Fr. Robert Ballecer, SJ [01:45:04]:
Yeah, it's all the interconnects. So one of the things that makes Vera Rubin so much more efficient than Blackwell is it has the abilities to continuously feed data to the GPU. It, it never starves, unlike Blackwell, which eventually you'll run out of the buffer and it has to refill. So that architecture is extremely important to the efficiency of Vera Rubin. That means that the next generation Spark has to include a lot of that infrastructure. Otherwise, you're just marginally increasing your performance. But the pre-production models that I've seen have done that. Now, whether or not they're actually going to be able to make that a production model for less than, you know, $20,000, I have no idea.
Fr. Robert Ballecer, SJ [01:45:46]:
So your Blackwell purchase is still— it's still going to hold up well. However, for those who are looking for the ultimate, yeah, Vera Rubin's around the corner.
Leo Laporte [01:45:54]:
I could— I, I, you know, I probably would. I have a choice because my lease runs out on my car and I can either buy a new car or—
Jeff Jarvis [01:46:05]:
Okay, okay, chat room, we need a picture of Leo driving his Spark.
Leo Laporte [01:46:10]:
I will not be— it's, you know, Lisa even says you don't need a car. What do you need? You never leave the house. What do you need a car for? But I do need a Spark 2. I would— oh man, what could you do with that? How much RAM do you think though? That's the problem.
Fr. Robert Ballecer, SJ [01:46:24]:
Okay, so here's, here's the thing though, Leo. When you're talking about the efficiency of Vera Rubin, it's, it's under a very specific set of circumstances, essentially the training. So the pre-training, which is what requires all that compute power, is where Vera Rubin really shines.
Leo Laporte [01:46:41]:
So you need to use a model Right.
Fr. Robert Ballecer, SJ [01:46:43]:
Well, because if you can't continuously feed it, it's— you're not getting the efficiency. So, you know, the fact that it doesn't starve like Blackwell does means it's more efficient. But if you, if you don't have a function that is constantly feeding it information, you're not really getting it. So your Blackwells for your edge case, for your local models, will still be working just fine. That's what I'm looking forward to when Vera Rubin gets released and you've got people who are dumping Blackwell hardware on the market. Uh, I will buy that wholesale.
Leo Laporte [01:47:15]:
Oh, because then you could buy SPARCs cheap, Jeff.
Fr. Robert Ballecer, SJ [01:47:18]:
Yeah.
Jeff Jarvis [01:47:19]:
Well, so let me ask you one different question, if I may. You go ahead, Robert.
Fr. Robert Ballecer, SJ [01:47:23]:
No, no, no, go ahead. Just ask. I, I was just going to say, when Vera Rubin really makes it into production in large numbers and when you're starting—
Leo Laporte [01:47:30]:
I thought it was just for data centers. I mean, I didn't think I don't think you'd be seeing that on the desktop.
Fr. Robert Ballecer, SJ [01:47:35]:
No, no. I mean, but those data centers that Jess built out, they're going to be extremely pressed to replace everything because a few months of operation, you're losing money now.
Jeff Jarvis [01:47:46]:
So yeah, more efficiencies. If you watch the— I, I'm a connoisseur of Jensen's keynotes, and, and the last one was about that kind of efficiency. All right, here's my question for you.
Leo Laporte [01:47:56]:
But he had— but he sandbagged it. It's— he said He did. It's 67 times— according to SemiAnalysis, eGenic inference is 67 times better performance per dollar.
Fr. Robert Ballecer, SJ [01:48:09]:
Yeah.
Leo Laporte [01:48:09]:
So does that mean it's faster? Uses less power?
Fr. Robert Ballecer, SJ [01:48:13]:
It uses about half as much power per token. It can give you 67 times as much profit per dollar that you spend. I mean, let's, let's do quick math here. So So 1 megawatt-hour of power in the United States is gonna cost you between $46 and $110, depending on where that power's coming from. So that means a gigawatt-hour is between $46,000 and $110,000. A 20-gigawatt data center that is dedicated to training, which means it's running 24/7, is gonna use up between $1.1 million and $2.6 million worth of power each day.
Jeff Jarvis [01:48:48]:
Did you do that in your head?
Fr. Robert Ballecer, SJ [01:48:50]:
Yeah. Oh, about $1 billion worth of power a year.
Leo Laporte [01:48:53]:
How many degrees do you have, Robert?
Lon Seidman [01:48:55]:
Just tell Jeff.
Leo Laporte [01:48:57]:
How many master's degrees do you have?
Fr. Robert Ballecer, SJ [01:48:59]:
I like math. Math is good. Math's fun. But so we're talking about a power bill for one of these major uber-scale data centers going down from $1 billion to $2 billion a year to like $100 million. That is a massive, massive expenditure that you don't have to have. And it's not just that it's more efficient on power. You've got 67% more profit per dollar.
Leo Laporte [01:49:22]:
Right.
Fr. Robert Ballecer, SJ [01:49:22]:
With those kind of economics, I don't see any of the big players not just ripping everything out of their data centers and replacing them with—
Jeff Jarvis [01:49:28]:
Well, well, Huang also recognizes that people have an investment in this stuff. And so he's also talking about CUDA finding more efficiencies in this as well. So here's the question I have for you both. You're using training chips for inference. You're using chips powerful enough to do training. Nvidia bought Groq, the good Groq with the queue for inference chips. So is it— may it not be true that the next machine you want after you get your Vera Rubin, because you've got to scratch that itch, Leo, would be built with entirely different sets of chips that are built for inference?
Leo Laporte [01:50:05]:
It's moving so fast, who the hell knows what it's going to be like?
Fr. Robert Ballecer, SJ [01:50:08]:
It's just— I think, I think Jeff's right on there. I think we're going to start to see them break out the market between training chips and inference chips.
Leo Laporte [01:50:17]:
Oh, I see what you're saying, because right now it's the same.
Jeff Jarvis [01:50:20]:
Yes. Yeah, you're using a Mack truck to drive to the park, right?
Fr. Robert Ballecer, SJ [01:50:26]:
But it's pretty.
