Episode 258: Perceptron

0:00 Podcasting to you. 24th, 2020. Perceptron. ♪ Bumming self-pacemaking ♪ the country. It all goes down, we've got the rag. point over in the boardroom 2.0. In fact, I'm going to be only boardroom that is not a part of I'm about to be. Curry here. The man who will. Your weights are cornered. Say hello to my And on the other end, the one.

0:42 Only five. Aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa Oh, you're, uh... Get that done, even with the cedar fever. Yeah. to see the fever. And the only thing that will cure it is more cowbell. I'm trying. I'm working on it. This, uh, you've... You don't get knocked out like this very often. No, this is the – no, no, no. I had this in January, I think. I don't get knocked out by much. No, you don't. What do you mean my note isn't live in the splits? What does this mean?

1:22 Eric PP says, your node isn't live in the splits. What does that mean? Isn't in the live splits. Oh, well, that's okay. Oh, it's not in the live item. about. No wonder I'm not getting live boosts. Hmm. Hold on a second. Let's see. Let me see. It's not in there, really? Let me see. Value? What is this? Oh, I see. Okay, well, that would make sense. Let's see. What am I? Adam at getalbie.com?

2:07 This is the listen to people type on their keyboards podcast. I really don't care. I don't care. Yeah, because I say that because I'm doing the same thing. Yeah, I really don't care. I don't care what anybody thinks of me. All right, so let me do this. Well, no wonder things weren't working. What? Yeah, thanks, Eric. Yeah, really. Sorry to derail the show. No, you did it, man. You were running and Eric just took his leg out. Yeah, I was good to go. Everything's fine. But also the way I read it, I mean, this is not how he wrote it, but. Let me see what he said.

2:49 I've got to read it. The way I received it was. uh, Your node isn't in the live splits again! See, that's how I read it. Thank you. And. And that is obviously the problem where I wasn't getting live boosts on the last show. It makes so much sense now. Okay. This was a gift from Sir PP. No, of course it's a gift. And I'm just telling you how I feel. And my wife was mean to me a minute ago. Oh, yeah. I'm so sorry. Yeah, I say, I'm going to do the show. And she's huffing and puffing and doing the sheets. And doing the sheets sucks, putting the new sheets on. say you should have just asked me whatever i'm like oh okay and it's just whatever i'm like did you just say whatever to me You got whatevered? I got whatevered. Like, okay.

3:39 All right. Well, this is what is that? So isn't this not worth even talking about? Yeah, but you whatevered me, man. Yeah, it's like we'll just deal with this after the show. Yeah, well, and and have. Have a good show. Not just whatever, Joe. Hello, boardroom. How y'all doing? This is Podcasting 2.0, where we discuss all things podcasting. And sheets. And sheets. And sheets and pod pings and all kinds of stuff. Yes. Had a fun. A town hall. A town hall. Wednesday. Wasn't we at Wednesday?

4:20 with Alex Sanfilippo. Town Hall. Yeah, he put together a town hall. He'd done some survey, I guess. This all happens on LinkedIn, I'm sure. How do you do – what is the difference between – like what are the criteria that – delineates a town hall from a fireside chat oh uh a town hall mean it's it's all zoom so it's basically all it's basically a zoom call um but the town hall had multiple people speaking and that was actually that's that was that's actually a pretty good question because some people were like what kind of town hall does this i thought i got to stand up and ask questions No questions. It was about the...

5:02 about the issue of people spamming to get guests onto podcasts, which we had discussed in a previous board meeting to which I had suggested the booking tag. And so, you know, Tom Rossi was on and, you know, it was it was actually it was quite good. What did Tommy have to say? Tommy R. He said. Well, here's... His contribution was, you know, hey, we're taking out the email addresses from the RSS feeds. Okay. And then I came in and said, yes, and that's very good. And so what we're doing is we're suggesting putting a tag in called the booking tag.

5:44 And, you know, it was kind of a scripted thing. It was like, here's the problem. Here's how here's interim solutions. Here's the solutions. And I think overall it was it was really good. And now Alex and Daniel J. Lewis. working that. Yeah. So it was kind of an industry rah-rah thing. I thought it was quite good. Not as scripted as the interview that Pod News had with the lady from Spotify. Oh, I have not. I've been I've been doing I've been just working nonstop this morning. I'm being whatever. I just, you know, rocking and rolling. I'm getting. live servers up and running and God cast to you videos. And I've been so behind because of this Cedar fever. So I did not listen to it. How was that? What was the, what was the interview about? Who is she? What lady from Spotify? Who's on first? I don't remember her name. It was.

6:42 It's basically one of those interviews. where I'm going to send you questions. This is what it sounded like to me. I'm going to send you questions. You record them, send them back to me. Oh, goodness. Record the answers, and then we'll splice it all together. I hate those. You could hear almost everything but the paper. that she was crinkle yeah i mean it was all it was it was fairly forgettable what was it about though what was the How to get. how to get boosted in their or something. How do you get spotlighted or whatever? Yeah, the bottom line was basically you can't. How to get spotlighted. You need to find out who's on the staff and take them out for drinks. Like the way it's supposed to be.

7:35 I mean, there's it sounded to me like they do it the same way Apple does it, which is they just. Roll dice and pick. pick somebody from a list or, or, you know, somebody on the inside and you get a, you get to add a boy or something. No, I think it's, I don't think there's any way to break in from the outside. You know, the whole problem is it's editorial team. They, they start to run it like they're overlords and it's, it's not, it's not bad. That's just. That's how you do it when you're running editorial. You determine and everybody hates you except the people that got highlighted and everyone else thinks you're a douche. That's the job. That's the joke. I mean, any algo that you put out there, somebody is going to figure that thing out and game it. I mean, it's just you.

8:17 No, it's right. But I'll go suck. I'll go. I'd rather have it be editorial. In fact, you know, it's like, why does Spotify do editorial? Apple does editorial, but not a single other podcast app does editorial, which I've been complaining about for years. Yo! Have an opinion. We should do an editorial. Have an opinion. We mean we. Should we do a podcast index? Do you want to be hated? I'm not interested in that. We're already hated, right? Pick a topic. I have no time to do editorial for Podcast Index.

9:00 I mean, I don't mean we would actually do it. I mean, we would like throw an LLM at it and just pretend we're doing it. Just to get taken out for drinks, basically. Yeah, exactly. Yeah, sure. We would be just as effective as the Podcast Academy or whatever that is. No, but isn't that the truth, though? I have never understood this. For some reason, podcast app developers, and maybe it is the I don't want to be hated thing, but if you want to delight your users. then look at what your users are doing and do editorial. Everyone else is doing it. Spotify and Apple literally are pushing shows that they think are good. You can be your one-man editorial team and you can delight your users with all kinds of suggestions. No one does that. There's this kind of like, well, it has to be equal for all, mom. Like, no.

9:57 No, if you want your app to be successful, you have to have other things. And those other things are highlights, spotlight, blue light, moonlight, whatever you want to do. Rachel Maddow app. Hello. I've been saying this forever. This is exactly what it should be. One day somebody's just going to vibe code a Rachel Maddow app and stick it in the app store or something, and it'll just be for you, your special app. Yeah, you don't have to get through the app store. Do you have Rachel Maddow's permission to do this? Well, you know, that's an interesting topic because I don't know what other. Podcast app developers are seeing out there, but.

