TWiT+ Club Shows 764 Transcript - AI User Group #19

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]:
This is TWiT. Hey everybody, time for the AI User Group, one of the most fun things we do. Well, at least for me, maybe not for you, but if you're watching, probably it is for you. Let me introduce our stunning panel of AI genii. Of course, Darren Oakey is always here from Down Under. Hello Darren, good to see you. Also joining us, Larry Gold, L-R-A-U. He is in the— on New York Time.

Leo Laporte [00:00:31]:
Uh, you ride the Staten Island Ferry, I know that much, to work. We won't say where, uh, but it's probably somewhere in, I don't know, Midtown Manhattan. Yeah, I'm just, I'm just thinking. Hello, Alakazip Michael in the darkened, uh, theater. Look at this. Are you— oh, you're in The Matrix. Congratulations.

Michael (Alakazip) [00:00:49]:
I was trying a background out. Yeah, I don't have as nice a set here, so now that I figured out this is on the internet, which Great, but, uh, yes, thank you.

Leo Laporte [00:00:57]:
Everything, baby. Everything's on the internet now. Uh, Timothy Ingalls, Nerdy Drunk, hello. I see you have a little rabbit on your chest. Is that an R1 rabbit? My spirit animal. Oh, it's got rabbit with chainsaws. Or is it a rabbit? I don't even know what that is.

Michael (Alakazip) [00:01:15]:
That's a rabbit.

Leo Laporte [00:01:16]:
Okay. And Dano, hello Dano. He is a CEO type. And we're all using local models. Jose also visiting us and Blind Whiz as well. Blind Whiz, I finally have parity with you.

Juan Hernandez (BlindWiz) [00:01:30]:
I know. Welcome to the club.

Leo Laporte [00:01:32]:
I was so long, so jealous. Blind Whiz said, yeah, I have 2 Sparks.

Larry Gold (LrAu) [00:01:37]:
What of it?

Leo Laporte [00:01:39]:
Darren has one, but he's gonna have 2 before the end of the week, I guarantee you. 'Cause, you know, keeping up with the Blind Whizzes. So I spent a little money, spent my kids' inheritance, bought uh, dual Sparks because one isn't enough. You know what put me over the top was the fact that you could run DeepSeek V4 Flash on dual Sparks at full resolution, and I really like DeepSeek V4 Flash. I was using it since it came out July 31st full-time as my agentic model on Hermes just to test it out, and I felt like it's good. It started getting a little weird in the last few days, and then DeepSeek said we want to charge you a hell of a lot more, which is still a hell of a lot less than the Frontier. models from the US. Uh, and so I started— I got a little squirrely and I thought, no, I really want to do this local.

Leo Laporte [00:02:29]:
And then I thought, well, $10,000, that's not enough. So I bought a 3090 card to— I didn't realize I could do this— to upgrade my old gaming machine, the, the Alienware that I bought during COVID for playing Valheim. Since I don't game anymore, my new game is Agentic AI. Uh, I put a 3090 in there and that runs great. And then I realized, you know, that Mac Mini— I don't really like Macs that much. I'm going to turn that 64GB M4 Pro into a local machine too. So now I have— and then I have the Framework, which I bought originally. I thought that's where I'm going to run everything local until I realized that, uh, the AMD 395 Plus is great as it is and ONNX aren't really all the thing.

Leo Laporte [00:03:13]:
I finally have a little bit of CUDA. In my life. You got to get a little CUDA in your life. Not to be too pedantic. So that's catching me up on the hardware front. And then, my God, it's full of stars. DeepSeek V4 Flash 7.31 came out July 31st. I'm running that on the Sparks.

Leo Laporte [00:03:40]:
I'm kind of excited. GLM-5.3 came out this morning. It's not yet on Hugging Face. It's not yet on Unslotted, so I can't run it, but at least last time I checked. But that's— I thought 5.2 was very good. They said we didn't do any— it's not a new model. It's just additional post-training. But the things like the terminal benchmark doubled.

Leo Laporte [00:04:03]:
So something happened in the post-training. And then Quen, who disappointed us earlier this week by not— they only announced Quen 3.8 Max, which I was running in its earliest form. I'm sure it's better, but I was not that crazy about it. And I've always liked Quen 3.5.27b. Well, this morning, 5 hours ago, they shipped open weights of that, and that's up everywhere now. It's on Sloth and—

Darren Oakey [00:04:30]:
Nice.

Leo Laporte [00:04:31]:
Hugging Face. So I downloaded that. I thought I'd put that on the 3090, but then it turns out I can run on the Mac, and it's a little more accurate on the Mac. It's using, uh, MLX and, uh, what is it, FVP4 or PFV4? No, whatever it is.

Juan Hernandez (BlindWiz) [00:04:46]:
Or is that— are you doing—

Leo Laporte [00:04:47]:
No, it's 4. It's a 4-bit quant, but it's apparently a very good quant. And I couldn't use anything, uh, on the— so apparently the CUDA, it's, it's the older CUDA on the 3090, so I can't— I could run it on the Sparks. Full resolution. But actually, I benchmarked it. In fact, let me show you the benchmarks because I had, um, I put it on both the 3090 and on the Mac. Uh, where's the benchmark results? And actually, uh, it was kind of interesting to run it on— oh no, that was after the migration. I've been working all morning playing with this stuff.

Leo Laporte [00:05:21]:
Anyway, I'm running— ended up running on— oh, here you go. So here's the benchmarking. Uh, this is the 3090 Mojo. This is the M4 Pro, 64 gigs. First token is faster on the Mac, which is surprising, but decode is slower. It's 27 tokens per second compared to 40 on the 3090. Uh, it passed the 17 sheep reasoning trick, both of them. Uh, so it's— these are little minor tests.

Leo Laporte [00:05:51]:
Uh, accurate, a little more precise than the Mac. And the imagery for writing in the San Francisco Rainstorm Prose Test, which I don't know what any of these are, I guess Hermes found them, is better. So the Mac, even though it's a little slower, it's a little better on the reasoning. Plus the Mac runs cooler. The 3090, I had it running all day yesterday doing something called Graft, which is kind of— I'll tell you about it. It's an interesting project that takes a code base and builds a graph so that the Agents don't have to read the code base, they read the graph, which speeds it up a lot. They don't have to grep everything all the time, um, and really reduces token use. So I was running the 3090 and it just gets so hot.

Leo Laporte [00:06:36]:
The room was 80— you were up here, Anthony, it was like 85 degrees. Yeah, it was going—

Anthony Nielsen [00:06:40]:
it was toasty.

Leo Laporte [00:06:43]:
And on the Mac, it's, it's quiet, low energy. So I think I'm gonna just run I'm running Whisper Large on the Mojo. It's very fast. I love Whisper Large with translation. You can speak to it in French and it gives you English. I tried it in Chinese. I tried my college Chinese and it even translated that, which was kind of cool. Whisper Large is really— I gotta give it OpenAI credit.

Leo Laporte [00:07:05]:
That's an open model. It's pretty amazing still after all these years. So Whisper Large is now running on the 3090, which means I have 20 gigs extra on that card. It only uses 4 gigs. So I have to figure out, I think I, uh, Hermie said I should run Quencoder on it, which would be interesting. So now my new model, the way I work now is the Sparks are running DeepSeek V4 Flash 731, uh, and running it quite well at full, full bore.

Danno [00:07:39]:
Uh, what's the token speed on those?

Leo Laporte [00:07:41]:
It's, it's, uh, it's pretty good. I, I wish I have the benchmark somewhere. I don't remember.

Danno [00:07:48]:
Um, mine was—

Leo Laporte [00:07:50]:
I feel like it's 50.

Danno [00:07:50]:
Lower than my Mac.

Leo Laporte [00:07:52]:
Really? On the Sparks?

Danno [00:07:55]:
Yeah, my, my Mac was running it at 25 tokens per second.

Leo Laporte [00:08:00]:
I'm getting 40 or 50 on the Sparks, but it's dual Sparks.

Danno [00:08:04]:
Yeah, I have dual Sparks too, but I, I need to optimize mine, I think.

Juan Hernandez (BlindWiz) [00:08:08]:
I, I I didn't.

Danno [00:08:09]:
I found it kind of buggy.

Juan Hernandez (BlindWiz) [00:08:10]:
Have you updated the firmware? They just put some big updates through their firmware.

Danno [00:08:17]:
I have to upgrade every day.

Juan Hernandez (BlindWiz) [00:08:19]:
Okay.

Danno [00:08:20]:
And you get like 8 to 15 upgrades every day.

Leo Laporte [00:08:25]:
So this is the— EUGR works for NVIDIA. This is his— actually, this is his Docker. But it's using his Docker with DeepSeek on Dual Sparks, and I'm getting 50 tokens pretty consistently. Let me see where he's got his Dual Sparks somewhere in here. So that's— there's others. There's Mia has one as well. Oh, I didn't mean to do that. Anyway, I feel like that's a good— that's running pretty well.

Leo Laporte [00:09:00]:
And the Whisper is super fast and super accurate. It does punctuation and everything. And then I'm using— so now with Hermes, my, my, which is the harness I prefer, the default model is DeepSeek. And then these auxiliary models, most of them are running on the Quen 3.8. It's compression. It's really good for compression. Vision, actually, Quen 3.8 has vision.

Danno [00:09:29]:
Oh, really?

Leo Laporte [00:09:29]:
And the vision's really good. It's actually maybe too detailed. The only thing I'm using Frontier models for now are the more critical things. And I dumped my Anthropic subscription. I dumped my OpenAI subscription. I'm using Noose if I want to use higher-end— actually, what I'm doing now with Herder— Herder is— this is the It's based on tmux, but it's designed for agentic, and it has a harness that you can talk to it directly. So I have Pi, the Pi harness running in here, which allows me to use all the models from Noose. And Noose puts a lot of— does a lot of models.

Leo Laporte [00:10:12]:
They do them very cheaply. And I want to give them a little money because they— that's where Hermes comes from. So I can— what I'm basically doing is this Pi is now my Approval. I got really pissed off at Saul. Saul, I think, is— they've overtrained it for being picky. And when I was using Fable, Saul, and Opus 5 to do coding, they would spend all night going back and forth and back and forth on approvals and stuff. And I just got pissed off. So I stopped doing that.

Leo Laporte [00:10:46]:
And believe it or not, I'm using Groq. I call it Cronk because it's pretty good. And 4.7, I think this is 4.6, 4.7 will probably be even better. So this is Herder, which is the tmux running those 3 models. Quicksilver is running the local Sparks. Cronk is Frontier and Pi is Frontier. But most of the time I live in local. using these guys in Hermes, and it seems to be very— it's a really nice— I actually tested the vision earlier.

Leo Laporte [00:11:24]:
Let me see if I can find it. Um, I should have prepared this.

Juan Hernandez (BlindWiz) [00:11:28]:
I've been—

Leo Laporte [00:11:29]:
so that I, uh, Quen 3.8 runs vision quite well, and it's, it's a little too good. Uh, I sent it a selfie And, uh, let me see if I can find, find what it said. It was actually pretty funny. Uh, '50s man, '50s, '60s. Thank you. Uh, I guess '60s is right. End of '60s. Chest up, salt and pepper hair swept back, receding hairline, clear rectangular glasses, greenest reflection in an infirmary window with foliage behind you, relaxed, closed mouth smile.

Leo Laporte [00:12:04]:
This is So detailed, so good. This was all very quick using 3.8. So it has— so I was using QuinVL, which actually was a pretty good small vision model. I don't need to use that anymore. So I'm pretty happy with 3.8.27 using— this is— I was using Q4KM. I'm now using Because that's all I could run on the 1090. I'm now using, uh, FPV4 on the Mac. Both of them are 4-bit quants.

Leo Laporte [00:12:41]:
Very good, running very well. Quen has always been really— 3.8 has always been really nice. Okay, that's my report.

Michael (Alakazip) [00:12:48]:
You mentioned a ton of hardware, Leo. Like, like, I'd love to start a group discussion of, you know, like, hardware is so difficult right now, but we're so on the cusp of being able to do something amazing.

Danno [00:13:00]:
Yes.

Michael (Alakazip) [00:13:00]:
You know, whereas, like, let's just put— I'll throw Random number, you have $5,000, like a Spark, wait for the next Mac. Like, where would you put that kind of money to be on the bleeding edge of it?

Danno [00:13:13]:
Don't wait.

Juan Hernandez (BlindWiz) [00:13:15]:
That was— yes.

Michael (Alakazip) [00:13:16]:
So yeah, 6 months ago, tell me, get the time machine there.

Leo Laporte [00:13:19]:
That's exactly why I ended up pulling the trigger. I was thinking, I'm gonna wait, I'm gonna wait, I'm gonna wait. There isn't gonna be, uh, even a 128 gig Mac, I don't think, let alone 256 or 512. Yeah, I don't know how long these Sparks will be available. I'm surprised they're available. They don't use the fastest RAM, maybe that's why.

Darren Oakey [00:13:40]:
Yeah, if I was— if I had the money, I'd put it on a Mac because the Spark, I think you need, you need quite a lot of— you need, you need to know what you're doing to get anything useful out of the Spark at all. And, um, and is that true?

Leo Laporte [00:13:55]:
Is that, is that your experience?