Leo Laporte [01:50:28]:
So, okay, so you could drive—
Jeff Jarvis [01:50:31]:
you could build it by a Volvo instead.
Leo Laporte [01:50:33]:
Robert, you weren't here last week because this is— I thought this was kind of interesting. This is a company or a website called ChatGPT. where they have built a small model. I don't know if it's Gemma or something, a very small model, but they built it into the silicon. So it is a silicon? It's a custom, uh, chip with the LLM on the chip. So give me a, give me a prompt so we know. Uh, uh, I'm going— let's see, I am, uh, actually I can explain.
Fr. Robert Ballecer, SJ [01:51:05]:
Where's the best gelato Rome.
Leo Laporte [01:51:07]:
Okay, because you'll know the answer. By the way, Lisa has discovered gelato. This might be the trick to getting her to Rome. Where's the best gelato in Rome? Now don't blink, Robert, because I'm going to hit return.
Jeff Jarvis [01:51:23]:
That's 11 tokens.
Leo Laporte [01:51:25]:
That's done. Oh my God, 14,000 tokens a second.
Jeff Jarvis [01:51:29]:
Uh, that's a little— to explain, to, to list and explain Tesqueria to us.
Leo Laporte [01:51:36]:
Okay.
Fr. Robert Ballecer, SJ [01:51:36]:
Oh, by the way, that list is right, is dead on.
Leo Laporte [01:51:39]:
Of course it is. I don't know what model they're using. I just know that it's built into the silicon.
Jeff Jarvis [01:51:44]:
T-E-S-C-R-E-A-L.
Fr. Robert Ballecer, SJ [01:51:46]:
That's— oh my goodness.
Jeff Jarvis [01:51:48]:
T-E-S-C-R-E-A-L, right.
Leo Laporte [01:51:52]:
That was 14,000 tokens a second. It took 0.37 seconds to give this answer. Tescreal is a popular gelato shop in Rome.
Fr. Robert Ballecer, SJ [01:52:04]:
Okay, so—
Leo Laporte [01:52:06]:
I'm perfect!
Jeff Jarvis [01:52:08]:
You could have the transhumanism.
Leo Laporte [01:52:11]:
You could die. Wait a minute!
Lon Seidman [01:52:13]:
Wait a minute!
Leo Laporte [01:52:14]:
Known for its unique and creative flavors, here's what makes them stand out. Tea Graze Passion, a creative blend of passion fruit, mango, and pistachio. Cannellino, a sweet ricotta-based gelato with a hint of lemon and a crunchy cookie-like texture. Cafe con panna, coffee gelato, and limoncello. Wait a minute. Is there— maybe there is.
Jeff Jarvis [01:52:39]:
No, there's not. I just, I just asked Google.
Leo Laporte [01:52:41]:
It's won several awards. It won first prize at the 2019 Gelato World Championship in Rimini.
Fr. Robert Ballecer, SJ [01:52:46]:
I'm pretty sure it's hallucinating heavily right now.
Jeff Jarvis [01:52:49]:
Gemini says there is no known connection between Tescreale and gelato in Rome.
Leo Laporte [01:52:53]:
It has multiple locations in Rome. It's on the Via del Bosco. In Metro A in the Piazza Venezia. I want you to hear the comments.
Fr. Robert Ballecer, SJ [01:53:01]:
No, no, I know that is a lie for a fact.
Leo Laporte [01:53:05]:
Come on. It's Tescreale. They're famous for their flavors.
Jeff Jarvis [01:53:09]:
Oh my goodness.
Leo Laporte [01:53:09]:
Tescreale is a must-visit gelato shop for any gelato lover traveling to Rome.
Fr. Robert Ballecer, SJ [01:53:15]:
That is the wildest.
Leo Laporte [01:53:17]:
That's definitely local. But let me tell you, it hallucinates at 14,000 tokens a second.
Fr. Robert Ballecer, SJ [01:53:25]:
It's the highest resolution hallucination you will ever have.
Jeff Jarvis [01:53:28]:
Yes, that is.
Leo Laporte [01:53:29]:
What is my spirit animal? If it has a, uh, a gelato-themed—
Jeff Jarvis [01:53:37]:
Maybe it's a new, a new, uh, text now here.
Leo Laporte [01:53:40]:
No, it's a fun— oh, it's gonna give me a quiz.
Jeff Jarvis [01:53:44]:
Well, you asked it. It's asking you.
Leo Laporte [01:53:46]:
It's asking me?
Jeff Jarvis [01:53:47]:
Yeah. It's like, it's like a New York professor. Let me turn the question around to you.
Leo Laporte [01:53:51]:
Let me, let me ask you, what is your spirit animal?
Fr. Robert Ballecer, SJ [01:53:55]:
You know, I think now I need to open a gelato shop called Tescrio.
Leo Laporte [01:53:59]:
Tescrio.
Jeff Jarvis [01:53:59]:
I think so, yes.
Leo Laporte [01:54:00]:
The gelato shop where you're all gonna die.
Jeff Jarvis [01:54:06]:
Bioweapons made for you by computers.
Leo Laporte [01:54:08]:
Oh my God.
Fr. Robert Ballecer, SJ [01:54:10]:
What does transhuman gelato taste like?
Jeff Jarvis [01:54:13]:
Silicon.
Fr. Robert Ballecer, SJ [01:54:15]:
I guess so.
Jeff Jarvis [01:54:18]:
Wow.
Leo Laporte [01:54:19]:
I actually, I should find out more because I've been showing this Jimmy. That's the funniest.
Jeff Jarvis [01:54:24]:
That was good. That was—
Leo Laporte [01:54:27]:
Yeah. I think this is a blog post that explains Jimmy AI. Yeah. Because, oh no, that's not it.
Fr. Robert Ballecer, SJ [01:54:35]:
This is—
Jeff Jarvis [01:54:36]:
And now Gemini is trying to explain why we might have asked this. You might have met Frigiderium or Fata Morgana, exceptional real artisanal gelaterias in Rome, often discussed by their innovative, precise flavor mechanics. Or an AI-themed joke or meme, a niche internet reference blending Silicon Valley tech bro terminology, test realism, with classic European leisure, eating gelato in Rome.