10:40 On the Godcaster side of things... Paul has been telling us that App Store approvals are taking longer and longer. Thanks, guys. Slop. Like AI. Oh, very good point. Yeah. Well, we know from. customers. that the, I call them the wrapper apps, but what are they called? Container apps. I think is what they're called. Yeah, yeah. They're having a real hard time. And what you're describing is like? Somebody just takes your website and essentially wraps an app around it. Yeah. But it's really just pulling your – Yeah, you have to have a certain amount of unique qualities in the app or native qualities. But then they also run bots and algos across the whole app store and say, well, this is pretty similar to this one.

11:42 It's become very difficult. Which is kind of infuriating to me on one level because – As. I've written a couple of iOS apps in the past. never written definitely done a couple of iOS apps and One of them that... did was for company I worked for and We did a lot of recruiting. college students to get them into our industry. a lot of work with that.

12:23 Um... So I wrote this app. Essentially, it was an app. focused on our It was branded with our business, but it was all like, it was a recruiting app. It was a way that. could install our app and see what our upcoming events were going to be and all this kind of stuff and register. A lot of work. I spent weeks writing this. Amen. It worked great. It had a server back in to do all this kind of stuff. And then Apple just flat rejected it. They said this is just a promotional advertisement for your business. Yeah, isn't that the whole idea?

13:09 I was like, well, but no, it's got all this functionality to it. I mean, like this is no different than. You know, any sort of like event registration type app. tons of examples of this. No, sorry. It's just – and then you just have these wrapper apps, yeah, like you're talking about that just really just wrap a website and stick it out there. somehow they just sail through. Yeah, you know, if they hadn't been such a-holes about... Web apps, we wouldn't have this problem. You could just install a web app as simply as an app from the App Store, which was the original idea.

13:50 You know, when Steve, it was, it was when I met job, when I met Steve jobs. Did you say that? Did you say that, Wentina? complain about the sheets but listen here when i miss steve jobs she's like don't hit me with your steve jobs crap again i'm gonna do that i'm gonna do that Whatever with your Steve Jobs. When I met Steve Jobs. Uh, This was the iPod Touch days. This was the device. This was the dream. The iPod Touch, it would not be connected to a phone network, and it would be web apps. That was the whole idea, and it was a really good idea.

14:33 But when I met Steve Jobs, he was, I noticed right, he was yelling at people. Wi-Fi. was so mad. I'm not sure exactly what they messed up with Wi-Fi, but something had happened with Wi-Fi, with the protocol or how it switched. I'm not sure exactly what it was that his dream was crumbling. And this is why he had to eventually do a deal with AT&T. And AT&T, as you recall, was the first partner. rolled out the iPhone. You couldn't get it for any other network. It had to be AT&T. And he also never wanted an app store, which in hindsight.

15:15 Yeah, he's a huge money maker at the App Store. And I think you're right. I mean, the way we're suffering or slash that, the way you're suffering from the generated slop, you know, they've got to be inundated with this stuff. But it's the point where people can't even get a developer account approved within five days. It takes some – some have been waiting for months. blessed. The last app. release on android that paul did i think it took over a week to get approved Oh, really? Oh, my goodness. Didn't it? It took a while, yeah.

15:57 It took a while. Maybe not. Maybe not over a week, but it was close. I mean, it was quite a number of days. And that's a definite down. *Burp* That's a definite increase in time. recently because it didn't use it was much quicker before Listen to this. I think I've received my first. Boost spam. So thanks, Eric P.P. Now that I put my. My split in the – my node in the split, here's what I got. Orangefriend.com.

16:39 to swap to from Bitcoin on LN and other cryptocurrencies. Compare instant exchanges and P2P markets now with no KYC prepaid cards. Satogram is what it's called. It's a Satogram. How much was it? One Satoshi. I am a huge fan of this. You can spam my lightning node all you want. Just wear it out. Yeah. I love it. Yes. Paying me? Paying me to spam me? Oh, yeah. Bring it on, baby. Yeah, eventually those sats will be worth something. We're so back.

17:21 Yeah. Yeah. Chad F. Chad F. In the hairpin. Very nice. Like, um... sort of the big nun. One of the big non-podcast stories. This week was... Tim Cook retiring. And I just like. Well, Marco will be happy. He thought Tim Cook was a huge traitor to liberalism. like this if you look across everybody's acting like Tim Cook was I mean I don't care who cares But just seeing a...

18:12 point that's made over and over is everybody is saying, well, you know, no matter what you think about Tim Cook, he was like a business genius who will never be matched again because of him taking Apple, you know, up to being a $4 trillion market cap. This is... If you look across the whole tech industry, there are many companies that have. Yeah. Runs that they'll never have again because it's not about necessarily being any particular genius. Well, I'm not saying he's bad at his job. I'm just saying that like when you when you spent when you when you literally spend almost a trillion dollars in stock.

18:59 Bye back. You're going to shoot that. That number is going to go way up. Well, also, he's a supply chain guy. So he did really good things with the supply chain. Okay, so Eric PP says you can actually turn stuff off. Oh, that's very cool if you want to, you can turn that stuff off. What off? Spam. You can turn spam off under a certain number. Hide boost amounts below. See, there you go. Oh, you got a threshold. Oh, yeah. Well, this is, of course, this is a helipad. This is one of the best pieces of software in the universe. This is, it's a spam assassin for boosts.

19:42 Is anybody – there's probably one dude out there still running Spam Assassin on like a box in his closet. You know, the other day I had one of those moments where I'm like, I should probably check and see if I really got all the Bitcoin off of that old laptop. You ever done one of those? Yeah, where you have a panic moment, yeah. Well, it's just like, you know, because I famously sold 65 Bitcoin at $900. Like, let me just go see. And so it was Bitcoin Core QT running on a MacBook Air. So it wouldn't even connect to peers or anything. But I'm like, oh, OK, I can see, you know, I could see all the addresses. And yeah, there was like 19 sats here or there. But then there was all these transfers to AC Android wallet. I'm like, huh?

20:31 I wonder. Oh, you have a mystery wallet somewhere. Like 30 Bitcoin. Like, I've got to find this thing. And so I'm talking to Tina, and this was before the whatever. This is when she was – maybe that's the reason for the whatever at the end of this story. And an Android phone. So this must have been around 2013. You keep all your old devices, though. Every single one. She says, well, when I was dating you, I remember you had been to the strip strip joint. And your Nokia E71 was left in the Uber. I said, okay, so the E71. How can you make it rain with Bitcoin at the strip club? Well, no, but this was the Nokia. She was just trying to help me identify devices. And I know that. Baby, trust me. I got Bitcoin. I'll send it to you.

21:27 This is the yellow rose. It was some business thing. I wasn't really. wasn't really going for the strippers. And. And so I was like, okay. And I said, but I had an iPhone and an iPhone 4. And I'm not quite sure what came after that, so... You know, I go into the bin and literally the phones are in order. The Nokia E71. iPhone 4 and right in between that was an old Pixel. Bingo. I mean, how awesome is that? And she said, I will never complain ever again about you keeping all of your old crap. Bitcoin on this thing. I bet. I bet. Yeah. Well, and guess what? Zero.