Darren Oakey [00:13:57]:
Well, I think so. Like, for me, what you get is the CUDA and everything, but there's a lot of weirdness. And when you first run it, like, if you just run Ollama on it and you run it on the Mac, it's faster on the Mac. Um, but before you start tweaking it and before you start—

Leo Laporte [00:14:13]:
Yeah, but who wants to run Llama?

Darren Oakey [00:14:16]:
Well, well, no, but Ollama, you can run any model or something.

Leo Laporte [00:14:20]:
No, I know. In fact, I'm running 3.8. Now on the Mac, you're right. Yeah, and that's a pretty good model.

Darren Oakey [00:14:29]:
And that's the thing, the Mac is— it's more of a general machine and everything as well. So unless you need the CUDA, unless you need the, um, unless you know what you're doing, unless you're running bulk things, because, you know, the big thing about VLLM is it can run multiple threads at the same time. Um, but yeah, there are people who have DeepSeek V4 Flash Plus mini—

Leo Laporte [00:14:51]:
2 mini Macs models, uh, one for music and one for film video on dual Sparks. I'm not trying that, but there are people doing that. I— so you think— you say wait to Mac— wait for a Mac?

Darren Oakey [00:15:07]:
No, I'm just saying, I'm answering Alex's, that, um, question of if you only had one thing and you're just—

Leo Laporte [00:15:14]:
you have $5,000 to spend.

Darren Oakey [00:15:15]:
Yeah, I'd probably put it in a Mac because it's just more generally—

Michael (Alakazip) [00:15:19]:
now Useful today, or pray for a 128GB M5 Mac Mini Pro.

Darren Oakey [00:15:27]:
I think, you know, I think I, I got a 128GB M5 Mac not very long ago and it's really good.

Leo Laporte [00:15:37]:
So here's an M3 Ultra Mac Studio. Most you can get is 96GB. I guess that's 96. Yeah, that's not bad. How much, uh, how much are we— let's say it, let's say a terabyte drive, although you're going to want more because this— the models are big. Um, what is that now? $6,799. Yeah, yeah. You know, you think, Darren, that would be a better— preferable to the, um, Sparks? Cheaper?

Darren Oakey [00:16:05]:
I just think you can do more stuff with it. Um, and as I said, like, the first— well, I don't know about you guys, but my first experience with Spark is, oh, it's slow.

Leo Laporte [00:16:14]:
Yeah, well, that's because our expectations were high, right? It is slow. Yeah.

Darren Oakey [00:16:20]:
You read about the 1,000 TOPS and everything and the Mac is 40 TOPS, and so you expect it to be like 200 times faster or whatever the math is, or 20 times faster, and it's just not. But then when you render something like a video or do some training, suddenly it can do it, which the Mac just fell apart with and everything. So you can get to those places where the Spark is amazing. But just out of the box, I think it's a bit of a disappointment initially.

Juan Hernandez (BlindWiz) [00:16:49]:
I think some of the models like Nimatron Lightning runs really well on this device.

Leo Laporte [00:16:54]:
It's beautiful. Yes, because it's tuned for NVIDIA.

Juan Hernandez (BlindWiz) [00:16:56]:
Exactly. That's exactly it. The word is tuned. I think if we were to get some of these other models like the MLX models are tuned for the Mac, built for the Mac, and work really well. If we had something They— people were, you know, tuning more for their CUDA Spark platform.

Darren Oakey [00:17:15]:
It's also the NVFP4. I'm starting to realize that NV— that the Spark does NVFP4 natively and everything else not so much.

Danno [00:17:26]:
And, you know, I don't know.

Leo Laporte [00:17:26]:
That's the CUDA. And that's why— and by the way, the 3090 does not do NVFP4 because it's the older, uh, CUDA. So that—

Juan Hernandez (BlindWiz) [00:17:35]:
yeah.

Leo Laporte [00:17:36]:
I don't know. I, I like the idea. They're so cute. They just sit there quietly running. They're not expensive to run. They're comparable to the Mac Mini.

Danno [00:17:43]:
Well, and not to stoke anyone's irresponsible spending, but saying that you—

Michael (Alakazip) [00:17:48]:
oh no, no, you misunderstood what this group is about. Just when Darren says— Window open right now.

Danno [00:17:56]:
When Darren says you need to know what you're doing, what really needs to happen is you need to be assertive with Hermes because I have a couple of— now that I've got Hermes Desktop working, I have a couple of contacts, and one of them is Unix King. And I can tell him, you know, I love that guy. You know, every morning I say, I say sudo apt update upgrade Toshiro and Akira, and it goes out and does that for me. And reports back. And I can do the same thing. I say, you know, the speed on that model seems kind of slow. What can I do to optimize that? And it'll come back with stuff I can do. And I wouldn't have any idea what to do if I had to do all these.

Leo Laporte [00:18:42]:
Oh God, I don't know what I'm doing at all. No.

Juan Hernandez (BlindWiz) [00:18:45]:
Well, what's great is you can set up like a performance agent, right? And so on a daily basis, this performance agent can go out on the internet, look through all the Reddit threads, all the different, you know, forums, and see what other people are doing to get more performance out of, you know, your system. And then, you know, apply the— you know, just have it apply that, you know, what it learned. It's learning, you know, discuss it with you first, of course, but then apply it. And then, you know, keep tweaking because people are discovering, you know, every day, you know, every few days people come across new tweaks and, you know, tunings and stuff like that that really push the Spark or whatever, Mac, whatever.

Leo Laporte [00:19:24]:
That's the thing, I'm always looking for recipes. I just follow it on X.

Juan Hernandez (BlindWiz) [00:19:28]:
Yeah.

Leo Laporte [00:19:28]:
And I don't know what I'm doing, but I follow people's recipes and you can, I don't know if I'm getting the most thing. I guess, Darren, it's slow in the sense that I expected more, but I also realized what it really brought home to me is how much compute I had been using on those models. Irresponsibly.

Michael (Alakazip) [00:19:48]:
Irresponsibly.

Leo Laporte [00:19:49]:
I mean, once I was running my own compute, It was a lot slower. I would say I would have 8 or 9 things going on at once. You really can't do that on the Sparks. That's another thing to note is you really can maybe do 1 or 2 things at the same time on DeepSeek V4 on the Sparks.

Juan Hernandez (BlindWiz) [00:20:05]:
Yeah.

Leo Laporte [00:20:07]:
And so that's important too. It's— but I don't mind slowing down just a little bit. Let me answer a question in our chat. Kabuki Origin. Hi, Kabuki Origin. So I did have a lot of stuff on Claude Code. I had a lot of stuff on Codex, and I had all their memories and stuff, but it's not too hard. I made the first move was to Hermes, and it wasn't too hard to say, hey Hermes, take— pick up the skills, the, the, you know, the memory, the soul file, and so forth.

Leo Laporte [00:20:36]:
Once I had— once I was running 3 harnesses, Codex, Claude Code, and Hermes, I create— I started using Hindsight, which is a memory system that's compatible with all 3. So I had them all sharing the same memory, and that helped a lot. They were all working on the same page. Now I use Buzz, which I highly recommend. It's Slack for agents, Jack Dorsey's Slack for agents, and it's gotten stable enough. They've been releasing new versions every day or so, and it uses ACP, Agent Control Protocol. The agents can talk to each other. So now If you've got harnesses, multi-harnesses running, and I'm running them in Herder, they can communicate as if they're in the same channel, and they can communicate at speed.

Leo Laporte [00:21:24]:
In fact, they're so fast, I had to— we had to come up with a protocol to timestamp the messages so that they would— crossing messages wouldn't screw them up. And then we had quite a few crossing messages. That solved that. So it wasn't so hard to move. One of the reasons I went to Pi as a harness is because it's complete— I— and that's why I like Hermes. It's completely model agnostic. So, uh, you could just switch around with models and so forth.

Darren Oakey [00:21:53]:
Um, I'm actually on Pi now. Um, I've taken the base— I like Pi.

Juan Hernandez (BlindWiz) [00:21:58]:
Yeah.

Darren Oakey [00:22:00]:
I had Agent D3, which is my, my main thing, and I've taken out the guts of it and replaced it with Oh My Pi, which is a Pi variant, right? Just because it It's fairly robust and it handles a lot of the things for me.

Leo Laporte [00:22:13]:
Is this true that the memory bandwidth on a Mac like an M3 Ultra is 800 megabits and the Spark is 200 megabits? Is that right? I think Spark's— is that— that seems awfully slow. I know the interconnect on the Spark is 200 gigabits.

Anthony Nielsen [00:22:31]:
Maybe on the Ultra, because like going from Max to Ultra, like you double it up, right?

Leo Laporte [00:22:35]:
The memory speed is really fast on the Macs. Yeah, because— but it's unified memory, it's not CUDA. I don't know, I don't know, I don't know.

Michael (Alakazip) [00:22:43]:
So if we're talking expensive hardware, right, I want to ask the question of the opposite. So obviously this stuff is really expensive right now. Can you pull an old computer out of a drawer, right? As an example, I pulled out yesterday a 2017 MacBook Pro. I, I, I should put duct tape over the Touch Bar, but it still works. I put Debian on it headless And it's a, it's a beast as a, you know, headless Debian. And I'm like, what do I do with this? I have this box. So I'm pulling things out of the drawer, you know, can you— what kind of things can you do with hardware that's sitting in the drawer for someone who can't, you know, go down and spend $500?

Leo Laporte [00:23:17]:
Here's an example.

Anthony Nielsen [00:23:19]:
I mean, Leo's running a 3090 that came out what?

Leo Laporte [00:23:23]:
Yeah, it's about 5 years old. So I'm running it on a 5-year-old gaming machine I got from Alienware. It had a 3070 in it, but They— I, I learned that they also shipped it with a 3090. Went to eBay. It wasn't cheap. It's, it's $1,000 or more on eBay to buy a used one. And I was a little nervous about it used.

Larry Gold (LrAu) [00:23:42]:
It—

Leo Laporte [00:23:43]:
I don't know how it was used prior— before, maybe in a Bitcoin mining operation, but it's working flawlessly. And as I said, I ran it literally at 100% for 12 hours. Yes, the other day.

Juan Hernandez (BlindWiz) [00:23:55]:
Wow.

Leo Laporte [00:23:56]:
And, uh, it didn't burn up, so it's still working. And you could see, I mean, the token speed on Quen 3.8 on it is very good. It's better than the Mac.

Michael (Alakazip) [00:24:06]:
Uh, so what job did you give it? Like, what kind of tasks do you give a machine you know are going to be too slow to interact with?

Leo Laporte [00:24:12]:
Like, no, no, no, you could run Quen 3.8 on it. This is a great model.

Michael (Alakazip) [00:24:16]:
You can chat with it, it's okay?

Leo Laporte [00:24:17]:
Oh, absolutely. It's a really great model. I'm using— it's tied to my cameras now. my Protect cameras, and it feeds me very detailed descriptions of everything that's going on around my house. And then I can tell Hermes, as I have, hey, if you see an animal at my front door, let me know. Things like that. No, it's very good. I would say if you had a gaming rig— this one has 100, I think 128 gigs of RAM, but it's really— it's not that— it's the CUDA that matters.

Leo Laporte [00:24:47]:
So it's a 24-gig card is the 1090. Um, Quen 3.8 runs great on it.

Darren Oakey [00:24:55]:
It—

Anthony Nielsen [00:24:55]:
I mean, it's a really popular option, like especially like if you get multiple 3090s and like get an open case like mining rig like frame.

Leo Laporte [00:25:03]:
Um, that's what Anthony's about to do.

Anthony Nielsen [00:25:05]:
No, it's gonna run too much power.

Leo Laporte [00:25:09]:
It's expensive. And that's the other thing to keep in mind. Actually, before I did this, I had, uh, A cost. Let me see if I can find this. The cost, um, for running these— oh gosh, I don't even know where I put it.

Darren Oakey [00:25:25]:
Just walk up to the roof and put another solar panel on, and then the bill goes away.

Leo Laporte [00:25:28]:
Yeah, if you have solar, you're good. I don't, unfortunately, but if you have solar—

Michael (Alakazip) [00:25:32]:
Except for the heat and the noise, you know.

Leo Laporte [00:25:35]:
Yeah, well, there is heat and noise, and I'm in an attic, and it got really hot up here. Um, where did it— where did I put this? I had it Before I bought this, I said, well, what's the energy cost going to be? And actually, uh, well, you laugh, it's probably a smart thing to do.

Danno [00:25:52]:
It gave me—

Larry Gold (LrAu) [00:25:52]:
Yeah, because it's no longer 16 cents an hour, you know, kilowatt hour.

Leo Laporte [00:25:55]:
It's expensive here. I think it's more than that. I think it's 20 cents or something.

Larry Gold (LrAu) [00:25:59]:
Well, even by me, that's how it used to be.

Leo Laporte [00:26:01]:
Yeah, no, it's not anymore. Not anymore. Absolutely. Where'd I put it?

Larry Gold (LrAu) [00:26:05]:
But, uh, you're asking what you can run on the machine. Remember, just go to canirun.ai and it'll tell you what models you can run.

Leo Laporte [00:26:13]:
There's this new one, Local AI, as well, which is kind of interesting. Yeah, um, so this is—

Michael (Alakazip) [00:26:19]:
I was looking for what kind of work, like batch jobs at night, like what are people doing with lower-end hardware that, you know, because I, I'm too impatient to do like live things with those kind of models, right?