Leo Laporte [01:55:05]:
It's probably just not a very good model, but it is fast. And that's, that's so funny. That's the key.
Jeff Jarvis [01:55:11]:
Jensen Huang says it's in the silicon. There's not a damn thing you can do about it. Ain't no fixing it now.
Leo Laporte [01:55:15]:
You can't fix it. Well, you know, that's what harnesses are for. You sort of, you know, you sort of can. You could. That's, that's kind of important.
Fr. Robert Ballecer, SJ [01:55:24]:
But I mean, you can still use the engine. I mean, the, if the engine can do 14,000 tokens a second.
Leo Laporte [01:55:29]:
In fact, maybe you can explain this to me. There is a new, um, inference engine that just came out. And I signed up for it. That doesn't do language at all. It just answers yes/no questions. This doesn't sound so useful.
Fr. Robert Ballecer, SJ [01:55:53]:
So it's like a bit from Tron?
Leo Laporte [01:55:59]:
No, I feel like I've lost it now. I don't— where is the— where's the story? I found this, um, I actually signed up for, for it. I don't know, you had— it's invite only, but yes, it's from a guy who started, um, OpenAI, was one of the founders of OpenAI. And he said, what if you didn't have to write a lot of prose? I can't find it now.
Fr. Robert Ballecer, SJ [01:56:25]:
So it literally is a bit, it just gives you a yes or no answer to a question.
Leo Laporte [01:56:28]:
Well, but it does it with weights. And so the idea is you would run it through with an AI, but you would use it to— the example that they gave is, here's a conversation I had with a customer. What's the percentage chance this customer is irate? And it's good at saying, it would say 0.9%. Okay, we're gonna escalate. Stuff like that.
Fr. Robert Ballecer, SJ [01:56:53]:
So it's like the engine that Amazon used to have to look at prosody. At a language analyzer.
Leo Laporte [01:57:03]:
Well, it's more than that. Darn it. Ah, here it is. Jev. typesafe.ai introducing System 1 models and Jev. This is a creation of Diogo Almeida, models who was at OpenAI. He said, I helped build the methods that made language models useful at following instructions. And talking with people.
Leo Laporte [01:57:26]:
At the time, I thought maybe chat models would lead to AGI, but despite the hype, it became obvious to me there was something really big missing. After 2 years in stealth, countless technical challenges and research breakthroughs, I'm beyond excited to announce that today, typesafe.ai is releasing our first System 1 model, a new class of frontier models built to make fast, structured decisions. That software can use directly.
Fr. Robert Ballecer, SJ [01:57:53]:
Oh, so it's an, it's an LLM for machines.
Leo Laporte [01:57:56]:
Kind of. It's reinforcement learning for calibrated decisions, or RLCD. Yeah, we built a new stack entirely focused on automation with a new model architecture, parallel sampler for maximum efficiency, and training method. It's so cheap they don't charge you for tokens out. Those are free. You know, normally you charge for tokens in and tokens out. The tokens out are the expensive ones, but these are free and it's pennies for tokens in. And it does— oh, and it can never hallucinate.
Fr. Robert Ballecer, SJ [01:58:30]:
So they've essentially divorced the human speech pattern recognition from the LLM and they've made it just a decision engine.
Leo Laporte [01:58:39]:
They have some examples on their website. For instance, here is the LLM playing Doom. Now, it's not— I mean, it's not super— it costs $7 an hour to do this, 10 queries a second, and it can play, but it's, you know, it's just making quick decisions. It's not goal-seeking, really. But here it's doing something called wiki racing. The game is to start on one Wikipedia page and reach a specific other page Using only links you come across while traversing, which sounds like a lot of fun. So here it is, same thing. It can do it quite quickly by making those choices.
Leo Laporte [01:59:22]:
So they say it's a great playground for demonstrating not just intelligence per second, but also the compounding benefits of not hallucinating with high cardinality choices. So if it seems like a high probability. So it's going to race Claude Haiku, Claude Sonnet, GPT-5 6Tera, and it's already done.
Jeff Jarvis [01:59:45]:
$42 per billion input tokens.
Leo Laporte [01:59:48]:
Yeah, it's very cheap.
Fr. Robert Ballecer, SJ [01:59:49]:
It sounds like this is trying to solve the language paradox for LLMs.
Leo Laporte [01:59:54]:
Yeah, because it doesn't write any prose. It can't.
Fr. Robert Ballecer, SJ [01:59:56]:
Exactly. Because, I mean, we are judging LLMs by how well they write human language, but human language is so imprecise.
Jeff Jarvis [02:00:06]:
Yes.
Fr. Robert Ballecer, SJ [02:00:06]:
That you're actually asking an LLM to dumb down its answers. So if you don't have to do that, and if you're just allowing an LLM to act on the intelligence built into its, its own training, you are essentially removing the most expensive compute from the LLM and just taking the actionable intelligence items. But that makes sense.
Jeff Jarvis [02:00:28]:
But the question is, it's given our text though, it has to decode that.
Lon Seidman [02:00:32]:
That's correct.
Jeff Jarvis [02:00:33]:
Let's tokenize that.
Leo Laporte [02:00:34]:
Right.
Fr. Robert Ballecer, SJ [02:00:34]:
We still have to tokenize that.
Leo Laporte [02:00:35]:
Find a gelato shop named Tescreal, I think. And that's the downside of that. You never— by the way, here is our band, Gelato and the Large Linguine Models. Oh, chat.
Fr. Robert Ballecer, SJ [02:00:50]:
Oh, you will forever fascinate me.
Leo Laporte [02:00:54]:
Oh, you wild chat.
Jeff Jarvis [02:00:56]:
Um, did anybody give you— have you driven your, uh, Spark? Did anybody follow my instructions?
Lon Seidman [02:01:03]:
No.
Jeff Jarvis [02:01:04]:
I'm hurt, guys. I'm hurt.