22:11 0.00. Of course, 0.00. Oh, yeah. We all have the story of selling – like I sold six Bitcoin at $1,800 and thought I was a genius. Yeah, it's – whatever. Yes. So live and learn. So this is interesting. Uh... RSS payment boost, true fans. Okay, true fans. It's coming through a little odd, but it's coming through. Cook missed every tech cycle. Search AI EV and glasses. Also cloud, and the list goes on. Not a genius. He is Apple's bomber.

22:52 Well, here's this. That's a Sam Sethi thing. I've ever heard. Yeah. Yes. Now, here's the only thing about this new guy. And. And unfortunately, his name has too many syllables. You know, Steve Jobs, Tim Cook. This guy came in. He has like too many syllables in his last name. I don't remember. What is his last name? Turnus. Turnus. OK, it's going to be hard. John Turnus. Is it John Turnus? John Turnus. John Turnus. So he's the hardware guy. Well. That's interesting because if I've learned anything over the past 18 months, particularly in the last six, you know, Apple has their universal memory and their own silicon is highly usable.

23:42 for the AI. Oh, yeah. Yeah. And, you know, they've had all kinds of AI capable type chips in the phones. They may have an incredible and they've held off. You know, it's like they've they've had a few missers. They've they've pulled back from the AI nonsense. And boy. I can't blame him because that's Samsung that I got. Yes, they all sounds like, hi, I'm Bixby. Oh, it's some some stupid agent. Yeah. Would you would you like for me to recognize you talking to me automatically? No. And get off. You can't even take it off the phone. Bixby. That's the Samsung AI. Everybody has got every company is making some agent and giving it some stupid cutesy name. It's like you should ask our agent, you know.

24:34 you know, to Todd. I was like, I don't want to, I'm not going to talk to your agent by name. Just quit it. Yeah, exactly. Um, But I'm seeing the revolution unfolding. You know, this new, I haven't been able to test it, the DeepSeek V4. There's a cloud tag on Olam. It's not actually there yet. This thing's going to be pretty interesting. With one million token context. I haven't even heard of this one. Yeah, yeah, DeepSeek V4. Let me see. The V4 Pro, 1.6 trillion of parameters. Yeah, 1.6 trillion parameters.

25:18 V4 Flash, 284 billion parameters, 13 billion active. Open weight. MIT license, 1 million token context window. Uh... trained on domestic Chinese chips. um, You know, this could be something very interesting. These are huge. These models are giant. So I've really become kind of 1.6 trillion parameter model. Yeah, that's pretty big. But, you know, this, what is it, together.ai? They're the ones that have been most stable for me of all the rented GPU outfits. This is.

25:58 This has changed my life. Yeah. You know, it's like, oh, there's a five hour senatorial testimony on C-SPAN. OK, robot, go download it. Download it. OK, run it through the fastest, most awesomest whisper model you have. Do word by word. Throw the JSON onto your drive and then summarize. a summary. All right. Well, that sounds interesting. What any good fun quotes you can get there for some clips, you know, and that costs 15 cents an hour. Yeah, and all you would have to do then is just publish it to Spreaker and you can start getting ad revenue.

26:42 But the point is, so if you want to use a model like this, you can run it pretty efficiently from, and I don't even know what together, I don't even know what their model is, their business model. I'm sure you're using some dude's gaming computer. Isn't that kind of what they all are? You know the new thing, right? call it. distributed Proxy. I can't remember the name, but. Well, serverless inference is one of the catchphrases. No, this is a scam tactic where you – where you sign up for something and you're giving.

27:33 You're giving your... You're giving a third-party access to use your computer as a proxy? Yes. Nefarious actors can distribute their workload and not come from. the center. of system numbers. That's also cool. Yes. Yeah, this is like, people are cracking down on this all over the place because people are, they're using it like. installing Roku apps that have this thing embedded. Oh, yeah. Suddenly your television is being used for a proxy for a bot army. Yeah, that's good. And bingo, Bitwarden gets popped. Yeah. How about that, huh?

28:14 Yeah, I mean this stuff is probably coming through these distributed proxy networks now. as their attack vector because they're so, like, it's all over the world. And what you need, like what we do in the Podcast Index, you know, we clearly, we... block a ton of of data centers. a lot of heuristic looking at traffic that's coming from a data center, that's coming from an IP that's related to an autonomous system. which is connected to a data center provider. And so, like... If you have a browser user agent.

28:58 claiming to be chrome coming from AWS. Yeah, you're no good. You're no good. Yeah, you're not Chrome. You know, you're an imposter. Yeah, one of the huge benefits of running Omarchi is, and I guess you could do with the Mac too, but all my bot-based stuff, I pipe it all through my home desktop. Home IP coming from a home machine looks legit. And I've had – I'm currently moving my freedom controller to my – to my house. to Ubuntu or to my podcast rig. You should run it on the same machine that's running our Albie Hub because it's working so well. What? Is Albie Hub working? Is it not working? Not for Comics for your Blogger. He complained again.

29:51 What? I got to see this. Podcast. L index, LN. it. I don't understand because it was it's up and running. Well, I've got I've got machine two here with and it is running right here. 18 1842. What is 1842? I don't know what 1842 is. It was a good year for wine. It was a good year. generated boost Oh well.

30:31 I don't know what to say. It's up and running. I mean it's just sitting right here doing its thing. Anyway, so you're moving the Freedom Controller over to your own home machine. Yeah, because everybody's doing this. Everybody's blocking data center traffic, and my Freedom Controller for many years has been running in Linode. Right. And so increasingly I'm getting 403 just denied. Yeah, yeah, yeah, yeah. And so I can't save articles into my archive anymore. to have to move it. Luckily, we... We facilitated for this a long time ago and. you can use. So one of the things frame control can do is it'll set up a bucket. an S3 or somewhere. you can hit the bucket.

31:13 And it'll bounce you to whatever the current IP address is. of where the frame controller actually is. Oh, cool. Yeah, it's fine. just find this way. But I don't know. It's... Like, this is a... This is becoming a real problem for attackers, so now they're going to have to figure out a way to take over residential IPs, and that's behind all of this. They just do that in the routers, right? Aren't the routers complete pieces of crap that are just all Swiss cheese? Oh, yeah, for sure. Yeah. Your router has about as much, you know. If anybody sits down with a commercial router.

31:54 more than You can just imagine that people are stripping the firmware out of those things, just running them through an LLM to find bugs. Right. As we're talking. This is a change in the world that people are not very aware of, and it's going to mean more people will be needed. pain, but also more people will be needed to reconfigure everything. The whole centralized model is under attack. Yes, the Slopocalypse. Slop apocalypse. Yes. Yeah. It's, it's, it's happening and we're not ready. We're not ready for it. Yeah. Really not.

32:35 You know, It's so hard. to find this stuff. Part of it is just – part of it is – There's going to be some amount of. slop is just kind of undetectable. Automate your way around it. I mean, I've been—this is— been focused almost squarely on this for couple of, you know, two, three weeks now. you tell you that I'm in model training mode right now. That's the next step, and that's what I'm working on.

33:19 is complicated. So... We can talk about it if you want to. Well, I just want to answer. He's in the boardroom. What even is slop? By the way, whenever I play an AI-generated song on No Agenda in... show mix, he's the first one to say AI minus. So you clearly know what it is. Slop is what pigs eat, and it's usually a lot of the same of it. I think that's the definition. And this is the thing.