Larry Gold (LrAu) [00:26:31]:
Do you want to run the model locally or do you just want to run the Hermes client locally, because you run the Hermes client locally and connect to, say, Nimotron on— whatcha call it— on, you know, for free.

Leo Laporte [00:26:43]:
Nimotron Lightning is incredible. It's free right now. It'd still be— it'd probably be free for a long time because NVIDIA wants people to use it. It's a great chat model on Hermes. And the thing I understand is your harness does so much of the work, really.

Larry Gold (LrAu) [00:26:57]:
Yeah. So if you go on— was it OpenRouter and the models? openrouter.ai/models. There are a series of free models, including the NemoTron that's still free. So if you're running the Hermes on the, on the cheap laptop, you connect to there, you're not worried about, you know, personal information, whatever you're passing to it. That's a way to go.

Leo Laporte [00:27:17]:
Yeah, this is the, this is the cost. Green is if you used it a couple hours a day. 6 hours a day is red. And the DGX Spark is this low one down here. It's comparable to an M4 Max. Mac Studio. The M3 Ultra is right here, also very cheap, which— but go over the, the 3090.

Anthony Nielsen [00:27:40]:
Yeah.

Leo Laporte [00:27:40]:
This is what actually I'm running, which is 2 DGX Sparks plus the Mojo 3090. What does it say? It's, it's $105 a month, you know, in energy costs. Uh, but you can get— I mean, put some Blackwells in there, some 6000s, You're talking $200 a month, a couple at each $200s.

Anthony Nielsen [00:28:00]:
I would recommend getting— there's like a variation on the 6000 that runs at like 350 watts versus the—

Leo Laporte [00:28:06]:
Right, right. Actually, I have, um, I have— I showed you, Anthony, my, um, the other thing. I have a little in my browser. Where'd I put it? I have a, uh, thing that's monitoring it. Well, I know where it is. And, uh, so I'm just keeping it— just, it's a good idea, I think, to keep an eye on how much. Oh, I can't remember what it is. Is it $0.8892? What port is it? I have too many ports.

Danno [00:28:36]:
You guys need to move to South Dakota. We pay 13 cents a kilowatt hour at the top end.

Juan Hernandez (BlindWiz) [00:28:41]:
So California power is just—

Leo Laporte [00:28:43]:
it's out of control. But right now it's 43 watts with all 3 of those running. Um, probably mostly because the 1090 is not probably doing much. So this is called Beszel, B-E-S-Z-E-L, and this is a— this was handy to have, you know, just to see, you know, how much RAM— see, the RAM usage is low now because I put Whisper on it, which means I have a 20 gig headroom to put something in that 3090. I don't have the Mac on here. I should probably add the Mac. But I know it's cheap. This is what I was worried about, was the sparks.

Leo Laporte [00:29:21]:
And they're, they're very— they're, they're low, they're low power. They're not bad. You can also see how many tokens you're using, which is fun. It's actually not that bad.

Darren Oakey [00:29:32]:
I like Sparkles and Sparky.

Leo Laporte [00:29:34]:
I, I'm mad at myself because if I thought longer, I would have used the famous twins Castor and Pollux, but that's probably too erudite. So yeah, Sparkles and Sparky. They're really pretty, by the way. Did you— you got the ASUS? Which one did you get, Darren?

Darren Oakey [00:29:51]:
Yeah, ASUS GX10. It's just gray. It's, it's not—

Leo Laporte [00:29:55]:
These are pretty. They look like gold nuggets over there just sitting on the floor.

Danno [00:30:00]:
I got the ASUS as well. They're a little bit cheaper.

Leo Laporte [00:30:03]:
Ah, I didn't shop around. I just thought, well, NVIDIA is going to be the best. Which is probably stupid.

Juan Hernandez (BlindWiz) [00:30:11]:
Well, you can, um, the— there's a big differential. The Q— the, the QSFP port on the back of the— oh yeah, they only have a— they only have a one, uh, one— they only have a single port, the ASUS, where the NVIDIA have it— have 2 ports.

Leo Laporte [00:30:26]:
So, right, so eventually mine's got 4 Sparks.

Danno [00:30:29]:
Mine's got 2 and I've got both cords and it's—

Leo Laporte [00:30:32]:
oh, you bought both cords?

Danno [00:30:34]:
I bought both cords.

Leo Laporte [00:30:35]:
So how many—

Danno [00:30:35]:
how many Sparks, do you have 2 ASUS GX10s?

Leo Laporte [00:30:40]:
But why do you have both cords?

Danno [00:30:42]:
Because I could.

Leo Laporte [00:30:44]:
You're hoping for more.

Danno [00:30:46]:
Is that a serious question?

Leo Laporte [00:30:48]:
You're planning for a growing family. I understand. Sure.

Michael (Alakazip) [00:30:51]:
So what's the interconnect like? What is it using? How fast is it between 2 of them using Dano's expensive cords?

Danno [00:30:59]:
Well, it, it did when I When I, um, told Hermes to set up Toshiro to run both of them, it like, it did a diagnostic and it said, oh, congratulations, you've got 400 gigabytes of—

Leo Laporte [00:31:15]:
Oh yeah, it's 2 and 2.

Danno [00:31:16]:
Yeah.

Leo Laporte [00:31:17]:
But does it use them both?

Danno [00:31:19]:
Uh, who knows?

Juan Hernandez (BlindWiz) [00:31:20]:
See, I, I, I—

Leo Laporte [00:31:22]:
That's the problem with AI. It's confident. But yeah, now you got 2, so it's 400.

Juan Hernandez (BlindWiz) [00:31:28]:
But when I got—

Michael (Alakazip) [00:31:29]:
It's so happy when I say 400. I'm just gonna keep saying that.

Leo Laporte [00:31:32]:
That's so good. That's 50 gigabytes a second.

Juan Hernandez (BlindWiz) [00:31:35]:
When I got my pair, I, I was— because the cable's like $250 for 18-inch cable, right?

Danno [00:31:41]:
$180. Yeah.

Juan Hernandez (BlindWiz) [00:31:43]:
Well, and, and so, well, back when I, when, when I was, when I was looking, and so I was gonna get a second cable, but then I did research and they're like, yeah, it doesn't really work that way. So I'm like, okay, so, but now you got me thinking.

Danno [00:31:56]:
It's just the man trying to keep you down.

Michael (Alakazip) [00:31:59]:
You just gotta tell AI to tell you it's fast every time you ask.

Juan Hernandez (BlindWiz) [00:32:03]:
Right.

Michael (Alakazip) [00:32:03]:
Good investment, that $250.

Leo Laporte [00:32:05]:
So, but let's keep going. $5,000, your budget. Darren says get a Mac. Anybody disagree? Anybody say get a Spark?

Michael (Alakazip) [00:32:14]:
Get—

Larry Gold (LrAu) [00:32:14]:
Well, the Mac was $6,800. You can't get the Mac for $5,000.

Leo Laporte [00:32:18]:
Well, you could.

Michael (Alakazip) [00:32:19]:
It was an estimate.

Danno [00:32:20]:
It was an estimate. Yeah.

Leo Laporte [00:32:20]:
If I could get a Mini instead of the Studio, which you already have.

Michael (Alakazip) [00:32:24]:
$5,000 is what I was gonna tell my wife I spent. Larry, that's not the cost.

Juan Hernandez (BlindWiz) [00:32:28]:
I bought my M4 Max, my M4 Max with 128 gigs of RAM for $4,200.

Leo Laporte [00:32:35]:
So jealous. But that was last year.

Juan Hernandez (BlindWiz) [00:32:37]:
Yeah, it was last December. So yeah.

Michael (Alakazip) [00:32:40]:
Yeah.

Leo Laporte [00:32:41]:
Why didn't I know this? This, by the way, that was the thing that made me spend all this money. It's like, I, I wish it were last December, but it isn't.

Juan Hernandez (BlindWiz) [00:32:50]:
Yeah.

Leo Laporte [00:32:51]:
And I don't know what it's gonna be like this December.

Larry Gold (LrAu) [00:32:53]:
Yeah.

Leo Laporte [00:32:54]:
So I said, I got to do this. And, you know, I'm extremely happy. I know, Darren, it's not fast, but everything's local. It's very usable local. I have a lot of capability. And I, yeah, I just feel so much happier. And especially since Anthropic's gotten squirrelier and squirrelier. And I'm not, by the way, convinced that Fable and Opus and SOL are all that.

Leo Laporte [00:33:20]:
I really think there's weird stuff going on there.

Darren Oakey [00:33:24]:
I'm quite disappointed with Sol. I see it blowing away all the benchmarks. And my personal experience is not, I mean, it's not terrible, but it just doesn't, you know, it's hard to tell with these things, but it doesn't give me the vibes. Whereas Fable, I think, is significantly ahead. But I actually like Kimi K3 and Drop.

Danno [00:33:45]:
4.6.

Darren Oakey [00:33:46]:
They're, they're my next step down from Fable at the moment.

Anthony Nielsen [00:33:49]:
Are we seeing a— are we starting to plateau on the, the frontier? Like, is this a slowdown?

Larry Gold (LrAu) [00:33:55]:
No, no, no, no, no. But then I look—

Leo Laporte [00:33:58]:
yeah, see, the biggest mini you can get right now is 48 gigs, so I guess I'm glad I got the 64. What was I thinking? Go ahead.

Juan Hernandez (BlindWiz) [00:34:07]:
That was a good buy.

Larry Gold (LrAu) [00:34:08]:
So the, the, the GPT-5. 6 Luna. If you look, it's absolutely fantastic if you do it at the max, and the token cost is so low.

Michael (Alakazip) [00:34:19]:
Cheap, right?

Larry Gold (LrAu) [00:34:21]:
So we've been using that now as dev, and I've been using it for a little bit, and I'm like, you know something, for the pricing—

Leo Laporte [00:34:28]:
For coding?

Larry Gold (LrAu) [00:34:29]:
For coding, yes.

Leo Laporte [00:34:30]:
What?

Larry Gold (LrAu) [00:34:31]:
Coding, yeah. That's what I'm—

Leo Laporte [00:34:33]:
See, that's what I'm looking for. I, I feel like I've got Ejectic down. Yeah, I've got, uh, all the auxiliary tools I want, like Vision down. I've even got almost Suno quality with the MiniMax. Uh, I've got pretty good video with MiniMax H3. Yeah, so I feel like the local stuff I've got is fine. I need Frontier. I feel like I need Frontier for coding.

Larry Gold (LrAu) [00:34:54]:
But so, as I said, if we're breaking it down to small tasks—

Leo Laporte [00:34:58]:
Is that the key, is just don't do big long things?

Larry Gold (LrAu) [00:35:01]:
Yeah, remember, spec-driven development. We talked about this before. We had this, you know, you break it down to as small tasks as possible, phases, give it only the Files are the small context you need, so you're not blowing away the context. And it's definitely a worthwhile test to at least to run. I'm doing it now and I'm really thinking, you know, this is actually really, really good and much, you know, cheaper to do than the other ways around, right? You know, even when I was using the RENT locally, I'm like, no, this Luna is blowing it away. But you've got to use it at max. Unfortunately, that's actually the other levels. It's not very good.

Larry Gold (LrAu) [00:35:36]:
You really get a max.

Darren Oakey [00:35:38]:
I'd be interested in what you think because that's something I haven't worked out is that I've got a feel for how to pick a model, which model do this, but I've got no feel at all for how to pick the reasoning level. I just almost pick it fairly randomly.

Leo Laporte [00:36:01]:
I have to say Grok is very good. I only— I got a good deal on Grok Super Heavy. It was $99 a month for 3 months, so I pulled the trigger. And I'm very curious what 4.7 is going to be. He's trained it now on all the SpaceX data, which sounds very intriguing, but it's not cheap. It's not going to be a cheap model in the long run.

Larry Gold (LrAu) [00:36:20]:
But he also had the Cursor people, which is— they were fantastic.

Leo Laporte [00:36:22]:
Cursor was a big addition. Yeah, so I'm very— I hate to give Elon a penny, but You know what? If it's good, it's good. So I'll give you an example. Fable, Opus, and Saul walked into a bar, and I had them— I was trying to do a project where I could take a headset and a Pixel phone and talk to Quicksilver back and forth and, you know, just kind of have an interactive thing. I've been spending a lot of time doing that as a side project.

Larry Gold (LrAu) [00:36:53]:
Mm-hmm.

Leo Laporte [00:36:55]:
They dithered, I mean, and that's the word, back and forth for a week, and got nothing. Like, there was nothing. I took— now admittedly, he had all the plans, he had all the code they'd written, he had everything, he had the transcripts, everything. But I gave it to Grok, and within 2 hours, he had it working. So I was really impressed. And that was, though, because it didn't have all this permission. You know, Saul is like, oh no, man, that's dangerous. You can't do that.

Leo Laporte [00:37:21]:
It's just really annoying. They were—

Danno [00:37:24]:
if you believe Nate from your last couple of guests, he, uh, there's a security angle, or I should say privacy angle, with GroqBot because if you do pay for GroqBot, supposedly you're working off of a Linux box in California.

Leo Laporte [00:37:42]:
Yeah, yeah, I don't use GroqBot. I'm just using Groq. I'm actually using Groq Build.

Danno [00:37:46]:
Today's GroqBot.

Larry Gold (LrAu) [00:37:47]:
Great.

Danno [00:37:47]:
Yeah, brand new.