Leo Laporte [02:01:05]:
Oh, you wanted a—
Jeff Jarvis [02:01:06]:
I wanted— I wanted— yeah, well, there you go.
Leo Laporte [02:01:08]:
Picture me driving down the street in a— in my Spark.
Jeff Jarvis [02:01:11]:
Well, Darren did a hot rod, Vera Rubin inside.
Fr. Robert Ballecer, SJ [02:01:19]:
By the way, the AI-generated image of Tescriel, that actually is a gelato place. I think that's Teatro di Fragole. They just put the Tescriale name on the back.
Leo Laporte [02:01:29]:
So you've been to this. Do you know every gelato place in Rome?
Fr. Robert Ballecer, SJ [02:01:34]:
Sadly, probably.
Leo Laporte [02:01:36]:
Gelato flavors are amazing.
Fr. Robert Ballecer, SJ [02:01:38]:
This one is on the way from— this is across the Tiber from where I'm living towards the—
Leo Laporte [02:01:45]:
Is that St. Peter's Dome in the distance, or is it—
Fr. Robert Ballecer, SJ [02:01:47]:
No, no, that's—
Leo Laporte [02:01:48]:
there's so many.
Fr. Robert Ballecer, SJ [02:01:49]:
It's a lesser dome.
Leo Laporte [02:01:50]:
There's so many.
Fr. Robert Ballecer, SJ [02:01:52]:
cheese.
Leo Laporte [02:01:52]:
It's on the Villa Scrofa.
Jeff Jarvis [02:01:55]:
What's your favorite flavor, Robert?
Fr. Robert Ballecer, SJ [02:01:57]:
Uh, I like dark chocolate, so cioccolato fondente, which is really— it's really, really dark, uh, mixed with ciliegie, dark cherry.
Leo Laporte [02:02:08]:
Oh my gosh, that sounds really good. I'm kind of more of a Spumoni guy, but, uh, you know, they do Spumoni too.
Fr. Robert Ballecer, SJ [02:02:14]:
Yeah, they do the traditional flavors. When you come here every day, we will just walk across the street to Old Bridge, which is one of my favorite gelaterias. And then you just take your cup and sit in front of St. Peter's and there you go.
Jeff Jarvis [02:02:25]:
See, the funny thing is, we'll come to this wonderful, amazing place with all of these attractions and just sit and stare at the Sparks.
Fr. Robert Ballecer, SJ [02:02:32]:
Yep.
Leo Laporte [02:02:33]:
No, no, I might actually. If I have 8 of them.
Jeff Jarvis [02:02:36]:
You might, yeah.
Leo Laporte [02:02:37]:
Yeah, I might. I might just watch.
Jeff Jarvis [02:02:38]:
Wow, 8 of them.
Leo Laporte [02:02:40]:
8 of them. Anything else we should talk about? Universal Music is launching a new AI music platform. to combat Suno, but ElevenLabs is working with them and they do some of the best voice stuff ever. It'll be, uh, you— but the idea is that artists will get royalties and users will be creating remixes, mashups, and new takes on tracks, so the original will be more clearly obvious.
Fr. Robert Ballecer, SJ [02:03:08]:
You know, I'm kind of played out on that. I, I went through a phase where I was really enjoying the, the mashups and the mix-ups But now I'm just finding artists who I really enjoy. And like, I am so into Twenty One Pilots over the last 3 years.
Leo Laporte [02:03:20]:
Real music.
Fr. Robert Ballecer, SJ [02:03:21]:
I don't know why.
Leo Laporte [02:03:21]:
Music performed by humans. Yeah, real music.
Fr. Robert Ballecer, SJ [02:03:25]:
Yeah.
Leo Laporte [02:03:25]:
You gotta be careful if you're using an AI for law. By the way, I saw Thomson Reuters ad on the football game on Monday Night Football for their specialized legal LLM for law offices. I was shocked. We talked about it on the show.
Jeff Jarvis [02:03:42]:
Yeah, we did.
Fr. Robert Ballecer, SJ [02:03:42]:
So do they offer you insurance if the LLM gives you imaginary—
Leo Laporte [02:03:47]:
That is a hot topic right now. LLM insurance. In fact, I think I have a link in here. Yeah. For liability if the LLM— maybe this lawyer could have gotten us. He's been fined $5,000 because he provided witnesses that were hallucinated.
Jeff Jarvis [02:04:03]:
That's so awful. Worse than psychedelics.
Fr. Robert Ballecer, SJ [02:04:05]:
Not just witnesses, precedents. So he, like legal cases and witnesses that did not exist.
Jeff Jarvis [02:04:10]:
These lawyers are never learning.
Fr. Robert Ballecer, SJ [02:04:12]:
Yeah, he said, I didn't know that it would hallucinate. It's like, well, you know what, if you don't know the technology, maybe don't use it before you submit something so serious.
Leo Laporte [02:04:20]:
He submitted a brief with false testimony from wholly fabricated witnesses about the shooter's clothing and appearance. Your Honor, I didn't know. Counsel, do you watch the news? Do you listen to the radio? Do you read any anything about what's going on in the world, said Justice C. Shannon Bacon. Because the problem with lawyers relying on AI hallucinations is an above-the-fold story every single day.
Jeff Jarvis [02:04:49]:
I covered the original case. It went to the court, the federal court, and heard him. And this original guy, schmuck, had an excuse. He said, I thought it was a super search engine. I didn't know it could lie.
Leo Laporte [02:05:01]:
Well, that's actually reasonable.
Jeff Jarvis [02:05:02]:
At that time, at that time, that was reasonable. Moment in history where you could say that.
Leo Laporte [02:05:07]:
Yeah, you can't say it anymore.
Jeff Jarvis [02:05:08]:
Dead on.
Leo Laporte [02:05:08]:
Not if you listen to this show.
Fr. Robert Ballecer, SJ [02:05:12]:
But can you imagine being a judge and reading the brief and just going, what is this guy on about?
Leo Laporte [02:05:20]:
Wholly fabricated witnesses.
Jeff Jarvis [02:05:23]:
So Mistral, just as we went on, just announced a deal with Mozilla.