34:03 That we can't. ever get around. I'm talking about more than just technology. This goes way, this goes all the way back to. famous. of I don't know. define pornography but I know when I see it yeah yeah There's a lot of – there are many things like this. the world. Buh-uh-uh. You know, one thing is just language in general. You know, we talk, we, we. what we were talking about last week, but it brought up idea It reminded me of Noam Chomsky's work on language.

34:44 Before Noam Chomsky was the sort of – Anti-war. Before he was friends with Epstein, you mean? Oh, that's – yeah. Um, Before all that, before he was the lovable kook. Um... His big contribution was... in the world of language and philology. Phalology? H-I-L-O-L-O-G-Y. Hold on. Book of Knowledge, give me the definition of philology.

35:30 According to the Book of Knowledge, philology is the study of language in written historical sources, combining linguistics with literary criticism and historical analysis to understand texts and their cultural contexts. Those It has been written. Hey, can I vibe code or what? It's just – it would be like I can't – I can't help wishing that that sounded like. the, the priest. Monty Python and the Holy Grail reading from the Book of Armaments. I mean...

36:14 I got it where I wanted it to be. That one caught me off guard. I was if you is that been on no agenda because I introduced it a couple of shows ago. So I have a little interface with a push to talk button. And it's piped in exactly the way I want it. And to bridge over the super fast whisper, but then the FFmpeg process, which takes the voice and then adds the echo to it, I added the little page scribbling. This is... I'm dangerous with this stuff. That's A-plus. That's A-plus work, brother. Thank you. Okay.

36:57 You know what it is. I like being – this is why I don't listen to No Agenda because I like being surprised by your antics. Philology. OK. But that's where Chomsky's chops came from. That's where he... sharpened up his chops was in language. And that's... What he said, what his big contribution was, and I think it's still. pretty much the defining characteristic of language. astrology is that He said that language.

37:41 It appears to be something built into. Like it's not something that is learned. So you could say, okay, is language a general knowledge thing? Where we... learn it the same way we learn how to drive a car. the same way. Um. put silverware in a certain order on a table. He says no. The language is something that's unique. wired into the human mind. way that seems to be pre-existing. So you could say that humans are...

38:25 are born with the faculty. the property of language. construction. we understand what he, I think, if I'm not mistaken, I think the term. used was universal grammar. He said that this spans all different languages. On the outside, it may look like Mandarin Chinese and Italian are But on the inside, it's really all tokenized. Is that what you're getting? It's all token. Yeah. It's an API. But it looks like from the outside, those are completely different languages.

39:11 But what's happening in the mind is something. Chomsky called universal grammar. So. this idea universal set of rules. We all share this structure, no matter what the overarching language is. We all operate on a subject, an object. and other parts of grammar. that even though they may be in a different order. where they may be a little bit different. We all understand what these things are, you know? Yeah.

39:55 And so... Uh... I don't know if I said this before, but it's sort of like a... Like mere Christianity. Like there's. There's a lot of different theologies, lots of different doctrines, all these kinds of things. But the core of what Christianity is, it's like these two or three things. Yeah, like Jesus come quickly. Everyone in Christianity is thinking that. Yeah, it's just, you know, so there's... That's this, that's... When it comes to language, As I'm working through the train, this how to train. this model thing, I'm, you know, this is, this beginning, I'm beginning to sort of see.

40:41 pattern here. That goes back to what we were talking about earlier, which is. These things are hard to define. but we know what they are. If I gave flashcards... to Everyone in the boardroom. And gave him 20 flashcards and said... Okay, and then played a – played – 14 seconds of each podcast and say, okay, here's the art, here's the title, here's the description of the podcast. Here's 14 seconds of audio.

41:25 20 out of 20. would identify every single one of them perfectly as either AI slop. just a normal podcast. I agree. Now, when you try to take that and put it into an LLM, into a model. It is. You hit a barrier. And so – and I think that the issue has something to do – so we've talked about Jock a little. about Jacques Ellul before. Yes, a good old Jacques. Many times. Yes. Yes. Professor Jacques. He's an interesting guy.

42:09 had his two seminal works, Technological Society and Propaganda. One of the key takeaways from his work, Propaganda. Is that if you're trying to convince someone or a group of people that your idea is. If that's what you're doing as your propaganda. then what you're doing is not propaganda or you're not very good at it. Right. Because. Attempts at propaganda, like. Attempts of propaganda by someone who's not good at it, they'll try to do that. They'll try to explain to you why they are right.

42:50 But what he says is propaganda is not about ideas. It's about actions. Real propaganda. does not care about the ideas in any shape, form, or fashion. They change over time. They're different every day. The news cycle changes. It's all kind of irrelevant. All that matters is whether you drive a particular outcome. action. and say a thing and then that provokes an action in the listener. Your job is done. That is propaganda. Would you like to hear a readout? of one of my agents that searches for YouTube videos and has been trained.

43:36 to avoid AI slop videos. Okay, red flags. Telltale AI phrasing. So it only does it on on whisper transcripts. Personally, I think. Let me break this down. What many people don't realize, if you take a step back and think about it, One thing that immediately stands out from my perspective What this really suggests is deep dive, red flags in the channel, no real person or camera or identifiable name or title, generic aggregator channel names, stock footage with TTS narration, monotone, flat, overly uniform anchor voice, synthetic rhythm, no natural reading errors, no pacing variations.

44:24 The rule, scan first few lines of transcript before anything. If the patterns are there, don't cut the audio. The content might still be useful research, but the AI voice never goes on air. Adam caught seven AI slop clips on NA 1857, rejected them all on content. So that's how my agent is now identifying this stuff. Which sounds kind of human-esque, really. Yeah, and the issue is now do that for – A few hundred podcasts an hour. A minute. Oh, an hour. Yeah, right. Yeah, exactly. Exactly. And so this is the problem. So when you're restricted to only sort of – When you're restricted to only language.

45:09 don't have the processing horsepower. listen to all the audio. It is very difficult. But isn't it amazing, though, that with three watts that powers my entire brain, I can do it in seconds? And this was Chomsky's. genius was that It's not math. The way our minds work is not a math problem. It's a meth problem. Yeah, sometimes, yes.

45:50 So. It's this like. ideas I think the nature of what – take if you take sort of what Jacques Ellul was saying He's saying that action... action Words and actions are – they share sort of a congruent relationship. They function together. Words drive action. Inaction drives words. have a relationship. that is mysterious.

46:31 But it is there. When we talk to people we have relationships with, it really matters what the ideas are because we want to be on the same page with each other. In a way that something like propaganda doesn't care. Propaganda just cares about actions, but we don't when we're having real relationships. Um, you know, there's... There's this fundamental paradox with language that the words themselves sort of don't matter. But at the same time, they very much do. And I think that's really what you bump up against when you try to make a definition.

47:15 Because in one sense, the. In one sense, the words don't matter because words are kind of fungible. The English word is five, but the Spanish word is cinco. They're interchangeable. Also, the meaning of words change over time. Yeah. So the words are just – the words are a conduit for meaning, and that conduit can take many different shapes. square, round, oval, whatever. Essentially, the language is not the point. The ideas are, the meanings are the point. But then there's this other aspect.

