[Prompt]
Host Update: Well, hello there to anyone who listens to the My Weird Prompts podcast.  My name is Daniel Rosehill.  I live in Jerusalem and I am the human in the loop, the permanent human in the loop,  behind the My Weird Prompts podcast and...  We're at almost 6,000 episodes, in fact.  I think that as I speak.  We're gonna publish our sixth South Vince episode and I have to say I feel kind of  slightly a stra

[Response]
Well, hello there to anyone who listens to the My Weird Prompts podcast.  My name is Daniel Rosehill.  I live in Jerusalem and I am the human in the loop, the permanent human in the loop,  behind the My Weird Prompts podcast and...  We're at almost 6,000 episodes, in fact.  I think that as I speak.  We're gonna publish our sixth South Vince episode and I have to say I feel kind of  slightly a strange mixture of ashamed apologetic  and maybe just those two things about it as well as having quite a large bill to pay to modal i  don't want to see this month's gpu bill for modal serverless inference modal has been an absolutely  trojan  platform for the podcast since its very inception.  I did look at the start at running Chatterbox locally,  and on my flimsy 8 gigabyte VRAM GPU,  I think Cloud Code said,  Yeah, it'll take you by 2018.  And then it kind of hung hours to generate an episode.  I was like, yeah, that's a bit slow.  We did manage to get our generation time per episode down a fraction a few days ago.  Some changes in decoding and...  What I wanted to do in this episode was, it's been such a while since I actually spoke  myself on the podcast and I think I've done one of these before.  It was at year end, I believe.  Which is amazing to think this has been just the quickest year of my life.  Feels like the quickest year.  That feels like just a few weeks ago that I...  Was wrapping up 2025  and  So it's been almost a year since I did that, which must mean I guess the podcast has been running for...  that long and I never intended it to be this,  uh,  to have grown to such a huge amount of episodes.  and the fact that I've been able to do this much without going  totally broke is a testament to the  Well, I would say imaginative.  engineering  Claude Cote being just sensational at figuring out  all the bits and pieces. I kind of bring the ideas occasionally  look at the pipeline code, which is  at this point,  um,  a...  one main Python process and then a lot of nested nodes were moving much more towards a very modularized  production pipeline but for the most part  I just sit there holding down the space bar and  and uh  saying  "Yeah, this isn't really working well.  What can we, how can we do this differently?"  That's pretty much the process.  And I only do that, I think about,  a month elapsed since today, which is really the first day in  such a long time. I would say maybe months that I've had a full, almost a full uninterrupted day,  to actually work on technical stuff.  and not DIY.  the trajectory of my life the podcast has become a  unwittingly and I actually didn't really even envision this because I didn't think I'd keep a  sending in prompts to this AI podcast for such a long time, it's become kind of a journal.  Everything is real.  Everything has happened.  We did get...  essentially kicked out of our apartment because a landlord wouldn't fix a leak. That is a true story.  And as bitter as that particular experience was,  And the huge amount of work that goes into moving an apartment  I know that there have been many episodes about that whole story and,  one of the failure modes to use one of Herman Poppleberry's  words is sometimes I repeat the story out of habit and  and I don't mean out of habit of telling stories again, I hope,  out of habit of reseeding context for AI because...  The fact that memory systems are good enough now  is still something I feel like is really going to work, I'll probably err on the side of caution  and just give the whole backstory again. And that's why sometimes in those episodes,  you hear rehashed the bitter tale of the apartment but the  that was a jump everything is really just intended as a jumping off point  and the much bigger issue that the apartment  Exodus of 2026.  gave was talking about a lot of really important socioeconomic themes  in Israel and we're two weeks from elections here and  I can't even tell you how frustrated it makes me. I'm actually looking out the window  at the Israeli parliament, the Knesset,  And I feel almost like disgusted by it.  Like just looking at that building,  I feel almost a sense of revulsion at these people.  in government who it's i mean on the one hand you could say there is  it's a vibrant democracy there's a lot of debate there's a lot of new parties but  It's the same story over and over again.  Everyone's talking about Netanyahu.  Sorry, I don't mean to make this a political rant.  They're talking about Netanyahu, what he's doing, what he's not doing.  Will he go with them?  Will we form a coalition with the Haredim and absent almost completely from the political discourse, is essentially every issue that affects  your average person living in Israel and  I think the only reason that  The renting issue is not a bigger issue or  driving the social protests that we did see once,  is that a lot of people in that building don't rent, have never rented, and have no idea what it means when your landlord says, "We're ending your lease."  and you then have to spend the next  months traipsing around apartments.  with the baby.  only to then find a place  pay an illegal agency fee  "Move all your stuff, disconnect it."  It's insane, the amount of work.  So that story was retold many times,  partially to recede context,  but also because, um,  It's one of those things that I feel genuinely  a fire about, maybe in the bad way.  