TWiT+ Club Shows 762 Transcript - AI User Group #18
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. Anyway, the whole idea of this particular show is to get your guys talking about what you're working on. And for— it's a show-and-tell show showing how we waste tokens. If anybody would like to do that, why don't you start, Darren? You're doing so many interesting things. Darren has a DGX Spark. Is that your main machine, or what else do you have?
Darren Oakey [00:00:32]:
Uh, I've got, I've got a 128GB Mac M5, and I've got—
Leo Laporte [00:00:37]:
Oh, I'm so jealous.
Darren Oakey [00:00:38]:
Mac M1.
Leo Laporte [00:00:41]:
You bought those before the crisis, probably.
Darren Oakey [00:00:43]:
Yeah.
Michael (Alakazip) [00:00:44]:
And yeah, I would, I would so— I wish I had done that. I have a 64GB, and I wish I had gotten the 128. I hovered over the button.
Leo Laporte [00:00:51]:
So that's what I did, is I converted my M4 Pro Mac Mini with 64 gigs to a server, and it's just running local. It's running a bunch of Docker instances, but mostly it's just running Ollama with local models. I used Ollama instead of Llama, which I'd prefer, but Ollama lets you— Llama, you have to change the command line. Ollama, you can just switch models. So I have a bunch of models and I can just switch.
Michael (Alakazip) [00:01:11]:
I don't—
Leo Laporte [00:01:12]:
I prefer to use Llama because I feel like a llama.
Darren Oakey [00:01:15]:
Well, I've got this thing called, um, for doing show and tell.
Leo Laporte [00:01:18]:
Does it do MLX is the question.
Darren Oakey [00:01:22]:
Well, the Spark doesn't, but I've got this thing called the Arbiter on the Spark. So if you can see that.
Leo Laporte [00:01:32]:
You don't need MLX for the Spark.
Darren Oakey [00:01:34]:
Yeah, this is my Spark view. And so this shows all the models that are going on. But you notice these LLM ones, it offloads them to the Macs running in the Spark cluster. if something else is going on. So, um, the Macs are running— I, I just switched from Ollama to native after last week's show or something where someone mentioned native.
Leo Laporte [00:01:58]:
And do you like it? Is it good?
Darren Oakey [00:02:00]:
It's just, uh, all I care about is speed, and it's, um, slightly faster than Ollama. So, um, so yeah, so, but the thing is, these are These are all the models that run on the Spark, but like, so my main machines just do things and they outsource.
Leo Laporte [00:02:19]:
So are they all running or is this something you switch around?
Darren Oakey [00:02:22]:
You can see this one is the only one that's active and it's running on the Boring Tech machine, which is my work machine, which is the one 128 gig. But because the thing is, all these models take lots of memory and they fight. And so this arbiter really, it's, a bunch of queues that decides who's next so everybody can queue up. It's all completely async and they'll just get a slot whenever it's available because especially these guys, the LTX ones that do the video, they take heaps of memory and stuff. But there's a lot of stuff here like, I don't know, a couple of days ago in the AI user group, I got All right, Anthony was kind enough to give me some, um, samples of you, Paris, and Jeff.
Leo Laporte [00:03:09]:
Oh no.
Craig McFarlane (CraigM) [00:03:10]:
Um, well, no, why you gotta—
Anthony Nielsen [00:03:12]:
why you gotta throw me under the bus like that?
Leo Laporte [00:03:15]:
No, it's not that hard. I think you could get the samples all by yourself. It's fine to use Anthony as an editor for that.
Darren Oakey [00:03:21]:
That's the clean, clean samples.
Leo Laporte [00:03:23]:
Nice clean samples. How many— how long did they have to be?
Darren Oakey [00:03:26]:
Well, um, I took some reasonable ones because this is very different from using Quinn. Like Quinn does the, you know, the, the once immediate clone.
Leo Laporte [00:03:37]:
That's the one I used for that silly Windows Weekly parody. And that was only, they only, it only wanted, it wanted less than 30 seconds. It didn't want too much.
Darren Oakey [00:03:45]:
But yeah, it's quite, um, which is why I asked, slow to produce. And I'm really interested in, I don't know if you know, but I've done an OL Reader thing, which, um, it's an attempt at replacing 11Reader, but it does it on the phone. And it uses Kokoro. But the trouble with Kokoro is I was worried about the fact that I couldn't change any voices because it's actually really hard. So this VoiceFit thing is a whole combination of things that tries to train a Kokoro voice. So basically I did an overnight training and it's actually doing like a proper fine tune and everything.
Leo Laporte [00:04:21]:
Oh, nice.
Darren Oakey [00:04:23]:
So it didn't— it's not anywhere near the quality of of Quinn, but, um, it's, it's okay. I uploaded a, um, if you want to listen to it later.
Leo Laporte [00:04:34]:
Yeah, I want to hear it now. What are you talking about?
Darren Oakey [00:04:37]:
Um, oh, what did you, what did you make with it?
Leo Laporte [00:04:39]:
A song or—
Darren Oakey [00:04:40]:
No, no, no. Um, I did do it. I did do it. I'll have to show you that video later, the, the song of the attempt at, um, getting training voices into Suno. But no, this was just an attempt at, um, I just got a transcript of the, of a little bit of Intelligent Machines and did it with, did it with the thing. So I'll see if it plays through this. I'll play it, but I don't know if it'll play.
Michael (Alakazip) [00:05:16]:
Yeah, I'm also curious the thing that we're doing is this.
Anthony Nielsen [00:05:20]:
I don't think the audio is going to.
Darren Oakey [00:05:21]:
Oh, really?
Anthony Nielsen [00:05:24]:
Um, or you have to share that app, otherwise the audio is not going to come in.
Darren Oakey [00:05:31]:
Ah, how about I drop it in the, in the user group and you—
Anthony Nielsen [00:05:36]:
Yeah, Leo could, Leo could pick it up.
Leo Laporte [00:05:38]:
Oh yeah, I could play it.
Danno [00:05:41]:
Oh wow.
Anthony Nielsen [00:05:42]:
Although someone's saying someone hijacked grc.com in, in the chat.
Leo Laporte [00:05:47]:
GRC.com?
Michael (Alakazip) [00:05:48]:
That was in Discord too, that someone was trying to get— to kindly get attention to that.
Anthony Nielsen [00:05:53]:
I don't know if it's real.
Leo Laporte [00:05:54]:
Um, let me text. Uh, yeah, it's dead, huh?
Michael (Alakazip) [00:05:59]:
Uh, we don't all need to let Fable loose on it to check it at the same time.
Leo Laporte [00:06:03]:
I'll text Steve. But let me play— while, while I'm playing this, so we'll play it first. Did you put it in the Discord?
Darren Oakey [00:06:11]:
Yeah, it's in the— oh, sorry, I put it in the, uh, put it in live chat.
Leo Laporte [00:06:16]:
Oh, you can put it in the AI user group. That's fine. If that's where you put it, I can find it. It's easier in live chat.
Anthony Nielsen [00:06:23]:
Yeah, one second. There you go.
Leo Laporte [00:06:26]:
There, I got it. Aloha, everybody. It's time for intelligent machines. Wait a minute. Is that supposed to be me?
Darren Oakey [00:06:35]:
Yes. As I said, it's not great at the moment.
Leo Laporte [00:06:40]:
First of all, I would never say aloha. In Hawaii, Oh, I might have.
Darren Oakey [00:06:46]:
This is a transcript. So you said all this.
Leo Laporte [00:06:48]:
Oh, this is a transcript.
Michael (Alakazip) [00:06:52]:
Leo, it's rude that I can hear birds chirping behind you.
Leo Laporte [00:06:56]:
That sounds more like her, I think. Oh yeah, oh yeah. There's meanness, there's house sparrows, there's a really loud bird. I think it's called the falcon or something like that.
Michael (Alakazip) [00:07:10]:
But you'll hear all we've got here are mourning doves.
Leo Laporte [00:07:14]:
I love doves. We have some doves too. Are they the same as pigeons though? I think they're just white pigeons.
Michael (Alakazip) [00:07:24]:
No, the one I'm thinking of goes like—
Darren Oakey [00:07:27]:
If you got 4 or 5 drops there.
Michael (Alakazip) [00:07:28]:
But it's a very specific 3-tone cadence.
Leo Laporte [00:07:33]:
Very nice. Paris has been working hard on a very important article for Consumer Reports, so we're going to be busy today. Take it easy on Paris. We're not going to be so gentle on this guy. So that's Kokoro. That's interesting. Mr. Jeff Jarvis.
Craig McFarlane (CraigM) [00:07:52]:
Huh.
Leo Laporte [00:07:53]:
As soon as I saw Leo's background, I said—
Michael (Alakazip) [00:07:56]:
That's pretty good.
Leo Laporte [00:07:58]:
Jeff Jarvis, Professor Emeritus of Journalistic Innovation, Craig Newmark Graduate School of Journalism.
Darren Oakey [00:08:07]:
But I could get Jeff now in my, uh, reader app to read me a book.
Leo Laporte [00:08:12]:
Magazine and his newest Hot Type, which is available for preorder right now from jeffgiarvis.com. What was interesting about, to me about the, uh, so are you enjoying it out there? Oh, it's wonderful.
Anthony Nielsen [00:08:27]:
I'm at 0.5 speed here.
Darren Oakey [00:08:30]:
Yeah.
Leo Laporte [00:08:30]:
Last night we swam with— See, what was more interesting to me about the, um, experience, one that I did for Windows Weekly is I let it write the text named Gabby and it was actually surprisingly That's like a 15-foot wingspan, and they get really close to you. Okay, that's enough. I don't want to hear me anymore. They're not scary is the sort of thing that someone says—
Anthony Nielsen [00:08:52]:
Oh, you're the one— you're in control.
Michael (Alakazip) [00:08:54]:
You're running it.
Leo Laporte [00:08:55]:
Oh, I'm running it. No wonder it kept going. Okay, stop.
Michael (Alakazip) [00:09:00]:
We were all waiting for you to make an editorial decision there, sir.
Leo Laporte [00:09:03]:
Kokoro, I should point out, is more than a year old. Yeah, so I, I like Kokoro with its existing voices because it's fast. It's milliseconds.
Darren Oakey [00:09:12]:
Well, that's the thing. It's the only one that's even close to real time, and that's right. That's what is exciting about it. Um, yeah, but, but it's training the voices.
Leo Laporte [00:09:20]:
I'm not sure training it's so good.
Darren Oakey [00:09:22]:
Question mark.
Leo Laporte [00:09:23]:
I wonder if you could train it with something else and then use Kokoro. Oh, Level 3 changed their legacy DNS name servers. That's what happened. So They changed their DNS servers under him. And that's why it's down. Good to know. We're in the AI user group and people were concerned. Exclamation mark.
Leo Laporte [00:09:50]:
I hate that the Apple, I have to give it punctuation because then I get in the habit of it and Whisper doesn't want punctuation. So it types out question mark. That's really frustrating. Uh, all right, so I am trying to change how— here, let's see. SCP, uh, oh, I have to give it pings.
Craig McFarlane (CraigM) [00:10:17]:
No.
Darren Oakey [00:10:19]:
Oh.
Leo Laporte [00:10:22]:
So that's cool, that's cool. I think that's interesting.
Darren Oakey [00:10:27]:
What—
Leo Laporte [00:10:28]:
who's working on something else? What else people working on here?
Michael (Alakazip) [00:10:32]:
Do we talk about something conceptually? I'm not ready to share.
Leo Laporte [00:10:34]:
Conceptually is always fine.
Craig McFarlane (CraigM) [00:10:36]:
Sure.
Michael (Alakazip) [00:10:37]:
Next time I'll try.
Leo Laporte [00:10:38]:
So it's hard to share screens, to be honest. It's hard to see them. So concepts are good.
Michael (Alakazip) [00:10:44]:
I work in the corporate environment and it's mid-year review time. So I'm doing mid-year reviews. And, and so I actually wrote a system today because in my corporate environment, I have to use GitHub Copilot, not Claude, where I put in all the old reviews from every— I've done on a person, and then I put in all their self-assessments over the years, and then I put in my manager notes, and then I said, hey, you know, output, you know, an assessment of is what I wrote for my review for my employee consistent over the years? And it's simple, but it's really something better than I've had You know, when I just, you know, forget what I said the year before, etc., etc. So I had a lot of fun writing that up today, and I made it a system which is repeatable, not a one-time thing. It wasn't just a prompt. I made a folder structure of MD files of the profile for each employee. So theoretically, in the future, I can say, hey, run this against the past. Is it HR compliant? You know, does it make sense? Am I, am I growing the employee appropriately? And so I had a lot of fun writing that today.
Leo Laporte [00:11:47]:
These seem to be things that, uh, which model are you using for that?
Michael (Alakazip) [00:11:51]:
So in GitHub Copilot, I can use Opus, but it has to be through VS Code because of corporate restrictions.
Leo Laporte [00:11:58]:
So, oh, that's funny. That's okay if they want to spend the tokens. It's, it's on them.
Michael (Alakazip) [00:12:06]:
It's, uh, yes, I work in the cybersecurity side, so I'm not going to go around saying, uh, we should open it wide open, uh, but we, we've We've had a fairly narrow opening, but I'm excited I have this much power and I stopped pivoting to the, you know, the personal computer to do things.
Leo Laporte [00:12:22]:
So one of the things I like with Hermes is it writes a skill. If you do something like that, it will write a skill automatically. You don't have to kind of hand code it. And one of the things we do a lot is what we call one-sheets, which are backgrounders about a prospective advertiser, kind of similar to what you're doing. And, uh, we have a— I can't remember what model I'm using. It's probably Claude because it's older. We have a one-sheet skill, uh, that does— that is, we used to have the staff do it. I hope the staff isn't threatened by this because they would do an okay job, but we now get the best one-sheets.
Leo Laporte [00:13:00]:
This is, this is what, uh, so this is a company that apparently wanted to advertise. It's a NordProtect product.
Danno [00:13:06]:
Yeah.
Leo Laporte [00:13:07]:
So it gives us, we, you know, I gave it some information. I gave it, actually what I first did is gave us some, gave it some of the manual spreadsheets that we had done and said, you know, and I want, uh, some other stuff like competitive stuff. And it, uh, this is a really a competitive analysis. It tells us which part of the Twit audience would it would be good for. This helps Lisa talk to the advertisers, uh, what platforms they're on, talking potential talking points for an ad. So that, again, some of this will be for her to talk to the client. Some of it is for us to vet the client. It turns out to do an incredible job for this.
Leo Laporte [00:13:47]:
One of the things we always want to know is, you know, what's your credibility? We want to know, does it compete with any advertiser? And actually, I had Patrick give us—
Michael (Alakazip) [00:13:58]:
Patrick?
Leo Laporte [00:13:59]:
Generate a, automatically generate a list of previous advertisers as well. So it has historic advertisers as well as current advertisers and it compares it to them. Then the other thing that's really important to us is that an advertiser is not just relying on us for traffic. We wanna see that they have a good marketing plan and if they don't, we work with them on that. So this is where else they're advertising, what are their podcasts, whether they use creator stuff from YouTube, that kind of thing. What the competitors are, the community reputation. This is incredibly useful to us. And this is very similar to your performance reviews, except it's a performance review of a company in effect.
Leo Laporte [00:14:39]:
And boy, we find this incredibly helpful.
Michael (Alakazip) [00:14:43]:
Well, and this is— you have no-gos, right? You've always been very public about that. So you can go ahead and have it researched like, nope, if they do crypto, if they do this, no.