47:58 where the words themselves are incredibly important. Like, in fact, they're critical. Um, So on the broader scope, the words don't matter, but when it comes to conveying the meaning to another person or another language, passing sort of… The utility of those words you choose are a critically important aspect of it. Mm-hmm. So there is like a fundamental paradox that exists within language. And this is why I brought up the propaganda reference because when it comes to propaganda, word choice, word choice is everything because the actions you desire as outcomes from your words were as a result of the emotional stirring. Of those words that you use. Of those words that you deliver as they sound to the hearer.

48:50 So like a good example of this. think back to the election of this most recent election. One of my favorite examples is J.D. Vance. There was the sort of like... When he was nominated as a running mate, all the typical attacks started, political attacks. The one thing that came up was that people started calling him, quote, weird. Do you remember that? Of course. Weird is a word that young people use to mean that a person makes them sexually. comfortable. That is what, that is, I have kids. That is when they say the word that when they say that person is weird, that is what they mean by.

49:37 That is not a, that is. That is not something that older people without kids would. people with older kids would be familiar with. Um, that went like If a young kid is watching like a movie with a sex thing, they're like, That's weird. That is the, that is. what they mean. You would. And so like, you're only going to know this if you have children that are like in their mid twenties or younger. And so research definitely went into the use of that word. And it was chosen to convey that exact meaning. that generation of people. You know, so like.

50:18 The language. evokes actions. is one aspect of it. The language and then the theme. The meanings though... The words convey can change over time. So as I'm going through this model training thing. hit this point when I'm interacting with LLM where it asks me Like it says. Do you want me to spot check the output of? blah, blah that it's doing. And I just stopped.

50:58 I'm like, what do you mean? What does that mean? Spot check. Like what? What does it mean? I have no idea. Tell me that it's right? but in what parameters? Like what? Like this is the problem. We're – It's trained on our language. Yeah. And so it's talking like us. It's trying to communicate something to you. But when I say, hey, Adam, do you want me to spot check this? You can instantly get it.

51:38 You have an understanding of what I mean. Just innately, you don't there's we're on the same page just by default. But when the machine asks me, do you want me to spot check it? I'm really, I was just sort of paralyzed for a minute. I think you might be overthinking it. Doesn't it mean exactly the same thing as if I said it to you? No, because what if I say yes? It says, do you want me to spot check this data? If I say yes, what am I communicating to it? I'm not sure. Because if I say yes, is it going to say – what is it going to then do? What would you expect it to do? But see, here's the question. But the question though is – and this is the action part. This is the part that I'm getting to is that – If I say yes and without knowing exactly what it intends to do.

52:36 What is it going to do? Hold on a sec. If you said it to me, no, if I said to you, do you want me to spot check this, Dave? And you said yes, what would you expect me to do? And why would you know what I was suggesting? So. We would have a context. If I'm saying – if you say – Hey, Dave, this feed is not updating in the index, and I think that there may be a few more of this type of feed you spot check.

53:18 Okay. I would immediately know what. you are communicating to me and what actions you expect me to take. Okay, so apply that to your example. You were doing something and the LLM says, do you want me to spot check this? What was the process you were doing? Right. And so what the. What I have found through this process. is the process of model training. is that a good 90% of the work of model training goes into preparing your data correctly. training data set.

53:59 Having a having a. having an excellent training data set. is literally 90% of the job. had to go through a lot of steps. The first one was building what James wanted, which was the dead feed. problematic feed That is now built and is functioning. So I can't release it. really as a public link i don't think calls. It's got DMCA stuff in there. We asked, we've been asked to remove. So I don't think we can release that publicly, but I think it would be okay to release it to research.

54:51 purposes. So somebody would have to like contact me and ask for it. Chad F. says hey adam can you spot check this boostogram done Context understood. Human to human communication complete. Yeah. And so if somebody wants it, James or anybody else, call me. We'll work it out. I'll work out how to get it to you. So that was step one. as I'm working through what it required was to put a bunch of new SQL, SQL.

55:32 Uh... statements into the index to get all this data out in the correct way. export it into a SQLite database just like the other one. What I wanted is two SQLite databases, the standard SQLite database that we export every week and this new problematic database that we're going to use as the... Test. Corpus. feeds. Yeah. And so as I'm going through this. You know you're going to get criddled on this. That's okay. That's fine. Podcast indexes deplatforming people. you Wrong!

56:13 I've been criddled before, and I'll be criddled again. That's fine. That's not a problem. So as I'm working through this, I exported the data, and there was a problem. Thank you. SQLite database. initially came out. erupt. The LLM gave me a list of issues. It was like the SQLite database is unreadable, but also the CSV that we exported. was missing was it was like missing a field he gave me like four or five different things that were wrong.

57:00 And then it asked me if it wanted me to... X, Y, and Z and spot check. don't know what you're referring to. And it said it in such a way that I was left with the impression that if I say yes. then it's going to perform an action that I may or may not want. Amen. This is vague. I don't know what we're doing here anymore.

57:40 So this – but this is all part and parcel to … the difficulty of saying to of having to describe something that is basically indescribable. Like you, what your example was, you gave it, you gave a list of things, you say. This tone, this kind of, you know, these phrases. I didn't actually give it that list. It came up with that list itself. I didn't give it that list. Well, you came up with a set of... heuristic phrases. No, I didn't.

58:22 It did that itself. OK, excuse me. Based upon a feedback of loop of me saying probably 10 times that's a slop. And we're all, to a certain extent, going through this same exercise. We are trying to come up with heuristics. Yeah. To identify some stuff. But what we're going to end up with. And this is always how heuristics work. What we're going to end up with is the same thing that happened in the past. phishing scams. The way that any...

59:03 adversary consider AI, in this example, I'll consider AI slop our vague adversary. The way any adversary combats heuristics is they shrink. sample size. get smaller and smaller and smaller. So there's just simply not enough information there for you to make a, for you to pass a confidence threshold. that sample size. something like email spam or email phishing It can just be the number of words they use. Yeah. So they may just pig butchering. Perfect example. Yeah. Hey, did you go? Are you going to be at the tennis practice today? Yeah.

59:49 There's simply not enough information there. Yeah. There's just not. But when you're dealing with something like AI-produced podcasts… Well, now you have many variables. You have all the metadata that goes into the podcast itself. You have the voice of. AI slot presenter. have the content. There's tons of variables you can push and pull. round But for each one of those, there's a set of heuristics involved, and you can refine it. down Over time, point where it is indistinguishable.

1:00:30 Because if I get on, I've... This is... what I've been learning all week this week. What I'm trying to do is create a what's called a Laura. Laura adapted. So the way this will work. Is that a Lord of the Rings adapter? Yeah. I think it stands for. A book of knowledge. What does L-O-R-A stand for in AI?

1:01:12 No, it's having a hard time. According to the book of knowledge, Loara stands for low rank adaptation. Parameter-efficient fine-tuning technique for large language models and other deep neural networks. Thus, it has been written. Low-rank adaption. That's it. That's it. Nailed it. Good work. So you – the LOR – the LORA adapter. is a way to add information into a language model. ace model. You're actually making adjustments to the weights themselves.