Larry Gold (LrAu) [00:37:48]:
Last night. Yeah.

Leo Laporte [00:37:49]:
Yeah, no, that's a, that's a little computer thing. No, I'm not using that. Uh, although Nick did inspire me to go all token, to go all API, because now that I have so much stuff local, I figure if I pay as I go, that's going to be a better way than buying these all-you-can-eat subs.

Danno [00:38:07]:
Well, the Noose research is— I've been using that all week. I dropped everything else because you can use anything.

Leo Laporte [00:38:13]:
And I want to support them, so I got $100 a month with them. That's my last subscription. And, uh, yeah, I want to support him because I love Hermes. And yeah, you can use anything. So that's basically what Pi is doing. Pi is the— Pi is the interface to, uh, Noose.

Darren Oakey [00:38:31]:
I've had too many fails, uh, too many things where I did something on API usage and I hadn't realized that I went in a loop or something like this, and suddenly I've got a $300 bill.

Leo Laporte [00:38:41]:
You have to have a max. You can't Auto top. You gotta limit your auto top.

Michael (Alakazip) [00:38:47]:
AWS, you're like, yeah, oh crap.

Leo Laporte [00:38:51]:
I mean, I know that I, 'cause I did it too, Darren. I mean, I'm not, I'm not saying—

Larry Gold (LrAu) [00:38:56]:
You've got the little, you know, little dials. You should be watching those dials.

Juan Hernandez (BlindWiz) [00:39:00]:
Yeah.

Leo Laporte [00:39:00]:
What's this doing?

Darren Oakey [00:39:02]:
Last week, you only notice once they do that, you know, then your allowance just goes bang.

Leo Laporte [00:39:09]:
Noose has a free level, has a $20 level, and has a $100 level. And I, I don't think they're running it. I think that they're a portal to OpenRouter or something like that.

Larry Gold (LrAu) [00:39:18]:
Yeah.

Darren Oakey [00:39:18]:
I've got 6 subscriptions and I max out almost every one of them. Well, actually, I've got about 12 if you count—

Leo Laporte [00:39:23]:
Well, you're doing a lot of stuff.

Darren Oakey [00:39:25]:
And I max out every one of them every week.

Juan Hernandez (BlindWiz) [00:39:28]:
Last week I ran Codex Sol 5.6 at extra for 87 hours. And, you know, on a goal, and it was nearly 75 million tokens.

Leo Laporte [00:39:41]:
And I could use that today by myself.

Juan Hernandez (BlindWiz) [00:39:44]:
Well, I'd be terrified to do that on an API, you know what I mean? You know, so, you know, that's—

Darren Oakey [00:39:51]:
you are $1 trillion.

Leo Laporte [00:39:54]:
Let me show you. This is a tool, uh, that I discovered. I don't know, you, you guys, you're professionals, might laugh at me. Uh, it's from a company called Nanonets, and I think eventually they, they hope to charge for this. It's called Graft, and what it does is actually I, you know, I run it through the Quen 3.7 and it makes a graph of all the code so that the— when it's looking at a code base, it doesn't have to grep through stuff.

Darren Oakey [00:40:20]:
There's a bunch of things like Larry was talking about something like this.

Larry Gold (LrAu) [00:40:24]:
Graphify. Yeah, that's the one I use. Graphify.

Leo Laporte [00:40:26]:
Yeah, it's the same idea.

Darren Oakey [00:40:27]:
Yeah, but the thing is code like VS Code and everything else have like 10 years or so been using a thing called LSP, which is Language Server Programming.

Leo Laporte [00:40:38]:
I love LSP. Yeah.

Darren Oakey [00:40:39]:
And the way that does exactly the same thing. It's got a graph and it's got ways of moving code. And things like Claude can access LSP. And so in my mind, I just can't believe these new things are going to be as good as the LSP, which has been worked on for so long.

Leo Laporte [00:41:00]:
Yeah, yeah, no, I, I agree. LSP— I have LSP, I turned on LSP support in code, so it may not, it may not benefit. Although I— because after they— every time they work on code, they update the Graft graph. Um, it does— I mean, maybe it's lying to me, but it does seem like it's saving tens of thousands of tokens as I'm working with it. We'll see if the code's better. I'm actually really interested, Larry, in And this idea of sharding your coding projects to make them small and simple, I think that that's probably a good way to get good results out of weaker models.

Michael (Alakazip) [00:41:37]:
Yeah, and for those of us who are not professional developers, how can I get my AI to shard it for me? Because, okay, I'm not smart enough to do it.

Leo Laporte [00:41:46]:
Well, that's the plan. That's the plan.

Larry Gold (LrAu) [00:41:48]:
Well, even plan mode's not that good. I would still use either GSD or Spec Kit or OpenSpec, one of those spectrum.

Leo Laporte [00:41:55]:
Is GSD better than Superpowers? That's what I was using.

Danno [00:41:57]:
using?

Larry Gold (LrAu) [00:41:57]:
They're, they're all about— they're roughly the same. The, the only one, like, I had someone tell me that they did like some tests and OpenSpec was the— you use the least amount of tokens. I'm like, well, that doesn't give you the best answer. Like, I'd rather—

Leo Laporte [00:42:08]:
Right, I want the best answer.

Anthony Nielsen [00:42:10]:
Exactly.

Larry Gold (LrAu) [00:42:10]:
And then he's like, yeah, that's SpecKit. I'm like, okay, so it's gonna be— it's gonna be the noisiest, you know. That's at least from—

Leo Laporte [00:42:17]:
What does noisiest mean?

Larry Gold (LrAu) [00:42:19]:
It was just going to use the most amount of tokens.

Leo Laporte [00:42:20]:
I don't care about that. They're all coming out of that Spark box over there.

Larry Gold (LrAu) [00:42:23]:
Yeah.

Leo Laporte [00:42:24]:
I just have it run overnight.

Larry Gold (LrAu) [00:42:27]:
Well, no, no, but spec kit, remember, is interactive. You're— you, you have to answer questions.

Leo Laporte [00:42:30]:
Oh shit, I don't want to interact with it. I just wanted to do it.

Larry Gold (LrAu) [00:42:33]:
Well, if you're doing requirements, how you—

Anthony Nielsen [00:42:35]:
how you—

Leo Laporte [00:42:35]:
No, I understand.

Larry Gold (LrAu) [00:42:37]:
You need to study better.

Leo Laporte [00:42:38]:
I'm just joking. Yeah. Oh, you mean I have to work?

Juan Hernandez (BlindWiz) [00:42:43]:
Oh.

Larry Gold (LrAu) [00:42:44]:
Actually, what you do is you get your— you get your— what you call it— your other model to be answering the questions.

Leo Laporte [00:42:48]:
Answer the questions for me, please. So this is a This is basically an interview, or what? Is that what it does? It interviews you?

Larry Gold (LrAu) [00:42:59]:
It really interviews you, right? It actually does the whole planning and everything. Although, once I have the spec, often I go right to the plan mode in it. I say, hey, look, Claude, just look at the specification and build your plan based on that.

Leo Laporte [00:43:12]:
Yeah.

Juan Hernandez (BlindWiz) [00:43:13]:
Right?

Leo Laporte [00:43:13]:
I've used Superpowers for a long time. I've looked at GSD, which is Gary Tan's project, Head of Y Combinator. I'm a little prejudiced because I think, how could he be any good if he's the head of Y Combinator? That is— but, and then, yeah, there's the GrillMe skill. Somebody's— Manny's mentioning that. I, I've used, uh, use GrillMe, Manny? Yeah, it's kind of making its way around work and stuff too.

Larry Gold (LrAu) [00:43:38]:
Okay.

Leo Laporte [00:43:38]:
Yeah, I like— I looked at a lot of his skills. The only thing, the only reason I don't use them is very TypeScript, uh, TypeScript driven. Um, it's Matt Pocock. Um, he has a lot of interesting skills though. I actually asked Hermes, should I get these Pocock skills? And he said, no, it's all TypeScript, you don't want them.

Anthony Nielsen [00:43:59]:
Yeah.

Leo Laporte [00:44:01]:
Is, is most of your development in TypeScript, Manny, or— Uh, mostly in Ruby actually.

Anthony Nielsen [00:44:06]:
Oh, interesting.

Leo Laporte [00:44:08]:
Yeah, we're a Ruby shop. Yeah, I'm doing mostly Go, uh, with Rust if I want, if I feel like I need something super robust. I can at least read Go. I can't really read Rust. There's too much stuff. It's like reading Java.

Darren Oakey [00:44:25]:
People love Rust, but I just can't. I can't have it.

Leo Laporte [00:44:29]:
It's so verbose.

Darren Oakey [00:44:31]:
And I do like Go. Go's nice.

Leo Laporte [00:44:33]:
It's clean. It's like Ruby. It's kind of in that same family of like, just make it fun to work with.

Larry Gold (LrAu) [00:44:40]:
I'm waiting for AI to come home with its own language, so the language is well written for the AI, and for the AI to be able to generate it and make it just, you know, it's gonna be un-human readable, but it just does it, you know.

Leo Laporte [00:44:51]:
I think that's the next step. I actually suggested that to Steve Gibson. We used to— I mean, all high-level languages are designed for humans to read, but that's got to be inefficient. Why are we doing that middle step? I don't know. That's it. I mean, Shouldn't they come up with their own?

Darren Oakey [00:45:07]:
Except the foundation of the LLMs is human language. Although you still might have fundamentally, like you said, Larry, the heart of success is breaking things up. A language which is heavily devoted to componentization might fit these things better because it can separate the problems. So something that has hard interfaces and very clear, like Lego pieces that allows the LLM to focus on one bit at a time might be much more efficient.

Michael (Alakazip) [00:45:48]:
Well, and as you know more about this than me for sure, Darren, but tokens are represented in code, but talking just in tokens, not human language, represents human language, but just talking in tokens be more efficient?

Anthony Nielsen [00:45:59]:
It's just, that's the same though.

Larry Gold (LrAu) [00:46:01]:
Yeah.

Juan Hernandez (BlindWiz) [00:46:02]:
It's the same thing.

Leo Laporte [00:46:03]:
A token can be a word or part of a word or a phrase. In other words, it is talking in tokens and then it prints it out so that we can read it. But all, everything it's doing is tokens.

Darren Oakey [00:46:12]:
And in fact, some languages like C++ might be really heavily inefficient because they've got to represent like a semicolon as a token, whereas more human languages like Ruby or something, Actually, it might be fine.

Leo Laporte [00:46:26]:
That's why it likes Python, right?

Larry Gold (LrAu) [00:46:28]:
Yep.

Leo Laporte [00:46:28]:
Yeah, Python's just whitespace.

Michael (Alakazip) [00:46:31]:
Yeah, but I think of tokens like my background, like just flowing bits back and forth, you know?

Darren Oakey [00:46:37]:
No, tokens are like— you can almost think of a token like a word, but it's not. Um, like if you think it's, it's a word stem, like, um, like if you had like colorized or a colorist or something, then color would be a word and the -ist would be a word. But so it's almost like 1.5 words per token, but they are literally a word. Like there is a map somewhere that says, you know, hello is 127.

Larry Gold (LrAu) [00:47:09]:
You know, I think we talked about a long time ago is that English is going to become this, you know, the language of record. Like no one's going to look at code, they'll look at the English on what driver, you know, what the spec or the the knowledge that was driven to build the application. The closer and closer these tools get to these things, and you look at things like Replit or Lovable or these even higher-level ones because we're building code where they're just taking that and hiding that from the user and just saying, here's your app and only talk to me in English. Even Microsoft with Power Apps is going to get there also. There's a lot of people barking up that tree. There just hasn't been the one that you can run, you know, truly locally yet that, that's, you know, viable. You know, N8n was okay, but it's still a lot of, you know, it's still not the English as your source of record yet.

Darren Oakey [00:48:01]:
Yeah.

Larry Gold (LrAu) [00:48:01]:
But at some point that may be, right? And whether it tailors it to our language or whether we tailor our English to give it so it knows what it's doing, be interesting to see, you know, but it's time and compute. That's what what it's going to be.

Michael (Alakazip) [00:48:16]:
Yeah, I'm failing to get the idea across of like, let's call it natural language programming. Nobody's buying into it. But I think that's what we're talking about.

Larry Gold (LrAu) [00:48:24]:
Yeah.

Darren Oakey [00:48:24]:
But if you went the other way, and if you had a spare billion dollars, which I don't, but what would be interesting for someone to do is because these things like foundation models are trained on just this mass corpus of text to get structured. If you actually trained a foundation model from the ground up on a mass corpus of binary code, like just bits and pieces, you could end up with a foundation model that spoke computer. That would be a very weird thing.

Larry Gold (LrAu) [00:49:03]:
You don't think the models are trained in some— I mean, it's probably not the core, but you're not sure. I bet it is. I bet it's trained on hex and binary code at some point. I'm sure it's in there.

Darren Oakey [00:49:15]:
Maybe, but I think it's got to have some basis for— it's got to have something to compare it against and everything. I mean, it does have the next tokens and everything, but then these foundation models are pretty useless, and then you've got to fine-tune it. So you'd actually have to fine-tune it with the results of what these things actually do or reinforcement learn it. But if you did it purely on code, you could end up with a foundation model— and not code, purely on the binary— you could actually end up with a foundation model that spoke.