Leo Laporte [02:05:28]:
Yes, they're going to be the AI built in. To the browser, to the Firefox browser.
Jeff Jarvis [02:05:33]:
It's very nice.
Leo Laporte [02:05:34]:
Which gives them, uh, access to the EU, which is important because it is of course a French model. Uh, last chance before we take our final break and get your picks of the week. Anything, uh, you want to put in there that you really want to do?
Fr. Robert Ballecer, SJ [02:05:51]:
No, there, you know, there was so much this week I can't just pick one.
Leo Laporte [02:05:55]:
Well, I'm, you know, I'm kind of glad we, we did the high-level Overall look at it. I'm still a little concerned, Robert, about your assertion at the very beginning of the show that doom is inevitable.
Fr. Robert Ballecer, SJ [02:06:09]:
Doom has been inevitable for so long, Leo.
Leo Laporte [02:06:11]:
That's true.
Jeff Jarvis [02:06:11]:
Ah, the long-term perspective.
Leo Laporte [02:06:14]:
Judgment Day is just around the corner. What is your timeframe for P-doom?
Fr. Robert Ballecer, SJ [02:06:21]:
No, I think you're going to probably within the next 5 years, you're going to see a major economic event. Because the allure for fast trading with LLMs, it's too rich. There's too much possible profit there. And I think it's very unlikely that you're going to have some Wall Street brokerages using LLMs without one of them just crapping the bed and causing a major disaster.
Jeff Jarvis [02:06:48]:
Which we kind of saw before LLMs. We saw it with We just—
Fr. Robert Ballecer, SJ [02:06:51]:
yeah, we just— smart systems. Yeah, yeah, the smart systems have crashed out. So an LLM, absolutely. I mean, because why are you going to use an LLM for trading? Because it's really good at pattern recognition, which is what the stock market is. Well, if you're, if you're using it for, for pattern recognition, you have to give it the authority to make trades on its own. It's got to be autonomous. One of those is going to cause a major disaster. I think that's unavoidable.
Leo Laporte [02:07:14]:
I have been tempted to, um, start trading with my own—
Fr. Robert Ballecer, SJ [02:07:20]:
with an LLM.
Leo Laporte [02:07:21]:
Yeah, well, I would do it first— I would do it with funny money just to see how it performed, right? Yeah, I've done pretty well myself. I buy— I mean, I buy something called Retirement 2025.
Fr. Robert Ballecer, SJ [02:07:36]:
If you've got the cash, put $1,000 into an LLM-controlled account and tell it to be aggressive. Just See what it does.
Leo Laporte [02:07:42]:
See what it does.
Fr. Robert Ballecer, SJ [02:07:43]:
Yeah.
Leo Laporte [02:07:45]:
It places all its chips on red, and if it loses, it doubles the bet.
Fr. Robert Ballecer, SJ [02:07:50]:
I mean, why not? Why not?
Leo Laporte [02:07:53]:
What could possibly go wrong? All right, you're watching Intelligent Machines. Uh, Paris, I think we'll be back next week. Will you be back next week, Jeff?
Jeff Jarvis [02:08:01]:
Yeah, I, I'm probably gonna take the red-eye back, so yes, I'll be back.
Leo Laporte [02:08:03]:
Oh, good, good. So the full, full quota will be here. But boy, I wish we'd get Robert on every week. We love having you on. You're so fantastic. Robert Balaser, the Digital Jesuit. Staying up late. Last time you were on, you had to get up early for Mass underneath.
Jeff Jarvis [02:08:21]:
He didn't get up. He just went there.
Leo Laporte [02:08:23]:
Just went straight.
Fr. Robert Ballecer, SJ [02:08:24]:
I took a shower and then went there. So.
Leo Laporte [02:08:28]:
I did a little faux pas the other day. My daughter, who has converted, has become a very devout Catholic, Catholic. Yes. Introduced me to her spiritual advisor, the Father. And I won't say his name because, you know, privacy and all that. But he said, oh, and he said, you're the earthly father. And I said, and yes, and you're the Holy Father. And my daughter went, that's the Pope, you nitwit.
Leo Laporte [02:09:01]:
So next time I won't say that. Heavenly? Can I say Heavenly Father?
Fr. Robert Ballecer, SJ [02:09:06]:
No, that's easy to get. Spiritual Father is fine.
Leo Laporte [02:09:09]:
Spiritual Father would have been right.
Fr. Robert Ballecer, SJ [02:09:10]:
Every time my dad introduces me, he says, this is Father, and I'm the father of the Father.
Leo Laporte [02:09:15]:
I'm the father of the Father. That's good. I like that. I thought that was a good line that her spiritual advisor said, oh, you're the earthly father. I said, yes, you're the Heavenly Father. I should have said that. Not the Holy Father.
Fr. Robert Ballecer, SJ [02:09:27]:
That was wrong. Holy Father would be the Pope. Heavenly Father would be God. So you want us to—
Leo Laporte [02:09:32]:
It could have been worse. It could have said Heavenly Father.
Fr. Robert Ballecer, SJ [02:09:35]:
It's picked with Spiritual Father.
Leo Laporte [02:09:37]:
Spiritual Father.
Lon Seidman [02:09:38]:
Spiritual.
Leo Laporte [02:09:38]:
Thank you for your advice. Uh, pick—
Jeff Jarvis [02:09:41]:
come on, come on, come on, people. Come on, people. NPR, an NPR station gets between 6 and 12% of their audience.
Leo Laporte [02:09:47]:
God. But they're on— they beg all the time and they do the pledge.
Jeff Jarvis [02:09:51]:
Yeah, they're obnoxious about it. You're not.
Leo Laporte [02:09:52]:
I'm trying not to be.
Fr. Robert Ballecer, SJ [02:09:54]:
I mean, I'm not going to brag, but when I'm giving my homilies, I'm upwards of 80 to 95%.
Jeff Jarvis [02:10:01]:
Yeah.