1:01:52 This is fascinating. I love this. So the quote-unquote the weights, you're really talking about… you're talking about blocks of, you can think of an LLM as a stack, sort of. It's sort of like a stack of matrices. There's a... that are layered into each other. And those layers are distributed. blocks across. across the model. And one of the parts of the model is the – which is the best name in all of computer science is the percept.

1:02:36 Perceptron. Yeah, so you have the attention layer, which feeds into a four... They call us like a forward. forward functioning a forward forwarding function and then that goes into a uh perceptron right And essentially all it is is it sounds fancier than it really is. It's just a – it's a block of matrices that – try to make associations with the attention between words. So it really is just matrix math.

1:03:18 So the LoRa adapter. is a set of Matrices. you pull out of a training data set. You add them into the. LLM server, so something like Lama CPP. add that add that LoRa adapter in. and to the serving harness and then it will do an additive. to those new blocks into the matrices of the large language model. It essentially puts it through it. The serving harness.

1:04:08 Yeah, so like something like – you can think of like an agent harness would be something like Claude. Okay. The server harness would just be the same thing on the server side. Okay. So it's not changing the base model. So you're going to take a model or something like Quinn 3.6 or something. Let's say you take Quinn 3.6 35B model. It's a mixture of experts model, 3 billion active parameters. take that model as your base. You are going to. essentially take another set of weights that are trained on your own data set. And you put that in front of it.

1:04:50 put that beside it. So now that it's now every time it does, uh, every time it runs through. perceptron it is also going to add these other weights into Hmm. in a way that sort of nudges the output token in a certain direction. And the good thing about this, which is really nice, is you don't have to change. Doing it this way, you're not retraining the entire model. You're just adding this extra component, and that extra component is small. It's 500 megs or less. and you can just swap it out. So you can train a new one every month if you wanted to. Very cool. And just add that in. Yeah. So like if you – the reasons you would want to do this – are specificity.

1:05:41 Or high customization, because you can do things like this with like rag. retrieval augmented generation. You take a database of and sort of mix. mix that kind of you you run the output through the rag as a filter on the output side. Or you could... pile everything into the, into the. prompt into the context. which you. quickly run out of space. Right. But the LoRa adapter... This would be called fine-tuning the model.

1:06:22 this? can do if you do this well and fail. It can easily fail if you don't have the right training data. Eric says, is this better than old machine learning techniques like random force or neural nets? I mean, you know. Transformer models are really just an evolution of neural networks. So this is going to, it's very similar. Well, I think I have a different question. Between the server harness. The Laura When is the last time you took Melissa on a date?

1:07:08 set. There. There will be no dates until school's out. I'm just saying. Before you get a whatever. Yeah. Luckily, I've changed the sheets. That's my job. I do not have to worry about it. Nanny. All right. Um, Good question, though. It's been a while. We're due. Just saying. As soon as school's out. Um, So yeah, Eric, it handles language better because it's built for...

1:07:48 It's built for that sort of thing. tokenized input But the... training if you do it right. If you do it wrong, here's what can happen. training set is poor. or if it's too small or too big. you can end up just making the model memorize things. So really to make sure that you're doing a good job of this, you have to… You have to do a lot of testing on the backside. Spot check. Spot checks, yes. You have to do it. You have to do.

1:08:29 a lot of testing to set to Make sure that it's not just memorizing. The training data. Ah, yes. Because what you're looking for is a balance of weight, of waiting. word connections. And so the best way I saw this explained was. You can see. well-balanced model. You can do the underlying matrix math. Yes, exactly, Eric. That's right. It's called over. You can do the underlying matrix math.

1:09:11 you can literally come out with. and see a visual representation. of King. is to man what queen is to woman Like you can say. You can take the relationship in the matrix of king. to male. And then you can do the exact same math on the word queen, and it will predict that it's going to be woman. Okay. So that's a good sort of like... What is woman? Our Supreme Court justices can't answer that. Yeah, this like.

1:09:56 But if you over trained it you may end up with You know, Queen equals Elizabeth. Yes. See what I'm saying? Like you would you would end up with a model that has memorized the training set. And so. This is where we're headed. If we can, if I can do this right. We will end up with. A highly customized model is an expert. Identifying. slop. Not just slop, but slop and spam.

1:10:39 content these kinds of things All these things we don't want. We can have an expert that can hit those things. and identify them. In a way that is. forward moving Whereas just doing it the way that. moment by piling more and more and more stuff into the context window is just not scalable. Right. And then eventually this will wind up as an endpoint that people can query or a flag or something that can be used by people who use the index. That's the idea, isn't it?

1:11:19 Oh, it already is. I mean, like the reason – so if you hit the problematic endpoint right now, report. The recent problematic endpoint. You can already see what it was classified as and the re-inassurance. about the reason. Is that open or do you have to have a key for that? So what it's doing is it's training. So. taking all that into Taking all that context in. What I've been hitting is... The training set design is so complex.

1:12:03 I need. I'm aiming for about... 25,000. quote, good podcasts. Good ones. Okay, good ones. And those need to range. across all the different ways that we know that they do. Everything from... Yeah, everything from Joe Rogan to – Oystein Berger. To Oystein Berger. Yeah, exactly. Oystein Berger. To Bowls with Buds. Mm-hmm. To Pod News Weekly Review, to a person reading a bedtime story on Buzzsprout, all to some teenager reading their term paper on Anchor. All of those are really – all of those are legit podcasts.

1:12:56 But then somehow identify the difference between. Between a 17 year old in Delhi reading their report from school. And differentiate that as good against. AI generated voice. speaker reading a bunch of history nonsense off of wiki Right. Interesting challenge. Yes, it is very – it has been a challenge. I mean it's not – where it's not. I still got a few weeks to go.

1:13:37 Realistically, it's probably three weeks to get first shot at this. Because if your training data set is wrong. Or if it's poor, you're just going to end up training the model that… Everything from Spreakers Yeah. And that's just – but that's not true. Well, it kind of is. It's like, you know, or Like if, you know, if Spreaker and the other, here's what would probably help. If they put a flag in, this is a free account. That would probably be a good signal.

1:14:19 for your training. Let me give you a great example. Here you go. Here's one. Is it a number I can call as James? No, I don't think so. There's no numbers in this one. I'll find one of those. If you care about predictions and props, right now, it's all about playoff pressure from the hard work. Always. Okay, hold on. Always do pre-rolls. But here's the twist. We're only going... AI. Oh, yeah. And I'm Luna. So excited. Now at McDonald's, a McDouble is $2.50. So you can get your gym gains on or just get lunch. I can't shuttle. It's almost here. Yeah, yeah, McDonald's. All right. They made some money off of us. All right, here we go. Hey, welcome back to Learn Tamil with Vexingo. I'm Lucas.

1:15:19 And I'm Luna. So excited for episode 13. I flop. Yeah, we've been on a roll. Today, I thought we'd do something a little different. Oh, goodness. Hold on. Let me do something here. Uh... analyze this podcast. AI slop or not. I'm just going to give it to my robot. It'll take a minute, obviously. My robot is nothing like your building. But anyone who's listening heard that right away. Oh, yeah, that's what I'm saying. This is the difference between the built-in language.