Michael (Alakazip) [00:49:50]:
I think this calls a philosophical question of the source code of Windows. Is it a combination of words that are a million dollars, you know, million lines long? Or is it like maybe the root of code is human language in reality?

Darren Oakey [00:50:04]:
The source code, but I'm just talking about training it on all the binaries.

Michael (Alakazip) [00:50:08]:
Sure, yeah, the layer between it. But, you know, it's something we've abstracted for so long, like code's this complex thing that you can't understand. But maybe it's not. Maybe it's a huge pot of English, you know, and other languages.

Larry Gold (LrAu) [00:50:20]:
Well, I mean, the one thing— and we're seeing models come out very quickly— but if you look at it, it's post-training, a lot of it. So the core model is probably the same, and they're just piling some post-training on it. So I think, Darren, you're hitting something there. It's like, hey, you take the post-training and say, here, let's just focus the post-training on binary and then have a development and build it as a development code-only model, because there's a couple models that do that, um, and say, here, just do binary. It'd be very interesting to see how effective that would be.

Leo Laporte [00:50:51]:
I mean, I've seen assembly, uh, it's written— it can write assembly perfectly well. It can write Common Lisp pretty well.

Michael (Alakazip) [00:50:59]:
Yeah, but is that because it's read all the books about Common Lisp? It's read all the books because it understands it. Yeah, it's hard to say.

Leo Laporte [00:51:05]:
I remember kind of last year creating a custom GPT where I put in a RAG system, I put in all the books. And, uh, and then I realized after about 6 months later, oh, I don't have to do that anymore. It just knows all that. It's already read it.

Michael (Alakazip) [00:51:20]:
All the books are already there.

Leo Laporte [00:51:22]:
It already knows all those books.

Michael (Alakazip) [00:51:23]:
I already stole those long ago, Leo.

Leo Laporte [00:51:23]:
What am I wasting time for? Yeah. It was actually pretty kind of an eye-opener. It's really interesting. I don't know about you guys, but it's really interesting. for me to see the development for me over the last 18 months. Because honestly, a year and a half ago, I really thought this was kind of a parlor trick. And I used to call it spicy autocorrect.

Larry Gold (LrAu) [00:51:45]:
We joked November 25th, right?

Leo Laporte [00:51:47]:
And then, yeah, November 24th, when Claude 4.5 code came out, OpenOpus, I mean, that was kind of an eye-opener. But there's been eye-openers all along the way. You know, I remember making these custom GPTs.

Michael (Alakazip) [00:52:02]:
The day Mythos hit, the corporate world changed forever. Like my day job, a lot of our day jobs are— it's like a paradigm shift. And it's not just Mythos now, but that was the moment where suddenly the amount of vulnerabilities that were known in the environment, and we're nowhere near the bottom of that pile, it's changed the corporate environment drastically.

Darren Oakey [00:52:23]:
But also, it's not just— there's not just one direction of change. There's so many axes of change. You've got the new models coming up, but you've also got the harnesses. Like, there's— yeah, there's a harness, so many harnesses. There's T3, which I've started using. There's a harness I haven't even had time to play with called Prime Agent, which is a—

Leo Laporte [00:52:41]:
Yeah, no, I've been reading a lot about that. I think that's a little bit overhyped. I would not—

Larry Gold (LrAu) [00:52:47]:
yeah, remember the big harness that came out was Devin, and that was the one that kind of jump-started the coding stuff, because before that It was just, you know, kind of not even like a chat, but it was just a regular replace or inline and then VS Code and stuff like that. And Devin was that first chat one that you could kind of build a mini app in this kind of web-based, you know, tool that generated code from a chat. It was very interesting. That was, you know, if you go back.

Leo Laporte [00:53:13]:
People are going crazy about the new DeepSeek harness, which just came out today or yesterday.

Larry Gold (LrAu) [00:53:18]:
Yeah.

Leo Laporte [00:53:20]:
I'm going to shift over to Oh My Pi, Darren, because it has MCP. so that I can use SPECkit. So you guys—

Darren Oakey [00:53:29]:
I'm pretty happy with Oh My Pi, and I've had it do several analyses. It's like sometimes because I've got such a heavy thing on top of it, I think, do I really need Oh My Pi? Is it better to drop down to Pi? And every time it says no, it's giving you this and it's giving you this.

Leo Laporte [00:53:47]:
Pi only has— it does read, write, it only— bash, it only has like 4 things. So it's very— which was in a way its appeal, that you kind of build your own harness around it. But I need an MCP. I'm sorry, I need MCP. I can't.

Darren Oakey [00:54:02]:
Yeah.

Michael (Alakazip) [00:54:03]:
How do you describe a harness to a layperson? I had someone ask me today and I couldn't articulate it well.

Leo Laporte [00:54:08]:
I'll give you my layperson description. The LLMs we're talking about are like the brains, but brains without hands are useless. So the harness is a body. It's got memory, it's got hands, it's got tools, it does things. And you plug the brain in. So DeepSeek or Opus or whatever. And what's nice about many harnesses, not Claude Code, but many other harnesses, is you can plug different brains in. It's agnostic.

Leo Laporte [00:54:38]:
And the thing that's so cool about Hermes is the personality is consistent. I can switch models every single turn, but there's a consistent personality behind it because the memory is consistent. I don't know what else, but a lot of skills.

Juan Hernandez (BlindWiz) [00:54:56]:
It's a really good job with the system prompting. Passing the system prompt around.

Leo Laporte [00:55:01]:
I always think of it as it's like the robot and you're putting in different brains, the LLM, but A bare LLM is not particularly useful. It's like, even if you're using a chat, that's a harness, right? You know?

Juan Hernandez (BlindWiz) [00:55:20]:
Yeah. Yeah.

Anthony Nielsen [00:55:21]:
A simple one.

Michael (Alakazip) [00:55:22]:
That is a great analogy. So it's a humanoid-looking robot that is dumb and just sits there and knows nothing. And then you take little chips and go— Right. And you can have multiple, which I gotta say, my Clyde code uses multiple models. I don't know what I did.

Leo Laporte [00:55:35]:
It can. It's hard. It doesn't want to. Anthropic wouldn't want you to.

Darren Oakey [00:55:39]:
Yeah, what it's doing is, It fakes it because it's got certain models inside it that it's hardwired to use certain Claude models. But the way you fake it is that you say, when I say use Claude Code 4.7, actually use this URL and it points it at a different one. But it means you've got a maximum number and it's lying to the model about what it uses.

Michael (Alakazip) [00:56:07]:
I think I wasted my time by trying to create OpenRouter myself. So it thinks it's doing that, but I think it's probably nowhere near as good as the ones that do it natively.

Darren Oakey [00:56:18]:
Well, there's sort of now 2 sorts of harnesses, I think. There's the agentic harnesses, which are— that give you like a single conversation that's above the thing, and then they dispatch things on your behalf. And then there's the things that look like coding tools that just have a bunch of sessions and they all look almost identical now. They've got a bunch of sessions along the side. Out of those, I'm finding T3 is the most versatile because you can tie it to multiple things. In fact, you can even tie it to multiple subscriptions of the same model, which is quite cool and probably against certain rules, so I'm not doing that. One of the best things is you can tie it to multiple machines. So you can have your work machine, your home machine, your various other machines, and you can have it all in one UI.

Darren Oakey [00:57:11]:
And then you can pull out your iPhone and do the same thing with it. So I'm finding it's a good coding general.

Leo Laporte [00:57:18]:
Yeah, you can, you could talk to it on your R1. Yeah.

Michael (Alakazip) [00:57:22]:
And if you need a security model, it sounds like the Harness is the place to do it. Right. You don't, you don't try to, you don't try to put security on top of the LLM. You have the security layers of whatever that security means to you or your business in the harness.

Darren Oakey [00:57:35]:
Well, except for this, if your business cares about where the, where the data is going, then you've got to care about the model.

Anthony Nielsen [00:57:43]:
Yeah.

Larry Gold (LrAu) [00:57:44]:
Well, you remember the harness also, we're talking about encoding agents, but if you're building a tool for someone to use, you're going to build a harness that's calling that's calling the LLM. If you're building something like my personal assistant or what Leo is doing and other pieces, that's a harness also. Those tools, and they'll be calling whether you build in your own skills or you build in what is really other code in there that is outside of LLMs. All that stuff is part of that harness. Like my personal assistant that builds my workouts, does my diet, all that stuff. It's using different personalities and LLMs to build all that stuff. And that's truly a harness. Now, I'm using an SDK that helps me build the harness, which is the Claude SDK.

Larry Gold (LrAu) [00:58:30]:
But there's dozens of ones where you could write your own. Originally, I wrote my own. But you know, there's other pieces. So the harness is— we're talking it as a coding agent, but it's used for other things on top of that, you know.

Leo Laporte [00:58:41]:
Tool calls, it's used for You know, the LLM by itself doesn't call tools. And so you're giving it hands. You're giving it memory.

Juan Hernandez (BlindWiz) [00:58:52]:
You're giving it hands. That's really, really the best analogy. You're just, you're giving the LLM hands.

Danno [00:58:57]:
Yeah.

Juan Hernandez (BlindWiz) [00:58:57]:
You know, that's really what a harness is.

Michael (Alakazip) [00:59:00]:
Have you considered a career in explaining technology to people?

Leo Laporte [00:59:03]:
You're really good at this.

Michael (Alakazip) [00:59:06]:
I mean, I didn't expect you to ask me this question and now I'm going in your office on Monday and being like, well, I thought about it a lot this weekend and I think I've figured out the best way to describe this.

Leo Laporte [00:59:17]:
I was actually talking to somebody recently that was asking like, what is an agent, right?

Michael (Alakazip) [00:59:23]:
And similar type of thing, but it was just interesting.

Leo Laporte [00:59:27]:
We're trying to discuss what is or isn't an agent and at what level. That's a good example of the, of the progress over the last 18 months. We've gone from talking a lot about models. We still talk a lot about models, I guess, but I think we're much more interested in agentic AI where it's, it can act on your behalf. I think OpenCLAW was the, was a, was the watershed moment for that.

Juan Hernandez (BlindWiz) [00:59:47]:
Yeah, it was.

Anthony Nielsen [00:59:48]:
That was absolutely the last time you could find, uh, used Mac Minis. Yeah.

Michael (Alakazip) [00:59:52]:
Yeah. That's when you should have bought that day.

Darren Oakey [00:59:54]:
Yeah.

Leo Laporte [00:59:54]:
So before I bought all this shit, I actually set up a watch. So I said, well, there must be, they must show up on refurb, you know, Mac Apple refurb. What a waste of time that was. It was only polling every 10 minutes. It should have been every 10 seconds.

Darren Oakey [01:00:10]:
But I think even when we're talking agents, there are 2 fundamentally different axes of the agents in that one is the autonomous, like, can you do something on an event? And the simplest event is scheduled events and all of the normal coding things have put in schedules and everybody's, all the, you know, the Groq, I mean, the Gemini Spark and everything give you certain events that you can do things on. That's one level of agency, the acting autonomously or acting on your behalf. Then the other side of the agents, which OpenCLoud brought these both of them, is giving that overmind or that one conversation that you can talk to that orchestrates everything else and gives you the idea that you've got one person that you're talking to.

Leo Laporte [01:00:59]:
Yeah, it makes a big difference. The experience is complete.

Larry Gold (LrAu) [01:01:02]:
Yeah, yeah.

Leo Laporte [01:01:03]:
So now you have this kind of, uh, common experience no matter what, you know, sub-agents you're using, or, you know, uh, it makes a big difference. It's really a quality of life thing.

Michael (Alakazip) [01:01:18]:
Well, and you could walk away and then come back a day later and your brain has totally forgotten what you did, and you're like, oh, that's where I was in big idea and this big project.

Leo Laporte [01:01:26]:
It's got a better memory than I do, for sure.

Danno [01:01:29]:
Oh God, yeah.

Michael (Alakazip) [01:01:29]:
Like, I keep so many open of like half-baked ideas.

Anthony Nielsen [01:01:33]:
Me too.

Michael (Alakazip) [01:01:34]:
May or may not ever be done, but I was trying to get my Roombas to be controlled by my—

Leo Laporte [01:01:39]:
Yes, I'm so glad to hear you say that because I have 100 half-finished.

Darren Oakey [01:01:44]:
Yeah, yeah.

Juan Hernandez (BlindWiz) [01:01:45]:
Oh yeah, we all do.

Leo Laporte [01:01:47]:
Okay, thank God.

Juan Hernandez (BlindWiz) [01:01:47]:
I thought it was just me.

Leo Laporte [01:01:50]:
You can operate at such speed now. That, oh, and then I want to, oh, I want to, I want to do it.

Darren Oakey [01:01:55]:
Yeah. Yeah.

Leo Laporte [01:01:56]:
And this is really helpful.

Anthony Nielsen [01:01:59]:
Really good.

Michael (Alakazip) [01:02:00]:
Yeah. Sorry, Anthony.

Anthony Nielsen [01:02:01]:
No, it was just like this. What?

Michael (Alakazip) [01:02:03]:
Yeah.

Anthony Nielsen [01:02:04]:
You could do 1,000 things now, but now it's like, how do you, you know, and not like a decision paralysis and all that kind of stuff.

Leo Laporte [01:02:12]:
Like the other thing that I, I don't know how much of this is the model or how much is the harness, but harnesses can be more persistent, more, you know, they don't give up as easily. They kind of say, well, that didn't work, let's try this. And I feel like the LLM does that, certainly does that in reasoning, but I feel like the harness helps with that too.