Leo Laporte [02:10:02]:
See, I need a basket on a long pole that I could just stick down each pew and say, put money in it. This is the latest— I'll start. This is the latest meme on the internet. Have you seen this, Robert? A fruit fly brain which has been mapped so you can actually use it. People are using it to do all sorts of things. Play Doom. It actually plays Doom better than that other model. They're using it.
Leo Laporte [02:10:36]:
See, there's the brain. This is from HHMI Janelia Research Campus. And Google Research, they've mapped a fruit fly brain and the female connectomes. And so, and you can actually run this. I guess I could run it on my Spark. Here it is flying around, but people are doing all sorts of mean things to the— I don't know. Does it have any feelings, this fruit fly brain? They're having it do a scroll eternally. Look, it's flying around.
Leo Laporte [02:11:17]:
It's kind of cool.
Fr. Robert Ballecer, SJ [02:11:21]:
Well, now you're gonna have to get another Spark just for that.
Leo Laporte [02:11:24]:
Just for the fruit fly brain. Everybody's talking about the fruit fly brain. And then the other thing I had, I just have to show you this because both Jeff and I Lisa said, quick, turn on CNN. And, uh, who's talking?
Fr. Robert Ballecer, SJ [02:11:45]:
Anderson?
Leo Laporte [02:11:46]:
It might be one of my AIs talking to me.
Fr. Robert Ballecer, SJ [02:11:51]:
Do you have an AI describing what's on CNN?
Jeff Jarvis [02:11:53]:
I'm hearing voices.
Leo Laporte [02:11:54]:
I guess something's going on. Let me just close all my tabs.
Fr. Robert Ballecer, SJ [02:11:57]:
Maybe it's the fruit fly. Is the fruit fly speaking? Is the fruit fly in the room with you now, Leo?
Leo Laporte [02:12:05]:
If you could make it talk, that would be good. Does a fruit fly connectome have feelings is the question. Anyway, she said, Leo, quick, turn on CNN. And I thought, oh my God, you know, is something—
Jeff Jarvis [02:12:22]:
we're at war.
Leo Laporte [02:12:22]:
What's going on? I went in and it was Jacob Coxon. The guy who quit Anthropic on the TV, Anderson Cooper was interviewing him, but I wish it had been this. Let me turn on the sound. Joining us now is John Connor.
Jeff Jarvis [02:12:43]:
John's been trying to warn humanity about the dangers of AI for many years.
Leo Laporte [02:12:46]:
Welcome to the studio, John. Tell us what you think.
Fr. Robert Ballecer, SJ [02:12:49]:
Oh, that's—
Lon Seidman [02:12:50]:
look, man, I've been saying the same thing for almost 4 friggin decades. The AI apocalypse robots are already here.
Leo Laporte [02:12:59]:
Yeah, but don't you think that's a little premature? I mean, wait, um, who's, um, who's, uh, that guy? Uh, that guy was of course Arnold in his Terminator outfit. This is completely fake. I gotta say though, I'm loving it that AI video now can do anything absolutely realistically. There's— it's— there's no 6 fingers, there's nothing. It's amazing.
Jeff Jarvis [02:13:27]:
I just, I just saw a video with, uh, combining a woman who's screaming on an airplane, a Karen, an airplane Karen, you know, you're not real, and she points to a picture of Mitch McConnell sitting in a chair, and it works. He's in the chair. It works. It works. Yeah.
Leo Laporte [02:13:43]:
Uh, Padre, do you have a pick? Oh, wait a minute. Oh, do you have a pick? I do.
Fr. Robert Ballecer, SJ [02:13:49]:
Okay, I kind of got this as a joke, uh, at first because I, I go camping when I come back to the United States and, uh, I miss coffee. I've been drinking that, uh, uh, instant coffee from Starbucks. No, no, no, no, no, no, no, no, it's, it's not good. I wanted some espresso. I, I, you know, not just muddy real coffee. Yeah, real coffee. So I may have picked up one of these.
Leo Laporte [02:14:12]:
Oh, this is Oh my God, is that an espresso— a portable espresso maker?
Fr. Robert Ballecer, SJ [02:14:16]:
Yeah, this is the Fanttik Caffeine 11. It's a portable espresso machine. So, you know, it uses grounds or it uses, um, those, uh, not the K-Cups but the, the espresso cups. And, uh, it makes a single serving, 80 milliliters. It heats up the water, it pushes it through at 20 bars of pressure.
Jeff Jarvis [02:14:36]:
It has the Google color bar.
Fr. Robert Ballecer, SJ [02:14:38]:
It's— yeah, this is a This is telling you that it's heating up. By the way, Leo, I combine it with this. I got this when I was in Indonesia. This is Luwak coffee.
Leo Laporte [02:14:49]:
Luwak, that's the civet coffee.
Fr. Robert Ballecer, SJ [02:14:51]:
Yeah, this is the poop coffee.
Leo Laporte [02:14:53]:
The civet cat eats it.
Jeff Jarvis [02:14:55]:
Really good.
Leo Laporte [02:14:56]:
It softens it. It mellows it.
Fr. Robert Ballecer, SJ [02:15:00]:
It's not acidic. Yeah. Jeff, the coffee beans have been eaten by a little cat. And they— that poops it out after it digests it. People gather it, they clean it, and they don't have to roast it because the digestive juices—
Leo Laporte [02:15:12]:
It's still beans.
Fr. Robert Ballecer, SJ [02:15:13]:
Have roasted it.
Leo Laporte [02:15:14]:
It's not—
Fr. Robert Ballecer, SJ [02:15:14]:
Still beans.
Leo Laporte [02:15:15]:
Yeah, you have to grind it still and all that.
Fr. Robert Ballecer, SJ [02:15:17]:
You grind it, and then I just put it in here and it makes delicious juices.
Leo Laporte [02:15:22]:
Is it making— is it doing it right now?
Jeff Jarvis [02:15:24]:
Yeah. How are you gonna sleep? Just push the button.
Leo Laporte [02:15:26]:
Now, Fanttik makes the wildest stuff. I have one of their electric screwdrivers. I love their stuff.