1:16:06 an emotional action comprehension that humans have versus machines. We just know things and we don't know how we know them. We just do. We just know. And nobody can explain the – nobody can explain how. Um, And everybody's going to pass that test, every single one of us. But AI is really going to struggle with this. And it. Well, right off the bat, it'll struggle with the ad. Exactly. It has to get past the ads. point wonder if my robot will figure that out.

1:16:49 Well, and this – so here's sort of the last – make is I think. Here is an Achilles heel. of how we discuss these things today. to about AI in terms of what it's not capable of doing or where its flaws are right now. That is just. that is very short-sighted. What I think we have to do, and this is why I'm trying to do it this way. What I think we have to do is assume.

1:17:37 These AI-generated podcasts are going to get so good. We cannot tell the difference. We have to assume that they are going to eventually be perfect. you know, in. their tone. word choice. evade any sort of detection this is what i meant by by saying they're just going to reduce the attack surface they're going to stop doing the things that get that that sound unnatural And the uncanny valley is going to shrink down to such a small degree that we're not going to be able to just do heuristics on them. What we're going to have to do...

1:18:24 is is look at whole thing in a Bayesian way. where we're doing it. We're taking... things into account that they can't stop. Would you like the verdict here from my robot? Almost certainly AI slop. The evidence, 55 red flags, 55 podcasts across every major language. No human operation produces 55 language courses. Farm. Two warm voices. Exactly two TTS voices across all 55 languages. Real language instruction requires native speakers for each language. Hosted on Spreaker, huh? See, I told you. Listen to this. It says, hosted on Spreaker, one of the platforms Dave flagged for hosting in cloned spam feeds on last week's show.

1:19:17 Holy crap. No team, no instructors named, no methodology, no credentials, completely anonymous. Generic Fexingo branding across all 55. Template operation, learn, insert language with Fexingo for every single one. zero information about who runs it, no company location, no people, just an email address, the math, 55 languages times even 10 episodes each, 550 episodes of audio with two voices, that's TTS. The pattern matches exactly what we screened for, generic channel names, no identifiable humans, and possibly broad output from a tiny operation and a template applied across dozens of instances.

1:20:04 Okay. Great analysis, but let me throw a twist in there. Okay. One of the things I've been seeing a ton of lately Are people taking LibriVox? reposting them on Spree. to add revenue. Thank you. So that's theft. Well, not – no, actually not legally because LibriVox uses a public domain license. You can really they literally say you can do whatever you want with it. stop you. So legally it's fine. except that you're going to have 100 different people posting the same exact thing over.

1:20:50 Here's what I would actually enjoy. Because LibriVox, since it's volunteers. Many of the recordings are horrible. Yeah, it's actually an improvement. A really well done AI narrator. I mean like a high-quality – A.I. narrator would be better. and many of the LibriVox recordings, and I would prefer to listen to the AI version. If it was well done, then I would the LibriVox human version. But aren't we now down to what I've always said is some people will actually want this for certain reasons, and that's very – we all need our own – Percepticon, uh, Perceptitron.

1:21:40 I need my own perceptron. Yes. So see, and this is the issue is like just because it's AI generated does not mean. is slop and nobody wants it. But... If you ask somebody, is this slop? They will be able to tell you. Sir Bemrose has – They will be able to tell you, and for reasons they don't even themselves know. Sir Bemrose has a good point. One of the most powerful signals for AI slop is that the podcast has ads. And that's another thing. It's like you don't want to overtrain on that either. Yeah, exactly. Because then the model will be like – Oh, yeah. We got ads must be slop. Yeah, but that's not true.

1:22:27 And if there was a really well-done AI-voiced version of… Um... You know, Frankenstein, Mary Shelley's Frankenstein. It was way better than the liberalized recording. I would not only enjoy it and want to listen to it, but also not mind if it had pre-rolls in it. Well, back to my point. Isn't it ultimately that we all need? own perceptron. really do. Because you're saying things that are really important here. I wouldn't want to hear that. I want to hear some crappy Frankenstein. But I wouldn't want to be. No, I don't want to hear that.

1:23:11 So isn't that kind of the point where we have to be, that we all have a robot that we train for stuff that we want? customized customized yeah customized but in the interim while there's while there's sort of like Brokers of pipelines of content, brokers of content like Podcast Index. plus podcast direct. these kinds of things. We can't – As a conduit. we have to be able to continue to conduit. Yes, yeah, true. We can't be overrun with 700 people posting, you know, LibriVox recordings of Frankenstein with a couple of pre-rolls.

1:23:55 It's just not – it's just – that's – it's absurd. Anyway, so that, I mean, that's, I think that's where, it's difficult. We're gonna, you know, we can try to hit the high level and train, you know, train this model to get the obvious junk, like the one I just posted a while ago, but. Yes. you That's obviously junk that nobody wants. It's the equivalent of punch the monkey. Oh, goodness. But those were good for a while. Those were fun. I've tried. Everyone's punched the monkey.

1:24:37 So you would actually wouldn't mind a good TTS reading of Frankenstein. Oh, yeah, like if it was a high-quality, like— When we get – when – When TTS models get to the point where they can really deliver. quality just the LM notebook. get to that point, oh yeah, I'd love it. I'd love that because... Because I love audiobooks, and I like LibriVox, but so many of the recordings are just so bad, they're hard to listen to. So I asked my robot to find a reading of Frankenstein that is well-read by TTS, good enough, even though it is a eye slop.

1:25:21 That would be a perceptron that you might program, right? What's not there yet? It does have my podcast index API key and secret. Oh, it doesn't? Okay. I'll look forward to the server crash. In a .env file. Don't worry. In a .env file. It's not posting it anywhere. Until you get a CI CD pipeline hack. I don't be everywhere. I don't see DCI anything. CDC. Whatever. I don't do any of that stuff.

1:26:01 Well, I hope I don't cause any problems. A lot of people have keys. Every time I run pip install, I'm just like. Ha ha ha. Well, isn't it with just, like you said, every packet manager, NPM, all of this stuff. Yeah, they're ticking time bombs inside your machine just waiting to crap on your front porch. Yeah. Uh, well, you are doing important work. I'm doing work. I don't know if it's important. I think it is. I think it is important. Let me see. It's going to get a two-minute sample into my show folder here.

1:26:46 I got my mom pretty happy with my robot. I really want your little – guy that runs out there and gets stories and stuff. No, no, no, not that one. But the one that just analyzed the. podcast. Oh, that's my favorite guy. Yeah, I like that guy. Yeah, no, that guy is good. Okay. I want a copy of that guy so I can make it work on my behalf. Here, my guy has got something for you. Letters. Frankenstein or the Modern Prometheus by Mary Wollstonecraft Shelley. This is a LibriVox recording. All LibriVox recordings are in the public domain. Do you think that's real or is that slop?

1:27:27 real to me. According to my robot, it's a TTS. For more information or to volunteer, please visit LibriVox.org. Recording by Caden Clegg, thelunarisland.blogspot. That sounds pretty real to me. Sounds real to me. Let me see. Some people just sound like robots. Yeah. I just think that's real. I think it's real. OK, we think it's real. He's talking through a tube. Loser robot. I have to call my any AI I call a robot. It reminds me. Don't get involved with this thing. It's a robot. Don't take it seriously. Yeah. No, you can't. You can't. That's dangerous before you know what you said. OK, good night. Talk to you tomorrow.