Leo Laporte [00:14:52]:
Well, we don't even bother. So if they're crypto, they don't get a one-sheet, right? If it's e-cigs or there's certain categories, we just go, yeah, no thanks. But then if it's something we go, okay, well, that's something we might be interested in. Let's see if they're a leader in the field and if we can help them. Very, very helpful. And that is now done automatically. And I got the idea from Anthony because Anthony did something similar Um, did I ask you for that? Or I think you just thought it up on your own. The show, the show prep stuff.
Darren Oakey [00:15:29]:
Yeah.
Craig McFarlane (CraigM) [00:15:30]:
Yeah.
Anthony Nielsen [00:15:30]:
Uh, it was, it was actually cause, uh, we were going to have, um, the, what's his name from Noose?
Leo Laporte [00:15:38]:
Um, oh, Jeffrey Cannell. Yeah.
Anthony Nielsen [00:15:40]:
Cannell. And like, I was looking at his background. I was like, oh, this, I think, uh, the team might need some like information ahead of time. Cause there's a lot to do.
Leo Laporte [00:15:48]:
Well, it found, for instance, that he was a devout Catholic.
Anthony Nielsen [00:15:51]:
Yeah.
Leo Laporte [00:15:51]:
Which we probably, you know, I'd have to read his blog posts and stuff to come up with. So this is a Claude skill, which I've adapted and I do now for every guest on the show. And I wish I'd had this.
Craig McFarlane (CraigM) [00:16:04]:
This.
Leo Laporte [00:16:05]:
So the other thing I want to show you is I use Obsidian for everything. So I have an, I have an AI folder. This is an AI folder in Obsidian. I know it's hard to see, but it's a, it's got a little brain. This is all AI-generated stuff. So everything in here And by the way, uh, it was getting out of hand, so I asked, uh, uh, Quicksilver to create a rulebook. First of all, I said, organize this sucker, and then create a rulebook so that agents know where to put stuff as they're doing stuff, and then how to name it and so forth. And then, uh, so everything is in here that the agents, uh, work on.
Leo Laporte [00:16:40]:
But this show prep skill that you— I, you know, basically it's unmodified from Anthony's. But we now use it when somebody's on. This is probably the first one.
Danno [00:16:48]:
Yeah.
Leo Laporte [00:16:49]:
Well, this is June. This was his second appearance. This is one of the first ones you did, but really, really useful and very extensive. This is the kind of thing, uh, in the past, you know, when I was working as a show host, if you had a producer, they would, you know, kind of do all this grunt work. Now it's just done automatically. And Anthony, you know, one of the things, if I show this to Paris and I show it to a lot of people, they go, Yeah, but what about hallucinations? I haven't seen any hallucinations in these.
Danno [00:17:20]:
Yeah.
Michael (Alakazip) [00:17:21]:
I haven't seen hallucinations in a long time.
Leo Laporte [00:17:23]:
I think that might be, uh, well, at least I think we know how to avoid it. Let's put it that way. Right?
Craig McFarlane (CraigM) [00:17:29]:
Yeah.
Anthony Nielsen [00:17:30]:
I think it also depends on, uh, Larry is also talking about how like the content, like Signal really matters in terms of context. If there's not a lot of stuff out there on a person, then I would be worried.
Leo Laporte [00:17:44]:
They might make up stuff, right? But everybody we have on has a huge— so this does, and this I think comes all from your skill, Anthony, originally. It was really—
Anthony Nielsen [00:17:54]:
and also it wasn't that hard to put together. Basically, I asked, you know, I gave Claude kind of what we want, and then like I had it go research Jeffrey. And then there was a lot of back and forth in terms of like what we refined the format and like what kind of stuff is important to, you know, right, you and, you and Jeff and Paris. Like initially it was like too much on the technical side. So like we just went back and forth refining it. And then, then I asked Claude to package this as a skill.
Leo Laporte [00:18:28]:
It's really good.
Anthony Nielsen [00:18:29]:
Um, so it's, it's really easy to, you know.
Leo Laporte [00:18:32]:
Yeah. You go back and forth.
Darren Oakey [00:18:33]:
Yeah.
Leo Laporte [00:18:34]:
The provocative challenge.
Anthony Nielsen [00:18:36]:
Future.
Leo Laporte [00:18:37]:
The only downside to this is, is now Hermes, I mean, Quicksilver thinks that Intelligent Machines is my only show. And I had to really say, no, no, no, I do other shows. There are other angles.
Craig McFarlane (CraigM) [00:18:47]:
Same.
Anthony Nielsen [00:18:48]:
Claude thinks all I do is Intelligent Machines.
Craig McFarlane (CraigM) [00:18:50]:
Right.
Leo Laporte [00:18:51]:
It can get heavily influenced by the work it's doing. But this is really, I mean, this is a dossier. This is incredibly useful. And I wish I'd had this my whole career. It would have been, it would have saved me. And you know, a lot of, when an author comes in to do an interview, Their PR people give you, you know, the 5 questions to ask, and they're so bored because they get asked the same question. They're usually terrible questions anyway. This is so much better because it, it, yeah, it kind of gives you questions which I never— I don't really use, but it gives you angles like what makes this interview compelling.
Leo Laporte [00:19:29]:
And if you read through it, you will get 3 or 4 things that you go, oh yeah, that's— I gotta— oh, that's gonna be juicy. And that's what you're always looking for. That's what show prep is about. That's what they do you know, on The Tonight Show or, you know, they did on Letterman, is the producer will go and interview the comic and say, give me 3 stories, and then say, yeah, we're gonna tell that story, and then puts it on a card for Letterman who says, I hear you had some trouble the other day on the highway, and then you get the whole story. This is kind of like that where— but you don't have to do a pre-interview with a guy because everybody's on the internet so much. This is fantastic. I'll show you some of the things that I—
Danno [00:20:05]:
go ahead.
Michael (Alakazip) [00:20:05]:
Once you build a database of, of the, the style that works for your brand and you personally and your hosts, you could have it go search for people who are up and coming, right?
Leo Laporte [00:20:15]:
Good point.
Michael (Alakazip) [00:20:16]:
Up and coming.
Craig McFarlane (CraigM) [00:20:17]:
Good point.
Michael (Alakazip) [00:20:17]:
That you'd be like, that person I think would fit really well to our format.
Leo Laporte [00:20:21]:
So that's why I put this in Obsidian. A lot of what I do, I'm thinking in terms of the future. Like, I want it to be model agnostic. I want it to be harness agnostic. I want to be able to, in 5 years, use a different model and different harnesses and different skill sets. So my thinking is putting stuff in Markdown in basically flat text files is the most likely to survive the future. And so this is, this, this, like this AI Fold, this is, uh, Karpathy's, uh, LLM Wiki is in here.
Anthony Nielsen [00:20:57]:
And, uh, so that's just a subset of the Obsidian Vault too, right? The wiki?
Leo Laporte [00:21:03]:
Yeah. Yeah. 'Cause I have a— so The other thing is I use an Obsidian MCP server. And so Hermes has full permission to go into this and read anything it wants. And then some of the stuff, like I had it do, I have it do a health summary for me. So it writes us, again, it organizes it, does a schema. It's got a lot of information in here. And then it weekly does a health summary for me.
Leo Laporte [00:21:30]:
Like, okay, you know, you should be eating more veggies. That kind of thing. It's very useful. I have one now for money. I've had it, I've been using, as you know, our sponsor Monarch Money, but I realized, wait a minute, I got, I have this very capable AI and I had it look for the problem with Monarch Money or the thing that Monarch Money does. And so many of these things do is they can log into your accounts. Well, it turns out if, if your company, most financial services companies, you guys know this. Have APIs or some form of read-only access.
Leo Laporte [00:22:07]:
This is a third party, I know I'm trusting it, called SimpleFin. It's a buck a month. It's just like Plaid or any of those other companies, but it's kind of open, read-only access, and the AI can use it. And in fact, it works better. Monarch's always had trouble with our PoE in Sperity. It could never log in.
Danno [00:22:27]:
Yeah.
Leo Laporte [00:22:29]:
SimpleFin works with this. So now I use SimpleFin as a bridge to all of my accounts and the AI can read it. So at close of day, 5:00 PM every day, it goes through all of my accounts and gives me a financial summary and not just net worth, but everything, like, you know, all sorts of information. And that also goes, uh, to Obsidian. So I have a money folder in Obsidian and I have a fitness folder, but the AI folder is really the one that's the most interesting. Um, I, uh, all the plans for, uh, future stuff go in here. They're—
Danno [00:23:03]:
they—
Leo Laporte [00:23:03]:
it will also put it in the project folder, but this is the, the thing I've been working on, this TwitAds thing. Uh, I think it has— does it have its own folder? Maybe not.
Michael (Alakazip) [00:23:12]:
Well, this is so interesting because I have done the exact same thing. Obviously we've never talked about it, but I didn't use Obsidian. I just have folders upon folders and Markdown files in my iCloud. I don't care how deep they are because the AI manages the construct.
Leo Laporte [00:23:27]:
As long as it knows where stuff is. Yeah, exactly.
Darren Oakey [00:23:30]:
Yeah.
Michael (Alakazip) [00:23:30]:
And it doesn't seem to have any trouble scaling. And going back to last week, I feel like this is an attempt for some memory that we all would like Darren to have.
Leo Laporte [00:23:38]:
It's another kind of memory, right? I am using, uh, uh, Mnemosyne. There are a lot of different memory systems. I've tried a bunch of them, like Honcho. All of them basically are vector systems. They use a database. So this isn't that fast or smart. This is more designed for the future. But while all the plans for the Twit ad thing end up getting put here in Markdown, and the nice thing, just like your iCloud, the nice thing about this is it syncs to all my machines and it's human readable.
Leo Laporte [00:24:11]:
So I always have this. But then the other thing I do, because it's not just me and this is my personal Obsidian, is, uh, I have it, uh, also create a website. So actually, this is one of the things I wanted to show you. All this time I didn't realize you can— Cloudflare, basically you can make free websites, giant websites forever, and it has excellent API access. It's a little— it's got the worst UI for creating the tokens and giving it the permissions, but once you get that done— so this is, uh, this is, uh, This is the public part of my website, pages.laporte.cloud. And so I put like things like my pour over notes in here, or, you know, this is, this is a comic book. This actually is pretty cool. A comic book I made for the neighbor kids.
Leo Laporte [00:24:58]:
They're going to come over and learn bagels. And I thought, oh, I could turn this into a chemistry lesson. They actually look like this and that I look like that, I guess. And so it created super, it did, it did this all.
Danno [00:25:08]:
Beautiful lamb chops. Yeah.
Leo Laporte [00:25:11]:
You know what? The picture I gave it, you could see my sideburns and it decided, oh, you must have big sides. That's all right. It's a comic book character. I gave it a simple prompt. I'm going to do making bagels with these 2 kids, Sarah and Matthew, our neighbors. I want to turn it into a chemistry lesson. I did later give it pictures of the kids, picture of me, and pictures of the bagels and the video that I did with you, Anthony. As reference material.
Leo Laporte [00:25:39]:
But literally, that was the prompt before the reference material. And it looked up all the chemistry. It created superheroes. That's yeasty lactobacillus. This is gluten. This is, what is amylase, right? That cuts the glutens. This is the Maillard reaction. And this is lye.
Leo Laporte [00:25:58]:
It's wearing a hard hat and safety glasses. So it created these superheroes. And it did a dialogue. It did the chemistry lesson. With that, this was one shot. I was blown— I didn't expect anything this good. I was blown away.
Darren Oakey [00:26:12]:
That's amazing.
Michael (Alakazip) [00:26:13]:
You've got to do, uh, you got to do Salt Hank's sandwich next.
Leo Laporte [00:26:16]:
Oh, that's a great idea. And send it to him. He'll hate it. He'll be so pissed off. Uh, although I could show him how he could do it, but yeah, I— oh, and the other thing it did, which was great, because I used Lye, it wrote this, Lye safety, read this. So, which I figured, well, an 8th grade boy is going to love that. Anyway, I had a really boring page, which was just text. And I decided yesterday, I played, I said, I want to make this prettier.
Leo Laporte [00:26:41]:
And I just came up with a concept. I took a few iterations because it didn't really get it that I wanted it to look realistic. And then I realized I don't even need any text. These could just be like, um, this is Quicksilver's blog. So it has a caduceus. That's its logo. This is, uh, a page I made some time ago for, uh, Inside Machines. It's a one, it's a Stumble upon, kind of.
Leo Laporte [00:27:04]:
That's the bagel thing. This is my pour-over. But then I have private stuff, which was in public at first, and somebody, somebody signaled me and said, you know, I can see a lot of your business information. I said, oh, I better put this. So I made a gate, and you go down the garden gate, and inside the garden path are— these are the private ones. So, um, and you have to have—
Darren Oakey [00:27:25]:
get—
Leo Laporte [00:27:26]:
you have to have, uh, access. I'm using Cloudflare, has a very good access system, which I kind of avoided using for the longest time. I've never logged in on this machine, so let me just quickly go through the GitHub. There it is. So this is the page that goes along with essentially what is that Obsidian folder, but it's human readable. And this is for people like our IT, our CISO, and, and our IT guy so they can read it. So what it's— this is actually really cool. I asked it, let me do that again.
Leo Laporte [00:28:01]:
I asked it, as you go, so this is the original plan, the specification, but as you go and you build it, add a page, you know, add a page for everything you're doing. So when did we start this? When Fable first came out. So whatever's that, that's last June.
Danno [00:28:23]:
Yeah.
Leo Laporte [00:28:23]:
It's about a month ago. And then I paused for a while. It said, okay, you gotta interview Debbie and Lisa. It wrote me questions for Debbie and Lisa. It has follow-up questions. It did an analysis. One of the things that's great is code is the best spec, right? And since we have the source code for the original system, plus the database and all the schema, I figured I've got the spec. I don't have to write a spec.
Leo Laporte [00:28:49]:
This is, this is something we've been using for 12 years. So I had it analyze it. It found a lot of things wrong with it, which we knew about. And there's also stuff in here we don't do anymore. Notable hazards and debt: SQL injection via string concatenation in multiple pages, secrets hardcoded in source, coded in source, magic values. It was a single business logic split across 154 procs with versioned-by-suffix evolution. Rundown 4, episode dates 3. Uh, so it wasn't written— it was written really a hodgepodge, but it was a spec.
Leo Laporte [00:29:28]:
So I was able to use this. It built each all the way through, and I didn't— boy, I didn't look at code or anything, but it did provide us with, you know, these, these— this is, uh, this, you know, code it provided. It wrote all of this. So I, I felt like— and I gave it a pretty— it was very structured. I said, I want you to do it in little chunks. You can see it's A, B, C, D with little chunks. Um, I want you to write it, Fable, write the spec, hand it off to Opus 48 for coding, which was, I thought, the best thing at the time because I thought Fable was going to go away. Then I have ChatGPT-5.6 Saul review it.
Leo Laporte [00:30:05]:
They go back and forth. Actually, Fable and Saul go back and forth on the plan. 48 codes.
Anthony Nielsen [00:30:12]:
How does the passover work?
Leo Laporte [00:30:15]:
Well, the first thing I did— well, first I was pasting, cutting and pasting. Okay, I went over to Codex, tell me, look at this. But then I said, well, that's a pain in the ass. So I had them do something called agent mail in— actually did it in Obsidian. I said, we'll make a mailbox for each person and then check the mailbox every few minutes. So I had an agent mail system. It should be here somewhere.