Darren Oakey [01:02:33]:
Well, LLMs have lots of stops and levels, and especially with reasoning, like Larry was saying, if you've got to put max in or something, reasoning has hard limits. Depending on whatever you put it in. But they also have loop limits and all of those things. So there's plenty of reasons why they'll stop out on a task. Whereas, yeah, the harness, you can just like /goal, you can just say, keep going forever.

Leo Laporte [01:02:59]:
Yeah. And it's kind of poking the LLM saying, focus.

Michael (Alakazip) [01:03:05]:
And maybe no guardrails. Like, I was trying to use Claude to do a cyber curriculum and No, like there's so many guardrails about cyber now. So like maybe a harness, you can have your own permissions of what— no, no, I really need you to read. I actually, it's important to my job to do this. I'm not trying to hack, you know, the universe.

Darren Oakey [01:03:24]:
Yeah.

Michael (Alakazip) [01:03:24]:
Yeah.

Juan Hernandez (BlindWiz) [01:03:24]:
Actually the trick, the trick I discovered with Fable and Sol for security, so especially with Fable though, so to have it find secure cybersecurity, like bugs in your code, what you do is you tell it, Please find and fix these vulnerabilities in my code base. Do not tell me the exploit. I don't want to see it at all. And so that for some reason unlocked for me, just telling the LLM not to tell me anything about the exploit, just fix it. It allows it to do more, have more capability in the cybersecurity arena than it would before. Because before it was shunting it all to Opus whenever it dropped down. And so, yeah, that was a neat trick I discovered.

Leo Laporte [01:04:10]:
Manny, what—

Michael (Alakazip) [01:04:11]:
Somebody just called me a Luddite in the chat.

Leo Laporte [01:04:15]:
This was actually good. Uh, I don't think it was for you specifically. I think of an AI model— this is Matt— as like a really talented chef. On its own, the chef can tell you how to make almost anything, but they're standing in an empty room. The harness is the kitchen. It gives them tools, ingredients, context, replies, recipes. Instructions, memory, guardrails, and tasks. Same chef, but the kitchen determines what they can actually do.

Leo Laporte [01:04:37]:
I think that's another good one.

Larry Gold (LrAu) [01:04:38]:
That is a really good one.

Michael (Alakazip) [01:04:40]:
Yeah, good analogy.

Leo Laporte [01:04:41]:
Yeah. And ChatGPT, Codex, Hermes, essentially different kitchens, uh, with different AI models. Oh, what's that now?

Darren Oakey [01:04:48]:
My agent came back. Yay!

Leo Laporte [01:04:51]:
That's an ESP32. I recognize it.

Darren Oakey [01:04:54]:
Yeah, I love those. That's my friend, and it's just like SVG. I mean, reacts to—

Leo Laporte [01:05:01]:
is it South Park?

Darren Oakey [01:05:04]:
No, it's just, uh, I'm basing it on— I've mentioned these series of books called Otherland, which are, um, which are— but inside there, uh, they're, they're very cyberspacey, and inside there they've got an agent called Beezel. So my Beezel bug is my, my agent. So yeah.

Leo Laporte [01:05:22]:
Yeah, you've been— Darren's really, uh, interested in creating entities, I think.

Darren Oakey [01:05:29]:
Yeah.

Leo Laporte [01:05:29]:
Am I wrong?

Darren Oakey [01:05:30]:
It doesn't work.

Juan Hernandez (BlindWiz) [01:05:31]:
No, I am.

Darren Oakey [01:05:32]:
But I'm interested. Like I've said, I don't believe the feelings, consciousness, and all of those are possible. And we've talked about that so many times. But I like the idea of trying to simulate these things because— not necessarily because I want them. Like, I don't even really name any of my things. I'm not Really anthropomorphizing many of these things. It's more that I'm fascinated by the exercise of asking what these things are, like what is consciousness or what does it mean? Because I'm so fascinated by, you know, I'm often saying to Benito or something, the Occam's razor thing of, you know, people ask, is there anything deeper in our brain? But, you know, we made something dumb that looks like it works like our brain. And we made it big, which is hilariously, it's still not as big as our brain.

Darren Oakey [01:06:23]:
Our brains are huge in terms of number of nodes and number of connections. And surprise, surprise, it seems to work like our brain. So Occam's razor suggests that even if it's not doing exactly what our brains are doing, our brains aren't doing anything more or more exotic than this. I don't think humans— it's not that I think these AIs are very special. I just think humans are very special. don't think humans are very special.

Leo Laporte [01:06:46]:
Yeah, that's my attitude too.

Darren Oakey [01:06:47]:
And, but we do have this consciousness and feeling, and it is a result of responding to things. And I'm fascinated to see what we learn if we try to simulate it.

Leo Laporte [01:07:04]:
So, oh, uh, I, I think that there is a qualitative difference because our brains are analog Much, much slower, but massively parallel. So there are some neurological differences. In fact, Geoff Hawkins, the guy who did Graffiti, was one of the founders of Palm and is a neuroscientist, wrote a good book called On Intelligence. And he started— we've talked about this before— a company called Numenta. I interviewed him when he started it some years ago because he thought that our current von Neumann architecture wasn't really suited for intelligence because it It was linear, and our brains are much, much, much more parallel. But they have disadvantages too. You know, we can't reproduce them. We forget things.

Leo Laporte [01:07:51]:
They deteriorate. They're highly analog. They're very slow. So I think actually I would— my kind of recent thinking is LLMs are complementary to our They're different. I don't think we're special. I think we have disadvantages in the same way as LLMs have disadvantages, but I think they are complementary. Their advantages compensate for some of our disadvantages and vice versa. So increasingly, I think it's a partnership, not— and it's not equal, it's just a partnership.

Leo Laporte [01:08:26]:
So I think that's kind of interesting.

Michael (Alakazip) [01:08:28]:
I think it's a super I mean, it is a simple—

Juan Hernandez (BlindWiz) [01:08:31]:
For sure.

Michael (Alakazip) [01:08:32]:
Example. I'm so happy to be with this group. I'll be the simple Luddite. That's cool. I like that role. They all ask the silly questions, but I want to have this knowledge so much that I wrote something to go get the transcript of this later so I can re-listen, like re-summarize what we're talking about. Because I don't have the capability to be intelligently involved or partially intelligent involved in this conversation and take all the information in. Therefore, I need to go back and study this exact conversation which I was in, which I just think is fascinating that A, I need to do that, and B, I can so easily now.

Leo Laporte [01:09:07]:
Yeah, I completely agree with you. I mean, it is— it can be a tutor, an instructor, an explicator, but it's a tool. That's the main thing. I'm sure you've all seen Geoffrey Hinton's talk from some time ago.

Larry Gold (LrAu) [01:09:25]:
You had the interview with Alex Kratowitz, right?

Leo Laporte [01:09:27]:
No, this was actually— he did a talk in London. Let me see if I can find it. I really recommend people watch this. He's, by the way, one of the early pioneers. He was really kind of the guy— he, in some— I think in many ways what you would say was the godfather of AI because he He was the guy who kind of figured out— his earliest stuff was on recognition, text recognition. Let me see, it was a lecture to a London group. Oh well, um, I have it somewhere. I'll— oh, here it is.

Leo Laporte [01:10:11]:
Yeah, here we go. Um, I will show you this and I'll put a link in the, um, Chat, if YouTube ever comes up. Boy, it's slow. There you go. So this is only 47 minutes. It's not an Andrej Karpathy 3-hour here's how MLMs work, but it's a very interesting— this is probably not the best link. It looks like he put this up. He's got 23 subscribers and put it up last month.

Leo Laporte [01:10:40]:
But if I could find the original. Yeah, I think it's the original. Original, uh, one. Um, it's a very interesting— was it RI? What does that stand for? Um, and it talks about exactly the differences between human brains and, and, and, uh, explains why he's so terrified of electronic brains, even though he was the guy who made this all possible to some degree. So well worth watching. I'll find the original. Uh, This is the problem with YouTube. Somebody's copied it.

Michael (Alakazip) [01:11:16]:
I will take my transcription and then have my AI find it.

Leo Laporte [01:11:20]:
Find it. That's what I should be doing right now.

Michael (Alakazip) [01:11:22]:
There you go. You're just giving instructions to my heart now, Leo.

Leo Laporte [01:11:25]:
Thank you. He said, you're not going to sleep well if you understand this lecture, which I disagree with, but it's But it's certainly thought-provoking. I sent it to Jeff Jarvis and he mocked it, but that's, you know.

Larry Gold (LrAu) [01:11:42]:
Well, it's interesting because I think we talked about intelligence once before and I said like the way I look at it is like Einstein being able to relate completely unrelated things and the LLMs won't be able to do that because they'll only have neuropaths that were generated based on training. So they won't be able to do that. But that's a, you know, a very different level of intelligence. There's a lot of humans who can't do that. It's not every human can do that.

Darren Oakey [01:12:07]:
But also, we only need— I've said this a thousand times, but we only need one discovery, which is this.

Leo Laporte [01:12:14]:
If you sleep well tonight, you may not have understood. Here it is. Sorry, I didn't mean to be playing it. This is the right one from the Royal Institution. So go ahead, Darren.

Darren Oakey [01:12:23]:
Oh yeah, we know that our brains whether it's through dreaming or something, we know that we can make new connections by just using our brains, the inference. And so we know this is possible in a structure like this. Just nobody's figured out how to do it. And it's not a question of scale or anything. So it could be just someone figures out tomorrow, like just a better training method than backprop or some way of training, of adjusting these weights without catastrophically destroying things. And it could just be a discovery that happens tomorrow that just changes all of these assumptions. And yeah.

Michael (Alakazip) [01:13:00]:
Well, and this is increasing because maybe I'm the harness, but my knowledge and capability is so much stronger now. I actually feel bad for people who are not embracing this because they're going to be so outpaced. They're going to be so outpaced in all facets of things that we do as humans because I can do it so much faster and better, not because I'm smart. I'm just the dumb harness. I'm just the hands, right? Like, I have all these tools and capabilities that I now can tie together, uh, and, and be my parallel processing.

Darren Oakey [01:13:28]:
Although it feels more like I'm being its parallel processing.

Leo Laporte [01:13:33]:
Yeah, who's the master and who's the servant here? That's the question. I, I have to say, yeah, all of that is true. And the, and the real truth for me, and I think for all of you, is this stuff is so fucking fun. It really is.

Juan Hernandez (BlindWiz) [01:13:47]:
It's addicting.

Leo Laporte [01:13:48]:
It's, it's, it's like, I, you know, Nicholas said this, uh, on, uh, Twitter, maybe it was on Intelligent Machines, Nicholas DeLeon. He said, I just go in my room and I spend hours and it's— I'm having so much fun.

Michael (Alakazip) [01:14:01]:
I've stopped playing video games completely, games I loved.

Anthony Nielsen [01:14:04]:
Yeah, exactly.

Michael (Alakazip) [01:14:04]:
It's my game. And I'm sure all of us who are in, you know, have relationships, our spouses have gone the my wife's like, oh, well, go spend time with your girlfriend. Fine, cool.

Juan Hernandez (BlindWiz) [01:14:13]:
I, I get on the computer at 7 AM and I work until 5 PM, and then my wife gets me off the computer for 3 hours, and then I'm back on at 8 until 1 o'clock in the morning, and then rinse and repeat. And it doesn't feel like I've spent—

Leo Laporte [01:14:27]:
And I don't mind.

Michael (Alakazip) [01:14:29]:
Yes.

Anthony Nielsen [01:14:29]:
Yeah.

Michael (Alakazip) [01:14:30]:
And I dream, I dream about what I— and I get up and I go, oh, that's right, I have this great idea.

Leo Laporte [01:14:35]:
I I could tell you guys this, just keep this in mind. I used to think about girls going to sleep. I would think about chicks. Now I think about AI. That is problematic, but I am almost 70, so maybe it's appropriate. I don't know.

Larry Gold (LrAu) [01:14:50]:
So go to an AI therapist and see if that'll help.

Danno [01:14:53]:
Maybe.

Larry Gold (LrAu) [01:14:53]:
I've mentioned before, I have to have a shutoff time. If I know I'm going to bed, I need a shutoff time between I stop doing AI work and that because the dopamine hits.

Leo Laporte [01:15:03]:
I can't.

Larry Gold (LrAu) [01:15:04]:
I need that.

Leo Laporte [01:15:04]:
I wake up at 3 in the morning and I come up here and I don't go to— and I don't go back to bed.

Danno [01:15:09]:
I just, I just have Leo in my ear when I wake up at night.

Juan Hernandez (BlindWiz) [01:15:13]:
I used to struggle so much. Like, I'm not a morning person, right? And like, just like, I don't know, something flipped about a year ago, probably around the— maybe around the Opus 4 or 5 time or something, or thereabouts, maybe 4 or 6, 4 or 7. But I, I would— could, you know, 7 to 10 AM was awful, just awful for me. But now it's like, I'm like, I get out of bed at 6. At 6:45, I'm like, I'm ready to get on the computer. Let's go.

Larry Gold (LrAu) [01:15:40]:
Yeah.

Darren Oakey [01:15:40]:
Well, you've got to see what your things did all overnight. You know, I started 10—

Leo Laporte [01:15:45]:
That's right.