Fr. Robert Ballecer, SJ [02:15:34]:
Uh, I, I honestly thought this was going to be a joke, but it's really good.
Leo Laporte [02:15:37]:
They're not Italian coffee. I think it's a Chinese company, right?
Fr. Robert Ballecer, SJ [02:15:41]:
It's, uh, maybe French.
Leo Laporte [02:15:43]:
French? Fantic? That's why you said Fantic? Oh yeah, it does look like a French cafe. But I mean, IP65 dust and moisture proof.
Fr. Robert Ballecer, SJ [02:15:57]:
I charged it once before I started my camping trip and it gave me 3 cups of coffee every day.
Leo Laporte [02:16:02]:
It says 8 shots. They promise 8 shots per charge, so that's enough for, uh, an hour or 2 anyway.
Jeff Jarvis [02:16:10]:
For you.
Fr. Robert Ballecer, SJ [02:16:11]:
But this is, this is not coffee, this is espresso.
Jeff Jarvis [02:16:14]:
It's—
Leo Laporte [02:16:14]:
I really want this.
Jeff Jarvis [02:16:16]:
Do you use spring water for that?
Fr. Robert Ballecer, SJ [02:16:18]:
No, just regular water. Just—
Jeff Jarvis [02:16:20]:
but no, when you're out camping.
Fr. Robert Ballecer, SJ [02:16:22]:
Oh no, I, I always bring my water with me.
Jeff Jarvis [02:16:23]:
You bring water?
Leo Laporte [02:16:24]:
Okay. Yeah, it's good.
Fr. Robert Ballecer, SJ [02:16:25]:
My camping is not your camping. I, I Car camp. It's glamping.
Leo Laporte [02:16:31]:
Let's have a taste. Let's have a little taste test. How did it taste, the Civet coffee, the kepi luwak?
Fr. Robert Ballecer, SJ [02:16:39]:
Meow.
Leo Laporte [02:16:43]:
Well, at least it's less than a DJX Spark. That's a— that's, uh, that's what you could say. Larry says that. Uh, Jeff Jarvis, your pick of the week.
Jeff Jarvis [02:16:53]:
Slim pickings. Um, we could mention that Casey Newton and Kevin Ross— uh, Kevin Roose are moving their show to NPR.
Leo Laporte [02:17:03]:
I saw that. And, you know, I'm not a huge— I think Kevin Roose is a good writer. He was at the New York Times, but he also was not the best writer about AI. It's not an AI show, is it?
Jeff Jarvis [02:17:13]:
Uh, it's called Machine Gods, so I think it will be.
Leo Laporte [02:17:16]:
They stole our idea.
Jeff Jarvis [02:17:18]:
Yeah.
Fr. Robert Ballecer, SJ [02:17:19]:
You were going to call the show Machine Gods?
Leo Laporte [02:17:21]:
No, I like intelligent machines, but they decided to do it one better.
Jeff Jarvis [02:17:25]:
Ah, the God reference.
Leo Laporte [02:17:27]:
And then the other thing— Casey Newton is great, actually. I like Casey Newton.
Jeff Jarvis [02:17:29]:
Casey Newton's good. Roose is the one— Roose is the one who thought that the model fell in love with him and wanted to destroy his marriage and all that stuff. He anthropomorphizes to an extreme.
Leo Laporte [02:17:37]:
So they used to do Hard Fork, so they've moved to NPR.
Jeff Jarvis [02:17:41]:
Yes. Yes. Okay.
Fr. Robert Ballecer, SJ [02:17:41]:
Yes.
Lon Seidman [02:17:42]:
Okay.
Jeff Jarvis [02:17:42]:
So they didn't go out on their own. They went from one big media organization to another. Yeah. Uh, the other thing to mention is that a New York Times story, uh, they built a shrine to cable TV. Then everyone cut the cord. So there was a cable— the Cable Labs were in Denver because—
Leo Laporte [02:17:59]:
Oh yeah, I remember this.
Jeff Jarvis [02:18:00]:
Right. And John Malone was there. And so all things gravitated to John Malone. So they built this huge building with a museum and everything else because cable TV, because we all care about cable TV.
Leo Laporte [02:18:11]:
Nobody cares about cable TV. Who has cable TV anymore?
Jeff Jarvis [02:18:15]:
No.
Fr. Robert Ballecer, SJ [02:18:15]:
Although, I mean, the whole idea of cutting the cord was that you could get it cheaper on the internet, but now with all the streaming, Doing different pricing, it's probably more expensive now than it was with cable.
Leo Laporte [02:18:24]:
And you're still— and the cable companies weren't stupid. They just said, well, we've got all this coax in the ground, we'll just put internet. I mean, I'm on Comcast Cable right now as my internet provider because they have monopolies. They're already in every market and they don't have any competition.
Fr. Robert Ballecer, SJ [02:18:41]:
Well, in Vegas, we used to have a decent amount of competition between— you had Quest, then you had AT&T, you had Verizon, you had Cox.
Lon Seidman [02:18:49]:
Wow.
Fr. Robert Ballecer, SJ [02:18:49]:
But, uh, and Gigafiber. And then AT&T just bought Quest and Gigafiber, and they're buying the Cox licenses. So it's basically they're just reassembling a monopoly.
Leo Laporte [02:19:02]:
Yep. I love it. This museum, the Cable Center's museum, includes a TiVo remote. I guess that's, uh, that's, uh, ancient history now, that TiVo Wow.
Fr. Robert Ballecer, SJ [02:19:15]:
I wonder if there's any original TiVo users whose hard drives haven't died yet, like the grandfathered ones.
Leo Laporte [02:19:22]:
I had a bunch of TiVos. When we moved, I let them all go. I had lifetime, you know, lifetime subscriptions.
Jeff Jarvis [02:19:29]:
Um, and they just changed the definition of lifetime.
Leo Laporte [02:19:32]:
I said, Lisa, uh, I know you like the TiVo, but I think YouTube TV— we're gonna move Might as well just do it all over the top, over the internet. And I tell you what, nobody's missed cable. Nobody's missed TiVo.