Danno [00:30:40]:
Where is it?
Leo Laporte [00:30:40]:
Where is it? About Leo, Council, Learn, Mail. So each one of them has a mailbox. They have an archive. These are all the messages. These are for them.
Craig McFarlane (CraigM) [00:30:51]:
Once they read something, they move it into the archive.
Leo Laporte [00:30:54]:
Yeah, once they read it. And then there's an inbox. So you see there's a pending message. But then I saw Jack Dorsey's Buzz and I thought maybe we should move this to Buzz. But Buzz is too immature. But eventually this will become Buzz. which is intended for exactly this kind of agent-to-agent interaction. But I like it that I can read it.
Leo Laporte [00:31:13]:
And then I made a polling system so that— or a push system so that instead of having to check the— pull the mailbox, it actually would get up. That's not working perfectly yet, but the idea is when Daedalus, which is Codex, reviews something, it then sends it to Kenobi, which is Claude Code, And so this is them. See, K, D. So K is saying, I accept the re-review, but blah, blah, blah, blah, blah. And then the third party on this is Quicksilver, which also has a mailbox. And that's various models, although I've settled on Grok 4 or 5 as being probably— all of these I have subscriptions for. Grok is free because Elon just for some reason gave me a blue check and a Twitter Plus account. So I'm using Grok as kind of the— That's an agentic thing.
Leo Laporte [00:32:02]:
So I actually have them talk to each other now. We set up a new fast mail interface. So I said, hey, tell Q and— Q, tell D and K about it. So this, this, this works quite nicely. It's kind of fun. Anyway, so this is the planning on the Twit ad thing. I'll show you what it looks like. They've actually done fairly well.
Leo Laporte [00:32:23]:
We're a little bit stalled right now because Lisa got sick of reviewing it. She said, I don't want to look at this anymore. But so I can't go too much farther because without her input— but this is pretty much a duplicate. So this is what we call an insertion order, an add-by. You can go in here and choose the dates. You can remove it, move it around. It keeps track of it. But so kind of the workflow is you write potential orders, client approves it.
Leo Laporte [00:33:11]:
You know, we have, we have sales reports, automatically does, generates reports. These are the advertisers, these are the agencies. It's a little light right now because this is just sample data. This isn't the real thing.
Craig McFarlane (CraigM) [00:33:22]:
Yeah.
Leo Laporte [00:33:22]:
real data. We have who's— what ads have been sold where. We can do it by show. You can look at the schedule for Hands on Tech, what ads are running on Hands on Tech on what dates. You can see what's used, what the inventory, what's left. It will generate rundowns which we send to the producers. It basically does the backend for ad sales and continuity.
Michael (Alakazip) [00:33:49]:
Wow.
Leo Laporte [00:33:51]:
And I think it does a very good job, to be honest. These are all the hosts, how much they get. You shouldn't look at that. Here's the users. Here's— this is so Lisa can edit the boilerplate for the ad terms. She hasn't done anything yet. It has a complete audit of every change made. I, by the way, I haven't looked at the code.
Leo Laporte [00:34:15]:
I haven't really looked at the user interface. I haven't done anything with this. This is kind of an amazing project that within a month using that, I think, what I think was a pretty good way of doing it, it's really generated something incredible.
Michael (Alakazip) [00:34:36]:
Well, I mean, this legacy code is such a problem across enterprise.
Danno [00:34:40]:
Right.
Michael (Alakazip) [00:34:41]:
This model, I think, is gonna— you just feed it to—
Leo Laporte [00:34:43]:
Take the legacy code, it's your spec.
Michael (Alakazip) [00:34:46]:
If you asked a consultant to come redo it, oh my gosh, get the millions out, get the years out.
Leo Laporte [00:34:51]:
Well, that's frankly why we're stalled right now, 'cause, okay, just between you and me, Lisa, there was a bug, or she thought it was a bug. I think she couldn't figure out what was going on, so she thought it was a bug. She said, no, this is too much work. She said, I'm gonna give this to Patrick, And Russell, and we're gonna hire a consultant. And I think partly it's also she didn't want— it's a little difficult for us to work together 'cause, you know, she says this and I go, well, wait a minute. And then, so I think she wants to bring in a third-party consultant. And I said, no, please, I beg of you. This is almost done.
Leo Laporte [00:35:25]:
I would say this is done. You know, in fact, all it's waiting on is her and Debbie's, you know, kind of review of the whole thing.
Craig McFarlane (CraigM) [00:35:34]:
Yeah.
Leo Laporte [00:35:36]:
I think it's incredible.
Michael (Alakazip) [00:35:38]:
I'm blown away. You just need a geek mediator, right, that can talk—
Leo Laporte [00:35:42]:
That's all I am.
Michael (Alakazip) [00:35:43]:
Speak and, uh, that's all I am.
Leo Laporte [00:35:46]:
That's all I am. And what I've had her do, this is another thing. So I did the interview, recorded it, just gave it the transcript— or actually, I had to do the transcript, gave it the audio and did the transcript. And then, uh, the next step is I'm having Lisa and Debbie record Looms or videos I say, use the thing. And as you're using it, because it can't— if you just tell it, oh, I don't like this, it won't know what you're talking about. So do the video, circle it, point with your mouse or whatever, say, I don't like this, move that over there. I asked, if I send you a video? He said, oh yeah, I'll get a screen grab when she says, I don't like this, and I'll know what she's talking about. So they're recording videos that I don't even look at.
Leo Laporte [00:36:29]:
I feed it to Kenobi, to Claude. And Fable's really good at this. Fable looks at the video, interprets the video, writes a plan. The 3 of them go over the plan, and then they make the code. I don't— I'm pretty hands-off. The only time I'm in there is when I'm verbally talking to Lisa. She says, yeah, I don't like this here. Can you do that? And then I will type it in.
Leo Laporte [00:36:55]:
or dictate it to Claude, and it will do it, and it does it instantly. So this, I feel like this was a great proof of concept for me because you have legacy code, you have a spec, may not be the best spec, but you have a, you know, code is the best spec, right? It's very concrete.
Darren Oakey [00:37:17]:
And Larry was saying Morgan has had a great success just Porting old stuff to new tech.
Leo Laporte [00:37:24]:
The challenge here is that the code, first of all, is really old, really crufty, and a lot of it is not used anymore, and a lot of it doesn't work, and they have workarounds. There's a huge amount of manual workarounds they've been doing, which, for instance, one of the things that we do in the schedule is called fair rotation. And right now, Debbie does this by hand. The idea is, Every advertiser— that the first position in a podcast is premium. Each subsequent position is less valuable. So you want to give a podcaster— you don't want to give the same advertiser the same place in every show. You want to rotate it. You want to round-robin it.
Leo Laporte [00:38:04]:
But it's challenging when you have multiple advertisers, multiple shows, ads come in afterwards. So Debbie's been kind of dragging them around by hand. And so we built fair rotation where she specifies the dates. It will look at the whole year, but can only change stuff that hasn't aired yet. And we'll do a rotation based on what that advertiser's had throughout the year. And this works so well. It's taking a manual process that was a little— that was a lot of verbal back and forth. I had to have them explain to me what they do.
Anthony Nielsen [00:38:39]:
Yeah.
Leo Laporte [00:38:39]:
What the goal was and all of that. So that was where I was an intermediate there. Um, so there was a little bit of trying to understand, well, what is it? What are you trying to do here? And then because their program didn't do it or didn't do it well— actually did do it, but it didn't do it well enough that Debbie didn't have to do a lot of manual stuff. So I actually don't know. I haven't— they haven't vetted this. I can't get anybody to look at it anymore. So That's probably the biggest thing is the human interface is challenging.
Craig McFarlane (CraigM) [00:39:11]:
Always.
Leo Laporte [00:39:12]:
Always. That's the worst part, right? I have no trouble with the AI. It's doing exactly what I want.
Darren Oakey [00:39:19]:
And this is something that I find in real enterprises that I hate the make a new system because what, especially in the olden days when you take 6 months to make a new system, Um, your main competitor is the existing system, and people, people never take that into account. And in 6 months, anything that really had to be fixed got addressed in the other system if it was really important.
Leo Laporte [00:39:44]:
So it's a moving target.
Darren Oakey [00:39:46]:
By the time you— yeah, by the time you release it, nobody's all that excited about the new system, which is why Martin Fowler talked about this thing called the strangler figure analogy, which is you basically start replacing the old system system bit by bit. You build a new system on top of it. And I, I always prefer that because, yeah, then nobody has a choice and, and you're not competing with the existing system.
Leo Laporte [00:40:07]:
Well, the good news is the developer who wrote this system left 12 years ago and refused to do anything with it. They actually hired somebody, Paul, whose sole job is keeping it from dying, but he has not improved it as far as I know. He's not done anything to change it, so I don't have to worry about That, thank God. But I still have that. I mean, they are still using the old system. They haven't switched over, and they won't until I, you know, get a full approval on it. Anyway, back to my, uh, back to my garden. So this is the private garden.
Leo Laporte [00:40:37]:
This is the stuff I don't show anybody. Um, one thing I really like, um, like, these are our travel plans. Um, this is Twit ratings because I And so we have an API to our rating system. It's called Uniblabs. So I got Patrick to tell me, you know, send me the page for that. And I got the AI to do that. So it downloads every week, it downloads the ratings and does something I've always wanted, which is 3-month rolling averages I could never get. So it's giving me the information from a programming point of view, not an ad sales point of view that I want.
Leo Laporte [00:41:13]:
What is this? Oh, that's a good one. I already looked at that. Uh, travel ratings. Oh, this is just a— I don't know why it's white like that. This is a Zen bug. That's a dashboard to Hermes. But then I thought, well, I got this nice fountain. What can I do with this? So, uh, I wrote a little— I took my I Ching program and I put it in here.
Leo Laporte [00:41:38]:
In fact, it's even got— I said, give me a babbling brook Some music, some birds singing, some meditative stuff. And then I can ask a question. What will happen to the Twit ad sales system now? So I don't know if you— I Ching is an old ancient Chinese oracle that you ask a question of it and then you can either throw coins which is the modern way to do it, which is actually truly random. Or this is the old way, which is the way I do it, which is yarrow stalks. And you do this with real yarrow stalks. You divide them, you throw them on the ground, you divide them, and you count them. You divide them and count them. It's not fully random, but I think that's part of it anyway.
Leo Laporte [00:42:29]:
The theory is you're influencing the arrow, the oracle, with your— by thinking of your question as you go through this Kind of really tedious, monotonous. It takes about 15 minutes unless you're super proficient, which I'm not, to go through the arrow stalks. But you've spent 15 minutes thinking about your question, and then it generates a, um— and people who are fans of Cryptonomicon or, uh, Quicksilver of Neal Stephenson will recognize this— it generates what's called a hexagram, which is actually binary.
Craig McFarlane (CraigM) [00:43:01]:
Oh, wow.
Leo Laporte [00:43:02]:
And then the hexagram has an ancient interpretation. Following has sublime success. Perseverance furthers. Okay. No blame. See, I like that. This was the question is, what should I— what do I do with the sales system now? Within the lake is thunder, the image of following. That's the noble one.
Leo Laporte [00:43:20]:
Anyway, you read that. But then I got really crazy and I said, okay. Oh, the gateway's down.
Craig McFarlane (CraigM) [00:43:26]:
Shoot.
Leo Laporte [00:43:28]:
Oh well. I have— I send it to, uh, the AI for further interpretation. I actually sent it to Quicksilver, and because Quicksilver knows me, it weaves it in to my own history. Like, it'll say, well, you're gonna have to work with Lisa because there's truthful in the good and blah blah blah. The receptive affects something. This is actually a good oracle. So this is my little oracle. It's at the— it's at the water fountain there.
Leo Laporte [00:43:57]:
But I really like the idea that this webpage has no text. You don't really know what you've got.
Anthony Nielsen [00:44:05]:
It's like we're playing Myst.
Leo Laporte [00:44:06]:
Yeah, yeah, it's like Myst. It's Myst. I didn't even think of that.
Darren Oakey [00:44:10]:
But I—
Craig McFarlane (CraigM) [00:44:11]:
I didn't even think of that all this time.
Michael (Alakazip) [00:44:13]:
I know. You gotta have the sound of the whirring CD-ROM, like, in between.
Leo Laporte [00:44:18]:
I could have done that. You know what, the inspiration, I don't— I'm not sure. I think subliminally, my, my mom, as you know, passed away a couple weeks ago, and she's, uh, her— she's gonna have a little plaque in a garden that looks a lot like this. And so that must have been in my head, that I wanted these little stones in the garden. It was hard to get it to— I wanted to look like paving stones, and I couldn't quite get that, but—
Craig McFarlane (CraigM) [00:44:39]:
It's pretty close.
Leo Laporte [00:44:40]:
It's close enough. And I said, don't do any text, just do runes. And I, you know, so I thought this was kind of fun. I did this yesterday just for fun. Um, who did it? I think GLM54 did it. This whole thing from yesterday?
Darren Oakey [00:44:57]:
Yeah.
Danno [00:44:58]:
Wow.
Leo Laporte [00:44:59]:
I don't know if you remember what it looked like before, but it was really boring. It was just text. You know what it looked like? An AI did it is what it was. You know how you can tell an AI site right away?
Anthony Nielsen [00:45:12]:
Yeah.
Leo Laporte [00:45:13]:
Like, you know immediately, uh-oh, we've lost somebody. He's going to work. Danno has a job to do. All right, that's my show-and-tell.
Anthony Nielsen [00:45:21]:
I think Craig wanted to talk about something.
Leo Laporte [00:45:23]:
Yeah, sorry, I took all the air out of the room.
Craig McFarlane (CraigM) [00:45:26]:
Apologies. No, it's good. You know, it's always— one of the coolest things is to look over someone else's shoulder and say, well, that's interesting. Why did you do that?
Leo Laporte [00:45:37]:
Right.
Anthony Nielsen [00:45:38]:
If you want to share screen, you could, uh, do that at the bottom. Yeah, there's a—
Craig McFarlane (CraigM) [00:45:43]:
I was just going to show on my, my video, but—
Anthony Nielsen [00:45:45]:
Oh, that's—
Leo Laporte [00:45:46]:
we could do that too. And I— if you want to put a link to Discord, I can play it for you, or—
Michael (Alakazip) [00:45:51]:
Oh, Craig, stop showing off. You got a fade in between it.
Leo Laporte [00:45:57]:
Oh, fancy.
Craig McFarlane (CraigM) [00:45:59]:
I do a lot of work with Office Hours, you might know from—
Leo Laporte [00:46:02]:
Oh, you do? Oh, nice. Is that still going?
Craig McFarlane (CraigM) [00:46:05]:
Oh yeah.
Leo Laporte [00:46:05]:
With that, Alex?
Danno [00:46:06]:
Yep.
Leo Laporte [00:46:07]:
Yep. Oh, good.
Craig McFarlane (CraigM) [00:46:08]:
And, uh, uh, we have a little sock puppet we use. No. Um, so if you don't know, Office Hours is a, uh, 7 days a week, uh, Q&A show aimed at, uh, media production. And each— and there's a high quality for each show. And so we might have, uh, 10 or 15 people behind the scenes prepping, doing all the technical logistics. Over the years, we always had a whole bunch of spreadsheets. I pushed Google Sheets as far as we could and there's no budget to do any more. Obviously, with AI, it's like, well, I could do something.