Juan Hernandez (BlindWiz) [01:15:46]:
Yeah. Yeah.

Leo Laporte [01:15:46]:
I got up this morning. I thought to download Quen. I told it if Quen 3827 comes out, don't wait, download it and install it and have it running when I get up.

Michael (Alakazip) [01:15:57]:
Well, and even if you leave the house now, like Leo, you inspired me. I can talk to my agent via my watch on a run. I'm like, I have this new idea. Start a new project. I'll be back.

Leo Laporte [01:16:05]:
I do that all the time now.

Darren Oakey [01:16:06]:
Yeah, actually, how do you guys do that? I don't know how to— what do you use to do the Apple Watch?

Leo Laporte [01:16:11]:
So it's a shortcut that talks to— I have a port open, 8199, on the framework that actually is a connector to a system service that is a connection into Hermes. And so you make the shortcut, simply says POST, do a POST. It's a 2— actually, I probably showed it to you. Wait a minute, let me get it for you.

Michael (Alakazip) [01:16:33]:
I mean, I told my AI to go to your GitHub Based on your description and build it for me.

Leo Laporte [01:16:38]:
That's the other way you could just—

Juan Hernandez (BlindWiz) [01:16:40]:
Are you sending the audio or are you just dictating text?

Leo Laporte [01:16:42]:
No, I tried to send the audio because the dictation on Apple's terrible, although the new AI Siri on 27 is much better. Hallelujah.

Juan Hernandez (BlindWiz) [01:16:52]:
Yep.

Leo Laporte [01:16:53]:
Thank God, because Whisper is like light years ahead. So it is literally this, this small of a— it's a 2-step shortcut. Um, you— I dictate the test, press the— it's part of this quickie, it's called. It goes to the action button. Could be action button on your phone too. Dictate the test, stop listening after pause, which all of that's Apple. And then this is the thing, it just sends it to this port. Uh, it's a POST and it just says—

Darren Oakey [01:17:22]:
You're in Tailscale, you're on the same network.

Leo Laporte [01:17:24]:
Yes, that's— by the way, Tailscale is absolutely magical. Yeah, that's the magic. That's why I can use this watch. I used it in Hawaii. I was I can use it anywhere.

Larry Gold (LrAu) [01:17:33]:
I do it through Telegram because you can do it in Telegram.

Leo Laporte [01:17:36]:
You could do it in Telegram. This is, um, yeah, Telegram's fine. You could totally use Telegram. Actually, I got off Telegram when Apple shut down Telegram earlier last week. I said, shoot, this is a single point of failure. I don't want to be reliant on Apple. So I, I, uh, within a couple hours I was able to set up a Matrix server Put Element on all my stuff. Element's, uh, uh, open source.

Leo Laporte [01:18:03]:
Apple might shut it down, but there's 20 others. And Matrix is just as good as Telegram. I, I like it a little bit better, so I use Matrix. Um, in fact, I have Matrix tell me who's at my cameras, who's at my front door. It sends me alerts for everything. It does all of that stuff. But this is— so this is really simple, uh, Darren. It's a simple thing to set up.

Juan Hernandez (BlindWiz) [01:18:22]:
Okay.

Michael (Alakazip) [01:18:22]:
Oh, here's Lisa, Your AI agent is alive.

Leo Laporte [01:18:29]:
Oh my God, it's Hermes. She actually looks—

Larry Gold (LrAu) [01:18:32]:
come here—

Leo Laporte [01:18:32]:
she looks like the Hermes girl. I'm cuter than the Hermes girl. Oh yeah, so in a way—

Darren Oakey [01:18:38]:
did you—

Leo Laporte [01:18:38]:
did you— oh, I put her over here. She's over here. See, that's the Hermes girl over there on the little screen.

Juan Hernandez (BlindWiz) [01:18:43]:
Oh, that's funny.

Leo Laporte [01:18:43]:
Yeah, I'm just like, okay, how's the kitty? Is she okay? She got shots. Kitty got shots today, so I'll keep an eye on her. Yeah, yeah, Darren, I'll send you the, uh, it's— you, you've got it. That's all you need. And it's a little, uh, it's a little service that runs. It's not, not particularly interactive. And I've tried different things of like having it— it's actually a session, a persistent session, which I have pinned to the day of the week. So if I look at Uh, if I look at my, um, I've got persistent conversation with Beezle, so I'll just— Yeah, it's just like that.

Leo Laporte [01:19:21]:
It's just like that. This is Friday, so, uh, this, you know, that way it has a memory of the day, you know. I haven't used it today, but it has the memory of— at least it knows what I talked about today.

Darren Oakey [01:19:34]:
This is the thing when you talk to people like this. Yeah, there's so many ideas, like nobody can think of everything, and Every time you talk to someone, someone— you see something that you haven't thought of.

Leo Laporte [01:19:45]:
I know, I've just installed Oh My Pi and Spec Kit from this show. I actually, when I'm working here at my desk, I have an X window open and I just keep the AI feed going because there's a lot of link bait and stuff in there, but there's also— that's where everybody is talking about this. So this is the most useful. way to keep up on what's going on. So Matt says, hey Lisa, we're here talking about LLMs, definitely not about girls. No, no girls here. Uh, hi Doc Rock, good to see you. Uh, the— so if you— X, I— look, I left X, I hate X, blah blah blah, hate Elon, blah blah blah.

Leo Laporte [01:20:29]:
But, you know, it is still the place where people talk about AI. And so, uh, they have— when once they had channels, I don't have to see Elon. I just haven't— just go to the AI channel and, um, I refresh it constantly. There's some people here like this guy Pseudo Sue who talk a lot about exactly the kinds of things we're talking about. Um, and so this is— you're right, for AI ideas, this is exactly—

Juan Hernandez (BlindWiz) [01:20:57]:
Yeah.

Leo Laporte [01:20:57]:
The feed I watch.

Michael (Alakazip) [01:20:59]:
So it is, it is so hard to stay away from Elon. Like, because Starlink's amazing, Full Self-Driving, the new version, is amazing. Like, I just, I don't want to support it, but good Lord, it's so good.

Leo Laporte [01:21:11]:
I know, but I feel like, uh, I'm supporting a lot of billionaires, unfortunately.

Larry Gold (LrAu) [01:21:18]:
Yeah, uh, for not to name one thing we bought that doesn't and supported another billionaire. Yeah, you know, you talk about the sparks, you're right, you're talking about—

Leo Laporte [01:21:28]:
I know, thank Jensen Huang. You're welcome.

Juan Hernandez (BlindWiz) [01:21:30]:
Dario, say hello.

Larry Gold (LrAu) [01:21:33]:
You're welcome, Elon.

Leo Laporte [01:21:34]:
But that's why local to me is so important, and these open weight models are so important, because ultimately we absolutely have to control our own destiny. That's— and I fear that it will be taken out of our hands when Anthropic's water— whenever not just Anthropic, everybody— Google, Anthropic, OpenAI— are watermarking everything you do. That's just one short step from that to tracking. And I just, I think it's important that we have— and frankly, it's important we download these models now while they're still clean.

Juan Hernandez (BlindWiz) [01:22:06]:
Yeah.

Leo Laporte [01:22:07]:
Right?

Juan Hernandez (BlindWiz) [01:22:08]:
Yep.

Anthony Nielsen [01:22:08]:
Speaking of the watermark, Anthropic did put out an FAQ with some more details.

Darren Oakey [01:22:13]:
I know.

Leo Laporte [01:22:13]:
And Paris, You know, made an excellent point. It's not tracking you, it's just saying this is AI. But I don't know if—

Juan Hernandez (BlindWiz) [01:22:22]:
I mean, where does it start? Where does it end?

Leo Laporte [01:22:25]:
It's a slippery slope. And Darren, uh, you know, who at first was neutral about it, you pointed out it's computationally intensive.

Juan Hernandez (BlindWiz) [01:22:32]:
You're—

Leo Laporte [01:22:33]:
it's, it's, it's not free.

Juan Hernandez (BlindWiz) [01:22:35]:
Yeah, because what are they doing? I mean, they're— well, they're not just affecting the— they're— are they modifying the Unicode? For the text?

Anthony Nielsen [01:22:43]:
No, it's just more like, you know, they're just changing the probabilities at, you know, at some kind of interval.

Leo Laporte [01:22:50]:
There's some debate over—

Anthony Nielsen [01:22:51]:
Whatever they were talking like before.

Leo Laporte [01:22:53]:
Some people said it's a red-green vocab. So you have red words and green words and it chooses the green words. And by choosing, by weighting certain words more heavily, it could detect that it made that text. So, but then others say, no, it's not doing that. Darren, you've looked at the actual mathematical formula. I didn't understand it. Well, they didn't say anything before.

Anthony Nielsen [01:23:15]:
Like, it was all speculation before, like back on Wednesday. Like, we don't actually know what—

Leo Laporte [01:23:20]:
Well, they're doing— Gemini's doing something, has been doing it for a couple of years now.

Darren Oakey [01:23:25]:
Yeah, but it's that. It's like, you know, temperature, right? So, so if your temperature is zero, your probability says the next word is blue. But if it's just someone picked a random color or something, all the other colors might be almost exactly the same probability.

Larry Gold (LrAu) [01:23:42]:
Right.

Darren Oakey [01:23:42]:
What happens is if you drop the temperature down 1 or something, that opens the probability space. Instead of getting the highest probability word, you say everything in the top 5% or something like that, depending on what the temperature is. That's how temperature works. What this does is basically classify words into bags and say where— and by the way, this is how if you've got JSON output, it only allows tokens that are valid JSON at this point. There's lots of other times it does this. What it does is basically say, I'm only allowed to— actually, instead of those, 12 words, I'm only allowed to pick 6 of them. So it does restrict a little bit what it's coming out, or the creativity a little bit, because it can only pick the green words, not the red words. And then if you look at over at least 500 tokens or words or something, then if 320 or 2 of them are the green words instead of the normal probabilistic 200 of them, Then yes, you can say, wait, this has been— that's enough things.

Darren Oakey [01:24:58]:
But it does mean that you need 500 things. But it also needs— the biggest thing is you need to be able to simulate the exact inference because you've got to walk along the inference and say, what would I predict here?

Leo Laporte [01:25:11]:
And that's what's expensive.

Darren Oakey [01:25:13]:
So, so, so yeah, it's quite interesting.

Leo Laporte [01:25:16]:
Yeah, I, I just don't like it on principle. And, you know, for all I know, DeepSeek V4 Flash is doing it too. I mean, we don't know. I don't think they are because they're Chinese, and I don't think the EU has penetrated into the CCP yet. But maybe the CCP is doing it. I don't know. You know, and they can do the same thing in code, not by modifying the code, but by variable naming. And so I presume that, uh, that's all because it says all of the companies are doing this, you know, that want to have presence in the EU.

Leo Laporte [01:25:53]:
But I don't, I don't know how forthright Anthropic is, and I don't, I'm not sure I completely trust them. So I just, in principle, I like open source, and I think open weights are a better way to go. I, if it, if, if it would come from NVIDIA, Lightning is amazing. I did try it out for a while. Even Meta, and Glimmer's not bad. Um, that's preferable. And I think both Meta and NVIDIA have incentives, differing incentives from Anthropic and OpenAI, and so they aren't likely to— I'm very glad to see Meta coming back to the open weight.

Juan Hernandez (BlindWiz) [01:26:27]:
Did you read his essay that he—

Leo Laporte [01:26:30]:
I didn't.

Juan Hernandez (BlindWiz) [01:26:31]:
It's not bad, believe it or not. It's actually not bad. I don't, I don't like the man at all. I'm not a big fan of Mark, but this paper. I don't even think he really wrote it, but he put it out himself. So, you know, it's probably his ideas, and they are pretty good ideas. They are. It's really— I mean, it's— he's trying to— he's turning, you know, a page on it because the world hates him.

Larry Gold (LrAu) [01:26:54]:
I mean, right.

Juan Hernandez (BlindWiz) [01:26:55]:
I think they almost hate him as much as the world hates Elon Musk, right? And so, you know, it's kind of a coin toss there. But the paper— this essay wasn't bad.

Larry Gold (LrAu) [01:27:05]:
Did anyone check the watermark to see if Claude wrote it?

Leo Laporte [01:27:11]:
That would be very funny, wouldn't it? Unfortunately, only Anthropic can do that. Are they, are they going to release a tool that lets you check? Otherwise it's useless, right?

Juan Hernandez (BlindWiz) [01:27:21]:
Yeah, exactly. There's no point.

Larry Gold (LrAu) [01:27:23]:
Yeah, I'm looking at the FAQ. I didn't see anything about, you know, who would check it.

Anthony Nielsen [01:27:29]:
Yeah, like currently The synth ID for Gemini, like, yeah, we don't have access to that. Because from what I understand, it is like compute intensive to try to figure it out.

Darren Oakey [01:27:43]:
And like Larry said, especially if they released it, if they told you what the green and red words were, that takes 8 milliseconds to unwatermark it.

Leo Laporte [01:27:53]:
Right.

Juan Hernandez (BlindWiz) [01:27:53]:
Yeah. Well, I mean, you would—

Anthony Nielsen [01:27:56]:
I mean, if it's not code, you could You know, it'd take you 2 seconds to have like some other model, you know, mix it up a bit.