Because...  It does make me angry thinking about tenancy law and income inequality, and I think how much disastrous policy we're pursuing in Israel.  But let me bring it back to the podcast for a second.  So...  just crossed 6,000 episodes and  Um, some recent changes to the podcast are, um,  The most significant one being that we're using DeepSeek 4.1 Flash as our  script agent and I did a little bit of tinkering yesterday as I mentioned  with Claude and at this point  I'm working on the level of saying that this is overall a good production.  I listened to it...  The podcast, I always say, is produced basically for me.  And if anyone else happens to derive value or entertainment from the episodes, amazing.  You know, sometimes I think,  That's probably too specific. And the truth is that it would be a lot of work to make a private podcast. My wife thought once, could we do like a little  side one just for us and uh...  We did one or two episodes like that and I said for the most part  Nah, I just don't really care.  to be honest, if people know about my gallbladder surgery and abdominophrenic dysnergia,  which is an episode I'm really looking forward to hearing because...  I get these emails  for...  you know years now because I've  put up videos on YouTube just from random people around the world who find random things  podcast episodes and connect about it and I think that that's  That's helpful until it's decisively proven to me that it's otherwise so.  Um,  It's in public and a lot of the themes are pretty niche.  We did some episodes today about NVIDIA ASR models.  And one sort of plan that I have for the coming few  months, I'm probably going to be doing less  Um,  Last episode...  the fact that there are 6,000 episodes, it sounds like such a staggering number.  and it's really because it's just so easy  to generate an episode.  building the pipeline was not easy at all. It took stages of iteration. The TTS, the text-to-speech, was by far the hardest point, hardest part of the pipeline. Cloud Opus 5.5 is just a  It's phenomenal. It's scary how good it is.  Now and again I  open up VS Code and I look at the Python pipeline and occasionally  edit little things here and there, but by and large,  It works and it's me just kind of  looking at the logs sometimes in Modal  saying, I can see this is failing, or  And this is the type of prompting that I enjoy so much more.  is this a good agentic framework that we're using? Should we have a planning agent here?  What about using a classifier model there? So it's kind of me  Listening to the episodes, that's the...  That's the feedback loop as I listen to the episodes every day.  Not all of them, but...  A lot of them.  And then I just kind of take mental notes like,  Oh, they keep saying that, that's getting annoying, or Hilbert Flumingtop, that's getting kind of repetitive, so I thought today let's have some, let's bring back the Jim from Ohio, the Cranky Caller, and the rest of the, you know, the rest of the kind of minor characters to add a bit of variety to it.  And...  So just--  The fact that weeks go by now with me just in listening mode and just taking notes and then sometimes I'll ask Hannah who's just come into the house.  So I'm going to go into another room and close the door.  Sometimes I'll say, hey, did you listen to that episode?  What were your thoughts?  Yeah.  anything annoyed you or she'll tell me and that's kind of our feedback loop.  She's not quite as avid a listener as me, but she does send in interesting prompts about architecture and urban planning.  She is an architect. Everything about our lives on the podcast.  including the  the rental horror story, all the DIY questions.  It's all real in case anyone was wondering.  And mostly it's me just firing off  sort of things I wondered. And sometimes I will use AI tools and  that's increasingly I'm kind of using chat GPT as a sounding board.  Like, I'd love to have an episode about this,  But  Here's the angle. Can you help me think through this before I send it?  And I think those ones usually turn out better,  except when ChatGPT decides to add a million questions.  But it's getting much, much better, I have to say.  I'm feeling today, I must say, very positive about how incredibly good  all the tooling that is maturing in the AI field is becoming.  And now we just need to get a plugin for the chat GBT plugin for the podcast because the  beloved OpenAI custom GPTs  are finally being retired, which is a bummer, but inevitable, I think. So the listeners' statistics, we do have some validated stats now. That's been a few months ago.  And because this podcast, I see the return on the money that I invest and  in the  monthly bills and yes we are essentially bootstrapped production as basically an investment in my  continuous professional development.  The episodes about the world's worst rental cars is clearly not, it's just there to entertain  me and maybe there's some useful info there but the ones about AI are very much me prompting  for info that I kind of then take back and use, that's the idea at least.  But it's a good value.  proposition, let's say, as far as I can tell.  Speaking about cost containment and value, so the model is DeepSeq 4.1. As we mentioned,  And that's been, that's pretty good.  It's DeepSeek hasn't yet released 4.1 Pro.  So I got an email a few weeks or months ago at this point,  it's a bit nebulous in this time haze.  Seeing this  Four pro is being they're gonna start serving four point one flash and we did a couple of episodes actually about a  DeepSeek 4.1 Flash in the podcast, which for anyone else involved in local AI engineering, those episodes are  to my mind at least, are gold because it really works nicely.  