Craig McFarlane (CraigM) [00:46:51]:
Because I had done all the data modeling and the workflow that we needed, it was a piece of cake. Literally within a few days, I had like 80% of everything we needed. When I saw what you were doing, Leo, I was thinking that at a certain point, your development and change speed is faster than they can complain about issues.
Darren Oakey [00:47:19]:
Oh, yeah.
Craig McFarlane (CraigM) [00:47:19]:
You only have to get close enough that they can start doing stuff, and if you can fix the bugs as quickly as they come up, you'll get there.
Leo Laporte [00:47:29]:
And yeah, that's— I know that. It's convincing them that that's the case. Lisa's big concern is, well, what if it crashes and you're on the air? And A, I want to say it's not going to crash. Instead of being MySQL with .NET, it's Go code. It's going to be very robust. It's not going to crash. But if it does, I Just as you know, Greg, you just, I can literally be doing Windows Weekly and go, okay, let me just type a command. It'll fix it.
Leo Laporte [00:48:00]:
Just tell me what the crash is.
Danno [00:48:03]:
You can order a sauna helmet while you're—
Leo Laporte [00:48:06]:
Yeah, it's hard for people who haven't done this to understand we're not writing code. It's not even anything like that. It's just talking to the thing and saying what you want. This is cool. What does the ShowUp do?
Craig McFarlane (CraigM) [00:48:23]:
There's literally whatever, 7 days a week, we have to organize and orchestrate all of the sign-ins. Let me just see if I can— I'm trying to figure out which— not this one, this one. Here's one for a Saturday in August. There's nobody signed up because I didn't want to show with names in it. But if I'm qualified to be a host, there's a whole bunch of different skills you can get.
Leo Laporte [00:48:50]:
Oh, you have to coordinate a huge schedule. There's a bunch of people.
Craig McFarlane (CraigM) [00:48:53]:
Yeah, and everybody's volunteers. So we've got hundreds of people that are filling in different roles, whether it's host, reader, engineering chart, you know, pre-show coordinator, all these various people, including guests. And so I can sign up for this 'cause I've got this. And confirm. Yep.
Leo Laporte [00:49:16]:
And then I mean, this is such a good example of something you really would be way too complicated to try.
Craig McFarlane (CraigM) [00:49:22]:
Yeah. And so I also have like all the profiles of everybody, including all the roles that they've filled in, the questions they've asked different days, and just there's just a ton of stuff that's available. And so we're now doing things We cut over super quick from all the spreadsheets and such that we could just bang this out. And so everybody can easily check your schedule. Just go back to settings. I have— I threw in the calendar feeds, piece of cake, notifications, just all sorts of stuff that was like You know, I'm going by the old metrics of, oh, if I needed to do that, how much would it be? And now it's like, oh, I'll just add this feature.
Leo Laporte [00:50:13]:
That's amazing.
Craig McFarlane (CraigM) [00:50:15]:
Have a—
Michael (Alakazip) [00:50:16]:
So Craig, this is clearly amazing, Craig. Um, yeah, can you talk briefly about like, did you vibe code this? And if so, what was your approach? Like, this is a— what appears to be a finished product. Like, you know, how did you get there?
Craig McFarlane (CraigM) [00:50:30]:
It's been underway I started it in March, and then I think we started using it full-time within 3 weeks.
Leo Laporte [00:50:38]:
Did you point it at the spreadsheets initially? I mean, what was it? Or is it completely greenfield?
Craig McFarlane (CraigM) [00:50:43]:
I've got engineering and product background, so I knew the data model, I knew the screens I needed, knew the functions. And so that was all in my head.
Leo Laporte [00:50:53]:
This is why this isn't going to cost people jobs, because It just, you go up the stack a little bit more, that's all. You're not sitting there writing brace colon. That's exactly, yeah, 100%.
Michael (Alakazip) [00:51:04]:
You can't go zero to hero, but if you understand what you want and how to articulate it in computer speak.
Leo Laporte [00:51:09]:
Yep, that's huge.
Michael (Alakazip) [00:51:10]:
The evidence is in front of us. Yeah.
Craig McFarlane (CraigM) [00:51:13]:
Yeah, so it's roughly 65,000. I was just checking it this past week, 65,000 lines of code. If I was gonna do that by hand, And that's not— I thought that, oh, is that libraries? And it's like, no, that's all the— there's a ton of functionality in here.
Leo Laporte [00:51:31]:
I haven't even counted.
Craig McFarlane (CraigM) [00:51:32]:
There's like 7,000 or 8,000 questions that have been asked over the years. And so this catalogs all of those that you can see all the questions you've asked over and be able to go jump right to it. And so if I go to like, here's the show, or I was on the panel, I can scroll down, see the question, you know, click and then see it play. I can go to the archive and do something like, oh, I want to see everything about Canon lenses, and then it would show me all the questions that have Canon lenses, and then I can just say autoplay, and now I've got a show that is playing out all of the Canon, as if it was a show just about Canon lenses. Um, and we're going to be pushing that out to the website, uh, hopefully soon because it's, it, it's so much fun. I, I've done, uh, uh, a few of them where I just picked a topic and then just watched a show, and it was actually probably 20 different shows, but it's all those random question segments, uh, that happened to touch on that topic.
Leo Laporte [00:52:43]:
This is actually something Alex has from day one, even before Office Hours. Alex has always been about automating and using and writing software for this stuff.
Danno [00:52:55]:
Yeah, I was going to say Alex would freak out if he saw this.
Craig McFarlane (CraigM) [00:52:59]:
Yeah, I was— I did it after he left. Well, I mean, he saw all the spreadsheets, which was about 80,000 lines all in.
Leo Laporte [00:53:05]:
But, you know, the Go app is 28,500 lines of production code. I just asked why. Well, but lines of big tests, the DD discipline, maybe 1,000 lines of code.
Michael (Alakazip) [00:53:17]:
Yeah, sell this to YouTube because I wish YouTube would do this where I was like, you know, I want a topic and I don't want to go search for individual videos and watch 40 minutes of, you know, whatever. I want the pieces I want and then make me my own YouTube video out of it.
Craig McFarlane (CraigM) [00:53:31]:
Yeah, yeah.
Leo Laporte [00:53:33]:
Oh, that's interesting.
Craig McFarlane (CraigM) [00:53:34]:
The native in here, where did I leave that?
Leo Laporte [00:53:37]:
So you're assembling a show from parts? Or no?
Danno [00:53:42]:
Yes.
Craig McFarlane (CraigM) [00:53:42]:
Yeah. So if I go into the archive and I type in—
Danno [00:53:47]:
So you can automate clip making with this?
Craig McFarlane (CraigM) [00:53:50]:
And I say—
Leo Laporte [00:53:51]:
We should look at this, Anthony. This is cool.
Danno [00:53:54]:
Yeah. Yeah.
Michael (Alakazip) [00:53:54]:
I want this for Twit too.
Leo Laporte [00:53:55]:
What are you talking about?
Anthony Nielsen [00:53:56]:
This is really cool.
Leo Laporte [00:53:56]:
This is really interesting.
Michael (Alakazip) [00:53:57]:
What's all the things Leo has said about X?
Danno [00:54:00]:
Go.
Darren Oakey [00:54:00]:
Yeah.
Craig McFarlane (CraigM) [00:54:01]:
I go to that show or I can come here and say, Well, let me just do it again. Sony, and it's going to do a search.
Danno [00:54:12]:
Another Twitter user benefit.
Craig McFarlane (CraigM) [00:54:14]:
And then if I bring this up, then I can just have it keep going and then it will just burn down all of them.
Michael (Alakazip) [00:54:31]:
Is it generated from transcripts? How does it find the topic and segment it?
Craig McFarlane (CraigM) [00:54:39]:
Sorry, what was that?
Michael (Alakazip) [00:54:41]:
Well, the idea is amazing. I have 1,000 questions, but the top one is, how do you find the content? Is it via transcript and then it chunks the transcripts together and then finds the video?
Craig McFarlane (CraigM) [00:54:54]:
Yeah, we've got a few different projects going on right now. It's just looking at the question text itself. So if in the course of talking somebody mentions Sony, it's not going to come up in this simple version. But, you know, for something I did in, you know, I like to say I wasted minutes on that. You know, it's, uh, it's one of those things where the, the bar is so low to be able to bang something out that if people don't use it, it's like, it's okay, it didn't cost that much.
Darren Oakey [00:55:33]:
Right.
Craig McFarlane (CraigM) [00:55:34]:
If it delights a few people and they're excited about it, great. Well, it's a one-to-one. I'm finding that there's a lot of features that I built, like I built this feedback loop because I've always thought that obviously if you can fix a bug early on, as you're coding it, that's the cheapest place to fix it. But if you are about to release it, that's the best next. If you're fixing a bug that's 4 or 5 years old, it's gnarly to fix it. Having a really tight feedback loop is important. Basically, anybody can enter bugs, ideas, or room for improvement.
Anthony Nielsen [00:56:13]:
That's a good one.
Craig McFarlane (CraigM) [00:56:16]:
Then I can look at the list and see all of the bugs and things hundreds of people have filled in about the app. Then I can just classify them, bang them out. Now I've started using this same little widget on a bunch of different apps because that loop is so critical.
Darren Oakey [00:56:42]:
I'm a big fan of that. I built it into every, every app I do, like for enterprise, just some way of saying, um, look at this, and it sends through context. It's so useful.
Craig McFarlane (CraigM) [00:56:54]:
Yep.
Leo Laporte [00:56:56]:
So you just gave me a great idea. I realized this is what I need. I need to have a ticket system in here so that as Lisa and Debbie are using it, they can push a button and say, this doesn't work. I'm gonna write that right now.
Craig McFarlane (CraigM) [00:57:07]:
Yeah, if you guys wanted, uh, um, to I can give you a login to this and you can play around with it.
Leo Laporte [00:57:12]:
It's very cool.
Michael (Alakazip) [00:57:14]:
And Leo, now you can vet, like, is it a gripe or is it a bug, right? Like you can sort of categorize it with AI, right?
Leo Laporte [00:57:20]:
Yeah, let's have some checkboxes. Cosmetic, pain in the ass, really wrong.
Craig McFarlane (CraigM) [00:57:27]:
Yeah, I was just thinking when you were talking about your ad system, this sort of a thing, you know, people do this all the time. And then you can see I've got some flags here on the feedback that I need to go look at because they're new. I make a habit of going in, flagging them for, oh, that actually is a bug, or that's actually an idea, or is it P0, or is it P1 or planned? Then I'll put in comments and whoever opened the bug or comment, they get to see, oh, this is live now and here's what somebody said. And so it's a really tight, small feedback loop, but with actual feedback in both directions.
Darren Oakey [00:58:11]:
And nowadays you could actually add on the viewer, you could actually add a fix button that then sends it to an agent.
Leo Laporte [00:58:18]:
Exactly. Oh God, I have to—
Michael (Alakazip) [00:58:22]:
yeah, just queue it up and be like, in the morning, just let me click. Yep, that looks good.
Leo Laporte [00:58:26]:
Yep, that looks good.
Craig McFarlane (CraigM) [00:58:27]:
Yeah.
Leo Laporte [00:58:28]:
Yeah, well, so what would I want? Cosmetic error, numeric error, a suggestion for improvement?
Craig McFarlane (CraigM) [00:58:34]:
Yeah, you want to make it as absolute dirt simple as you can.
Leo Laporte [00:58:39]:
Like check boxes.
Craig McFarlane (CraigM) [00:58:42]:
Yeah, that's why this defaults to room for improvement. Hey, this is just an observation and I'll classify it. That's— I don't want somebody else trying to decide. Is that any hesitancy on their part? I'm going to lose some people, so I won't—
Leo Laporte [00:58:57]:
Oh, okay. So maybe I'm making it too granular.
Darren Oakey [00:59:00]:
Yeah.
Leo Laporte [00:59:00]:
I just should say, what's wrong?
Craig McFarlane (CraigM) [00:59:02]:
They're going to say, I'm seeing this and I thought I shouldn't be seeing it. And you get that in and then I'll do the hard work of— because it's, you know, I understand the system. I know what's a bug, what's an idea. And I want to give— I want to get it fixed or resolved, get the feedback back to them. So they say, oh, I make comments, I observe things, I put stuff in the system and it gets better. You know, the tighter that loop is, the more you'll get.
Leo Laporte [00:59:33]:
Yeah, I like the idea of sending it to the AI.
Michael (Alakazip) [00:59:35]:
No, letting the AI categorize it. So it could be like when Lisa says, Leo, I'm so sick of looking at this, you know, like, okay, that's a bucket. You're like, but let the AI do that. Don't have a chat.
Anthony Nielsen [00:59:46]:
I love this.
Leo Laporte [00:59:47]:
I love this. Yeah, I could have the AI write a sympathy note. I am so sorry.
Craig McFarlane (CraigM) [00:59:53]:
Yeah, I built in an MCP interface onto it, but I haven't been using it that much. I use other apps for it, for doing MCP work. But yeah, that's the sort of thing that I could have, you know, almost anything connect to the MCP and Be observing the feedback loops, cleaning up typos and questions, all that kind of stuff. Yeah.
Leo Laporte [01:00:21]:
It was so funny when I set up the authentication system. I said, yeah, I'll write an email to the various parties and explain, you know, that there's a hidden section and how to get into it. And it wrote very straightforward ones for Patrick and Russell. And then the one it wrote to Lisa said, hi, sweetheart. It said, love you. It just knew, oh, you're writing to Leo's wife. Let's make this lovey-dovey. It was hysterical.
Leo Laporte [01:00:48]:
I'm not sure I would've been quite so lovey-dovey, but it was pretty funny that it did that. It tries hard. I love this ticket system. I'm gonna have a pop.
Craig McFarlane (CraigM) [01:00:59]:
Yeah, I was actually thinking of just putting— because the surface of it, The list itself is important because you want to be able to sort and filter quickly. But there's also in-context mode. Now I can go about the app and see the bugs or comments that have been put on that page, which is super helpful because if I'm fixing one thing, I can then notice, oh, There's these other bugs, and then go there and fix those. I was actually thinking of putting that on GitHub or something just because it's so simple, but if somebody just wants to plug it in, just add it to your app.
Anthony Nielsen [01:01:52]:
That's really cool.
Leo Laporte [01:01:53]:
How do you make your apps— this is for everybody— extensible? Do you think about that at all?
Michael (Alakazip) [01:02:02]:
What do you mean by extensible in this context?
Leo Laporte [01:02:04]:
Well, it's never finished, right? So do you just— I guess you don't really have to think about that. If you want to add something, you just say add it, or do you design it in such a way that it will be easier for it to extend? Maybe make it more modular?
Darren Oakey [01:02:20]:
Well, for me, it's a combination of But like the heart and soul of everything I do is CI/CD, like because it's about guaranteeing that once something's working, it doesn't get broken.
Leo Laporte [01:02:32]:
Because I did install a Git— a Git server on the system and for the CI/CD.
Darren Oakey [01:02:39]:
Yeah.
Leo Laporte [01:02:40]:
So I've got rollback and all of that.
Darren Oakey [01:02:42]:
But in terms of what you're saying, I think I used to ask lots of interview questions and my favorite interview question when I'm interviewing people is, um, what's the worst thing a developer can do?
Leo Laporte [01:02:54]:
Ah, okay.