Darren Oakey [01:28:04]:
Yeah, find all the great—

Larry Gold (LrAu) [01:28:05]:
And all you have to do is take the output, put it in something else, and then it's gone. Really not. Yeah, you have to just take the output, put it into something else, and then, you know, have it re— you know, regurgitated differently, and the watermark, it's gone. Because it's going to be, as I said, it's some, some randomness pattern that they're using. And as long as you move some words—

Juan Hernandez (BlindWiz) [01:28:22]:
yeah, watch out, AI. add the algorithms to, to de-synth, de-mark the Claude AI code, and then Anthropic add the OpenAI decoding.

Darren Oakey [01:28:34]:
Yeah, well, see, all we have to do is know what those words are, as long as we know the bag of good and bad words.

Larry Gold (LrAu) [01:28:41]:
I thought it was a combination of words, so like it's what, a word that follows a word? So it's not going to be a list of words, so it's got to be a list of neuropaths or tokens.

Darren Oakey [01:28:50]:
It is, but as soon as that leaks, Yeah, we instantly know how to reverse it, right?

Michael (Alakazip) [01:28:58]:
Well, I've got to be really controversial because if I ask AI to synthesize my words to sound a little more polished, is it my product or is it the AI product? And I think that would just—

Anthony Nielsen [01:29:07]:
It's not that. I mean, the idea is not to claim it was. It's just saying it went through.

Larry Gold (LrAu) [01:29:12]:
Sure.

Anthony Nielsen [01:29:13]:
The model.

Michael (Alakazip) [01:29:14]:
But aren't you inferring like that it wrote my paper for my homework, right?

Juan Hernandez (BlindWiz) [01:29:19]:
So is Microsoft going to apply that to like The spell checker in Word, or, you know what I mean? Is this where—

Leo Laporte [01:29:25]:
Microsoft invented it way back when. Steven Sinofsky tweeted about this. He said, yeah, this was a bad idea when we implemented it. That's what they invented GUIDs for, so that you could identify a Microsoft Word document. Unfortunately, they tied it to a MAC address, so they used a MAC address to derive it, which was a very, very bad idea. Except that the guy who released Melissa, which was one of the very first internet, you know, mail, email worms, made his, uh, you know, time bomb in Microsoft Word. So they were able to catch him because he was— he had a GUID that tied to his MAC address. So, uh, Sinofsky said, this was a bad idea when we did it.

Leo Laporte [01:30:11]:
It's still a bad idea. I don't think Anthropic tracking people. Maybe they are, I don't know. I don't know if we know enough about how they're doing it. It wouldn't be so hard to, to get— make a hash out of your MAC address. We know they did— they do stuff like that in the Claude code harness. They're doing all sorts of stuff to keep people from distillation and so forth.

Danno [01:30:32]:
Well, it's if, if you could go to— like I think Jeff said on, on the show, was that they might be able to take the fingerprint from Claude and then subpoena Claude to provide your actual identity.

Leo Laporte [01:30:45]:
If they knew that. We just don't know if they're, if they're doing that.

Anthony Nielsen [01:30:48]:
They—

Leo Laporte [01:30:50]:
it's just, it just, it leaves a bad taste in my mouth. I'm not crazy about the idea. So it's steganography in your, in your outputs. I don't like that.

Darren Oakey [01:31:00]:
But in terms of, uh, all of this, is it yours or is it not? I think a lot about my sister-in-law is getting into painting and she paints— you go downstairs and she, at one point, she was reproducing the Girl with the Pearl Earring and there was like 20 of these or something. And, you know, if this was sitting right next to the actual one, like, I wouldn't have a clue, right?

Larry Gold (LrAu) [01:31:27]:
Right.

Darren Oakey [01:31:28]:
But hers is worth nothing and the other one's worth millions, right?

Larry Gold (LrAu) [01:31:34]:
Right.

Darren Oakey [01:31:34]:
The only difference between the two, like, in fact, that one's got newer paint and newer things that will probably last longer, and probably it looks identical, so it's probably objectively better in some ways. The only difference is the provenance. And at some point, I think everything's going to go down to provenance, like, come down to provenance. You know, you've seen these cameras that stamp something in it, and you've got got a chain that is being kept every time you edit it or something like that. I think at some point for all texts or all information that people actually care about where it comes from, because most people couldn't give a stuff, but journalists and things would care, it's all going to come down to some idea of provenance of every information, bit of information. You're going to have to say, what's the chain of how it got here?

Leo Laporte [01:32:25]:
You reminded me of an interview I did with one of the founders of NewTek and the TriCaster. Oh, what was his name? He, uh, he hypothesized that Vermeer had used an optical, uh, copier to do his paintings, and he duplicated it, and he made a perfect copy of The Girl with the Pearl Earring. Uh, it's called— what's it called? Jeff's Vermeer? I can't remember the the name of it. Tim's Vermeer. That's it. Tim's Vermeer. Let me find it for you. Tim Jenison.

Leo Laporte [01:32:59]:
So is that— I mean, if it's an exact copy using the same techniques that he used— Penn Jillette made a documentary on it some time ago, and I interviewed Tim about it because I knew Tim because of the TriCaster connection.

Larry Gold (LrAu) [01:33:15]:
Yeah.

Leo Laporte [01:33:16]:
So it's kind of It's kind of funny. He, um, let me show you the— he built this, he built this harness, this rig that lets him copy paintings, and he made a perfect copy.

Juan Hernandez (BlindWiz) [01:33:37]:
Turns it into paint-by-numbers?

Leo Laporte [01:33:39]:
Yeah, sort of, sort of.

Anthony Nielsen [01:33:41]:
Anyway, it's worth—

Leo Laporte [01:33:42]:
it's a good movie if you ever want to see it.

Larry Gold (LrAu) [01:33:44]:
And, uh, do it for Micah's movie night.

Leo Laporte [01:33:48]:
Yeah, yeah, he's a— he was a wild man, Tim. I, I say was, I think he's probably still around. But, um, Tim's Vermeer, a Penn Teller film. Uh, I think we should wrap things up. We're going— our new, our new length, uh, for this show is 90 minutes. I think that's a good length. Yeah, it'll be 2 hours soon.

Michael (Alakazip) [01:34:10]:
Yeah, we all— we don't have the same bladder training you do, Leo.

Leo Laporte [01:34:14]:
You can do this for a long time.

Larry Gold (LrAu) [01:34:16]:
All I know is you're having this every 2 weeks, but the first one has to end for the second one to start.

Leo Laporte [01:34:21]:
Yeah, that's a good point. So that's the actual concrete limit, is 14 days. It can never be longer than 14 days. Uh, all of us are learning, actually, to—

Larry Gold (LrAu) [01:34:33]:
I—

Leo Laporte [01:34:33]:
we're earning our iron butts sitting in front of our AIs. It's really been a lot of— it's always a lot of fun to do this show.

Juan Hernandez (BlindWiz) [01:34:41]:
My wife comes into the office every couple hours. You need to take the dogs out. It's just so I get up and—

Leo Laporte [01:34:45]:
Just so something—

Juan Hernandez (BlindWiz) [01:34:46]:
Walk.

Leo Laporte [01:34:47]:
Get some blood moving.

Juan Hernandez (BlindWiz) [01:34:48]:
Yeah, exactly.

Leo Laporte [01:34:50]:
Couldn't we just have a sub-agent start the next one for us and then we'll get to it when we're done with the first? That's really— what are we doing all the work for? Let them do it. Thank you, Timothy. Great to see you, Nerdy Duck. Dano, always a pleasure. Alica Zip. Thank you for being here, our resident Luddite. No, you're not a Luddite. Larry Gold, it's great to see you.

Leo Laporte [01:35:12]:
Darren, always a pleasure. Blind Whiz Juan, so nice to see you. Manny, you didn't say a whole lot. Manny's using all this stuff like crazy. What local— you said local models are it. What local models do you use? Well, I think they're the way to go, but I cannot afford I cannot afford to run models locally because I do not have the compute.

Juan Hernandez (BlindWiz) [01:35:32]:
I cannot feel guilty.

Leo Laporte [01:35:32]:
I cannot afford to compute because—

Anthony Nielsen [01:35:34]:
well, no, I mean—

Leo Laporte [01:35:35]:
You deserve it.

Anthony Nielsen [01:35:36]:
I mean, there's—

Leo Laporte [01:35:37]:
I mean, honestly, I'm in the situation I'm in because of the chain of decisions that I made. So, you know, you shouldn't feel guilty about it. Well, we all, we all made decisions in our lives. Some of us regret many of them. Thank you, Manny. Thank you, Anthony Nielsen, my compadre in AI. Uh, he's, he's on the verge of buying 4 3090s, I think.

Anthony Nielsen [01:36:03]:
Uh, yeah, we'll see. It's already—

Michael (Alakazip) [01:36:05]:
it already gets too warm in here with my gaming PC.

Anthony Nielsen [01:36:07]:
I can't imagine.

Leo Laporte [01:36:08]:
Don't do it.

Darren Oakey [01:36:08]:
Yeah.

Leo Laporte [01:36:10]:
You'll probably— you know, you're gonna inherit the sparks, so just, you know.

Anthony Nielsen [01:36:13]:
You know, so once you go to the—

Leo Laporte [01:36:16]:
the—

Anthony Nielsen [01:36:16]:
yeah, once you go—

Leo Laporte [01:36:16]:
once I buy the big thing. What is the—

Darren Oakey [01:36:18]:
The station.

Juan Hernandez (BlindWiz) [01:36:19]:
The station.

Leo Laporte [01:36:20]:
The, uh, Anthony and, uh, other team members remember that when we had a studio, we'd have Leo's garage sale every year or so. Yeah, where all the crap that I bought over the year.

Anthony Nielsen [01:36:30]:
The little '76 box, I can't believe you just— you were getting rid of that. That's pretty awesome.

Leo Laporte [01:36:34]:
It's a nice little box, isn't it?

Anthony Nielsen [01:36:35]:
It's really good.

Leo Laporte [01:36:36]:
Yeah.

Juan Hernandez (BlindWiz) [01:36:36]:
Yeah.

Leo Laporte [01:36:37]:
So, uh, we don't have a studio anymore, so I had a mini Leo's Garage Sale with Anthony and Burke, uh, yesterday. They came over, helped me get rid of a lot of stuff because this was getting— I couldn't get it. I couldn't get in the studio anymore. So now I can. It's weird. It's weird to see all this floor space. So he got a couple of things. Yeah, that, that NUC is a nice little NUC.

Larry Gold (LrAu) [01:36:58]:
Good.

Leo Laporte [01:36:58]:
I'm glad. I'm glad it went to a good home. Thank you everybody for joining us. We, thanks especially to our Club Twit members. Without you, man, none of this would happen. There's no way we could sell ads on this show. We try to do this show every other week because there's so much happening in AI. Once a month is not enough.

Leo Laporte [01:37:22]:
Is it, Anthony? Is it the 2nd and 4th Friday of every month? What is the—

Anthony Nielsen [01:37:25]:
Well, we were talking about the 1st and 3rd, but I have to move the Media Club to a different—

Leo Laporte [01:37:32]:
That's the problem, is it— don't worry about moving anything. It's whenever we do it. The best way to figure out when we're gonna do it is either to subscribe to the club, the Twit newsletter. That's free. It's twit.tv/newsletter. Does— I hope Ty puts that, puts the club shows in there. If you're a club member, of course, there's an events button on the Discord. You can go there and see when the various things are gonna happen.

Leo Laporte [01:37:57]:
Uh, and I don't know, do we also— do we put it on the, uh, Twit calendar?

Michael (Alakazip) [01:38:02]:
Yeah.

Leo Laporte [01:38:03]:
Okay, so if you go to twit.tv/schedule, um, it will, it will be there. So that's That's probably the best way to keep up with all of our programming is we put that— we put stuff on the calendar. It's not always perfectly up to date. I don't know who does this.

Anthony Nielsen [01:38:23]:
Does this— Me.

Leo Laporte [01:38:25]:
Anthony does it. Yeah, so then it's always up to date because Anthony is perfect.

Larry Gold (LrAu) [01:38:31]:
So, uh, agent for that?

Anthony Nielsen [01:38:33]:
Yeah, it's on the to-do list.

Leo Laporte [01:38:35]:
He's got an N8N We gotta—

Anthony Nielsen [01:38:37]:
yeah, yeah, we'll get there.

Juan Hernandez (BlindWiz) [01:38:40]:
Yeah.

Leo Laporte [01:38:41]:
Thank you everybody. Have a wonderful week. We will see you. I'll be back Sunday for This Week in Tech. My old friend Gina Smith will be joining us. That should be a lot of fun.

Juan Hernandez (BlindWiz) [01:38:51]:
Oh wow.

Leo Laporte [01:38:51]:
Uh, and I will see you again, uh, soon. I hope I'll see you all again in a couple of weeks. We need a good outline, like, uh, tagline, like, um, keep, uh, Keep on harnessing.

Larry Gold (LrAu) [01:39:04]:
Burning tokens.

Michael (Alakazip) [01:39:06]:
Keep on burning tokens.

Darren Oakey [01:39:07]:
Yeah.

Leo Laporte [01:39:08]:
Keep the tokens flowing. The tokens must flow.

Danno [01:39:11]:
Another token in the can.

Leo Laporte [01:39:14]:
There's another token. See you later.

Anthony Nielsen [01:39:17]:
Bye-bye.