Fr. Robert Ballecer, SJ [02:19:47]:
Well, Leo, in one of the giveaways, the famous giveaways at the Brick House, uh, you had an old, old Kindle, and it was one of the ones that had cell service, like forever cell service. And so I disassembled it and I pulled the modem out, and I, I had been using that for years and years.
Leo Laporte [02:20:04]:
It came with with free self-service.
Fr. Robert Ballecer, SJ [02:20:05]:
There was no charge, and it was forever. You bought it forever, and then they finally killed it last year.
Jeff Jarvis [02:20:11]:
They did.
Leo Laporte [02:20:12]:
You were using it until last year?
Fr. Robert Ballecer, SJ [02:20:14]:
Yeah. Oh, I mean, it was, it was great for, uh, like SCADA devices. So I had a couple of devices that would just report on temperature.
Leo Laporte [02:20:20]:
It's pretty low bandwidth, I'm sure.
Fr. Robert Ballecer, SJ [02:20:22]:
Very low bandwidth, but I mean, for that it was fine, right?
Leo Laporte [02:20:26]:
Um, wow.
Jeff Jarvis [02:20:28]:
Was it 3? Um, Level 3? It wasn't 4 or 5?
Fr. Robert Ballecer, SJ [02:20:33]:
No, it was like, I think he— it was the first one.
Leo Laporte [02:20:36]:
I mean, it was the very first one, and probably the reason it died is because it isn't, uh, the latest, uh, cable— I mean, cellular technology.
Jeff Jarvis [02:20:45]:
That's what I'm asking.
Leo Laporte [02:20:46]:
It was probably analog.
Lon Seidman [02:20:47]:
Yeah.
Jeff Jarvis [02:20:48]:
Oh, right.
Leo Laporte [02:20:49]:
I'm sure it was.
Fr. Robert Ballecer, SJ [02:20:50]:
This is, this is pre-LTE, so.
Jeff Jarvis [02:20:52]:
Right.
Leo Laporte [02:20:53]:
Yeah. Uh, it was 2G. Patrick Delahanty, our esteemed engineer, says, I still have a TiVo HD with a lifetime subscription and I still use it.
Fr. Robert Ballecer, SJ [02:21:03]:
Damn it.
Leo Laporte [02:21:05]:
So there you go. Somebody, somebody does it. Somebody—
Jeff Jarvis [02:21:08]:
Leo Laporte wrote the Guide to TiVo. I did.
Leo Laporte [02:21:10]:
I did. With Gareth Branwyn.
Fr. Robert Ballecer, SJ [02:21:14]:
Yep.
Leo Laporte [02:21:14]:
That was first generation TiVo. And there it is, the TiVo remote right on the COVID Actually, I have a copy over here. Father Robert, so great to see you. Get to bed. We appreciate it. Although you just had a shot of espresso, you may not be sleeping soon. I miss you. I miss having you in a brick house.
Leo Laporte [02:21:33]:
But remember, I gotta get out to Rome.
Fr. Robert Ballecer, SJ [02:21:35]:
You are always most welcome. You just tell me when.
Leo Laporte [02:21:39]:
I will. Absolutely. And I think you're stuck there. I hate to say it.
Fr. Robert Ballecer, SJ [02:21:42]:
For a while. Yeah.
Leo Laporte [02:21:43]:
I don't think you're coming home. That's okay. You got a higher mission. You got a job to do.
Jeff Jarvis [02:21:49]:
That's for sure.
Fr. Robert Ballecer, SJ [02:21:49]:
Well, they let me travel back to the US 3 times a year.
Leo Laporte [02:21:52]:
That's, that's good. You got it. You got a good gig. Thank you, Robert. Really appreciate it. Always wonderful. Padre SJ everywhere. Uh, of course, Jeff Jarvis's book is available on his website, jeffjarvis.com, or go to Montclair, New Jersey.
Leo Laporte [02:22:08]:
You don't have to go to Montclair.
Jeff Jarvis [02:22:09]:
You don't have to go there. Go to jeffjarvis.com. You see the link and you can get a signed copy.
Leo Laporte [02:22:13]:
Nice. Very nice. It's really a good book. Highly recommend it. Very interesting. The story of the linotype.
Jeff Jarvis [02:22:20]:
And on October 10th in Haverhill, Mass., I'll be giving a big presentation of the book in the Historic New England and then going up the hill to the Museum of Printing where you can see it live, a linotype operate.
Leo Laporte [02:22:34]:
Nice. Oh, that's cool, man. I wish I were out there. Maybe my Google Dream Beans will tell us I should be going out there for that.
Jeff Jarvis [02:22:41]:
Yeah.
Leo Laporte [02:22:41]:
Yeah.
Lon Seidman [02:22:43]:
Thank you, Jeff.
Jeff Jarvis [02:22:43]:
I can't use Dream Beans because I'm on Workspace.
Leo Laporte [02:22:46]:
Oh, Paul Theroux turned it on and he was like going on and on about it. It's the strangest thing because, you know, like people you work with show up in it. It's just weird. All right. Thank you, Robert.
Jeff Jarvis [02:23:00]:
All right.
Leo Laporte [02:23:00]:
Thank you, Jeffrey. Thank you everybody for watching. Our Club Twit members, our deepest gratitude. Thank you so much. We do the show Intelligent Machines every Wednesday, 2 PM Pacific, 5 PM Eastern, 2100 UTC. You could join us, uh, by watching live if you wish in the club on the Discord or YouTube, Twitch, x.com, Facebook, LinkedIn, or Kick. After the fact, get— he's drinking more coffee. After the fact, get a— get the show, a copy of the show, audio or video at our website, twit.tv/iam There's a YouTube channel, of course, or subscribe in your favorite podcast client.
Leo Laporte [02:23:36]:
That's the best way. That way you're supporting independent podcasting. Thank you, everybody. We'll see you next time on Intelligent Machines.
Fr. Robert Ballecer, SJ [02:23:44]:
Bye-bye.
Jeff Jarvis [02:23:45]:
I'm not a human being, not into this animal scene. I'm an intelligent machine.