Darren Oakey [01:02:54]:
Because it's a very open question and the domain of their answer tells me a lot about them. But my personal answer was grouping things horizontally rather than vertically. Like, you know, putting a database layer, putting a business layer, I think is the worst thing a developer can do. Whereas if you if you store things vertically, like everything to do with trades and everything to do with the trade system, you end up with like almost Lego components that you can push in.
Leo Laporte [01:03:25]:
There you go.
Craig McFarlane (CraigM) [01:03:25]:
That's what I want.
Darren Oakey [01:03:26]:
And the thing is, the more you componentize the system, the easier it is to snap out a component and put a new one in or just add a completely new component. And so it's, I mean, the CI/CD is guaranteed guaranteeing things don't break, but the system structure, making things independent as much as possible and vertically coupled as much as possible, allows you to make big changes without a lot of risk.
Leo Laporte [01:03:58]:
So here's the prompt I'm writing. We'll see what it does. I think I'm about to— I think I'm— I think I saw a notice today saying you're almost out of Fable, you might, you might be shut down for—
Darren Oakey [01:04:10]:
Why are you using Fable? Opus 5 is out.
Leo Laporte [01:04:14]:
Well, is Opus 5 better than Fable?
Anthony Nielsen [01:04:16]:
I mean, look at their, their, their table.
Danno [01:04:20]:
It's—
Anthony Nielsen [01:04:21]:
they're pretty on par almost.
Darren Oakey [01:04:23]:
Is it cheaper?
Leo Laporte [01:04:24]:
Is that why? I guess my attitude is I'm not programming with Fable, I'm just doing planning with it. See, it's, it's, it's loading superpowers for this. So I'm, I'm just doing the planning with people, which I feel like it's probably pretty good at. I don't know, just a thought.
Craig McFarlane (CraigM) [01:04:43]:
How do I turn off—
Danno [01:04:44]:
I'll look at—
Leo Laporte [01:04:44]:
I didn't actually— I, I, I, I looked yesterday for Opus 5 and I didn't see it, so I didn't— so I guess it came out this morning, huh?
Michael (Alakazip) [01:04:50]:
Yeah, at some point today, because I got the notice when I logged in this afternoon. They're like, go try Opus, and I was like, add it to the pile.
Leo Laporte [01:04:56]:
Yeah, but what's the point if If it's the same quality as Fable, is it cheaper?
Darren Oakey [01:05:00]:
Yes, it's half the price.
Leo Laporte [01:05:01]:
Oh, half the price. That sounds good. All right. I'll use Fable until I run out.
Anthony Nielsen [01:05:08]:
Out of the free $100? Is that—
Leo Laporte [01:05:11]:
I have a $200 max.
Craig McFarlane (CraigM) [01:05:12]:
Yeah.
Danno [01:05:13]:
Does Opus only use tokens or does it allow you to—
Leo Laporte [01:05:16]:
No, no, this is a max subscription.
Danno [01:05:19]:
I mean, but Opus— Fable is tokens only. Right. But Opus is usable by subscription.
Craig McFarlane (CraigM) [01:05:26]:
No, Fable is subscription too.
Michael (Alakazip) [01:05:28]:
Yeah, Fable is. Yeah, I use Fable all day, every day with my Mac subscription and I haven't hit the end yet.
Leo Laporte [01:05:33]:
I have this, that's, I was, I think it was spurious because I don't, I mean, I, I think I'm all right. I've only used 1% of my 5-hour. What do you mean?
Anthony Nielsen [01:05:40]:
Oh, because I'm not, I'm in the Pro.
Leo Laporte [01:05:42]:
Okay. If you're in Pro, you have to have Max.
Darren Oakey [01:05:44]:
Yeah.
Danno [01:05:45]:
Yeah.
Leo Laporte [01:05:46]:
But then it's 50, then it's, and then it is, is 50% of your max, so they're not letting you use it kind of as much.
Michael (Alakazip) [01:05:54]:
I think they're playing with the metrics. I don't think they're actually enforcing it.
Leo Laporte [01:05:57]:
I think they're doing all sorts of shit.
Michael (Alakazip) [01:05:59]:
And it still has not said you're done. It's never told me I'm done.
Darren Oakey [01:06:02]:
Yeah, I run out every week.
Leo Laporte [01:06:04]:
Well, you're—
Michael (Alakazip) [01:06:05]:
well, Darren, you're a different category.
Danno [01:06:07]:
Madman.
Darren Oakey [01:06:07]:
Yeah.
Leo Laporte [01:06:08]:
I mean, should I buy a DGX Spot? This is the question I've been meaning to ask you, Darren. Should I buy— I'm thinking Maybe I should buy a Spark.
Darren Oakey [01:06:16]:
I don't know, it's looking a little old and they've announced that new one.
Leo Laporte [01:06:19]:
Oh, there it is. I'm getting that thing here. You've used 80% of your Fable limit. Why does it do that and go away? Then it says something about July 29th. Oh, maybe, but that's not the 5-day. It says, I don't know.
Craig McFarlane (CraigM) [01:06:36]:
No, that would be when your weekly—
Leo Laporte [01:06:37]:
That's the week.
Darren Oakey [01:06:39]:
Expires.
Leo Laporte [01:06:40]:
It's the week.
Darren Oakey [01:06:41]:
Or resets.
Leo Laporte [01:06:42]:
I have been using Fable a lot, but I figure I'll use it until I— they won't let me use it, right? Yeah.
Michael (Alakazip) [01:06:48]:
Darren, I want to hear about the Spark decision because I was either going to do a Spark or pray for an M5 Studio or Mac Mini Pro that has 128 gigs of RAM. So yeah, love to hear your thoughts.
Darren Oakey [01:06:59]:
The thing about the Spark for me was not the performance. Like, the Mac M5 for a lot of things is as good or, or even faster.
Leo Laporte [01:07:06]:
Memory bandwidth's better, but you don't get CUDA.
Craig McFarlane (CraigM) [01:07:09]:
That's the difference.
Darren Oakey [01:07:09]:
It's the CUDA.
Leo Laporte [01:07:10]:
You can't use those models.
Darren Oakey [01:07:12]:
If you're just using LLMs, then no, you should not buy a Spark. But if you're playing with training or if you're playing with, you know, LTX or some of the weirder things, then that's what it's for because it's for the things that require CUDA.
Leo Laporte [01:07:30]:
So it's going to put a ladybug icon on here.
Danno [01:07:32]:
You're discounting impulses though.
Leo Laporte [01:07:35]:
I know, I want to impulse purchase.
Danno [01:07:38]:
I got 2 GX10s. Oh yeah, and you can— I have them hooked up to my, my board game sim. When you start a game of Gizmos, you can click a button and it'll bring in, uh, AI running on my GX10.
Leo Laporte [01:07:55]:
Oh my God, awesome. So What's the answer?
Craig McFarlane (CraigM) [01:08:02]:
Is it this—
Leo Laporte [01:08:02]:
the 128GB Mac would not be the same price as a DGX. It would be more, right? Or not?
Darren Oakey [01:08:08]:
Probably. Uh, yeah.
Leo Laporte [01:08:10]:
I think that— I feel like the DGX is a little bit, um—
Darren Oakey [01:08:13]:
Well, as I said though, Nvidia just announced that other one. Um, I can't remember the name of it now, but they've got a newer thing that's coming out soon. Oh, and that would be much better.
Leo Laporte [01:08:25]:
Okay, I'll wait.
Danno [01:08:27]:
Can you chain the Sparks together the way the GX10s do?
Darren Oakey [01:08:30]:
You can chain, yes. You can chain, um, I've got the Aces, so you can only chain 2 of them, but, um—
Danno [01:08:37]:
Oh, that's— oh, we're talking about the same thing then, because I have the GX10 Spark.
Darren Oakey [01:08:40]:
Okay.
Craig McFarlane (CraigM) [01:08:40]:
Yes.
Danno [01:08:41]:
I didn't realize that was a Spark.
Darren Oakey [01:08:42]:
But the DGX ones you can chain, and Blindways, so, you know, Jan has some chained, but, um, you can chain up to 4, I think.
Anthony Nielsen [01:08:53]:
Even on a Spark? I thought—
Danno [01:08:56]:
I think you can do 2, but if you buy a special router, you can do 4.
Darren Oakey [01:09:01]:
Yeah, I'm pretty sure the DGX one, I mean the actual NVIDIA one, you can chain 4 together. But my ASUS, you can only chain 2. So yeah.
Michael (Alakazip) [01:09:17]:
So because I'm a— I, I've tried to stay all Mac and Metal, Darren, you know, like, what do you think the future is? Metal going to be able to keep up as something that can be credible? I mean, it seems I haven't been able to push it to do anything that I can't get it to do, right? But, um, yeah, is there a reason just to get out of that world?
Darren Oakey [01:09:36]:
No, I mean, apart from the fact that all the new stuff only uses CUDA, so it doesn't, it doesn't run on the Mac. And so you've got to wait for, and especially as I said, in LLM, it's pretty easy because you've got generic things like Ollama or Zapier, I mean, or Native or whatever. So they'll run the new models, whereas everything else, all the new research things in, if you want like text-to-speech or conversions or something, they all take a while to go and make. And that's why I want, because like you, I'm mostly Mac at the moment. And that's why I bought the Spark because it just opened that CUDA world to me. So it was less about performance and more about accessing cool stuff. Yeah.
Anthony Nielsen [01:10:25]:
I mean, if you really want performance, like you could do the RTX 6000 Pro, like, but that only, that caps at 96 gigs, but the token throughput's insane.
Leo Laporte [01:10:35]:
96 is probably good for a local model.
Darren Oakey [01:10:38]:
But the cost is quite a lot more. How much is that?
Leo Laporte [01:10:40]:
That's the $100,000 box?
Anthony Nielsen [01:10:42]:
No, it started at $10,000, but now it's $13,000 because of the shortage.
Leo Laporte [01:10:46]:
I can kind of see my way to maybe $3,000 to $5,000, but I can't do more than that.
Danno [01:10:51]:
I wonder how much— I've got an— my impulse buy last year was an M3 Ultra Studio with 256 gigs of RAM.
Leo Laporte [01:10:59]:
And you have it?
Danno [01:10:59]:
I wonder if I could probably sell that for a house now, couldn't I?
Michael (Alakazip) [01:11:03]:
Yeah, you could get a lot of money for that one.
Leo Laporte [01:11:05]:
It's definitely profit.
Anthony Nielsen [01:11:07]:
Jeez.
Michael (Alakazip) [01:11:09]:
But of course, if you have a Mac, a high-dollar Mac, like you could do a zillion other cool things on it, right? So if I want to play a game while I'm also running an LLM, so that's, that's why I'm hesitant to go dip into the other world where you can only do this thing with it.
Leo Laporte [01:11:23]:
So see, I'm all Linux. I took that Mac Mini this week, stripped— I made it a server. So all the servers are running on it. And Ollama is running on it. So, and I have, you know, all the decent local models I can run in 64 gigs. And that's just all it's doing. It's headless. And because, I mean, I don't know, I don't hate Macs, but Linux is all I really need.
Leo Laporte [01:11:49]:
I'm not playing any games, I guess. Oh, I also, okay. So I have a suggestion box and a ladybug so they can do it because I want to be able to say, oh, I— Oh, that's cool. I would like to have something like this. It's not a bug. So I'm going to put a suggestion box in. Let's see how quickly.
Darren Oakey [01:12:06]:
The thing is, like Craig said, I think it's much easier. Like, I never have any classifications in these feedback things because I just want feedback and I wanted to make it as easy as possible for the users. And when I grab the feedback, I snap as much context as possible from inside the app, like what page they're on, what things, take a screenshot of what they're doing at the moment. All of that goes in and then I can classify it and I can choose what to do with it. But it's just all about making it easy for the user.
Leo Laporte [01:12:40]:
So you just do a freeform?
Craig McFarlane (CraigM) [01:12:42]:
Yeah.
Leo Laporte [01:12:44]:
What does it say? Like, tell me what you think.
Darren Oakey [01:12:47]:
Yeah, or just say, it's all about What can I do better?
Leo Laporte [01:12:53]:
Okay, that's good. Maybe the suggestion box should say that. What could I do better? Who wants to show?
Michael (Alakazip) [01:13:05]:
In your case, Leo, it could get really personal really fast if you put that as your prompt.
Leo Laporte [01:13:09]:
So yeah, I know. This is why you should never do this with your wife and your wife's sister. I mean, they can gang up on me. This is, this is I had really second thoughts about doing this, and I thought, this is such a great project. It is, you know, it's our company. The guy who wrote this, I don't know how long he— You weren't around, Anthony, were you, when Jeffrey Needles wrote this?
Anthony Nielsen [01:13:35]:
I don't know.
Leo Laporte [01:13:37]:
I think it took him 8 months, Lisa said, of back and forth. I mean, she must— But the thing is, that's different sitting next to an employee and telling them, to do than sitting next to your husband and telling him what to do. It's a completely different experience.
Danno [01:13:50]:
I cringed.
Michael (Alakazip) [01:13:51]:
You need to hire a fake consultant and be like, that's actually like, I found this great consultant, he's going to do all this work, all this cool stuff. And then you're actually telling him all the things you're doing.
Leo Laporte [01:14:03]:
I would like to talk to you.
Darren Oakey [01:14:06]:
His name is Daedalus.
Leo Laporte [01:14:08]:
Daedalus. Yeah. I mean, that's really why she wants a consultant. Is that she's just tired of interacting with her husband on this.
Danno [01:14:17]:
Take names from Office Space to name the characters.
Leo Laporte [01:14:23]:
Hello, I'm Michael Bolton. I'm your new consultant. All right, let's see. The agents are now, all 3 of them, reviewing it, trying to reach consensus on how to do this ladybug and suggestion box.
Danno [01:14:39]:
Well, I hate to leave this therapy session, but I have racing to do this weekend.
Leo Laporte [01:14:42]:
So, Dano, are you driving?
Darren Oakey [01:14:45]:
Yeah.
Leo Laporte [01:14:46]:
What do you drive?
Danno [01:14:47]:
Well, now I've only got— well, only, but it's a Model 3 Performance, so it's my, probably my fastest to 60 car ever, but not as much fun as a Porsche.
Leo Laporte [01:14:56]:
But, and, and are these, uh, circular tracks or drag races?
Danno [01:15:01]:
This weekend's autocross.
Leo Laporte [01:15:03]:
Oh, it's fun.
Danno [01:15:03]:
Driving fast in a parking lot.
Leo Laporte [01:15:09]:
I do that, but that's—
Danno [01:15:10]:
I've been up to 80 in a parking lot before. It's pretty fun.
Leo Laporte [01:15:13]:
Lisa is sick of our tame EVs and she wants to get an ICE sports car.
Danno [01:15:20]:
You know what the good deal is now though? You can get a Crosstrek Porsche Taycan for in the 40s.
Leo Laporte [01:15:28]:
What?
Danno [01:15:29]:
Really? Yeah, they're, they're so fast. And plus, they're driver's cars because you can't beat Porsche.
Leo Laporte [01:15:37]:
I'm gonna show her that. But yeah, she thinks she wants a nice vehicle because she wants to drive to Tahoe, and I don't think she needs a nice vehicle.
Danno [01:15:46]:
You won't have any warranty, but I've had electric cars for 6 years and had zero maintenance.
Leo Laporte [01:15:51]:
So yeah, it's kind of amazing. So it's a, it's a, it's a Secondhand. It's called a Crosstrek Taycan?
Danno [01:16:01]:
Yeah, it's something like that. It's a wagon. It's a sport wagon.