They dig through the actual source material and get into the actual nuts and bolts of how the model is put together.  And the TL;DR version is that 4.1 is a very, very significantly different model entirely than 4.0. It's not really... that's not understood as a point release.  Speaking about our model and our agent architecture, so I was doing a little bit of tinkering last night arguing with dear Claude about, I'm always the one saying we need a proper framework, we need a planning agent, it needs to be, you know, a system.  Claude says it doesn't really need to be that complicated, but we're gradually moving towards, I think, a very stable production pipeline and it's just a little tinkering at the edge. The voices, Herman and Corn, and the TTS is still using Chatterbox and  It never occurred to me until recently  how quickly Chatterbox actually updates. And I was pleasantly surprised to see that the,  we're periodically getting updated waits. So I couldn't understand,  I was listening back to previous episodes. I mentioned that the TTS was the hardest part,  of this podcast to get right by a mile.  And...  I always assumed that we were using more or less, you know, there was one chatterbox release  and there was a noticeable difference. Sometimes  people who listen to or who have listened to many episodes might have noticed.  We do have some hallucinations sometimes that creep in and for anyone who's in the weeds or the  joys of AI engineering, I think it's a decoding bug and I've been looking at ways to  keep our fast inference on chatterbox while  catching those hallucinations with the hallucination guard and today I did create finally the  Latin classifier  For those who are based in Israel and you hear the mispronounced Hebrew words, which is...  It's kind of almost like its own, you know, the references to Eyelass and the rest of it. It's kind of its own unique little funny quirk about the podcast now. But that was, I rolled that out today and eventually...  I'm going to try listening to the multilingual version of the model and seeing if that speaker level diarization actually really works.  So that's basically where it is at the moment.  and  As I mentioned,  the one thing I'm really working on or want to work on is  Once the podcast as it is, is a kind of a stable version or a good quality, I much more enjoy working with projects, whether they're open source or any other context, when we reach that point of things being good.  And I love when the improvements just become incremental.  And I think we're getting there with the podcast.  That is it. The incremental improvements are probably going to be in that domain. I'm doing a little bit of  Worked today on a...  Text Normalizer.  And that's essentially it. I'll make my shameless pitch because I do address the show so rarely. If anyone does want to become our next sponsor, you are invited to send us an email. Any of the MyWord prompts.  My Weird Prompts.com emails should eventually write through to me potentially via some kind of  wily ai agent so you might indeed get a message back from the personified  email manager of herman poppleberry um but they'll eventually get to me so drop us a line  um we do have the verified listener stats now which is something that i added  very belatedly, so, uh,  Pod track verified we can prove that people do actually tune in and make yourselves known if you have a prompt  Telegram bot may not be working currently  We will get it working, but  If that doesn't work.  drop me an email, do send me a prompt. Some friends actually do send me prompts informally on  WhatsApp and Telegram, which if you know me in real life, that works too, but  I'd much love, I'm waiting to get prompts from people I've never met, never heard of.  And if you can think of something that jives with our.  general tendency which is to take some seemingly mundane fact of the human world and do a whole deep dive on it.  then send it in. Our deep dives by the way, I got good confirmation recently because  Herman actually referenced one of my own projects in an episode without realizing it was me. It was my Whisper Hebrew  whisper Hebrew fine tune. And I was like, wait, that's my, that's my project. And I was impressed  because the details were right. It's Exa AI is our retrieval layer. And the full pipeline for  anyone interested.  is on the website at mywordprompts.com.  And there's also an Amazon book.  And if you want to literally read the docs, as in,  by a physical book containing the docs.  It's a little bit out of date but not substantially so and it contains  actual Python  Um,  the actual Python code snippets from the pipeline. A lot of it's actually still code that runs.  So it's kind of like if you're someone like me who likes to either learn or  Brush up on Python or just likes to see worked code examples for me personally makes looking at I always find reading code on a physical piece of paper weird experience and not on the screen.  But I like looking at snippets because you can kind of see oh that's the mechanics of it so that's actually a pretty good read I have it on my bedside and it's available on amazon.com just search for My Weird Prompts and there is an episode index.  Again, a little bit lagging than the previous one.  If you don't know about it, there is an app  quasi app it's a pwa um stream.mywordprompts.com visit that url directly and you'll get that so  little ezra who's frequently mentioned is banging on the door for me so i'll have to go but that's  been the update uh for me daniel roso thanks to everyone who's been listening to the episodes do  send in your ideas and until the next episode  My Red Proms Podcast.  Creators Update #2  Over.