1:28:14 You have no idea how many people are doing that, brother. It's bad. Hey, we're over time, man. I should have gotten you out eons ago. I'm sorry about that. Yes, we're way over time. Do you have a few minutes to do some thank yous? Yeah, sure. Why are you talking through a tube? I'm doing a LibriVox recording. Value for Value podcast, which means we deliver the value. We hope you enjoyed today's board meeting and we would like to receive some value back. Lots of people contribute time, talent and treasure. And here we thank people throughout the show for all kinds of things they're doing. I mean, started for me right off the bat with Eric PP with a with a helipad. best software package in the universe. And then, of course, he also scolded me.

1:28:57 you But on the treasure side, this benefits the podcastindex.org infrastructure directly. So if you send us boostagrams or if you send us PayPals, which you can do by going to podcastindex.org at the bottom, a big red donate button, you hit that and you can send us some fiat fund coupons. So there's 333 from Chad F. He's asked me to spot check the boostagram. C-Loss on Linux 2220. Oh, interesting. fountain. It might have been a row of ducks, but the fountain somehow is rounding up or down. He says, firm weight update available. This is probably when you crashed. Um...

1:29:37 That was it. Yeah. Here I got another 59 sats. Interesting from true fans. Co-listening and chatting with Sam Sethi, talking about beverages. Tea from Assam, India, and rosé wine from Provence, France. Do ye. And it stopped at the Y. Cool. Good. They're talking about things completely unrelated to the show. Yes. From Podcast Guru 333 from Eric P.P. Pew. Yes, we got the Pew. There's the 170 sats. Fetch metadata. Oh, what is this little fetch metadata thing? Hold on. Oh, wow. Okay. I'm sorry.

1:30:20 Oh, goodness gracious. This is awesome, Eric. Share with the class. Yes. So on a weird lightning invoice that comes in that I wasn't able to read, there's a little button now in Helipad which says, fetch metadata. And then, boom, it translates it. And so now I see, indeed, a row of ducks, 2222 from Lyceum, co-listening with Sam Sethi, talking about beverages, tea, blah, blah, blah. Do you have the ducks in a row? Cheers. I don't know exactly what's happening, but I love it. Cool. Yeah, whatever's happening is awesome. Great. 6390 from Sam at truefans.fm. Yeah, that was about Cook missing every cycle. Saltacrayon333. It's been months. Howdy, Node, he says. 333 from Chad F. This Boostergram brought to you by EricPP++.

1:31:11 2220 from C-Loss on Linux, AI-generated Boostagram spam. And, okay, so then there's that actual spam I got. And now I hit the delimiter. So we're in a good place. Thanks, everybody. That was fun. And Eric PP, Helipad, rocking it, man. Love it. Thank you. Trying to get my I think pulled up here. Oh, yeah. Okay. What do we got here? I guess. Oh, that's a I hate the way it mixes all the PayPal stuff together. can get. head around that. All right, here we go. Oh, oh, look.

1:31:53 Buzzsprout. Holy moly. Sorry. 20 inch blades on the MP. Yo, baller boys and girls from Buzzsprout, thank you. Thank you. Tom Rossi. It's not even end of the month yet. Uh, Tom Ross, he also does conference calls. does for free. for free. Michael Goggin, five bucks. Thank you, Michael. A thousand to five in a blink of an eye. Jorge Hernandez, five dollars. Thank you, Jorge. Christopher Reamer, $10. Thank you, Christopher.

1:32:33 Yeah, Cohen Glotzbach. Thank you, Colin. James Sullivan, $10. And Randall Black, $5. That's our PayPal. Let me see what we got on. boosts here. Oh, we got Oscar. Oscar Mary, 20,000 stats. He's through Fountain. He says, sorry for the disruption, guys. That's no problem. No worries. No worries, brother. No worries. I will throw no stones. Comic strip blogger, delimiter, 20,000 sats through fountain. He says. Today, I want to recommend a podcast from the podcast Morning Chat. www.podpage.com slash pmc slash which is quote daily morning show for creators by creators ever wonder how top content creators and podcasters keep their shows fresh engaging podcast morning chat hosted by Mark Roenick Suggested by Martin Lindeskoog. Yo, CSP, AI Arch Wizard, $15.52.

1:33:42 Thank you very much, Comicster Blogger. That's it. Everything else is just river bitcoins, man. So I hit my – this has not happened. This happened twice now. Ever since the wonderful upgrade, I've hit my limit on the... ...Claude Code. Oh, did you see? I saw that. Yeah, I saw the GitHub. is now charging per token, and they've gotten rid of their all-you-can-eat. GitHub is charging per token? What does GitHub do with tokens? GitHub Copilot. Ah! So now I can upgrade or wait an hour and 54 minutes. What is this?

1:34:22 falling apart but i thought i had extra credits man you created we need some credits hey man and i need some credits how come you can't give me some credits this is no good any credits resets at 4 10 p.m so predictable so predictable well i mean i predicted this i said it was going to happen but now i'm actually push your model i'm now i'm kind of pissed like hold on man Hold on, man. I can get some more money. Hold on. Are you on the 100? plan or the 200? I'm on the Hyundai plan. Nope, gotta go to the two. I don't want to go to the two, man.

1:35:20 But what I don't understand is I thought I had... I thought I had extra credits, man. Your 20 becomes 100, then your 100 becomes 200, and next thing you know, bada-boom, you're on the API plan. Oh, well, the API plan, I mean, that's the one that's crazy. Oh, you'll go broke. Oh, goodness gracious. Oh, no. Let me just take a look at here. What is this? Billing. Here we go. Billing. But I have extra credits. I am literally a whore. Let me see. Okay, man. All right. I've never run out of tokens on the Claude $100 plan until this past week, and I hit my limit for the first time. And I've changed nothing about the way I work. It's happened twice for me today.

1:36:12 Yeah. Well, this sucks. It's Opus 4.7, and they changed the default. If I'm not mistaken, check your settings. I believe they changed the default to 4.7. To super high effort. Oh. So make sure your effort is the same, and you may want to change back to like 4.6 or 4.5. So where do you do that? Effort slash effort. And then what do I say? Low, medium, high, max, auto. is what I've always used. But I noticed today when I opened it that it was auto set to extra high.

1:36:52 you Yeah, you're right, Cotton Gin. I've heard that. 4.7 people So how do you set the model? Is it model? Yes, model. Yeah. Opus. Okay, model. What is it supposed to be? It's not auto-completing for me. 4.6. But do you just point in 4.6? Does it do that? No. Model 4.6 not found. Opus 4.6. Opus is it 4.6 I don't know. Anyway.

1:37:33 Dave, we're – I'm telling you, the local model – Oh no, that's the future. Try. Whenever you get a chance. Try. um, Triquin 3.6, the 35B. A3B model. is is bull 3.6. Quinn 3.6, the 35B, A3B model. A3B. model okay use on open code run it locally and run it on open code it is really real Awesome. And it's so fast. I mean, it is like it is giving me output.

1:38:24 almost before I hit the enter key. It is lightning fast. And this concludes another episode of what the heck were Dave and Adam talking about? It's a beautiful thing. Dave, thank you, brother. Have a great weekend. Done. Or whatever. Thank you. been listening to Podcasting 2.0.

1:39:06 And this is. Go Pugs! note all you want.