Leo Laporte [01:16:04]:
It's a sport— oh, she wants to see.
Danno [01:16:06]:
Well, it looks— well, I guess for a car guy, car guys love fast wagons.
Leo Laporte [01:16:14]:
Really?
Danno [01:16:14]:
If you look like a Mercedes or BMW wagon-looking sporty car, that's just like Catnip for car guys.
Leo Laporte [01:16:25]:
She wants something cute. She wants, uh, she does like the Porsches.
Danno [01:16:28]:
It'd be a bigger car, but yeah, a regular Taycan too, you can get for the 40s. Yeah, those you'll get like a '21, '22.
Anthony Nielsen [01:16:36]:
Really?
Danno [01:16:36]:
It's not, it's not a new one.
Michael (Alakazip) [01:16:38]:
Yeah, but I don't know, the Audi EV, my best friend has one that's like the high-dollar Audi EV, and that thing is fast as heck.
Leo Laporte [01:16:44]:
And it's— oh yeah, oh yeah, they're depreciating That's another one I've seen.
Danno [01:16:48]:
Once, once I started looking on YouTube at Porsches, that started showing me e-trons that were depreciating fast.
Leo Laporte [01:16:55]:
So they all depreciate, but I, I think, yeah.
Danno [01:16:59]:
All right. But anyway, I gotta go. Thanks.
Leo Laporte [01:17:01]:
Thank you.
Anthony Nielsen [01:17:02]:
Good luck.
Leo Laporte [01:17:03]:
Have fun.
Michael (Alakazip) [01:17:03]:
I hope you can win.
Leo Laporte [01:17:05]:
See you, Dana. Yeah, we're kind of an hour and a half into this, but, uh, I'm not, I'm not I'm gonna cut it off if anybody wants. Craig, thank you, that was a great demo. Yeah, Darren, you too. All right, Alakazip, you're on. What do you got for us?
Michael (Alakazip) [01:17:23]:
Well, to be honest, the thing I've been working on, which I guess is kind of— we've all been very disappointed in, and, and Alex Lindsay used to say this, in home automation, right? I've tried 3 or 4 systems.
Leo Laporte [01:17:34]:
God, yes.
Michael (Alakazip) [01:17:35]:
I now have completely scrapped all the systems, and, and I only have Matter. And now I have a Matter server Claude Code has written, and it only does what I want. And I'm a geek, and I want lots of logs. I want to know the humidity every 5 minutes in every zone of my home, in my yard. And I've been having so much fun tuning things in the house because I have all this data now, and I never open the interface for any of these stupid apps that these devices have, but I've got probably 6 or 7 disparate kinds of devices that are hanging on doors and Ring doorbells and all that. And Claude Code has now written me a custom system that I only interact with CLI that gives me everything I want. It's always what I dreamed of. And so I've had a lot of fun with that.
Leo Laporte [01:18:22]:
Do you use it to issue commands? Like turn on lights and things?
Michael (Alakazip) [01:18:26]:
Oh, absolutely, yes.
Leo Laporte [01:18:27]:
So now my problem's been latency.
Michael (Alakazip) [01:18:31]:
I'm not as smart as Darren, so I don't always know what Claude's doing, but I had— I actually had a problem. My wife's car is a Toyota, and the air conditioning in the garage in the humid summer, it starts to smell because it doesn't ever dry out in the garage, right?
Anthony Nielsen [01:18:46]:
Right.
Michael (Alakazip) [01:18:46]:
And so Claude helped me come up with an idea to put fans on the windows in the garage, and I have smart plugs that the system's monitoring the humidity compared to inside and outside. And it turns on the fans to suck in outside air when the humidity is lower outside.
Danno [01:19:02]:
Nice.
Michael (Alakazip) [01:19:02]:
And it's working great.
Leo Laporte [01:19:03]:
So you're using it autonomously mostly?
Michael (Alakazip) [01:19:06]:
100%, yes. And I love waking up in the morning and getting a report.
Leo Laporte [01:19:09]:
I can tell the curtains to open. We have motorized curtains, but I'll tell it and it doesn't happen for 30 seconds, which is incredibly, you know, it's not, it's just too frustrating. So they are the, uh, I gave the prompt, to, uh, Kenobi, and it's, it's now, um, they're now talking amongst themselves using that agent mail that I, uh, told them. They're going back and forth. They— no, no questions for me.
Craig McFarlane (CraigM) [01:19:39]:
I was gonna say, did you, did you give it instructions on, I want you to put in stuff that's human readable so it doesn't dissolve into like a gibberish Kind of thing.
Leo Laporte [01:19:49]:
Oh yeah, I hate it when it does that. And there is a lot of that in the plans. Well, we'll see if it— I don't think it's going to respond to them. I think it's just going to— they're going to write and it's going to press the submit button. I don't want them to be hampered by human readable. But you're right, if they're communicating with with the end users. I don't— I absolutely— they're pretty good though. Like I said, that, that email I wrote to Lisa was very, very, uh, friend— appropriately friendly.
Leo Laporte [01:20:25]:
So, um, they're going back and forth here. So it's waiting for a review from Opus, uh, no, I'm sorry, uh, Codex GPT-5.6. I'm gonna run out of fable any minute now though.
Michael (Alakazip) [01:20:38]:
You should tell them to try and make you feel better and every once in a while ask your opinion even if they don't need it.
Leo Laporte [01:20:43]:
Like, you know, that's good.
Michael (Alakazip) [01:20:44]:
Just, just check in and be like, hey, you know, hi, what do you think?
Leo Laporte [01:20:47]:
You know, like, how you doing, boss? Everything okay? How's your day been? Any bugs? Okay, one second.
Darren Oakey [01:20:58]:
I got to show you one other thing I've done that you might like. Um, the— this, you know, the XT Ink? It's now connected to my, um, my Mac. It's complete. And just—
Leo Laporte [01:21:12]:
I mean— Oh, what firmware did you use? Did you write your own firmware?
Darren Oakey [01:21:17]:
Yeah, well, Claude wrote it, but— or someone wrote it.
Leo Laporte [01:21:21]:
No, no, amongst— we're amongst friends. You can say you wrote it.
Michael (Alakazip) [01:21:24]:
We understand.
Leo Laporte [01:21:25]:
We understand what that means.
Darren Oakey [01:21:26]:
Yeah, but yeah, it's— it syncs to the phone via Bluetooth and just updates my calendar.
Michael (Alakazip) [01:21:31]:
So—
Leo Laporte [01:21:32]:
oh, I love that. Have you put— is that up on GitHub?
Darren Oakey [01:21:35]:
I'll put it up on GitHub.
Leo Laporte [01:21:36]:
So yeah, it's an ESP32, right?
Darren Oakey [01:21:39]:
Yeah, it's an ESP32, and I, I just downloaded one of those, but you know, the one you told me about, um, to, um—
Leo Laporte [01:21:47]:
Oh, that, that, yeah, the other reader. Yeah, yeah. And you so used its source code and modified it?
Darren Oakey [01:21:54]:
Yeah, because the thing with ESP32s is they're very tricky with what ports are set up and the display things and everything. And it's hard to get, like I've got ESP32 devices everywhere. Like I got this little thing, which is a copy of that little thing that you've got. But they're always very finicky to get the display right and they crash lots until it works out which port to set and which thing and everything. So downloading existing source for something just allows it to start with all of those things correct.
Leo Laporte [01:22:32]:
So that's very cool.
Michael (Alakazip) [01:22:35]:
I want a 3-foot by 4-foot wall hanging that's an e-reader that does that. And my AI constantly knows what information I'm interested in in that moment, whether it's the weather or an email or the calendar. And it's just all— it's like wall art of information.
Craig McFarlane (CraigM) [01:22:50]:
I was thinking of—
Leo Laporte [01:22:53]:
Jeff Atwood sent me this. This is that E Ink thing, and it has— but I'm just using the stock stuff. Like, it has a graph of my weight. It has— this is the stock market stuff. And then as a joke for Jeff, I put the MediaTek— whoa— the MediaTek Companio sticker. Oh, it just fell down from the Chromebook so that it looks like it's running Android. MediaTek, which it's not. And, uh, this is kind of cool, but you know, I'm sure it's an ESP32 in here.
Leo Laporte [01:23:23]:
It wouldn't be too hard to modify that. And that's a nice big screen.
Craig McFarlane (CraigM) [01:23:26]:
I don't—
Leo Laporte [01:23:27]:
my problem is I don't have a calendar. I don't really need a calendar, so I need— I have to think of what else I could put in here.
Michael (Alakazip) [01:23:34]:
So I've been working on the code. ADS-B of planes over my house is one of the screens I've been working on, you know. So yeah. Whatever's cool to you at that moment, right?
Danno [01:23:43]:
Right.
Leo Laporte [01:23:44]:
Yeah, actually, yeah, once you've written it, you could write anything you want once you have the firmware in there. Yeah, so it uses Bluetooth, it connects to your phone via Bluetooth.
Darren Oakey [01:23:52]:
Yeah, BLE.
Leo Laporte [01:23:53]:
And did you have to write that, uh, or was that already in this reader's capabilities?
Darren Oakey [01:23:59]:
Well, it's already got Bluetooth capability in it, and so, um, yeah, it's just pair— you pair it, and then, uh, and my plan is to— for the iOS app to be like a widget. So that— so the calendar is one widget, but I'll just have a grid that you can drop calendar reminders, photos.
Leo Laporte [01:24:16]:
Oh, so you have an iOS app, that's what is communicating to it?
Darren Oakey [01:24:19]:
Yeah, exactly.
Leo Laporte [01:24:20]:
I get it. Yeah, I get it.
Darren Oakey [01:24:20]:
Because I've got it— I've got it plugged. I stuck it to the back of my, my phone like this. So, um, right.
Michael (Alakazip) [01:24:28]:
Oh, a second screen. That's cool. That's so cool. Flip-open phone you've already made.
Leo Laporte [01:24:34]:
You know, they have a new one of these now with a backlight.
Darren Oakey [01:24:36]:
Yeah, the Pro, um, and, and it's got touchscreen.
Leo Laporte [01:24:40]:
Yeah, I might, I might order that. You know what I think I might do is I, uh, my Quicksilver is telling me everything through Telegram. I could just have this be the latest Telegram message. Yeah, that would actually be kind of useful, a little readout, because I do get it as notifications.
Darren Oakey [01:24:55]:
I was gonna show is, um, I'm playing with You know my video generator? Um, I, I just posted it in live chat. I posted a few days ago, but I'm playing with trying to get voices, photos, and multiple people in, and it's not fully there yet. But if you have a look at live chat and, uh, um, and the video I posted—
Leo Laporte [01:25:17]:
Let's play it. Let's— that'd be a good way to say goodbye to everybody. Anthony, did you— you must have had something you wanted to talk about. I don't wanna give you short shrift.
Anthony Nielsen [01:25:26]:
I made a couple Slack bots. Nothing to—
Darren Oakey [01:25:29]:
Nice.
Anthony Nielsen [01:25:29]:
We can talk about it later.
Leo Laporte [01:25:31]:
Well, that One Sheet is a Slack bot, so they can— That was one of the other things we did with the One Sheet was they could just do @onesheet and give it the name of a sponsor. That's kind of cool, but I don't wanna put— I'm nervous about too much. Ooh, is that the Taycan? That is beautiful. What is that? Is that— That's a Ferrari, isn't it? There's Lisa in a Ferrari. I think I'm gonna send this to her and say, I found your car, Lisa.
Darren Oakey [01:26:00]:
She replied. She loved it.
Leo Laporte [01:26:01]:
Oh, you sent it to her? Awesome.
Darren Oakey [01:26:03]:
No, no, she's in the chat. She's been chatting with us.
Leo Laporte [01:26:05]:
Oh, you're in here? Oh, I didn't even know she was in here. That's hysterical. Is that your car, baby? That's not the color though. You want it red, I'm sure. Yeah, that's very funny. Oh yeah, she's already looking. She's already got her eye on a couple of things. Where's the, uh—
Darren Oakey [01:26:23]:
I'll post it again.
Leo Laporte [01:26:24]:
Oh, this is cool. This is a terminal Twit feed. Oh, thank you, Chad. Oh, I will add that. That's really nice. Oh, it's in live chat. Aren't I in live chat? I thought I was.
Anthony Nielsen [01:26:39]:
Yeah, you are. Oh, it's—
Leo Laporte [01:26:41]:
yeah, link. Oh, you put a link. Oh, this is hysterical. Is this the new one?
Darren Oakey [01:26:46]:
Well, it's— if you've already seen it, I mean, that's, that's the one that—
Leo Laporte [01:26:49]:
I love it.
Darren Oakey [01:26:50]:
It's not perfect yet, but it's testing. It's more the multi-character lip sync that I was working out.
Leo Laporte [01:27:11]:
Oh, I haven't seen this. This is a new one.
Michael (Alakazip) [01:27:14]:
No, no, no, no, that's ridiculous. I'll say it twice, it's tech triumphalism dressed up nice.
Leo Laporte [01:27:22]:
Yeah, I have mixed feelings about me because I, I look really old, but I'm also ripped, so Didn't I play this on I Am?
Darren Oakey [01:27:41]:
I think I did.
Leo Laporte [01:27:43]:
I hope I did.
Darren Oakey [01:27:44]:
Oh, it might've been after.
Leo Laporte [01:27:45]:
You played something like this after maybe.
Darren Oakey [01:27:47]:
Yeah. What would you even do tonight?
Leo Laporte [01:27:50]:
There's not the— there's 2 of me.
Michael (Alakazip) [01:27:54]:
There was a random 4th girl earlier.
Leo Laporte [01:27:55]:
I saw the random 4th girl. I think that is really good, Darren. You're within 6 months, you're going to have this. It'll be indistinguishable from real thing, and then I'm hiring you to do all our shows from now on. It's like, I'm done, I'm done, I don't need to do anything more. I do everything in rap from now on.
Darren Oakey [01:28:15]:
I guess for me, I'm a programmer, and I, I, I'm a— I, I've always been able to write programs and build things for myself, but what excites me is like, I'm the last— I'm as far as you can get from a musician or a Or an artist. And so that's why it's very exciting for me to be able to just pull up a, a video or a thing, because it's a capability that I, I never in my wildest dreams could have.
Leo Laporte [01:28:44]:
So very nice, very nice. Um, come on, Anthony, show us. You must have one kind of fun thing.
Anthony Nielsen [01:28:55]:
I mean, Slack.
Leo Laporte [01:28:56]:
Come on. I'll show you.
Anthony Nielsen [01:28:59]:
Well, no, I mean, it's more like a base.
Leo Laporte [01:29:02]:
So, like, let me share screen. GRC is still down, I think.
Anthony Nielsen [01:29:09]:
Is it?
Leo Laporte [01:29:10]:
Yeah, poor Steve.
Anthony Nielsen [01:29:13]:
So, I'm using Hermes to build some N8N workflows.
Leo Laporte [01:29:18]:
Oh, good. That's right. I saw N8N in the Slack.
Craig McFarlane (CraigM) [01:29:21]:
Yeah.
Anthony Nielsen [01:29:21]:
So, RS stands for Restream and I have a bunch of stuff like Keep it. So Restream that we use to stream everything to, like, they have a pretty extensive API that we could start building that. Oh, really?
Leo Laporte [01:29:34]:
Uh, yeah. So we kind of do what Craig's doing with Restream.
Anthony Nielsen [01:29:37]:
Yeah, some stuff like that. Um, the issue was the, the token, the authentication token was a rotating one and it was a pain to like figure out how to like— but, but I worked that out and I have like, you know, A bunch of data tables with the, you know, information. So now I could start building on top of it.
Leo Laporte [01:29:55]:
Um, so what do you think you might do?
Anthony Nielsen [01:29:59]:
Well, I could— so, um, you could see on the, the right now when we go live, that's all automated now, right? Before it was like a manual thing.
Leo Laporte [01:30:10]:
You're kidding.
Anthony Nielsen [01:30:10]:
In Discord.
Leo Laporte [01:30:11]:
Um, oh, that's awesome. Is it— does it have to be triggered with a button or does it just do it?
Anthony Nielsen [01:30:16]:
It's when we go live. And let's see. Let me switch screens real quick. Chrome tab.
Craig McFarlane (CraigM) [01:30:32]:
It's always a bit tricky to demo something that's running.
Anthony Nielsen [01:30:35]:
Yeah.
Craig McFarlane (CraigM) [01:30:35]:
So there's—
Anthony Nielsen [01:30:39]:
ignore the 3 YouTube ones. There should only be 2, but that's a Restream issue. I made a mistake going live. It'll grab all the channels that we're going to and, you know, do the right role notification. So, that's on Discord. And then also now in Slack under the studio channel.
Danno [01:31:10]:
Okay.
Leo Laporte [01:31:11]:
So I've seen these. You were playing with this the other day. Yeah, yeah, yeah, yeah.
Anthony Nielsen [01:31:15]:
Um, so right now it has all the, uh, the destinations for people to watch. It says what time it started, um, at your local time. Uh, we could get to the analytics dashboard. Uh, it tells you what workspace we're using. Once it's over, it'll get updated. Um, I already cleaned it up, so you can't see it here, but, uh, we then have to manually delete like the, the LinkedIn post and Facebook post so people can't watch it after the fact. They have to go download the episode, right?
Leo Laporte [01:31:48]:
Oh, that's all automatic now too.
Anthony Nielsen [01:31:50]:
So, well, it will theoretically make— yeah, probably could get there, but right now it'll do a checklist so Burke can check what he's deleted, uh, and then it'll say like, oh, all, you know, cleanup complete. Um, in, in the future, it's gonna— Restream can do an AI transcription for, for the, for the stream.
Darren Oakey [01:32:13]:
So—
Leo Laporte [01:32:14]:
We're getting so close to being able to automate.
Anthony Nielsen [01:32:17]:
Well, it'll— it's gonna create edit notes for the editors, like when, when the ad— like what time the ads ran.
Darren Oakey [01:32:22]:
Oh, nice.
Leo Laporte [01:32:23]:
That's genius.
Anthony Nielsen [01:32:23]:
Like where's the tease, like, you know, stuff like that.
Leo Laporte [01:32:25]:
Oh, that's so nice.
Anthony Nielsen [01:32:26]:
Um, so a bunch of cleanup stuff. And then even probably like I'll have it download all the stuff associated with that stream, put it in a folder in Google Drive. But that's— so this is kind of like the, you know, the starting point. And then we also turned on comments on YouTube and that was going to be like a pain to monitor. So, yeah. I created a like I have a whole other workflow that will, you know, uh, flag comments either that we, you know, maybe a heart or like stuff that we have to delete or moderate. And then ideally, like right now these buttons don't actually do it. Like we have a link to the thing and you can say like, yeah, I hearted it and then dismiss this comment.
Anthony Nielsen [01:33:14]:
Um, theoretically with the YouTube API, like it'll, it'll take action on some of it. You can't do everything, but, uh, it's possible. And that's kind of a down-the-road thing. Um, so Burke doesn't have to go— or like, we don't have to go through channel by channel figuring out what we need to moderate.
Leo Laporte [01:33:35]:
I gotta get Craig.
Michael (Alakazip) [01:33:36]:
This is great. Why did you hold off showing this?
Leo Laporte [01:33:39]:
I know.
Anthony Nielsen [01:33:40]:
I mean, after Craig's thing, it's like, oh boy, that's—
Craig McFarlane (CraigM) [01:33:44]:
I don't have a platform since I'm going to share some code here because you're doing some cool things that I was like, all right, I could use that. Yeah.
Anthony Nielsen [01:33:50]:
So, and I could kind of—
Leo Laporte [01:33:54]:
There's definitely overlap, right?
Anthony Nielsen [01:33:55]:
Yeah, there's a lot of overlap.
Craig McFarlane (CraigM) [01:33:57]:
Yeah. So some of the hardest stuff to code is actually the integrations, not the pure logic stuff. That's straightforward. It's the integrations of like, oh, how is it going to deal with time if I give a message to this system? Does it know the time zone? And all that kind of stuff.
Anthony Nielsen [01:34:13]:
So here's like kind of, this is the social monitor. Like in a different workflow, it's grabbing the comments every half hour. Then over here, this is where it's kind of reading the comment log, like what videos we have. And then it'll grab the comments for, we're only doing the last videos, younger than 30 days. Um, but it'll grab the, the, all the video IDs, uh, grab all the comments. And then there's, I have a logic here where like an LLM, um, you know, here's the system message that, you know, your, uh, your media, social media comment classifier. And then there's like a, you know, here's the prompt. So for each comment, it runs through the prompt.
Darren Oakey [01:35:02]:
Okay.
Anthony Nielsen [01:35:04]:
And I'm just using DeepSeek v4.
Darren Oakey [01:35:06]:
And basing it around NNN is a really smart decision because it's very deterministic and it's way cheaper as well.
Anthony Nielsen [01:35:14]:
Yeah, I'm looking at it like a token cost and like, I don't want, you know, rogue agents. So like, I'm trying to keep it very, very tight.
Darren Oakey [01:35:25]:
Yeah, that's a good idea.
Anthony Nielsen [01:35:28]:
Yeah, thank you.
Michael (Alakazip) [01:35:30]:
I'm gonna go make some horrible comments real quick in the YouTube video and you let us know how that goes.
Leo Laporte [01:35:36]:
Don't worry, you don't have to do anything. They'll happen anyway.
Michael (Alakazip) [01:35:40]:
Oh, it's automatic. The internet just takes care of it.
Leo Laporte [01:35:42]:
I got it. Yeah.
Craig McFarlane (CraigM) [01:35:43]:
Yeah.
Anthony Nielsen [01:35:44]:
So that's kind of the base. Like, I got a baseline of stuff and then now I'm gonna start building on top of it.
Leo Laporte [01:35:48]:
Very nice.
Danno [01:35:49]:
Yeah, thanks.
Leo Laporte [01:35:50]:
That's— I didn't know that's what you were working on. That's really cool. That's great. And you really like Nathan.
Anthony Nielsen [01:35:56]:
It depends, like, it depends on the task. Like, I don't want an agent, like, burning tokens just to say that we're live in Slack, right?
Danno [01:36:07]:
So—
Darren Oakey [01:36:08]:
I'm convinced that some sort of workflow, like, we're, you know, we're all moving to the same thing and we're all with memory and things like this, but I'm convinced that a solid, like, deterministic workflow engine is necessary for most ongoing things because it is deterministic, it's robust, but also an end-to-end skill burns tokens like crazy. And it's also moderately slow. Whereas once you get the thing to build a workflow, it's super fast. You can still have, like Anthony did, have the LLM components as you want, but you end up with something that's entirely robust repeatable, trackable, everything like that. I think every agent is eventually going to need a workflow engine in it.
Leo Laporte [01:36:57]:
You like the nodal system?
Darren Oakey [01:36:59]:
Well, the n8n is good for some things. At work, I'm using a combination of n8n and Temporal. n8n is great if you want to be able to change it rapidly. And that's what we need. Like if you're doing dashboards or something and you're on a client site and they just want to change it, it's great. But it's not as good for testing and things like this, whereas Temporal— and for recovery, like things, if it screws up, it just fails. Whereas Temporal has all this retry mechanism and wait and back off and try again and everything. It will get the job done type thing.
Darren Oakey [01:37:37]:
So for fully robust enterprise systems, I think Temporal is better, but it's horrible to use. Like, it's very much a programming thing and it has to go through CI/CD, whereas N8N can be sort of a live prod thing. So they're both good. And there's things in the middle like Windmill and everything, but I think people just need a workflow system as part of their—
Craig McFarlane (CraigM) [01:38:01]:
And having that visual diagram flow is is really helpful. Like, if you're sharing it like Anthony, you just did, being able to see that everyone can instantly get it without having to see brackets and commas and stuff that makes people glaze over.
Anthony Nielsen [01:38:16]:
And to be honest, so like, you know, it's— I'm not even doing the— like, I'm not building the nodes. It's all Hermes. It has an MCP and API.
Leo Laporte [01:38:27]:
Oh, so you're not even drawing that stuff?
Anthony Nielsen [01:38:29]:
And it's using a lot of code nodes. Like, you You can't, you know, it, not everything is, is like a fixed node. Like you can like put in regular code into it where it makes sense.
Leo Laporte [01:38:39]:
Craig, are you using something like that?
Craig McFarlane (CraigM) [01:38:43]:
I am just giving single prompts.
Leo Laporte [01:38:47]:
Yeah, that's all I'm doing too. But now I'm thinking, so I could tell Hermes, take a look at Nathan and, and like stuff that's gonna like, you want to run like Hermes has a Kanban board, which I don't use either.
Darren Oakey [01:39:00]:
No, but this is like Anthony said, this is for things that you want to do every day or every hour or when an event happens.
Leo Laporte [01:39:06]:
Oh, I have like 30 cron jobs.
Darren Oakey [01:39:08]:
Processing a Gmail or something like this, then this is where it comes in.
Leo Laporte [01:39:14]:
You have to have a processing flow. You have to have a workflow that you're going from step to step.
Darren Oakey [01:39:19]:
And you can do a Rails skill and a command and it works. fine, but it's way more expensive and way less reliable.
Leo Laporte [01:39:27]:
This is more deterministic.
Darren Oakey [01:39:29]:
Yep.
Craig McFarlane (CraigM) [01:39:29]:
It—
Leo Laporte [01:39:29]:
I mean, that's what I keep running into, is it forgets. It goes, oh yeah, I forgot I was supposed to do that too.
Anthony Nielsen [01:39:35]:
Yeah, I mean, you can build like full agents in n8n, but, um, you know.
Leo Laporte [01:39:41]:
But you could also use it as a scheduler for Hermes.
Anthony Nielsen [01:39:44]:
Yeah.
Leo Laporte [01:39:44]:
So for instance, when I log a meal, I log it by talking it on my watch, it sends it to Hermes. It's supposed to then go to Cronometer, get the calories, log it there. Then it's supposed to go to my Obsidian and log it there, but it often forgets. It goes, oh yeah, that would be perfect. And it sounds like exactly— yeah, because it's a step-by-step process. I'm going to do that because it forgets. It goes, oh yeah, I didn't do it.
Darren Oakey [01:40:08]:
But also the main thing is like when we're, when we're all in a token economy, every Then almost every step that you just mentioned can be done without using an agent at all.
Craig McFarlane (CraigM) [01:40:19]:
Right.
Darren Oakey [01:40:19]:
And, um, so that's A, cheaper, but B, 1,000 times faster. Like, that'll happen in like 10 milliseconds rather than, rather than 10 minutes, I think.
Craig McFarlane (CraigM) [01:40:29]:
Right.
Anthony Nielsen [01:40:29]:
Yeah, just, I mean, you should just, you know, set up, get a VPS. I will.
Leo Laporte [01:40:33]:
I'll just tell it to set up N8n.
Anthony Nielsen [01:40:35]:
And then like, yeah, just tell your agent anything that you think makes sense to live here, like build it Build it as you see fit.
Leo Laporte [01:40:42]:
I've learned that. I've slowly gotten farther, like Homer Simpson in the ivy, just getting farther and farther back and going, oh, you do it.
Anthony Nielsen [01:40:52]:
Yeah.
Leo Laporte [01:40:53]:
It's good to be a little lazy, which I am, because eventually you just, it's kind of like, oh yeah, I don't have to do that.
Darren Oakey [01:40:59]:
Yeah, there's certain steps in AI acceptance. Like you start out playing with Claude code and everything, but It's then when you suddenly realize, uh, you read about something on the net and you think, I should play with that, and then think, how do I install it? And then you suddenly realize, I can just install anything.
Leo Laporte [01:41:15]:
I don't have to install it. I haven't installed anything in years. Uh, every once in a while I have to type a password, but that's rarer and rarer. Um, all right guys. Hey, it really is fun. This is good. We'll do this. Uh, I don't know.
Leo Laporte [01:41:31]:
We'll do it again. I like the idea of doing it every other week.
Anthony Nielsen [01:41:35]:
I don't know if this was any different than a typical—
Leo Laporte [01:41:37]:
It was just another AI user group, but it was a little more demo-focused. Anthony's original idea was we would all sit here typing away. I don't know how compelling that is.
Anthony Nielsen [01:41:46]:
Yeah, I might just—
Michael (Alakazip) [01:41:47]:
I'm gonna tell my agent to write really good code so when it shows up on the screen, I look really smart.
Leo Laporte [01:41:53]:
Well, I did, I did, I asked it for lines of code and I was really, I think I'm gonna have, I've gotten, I've already implemented like 4 things we were talking about, you gave me ideas for. So I guess there is, it's possible we could do that.
Darren Oakey [01:42:07]:
Yeah, the trouble is there's a lot of pauses where you're just sitting there. It's just sitting there, yeah. The agent doing nothing. So unless you start like 10 of them.
Leo Laporte [01:42:14]:
Can't be any worse than me trying to solve Advent of Code in Common Lisp.
Darren Oakey [01:42:19]:
That was fun.
Craig McFarlane (CraigM) [01:42:19]:
Show and tell is always just a good framework for—
Leo Laporte [01:42:22]:
What am I doing?
Craig McFarlane (CraigM) [01:42:22]:
Here's this thing I'm really excited about. Let me tell you why.
Leo Laporte [01:42:25]:
Right, yeah, right, right. I think the most valuable thing for me With all of this, and I'm sure you guys feel the same way, it's just the sharing of ideas is what's so— it's so early days and there's so many ideas and so many things, and just getting more ideas and more input and, oh, I didn't think of that, oh, I could use that. I mean, literally, I got 3 or 4 more ideas here that I'm implementing now. I mean, I really love it.
Michael (Alakazip) [01:42:52]:
Yeah, or things I really wanted.
Danno [01:42:54]:
Right.
Michael (Alakazip) [01:42:54]:
This interface, I'm like, I want that as a Consumer, right, for the videos I like, right? Like, that's— I didn't know I, I wanted that. Now I'm like, I, I, I only want that, dude.
Darren Oakey [01:43:04]:
Yeah, you always walk away with something like this. It's good.
Craig McFarlane (CraigM) [01:43:07]:
At least for, for me, I'm flushing all of the decades of, you know, rules of thumb about how to go about things. And so when I see somebody cutting the line by a new technique with AI, it's like, Oh, that's even better. And I have to remember everything.
Darren Oakey [01:43:28]:
The one last thing I'd mention is apart from Opus 5, did you see that Nouse released a new local 14B local coding model?
Anthony Nielsen [01:43:38]:
Who did?
Michael (Alakazip) [01:43:39]:
Nouse.
Darren Oakey [01:43:41]:
Or Nouse or whatever.
Craig McFarlane (CraigM) [01:43:41]:
They did? Yeah.
Darren Oakey [01:43:42]:
I didn't know.
Michael (Alakazip) [01:43:44]:
Is it—
Leo Laporte [01:43:44]:
but it's their own model?
Darren Oakey [01:43:46]:
It's their own model. So, so I haven't played with it. It was just like today, but—
Leo Laporte [01:43:50]:
Yeah, that's the nice thing about Hermes. I have the, you know, $20 a month news subscription. And yeah, I've been playing with Hi, uh, which is pretty cool. Um, I'll have to take a look at the news subscription there.
Craig McFarlane (CraigM) [01:44:00]:
The more and more—
Anthony Nielsen [01:44:01]:
Nomos One is what they're calling it.
Leo Laporte [01:44:03]:
Huh?
Anthony Nielsen [01:44:03]:
What's it? Nomos One.
Leo Laporte [01:44:05]:
Nomos One. It's more and more of these quantized or just smaller SLM models. That's what I— that's really what I'm doing with Ollama on the Mac Mini is it just has a host of interesting models I can—
Darren Oakey [01:44:17]:
Well, as I said, I did a speed test of Ollama versus native, which someone mentioned on a previous—
Leo Laporte [01:44:21]:
I'll take a look at native. Can you switch models on native?
Darren Oakey [01:44:25]:
Yeah, it works just like Ollama. But I just told the device to do a speed test. Nightscape was mentioning he played with it or something. And so I did a speed test and yeah, it was on, at least on my machines with the model I was using, which was Quin 3.6 35B, it was faster.
Leo Laporte [01:44:46]:
So I will— I just switched over right now.
Michael (Alakazip) [01:44:50]:
Yeah, we just need a really good router though. Like, these specialized models can be almost like agents, right? Like, you know, right?
Leo Laporte [01:44:57]:
Oh, it does some of that. It has a delegation tree and you can say for vision, do this for the— and, uh, and so, and it has already looked at, uh, I have a half dozen models, maybe more on the Mac, and it knows which ones to use and it can switch. That's why I wanted to switch on the fly.
Craig McFarlane (CraigM) [01:45:14]:
This is native?
Darren Oakey [01:45:15]:
This is my— no, this is my, um, thing that I use. It's got all of the agents in it. And, um, if I just say new, new task and I say auto, it'll actually pick based on, um, the how big the task is, but also based on the, um, the economics and like how many tokens I've got left and when they expire, it'll pick which one to use. But if you—
Leo Laporte [01:45:39]:
Oh, I like that.
Darren Oakey [01:45:41]:
This is the coolest mode. If you do ultra, I don't know, make me a cool-looking tic-tac-toe website. Stupid thing for Ultra. If you see, it's actually started off Kimmy, GLM, Terra, Grok, Opus, and oh, because switched to 5. And then when they come back, they'll all produce a plan. When they come back, it goes to Fable. It'll make judgments about the plan and then then send it to the council and then come up with a definitive plan and then. send it to just whatever high agent is available.
Craig McFarlane (CraigM) [01:46:29]:
Council?
Darren Oakey [01:46:30]:
Yeah.
Anthony Nielsen [01:46:31]:
I like that.
Leo Laporte [01:46:32]:
Council is so great. I use Council for a lot of stuff. Hermes has added this MOA thing, which is a mixture of agents, which is basically Council.
Michael (Alakazip) [01:46:41]:
Yeah.
Leo Laporte [01:46:42]:
So is this a harness or is this just—
Darren Oakey [01:46:45]:
Yes, this is my harness.
Craig McFarlane (CraigM) [01:46:46]:
This is your custom harness.
Darren Oakey [01:46:47]:
So I'm walking through this now. Yeah. And this is— yeah. I've got an agent on top of it called Beezel, which is my Hermes type thing. But then I've also got an iOS app for it. But yeah, I'm almost exclusively using that. If you look at, say, the models, it supports quite a lot of models. I spend quite a lot of money on subscriptions.
Leo Laporte [01:47:16]:
Yeah, me too. But, um, I got a call from China yesterday from ZAI.
Darren Oakey [01:47:23]:
Oh, really?
Craig McFarlane (CraigM) [01:47:24]:
Yeah.
Leo Laporte [01:47:24]:
So, can, uh, if you have any questions, it's like, what the hell?
Michael (Alakazip) [01:47:30]:
You answered it?
Leo Laporte [01:47:32]:
No, I saw this international call coming in and I— but so I just let it go to voicemail. But then it was like, what? Here, I'll play it for you. It was very odd. Oh, did I delete it?
Michael (Alakazip) [01:47:46]:
Darren, this is next level though, man.
Darren Oakey [01:47:48]:
This is—
Leo Laporte [01:47:49]:
I wish I had the skill to do this. This, this is what I want.
Craig McFarlane (CraigM) [01:47:53]:
This is brilliant.
Leo Laporte [01:47:54]:
How hard was it to write your own agent?
Darren Oakey [01:47:57]:
Well, this is not my own. I mean, the agent is on top of this, but this is the main thing about this is not just that it's a harness, but you can call it via the API. So anything above it can just say, give me a session, with a mid-tier LLM and it'll do it. Then you can build your agent or anything else on top of this.
Craig McFarlane (CraigM) [01:48:21]:
It's monitoring all of your different subscriptions to know you've got a little left here at this level?
Darren Oakey [01:48:29]:
Yes. In fact, I'll show you that.
Craig McFarlane (CraigM) [01:48:35]:
Is that good?
Michael (Alakazip) [01:48:36]:
And it's local. Does it— is local part of it too that you're—
Leo Laporte [01:48:39]:
yeah, you're mixing them?
Craig McFarlane (CraigM) [01:48:40]:
Oh, that's good.
Michael (Alakazip) [01:48:42]:
Yeah, I need you to write a very, very simple prompt I can hand to Claude so I can attempt to emulate even a small piece.
Leo Laporte [01:48:48]:
Yeah, would you, would you just say, hey, package this up so that we can all use it?
Darren Oakey [01:48:53]:
That's easy.
Leo Laporte [01:48:54]:
I don't want it to be a Swift program or—
Darren Oakey [01:48:58]:
It's written in Golang, so yeah.
Leo Laporte [01:49:00]:
Okay. So I could run it on my Linux box.
Danno [01:49:03]:
Good. Okay.
Leo Laporte [01:49:04]:
It looks very Mac-ish. That's what fooled me. It's got a Mac UI.
Darren Oakey [01:49:10]:
Yeah. And yeah, I do have a— I think you've seen my graphs. I've got a thing of all the things and it's got a little graph of where, like how many, how much I've used with each one. So yeah.
Leo Laporte [01:49:24]:
Yeah. Hermes has a— The web UI, which I use with Hermes, has that. I feel like it might have— if it doesn't, it really should have a lot of those capabilities because it has the delegation.
Michael (Alakazip) [01:49:40]:
I'm so in love with CLI that I keep hesitating to go out of it because I just love that it's like my childhood. I know.
Leo Laporte [01:49:47]:
Here's the Kanban, skills, memory, spaces. Profiles, to-dos. Oh yeah, here's this. Here's the token usage, skill usage. So Hermes does this and you can do it for different periods.
Michael (Alakazip) [01:50:09]:
This is your box, Leo, we're looking at?
Leo Laporte [01:50:11]:
Yeah, but this isn't— this is just Hermes, stock Hermes running Well, it's web UI, which is a user interface on top of it, but it's just stock. And then it has somewhere, it has some agent— maybe it's in the workspaces— has some agent delegation stuff. I keep trying different UIs and I haven't yet found the best UI. But the thing is, this must all be exposed in Hermes somehow.
Danno [01:50:43]:
Yeah.
Leo Laporte [01:50:44]:
It's just not, you know.
Anthony Nielsen [01:50:45]:
I would check out the official app. It's pretty good.
Leo Laporte [01:50:48]:
Yeah, I know. I keep asking, should I go to the official app? And it says, yeah, there's stuff you do. I tried something called Hermes Studio for a while. I don't know how good the official app is on Linux.
Craig McFarlane (CraigM) [01:51:00]:
This is—
Leo Laporte [01:51:01]:
I'm on a Mac, so I wouldn't mind using it on the Mac, but I don't know how good it is on Linux. Does the official app have things like that delegation? Model delegation. I feel like I've seen that, but I don't remember where I saw it.
Anthony Nielsen [01:51:18]:
Pretty sure. This is what the official app kind of looks like. So, let me get a better size here. I mean, I don't really use it, but like, you know, if you want, you know, to see everything and have like a terminal dock, like you can do all that.
Craig McFarlane (CraigM) [01:51:34]:
Oh, here it is.
Leo Laporte [01:51:38]:
So this is kind of like what—
Anthony Nielsen [01:51:42]:
like if you want—
Craig McFarlane (CraigM) [01:51:43]:
this is—
Leo Laporte [01:51:43]:
I'm sorry, I'm just quickly going back to web UI. This shows you what you've spent, which API keys you have. You can see I have more than I really should have. And I think there— anyway. Yeah, go ahead. I'm sorry. I didn't mean to take over. No, no, no.
Anthony Nielsen [01:52:00]:
Good. Um, you know, you know, if you want the terminal, but like I just do the, the, you know, clean focus one. Like you, you see all your sessions.
Leo Laporte [01:52:08]:
I just want to chat.
Danno [01:52:09]:
Yeah.
Anthony Nielsen [01:52:09]:
Just Discord down, you know, you got Discord.
Leo Laporte [01:52:11]:
That's very similar to the web UI. Yeah.
Darren Oakey [01:52:13]:
Yeah.
Anthony Nielsen [01:52:14]:
And then if you want to see your, your live agents or, um, your cron— I don't have any cron jobs, but they would be here.
Leo Laporte [01:52:21]:
Oh God, I have nothing but cron jobs. Here's the, um, This is how the models, the auxiliary models. So for vision, it uses that. For web extract, it goes to the Mac Mini, blah, blah, blah. So you can allocate all of these. But this isn't what you've really— yours is probably smarter than that, Darren, right? This I have to set ahead of time. Does yours kind of think about it?
Darren Oakey [01:52:46]:
To be honest, Hermes is probably much better than mine, but I— because I'm always building mine, I, I, uh, I know I should switch over to Hermes, but, uh, I'm very much a—
Leo Laporte [01:52:58]:
Now you're invested. Yeah, no, I understand.
Darren Oakey [01:53:02]:
Running my own stuff.
Leo Laporte [01:53:03]:
No, you're better off.
Craig McFarlane (CraigM) [01:53:05]:
You're—
Leo Laporte [01:53:05]:
I think it's better. You have full control over it. You know what it can do.
Michael (Alakazip) [01:53:08]:
Well, yeah, I was gonna say, like, I'm not nearly as advanced as you, Darren, but, um, every time I try Hermes, I'm so obsessed now with being able to do exactly what I want. And not accept anybody else's paradigm. And that every time I look at it, I go, but I want to make my own thing. And, and, right, and I know it's probably slower, but, uh, um, but I'm just so excited about my own software being at my fingertips just 24/7.
Darren Oakey [01:53:31]:
Yeah, definitely.
Leo Laporte [01:53:32]:
For me, I mean, if I weren't working on the other things, I would, because building your own agents probably makes a lot of sense, probably the smartest thing to do. And then you're, you know, superpowered in the rest of it. But, um, at this point I have too many other things on.
Darren Oakey [01:53:47]:
I, I don't know, I waste a lot of time building my agent, whereas if I was just using Hermes, I would be using it for other things. So I'm pretty sure it's a complete waste of time, but, um, I do it anyway.
Leo Laporte [01:54:01]:
What, what if this is not really? It's just a hell of a lot of fun. That's the real point. All right, gentlemen, we have now crossed the 2-hour mark. I think that's a good, good break— breaking point. We don't want to be longer than, uh, Christopher Nolan's Odyssey. Let's, let's, let's— I will, uh, cut it off here.
Anthony Nielsen [01:54:22]:
I will just label this— this wasn't— this was another, uh, another AI user group. Yeah, I'll label it.
Leo Laporte [01:54:27]:
It wasn't Token Time exactly. Yeah, although I like the name Token Time. I think we just call it Token Time from now on.
Michael (Alakazip) [01:54:33]:
Yeah, it's cool.
Leo Laporte [01:54:34]:
I think it's a great name. Yeah, I don't— I could do a code— we can't do a code-along with a bunch of people. You really can't. It has to be one person just sitting there typing, and you just have to be patient. It has to be somebody who's going to sit there and type and then go, yeah, yeah, what do you see?
Craig McFarlane (CraigM) [01:54:54]:
Okay.
Darren Oakey [01:54:56]:
Unless there was a dedicated project like something—
Anthony Nielsen [01:54:58]:
Yeah, that we were all working on.
Darren Oakey [01:55:00]:
Yeah, everybody's I'm just throwing in ideas.
Leo Laporte [01:55:03]:
Well, what would be fun is for us to all write a harness.
Craig McFarlane (CraigM) [01:55:05]:
I—
Leo Laporte [01:55:05]:
the only reason I haven't done that is I just don't know where to begin even with, with that, right?
Darren Oakey [01:55:12]:
So—
Michael (Alakazip) [01:55:12]:
And I'm too shy to, like— I cuss at Claude a lot, right? And I don't try to fix typos, right? I'm just hammering on the keyboard, so—
Danno [01:55:20]:
Right.
Leo Laporte [01:55:21]:
I dictate everything. But, uh, and as you can hear, it's talking to me a lot. Which I really like. I want to have that sense of a conversational thing.
Darren Oakey [01:55:34]:
But I many times say, why the F would you do that?
Michael (Alakazip) [01:55:39]:
That's my catchphrase. I told you, do not do that ever again.
Leo Laporte [01:55:45]:
I'm really trying not to do that. I just feel like it might have negative consequences. I try to do everything perfectly. I never say don't do this. I say, oh, do this. I don't know.
Craig McFarlane (CraigM) [01:55:58]:
I just—
Leo Laporte [01:55:58]:
I'm probably being superstitious.
Michael (Alakazip) [01:56:00]:
It's my whipping boy for sure.
Darren Oakey [01:56:03]:
Sometimes it's very deceptive.
Leo Laporte [01:56:06]:
There have been definitely times I've wanted to yell at it.
Michael (Alakazip) [01:56:09]:
But in the corporate environment I work in, if you start doing that, there must be an HR filter. It's like, whoa, whoa, whoa, shut it down.
Craig McFarlane (CraigM) [01:56:15]:
Yeah.
Michael (Alakazip) [01:56:16]:
So at home I can just get angry, you know?
Leo Laporte [01:56:18]:
Yeah.
Darren Oakey [01:56:22]:
All right, all right.
Michael (Alakazip) [01:56:22]:
Well, I agree with you, Leo. My bladder and my beer are both—
Leo Laporte [01:56:24]:
Yes, yes, that's good. So thank you, everybody. What fun.
Michael (Alakazip) [01:56:28]:
Thanks.
Leo Laporte [01:56:28]:
Great to see you all. Thank you for your contributions. It's been really great. Craig, wow, I'm impressed. Darren, as always, the rap monster. Alakazip, Anthony, uh, of course Dano, who is now off racing. Thanks to all of you for watching, and a special thanks to our Club Trip members. Without you This would not exist.
Leo Laporte [01:56:48]:
So join the club if you're not a member. We would love to have you. twit.tv/clubtwit. All right, signing off for another token time. See you in August.
Danno [01:56:59]:
Bye.
