Hey everyone. So this is Daniel Rosehill and this is going to be an episode with a difference of the My Weird Prompts because we're not going to hear from our delightful imaginary donkey, Herman Poppleberry, nor his brother Corn the sloth, nor the always cantankerous Jim from Ohio. What I thought I would do instead is talk about um just the origins of this podcast and what I'm trying to achieve and the limits of the of the podcast and the process that that's in place. So I would say firstly, this is this podcast, the concept of using AI to generate a podcast um is something that I've been looking at for um in and out of the past year. And my motivation has actually been using this as a tool for learning. Um I know that might raise some sort of eyebrows and uh furrowed looks. Uh you can't really trust what AI says. And I have had overall, I mean, we've talked in the in the a lot of these podcast episodes are are talking about very uh the technical uh parts of AI and AI engineering.
And uh we covered hallucinations in a recent episode, in fact. Uh but I found that the models over time have gotten very, very good at synthesizing information. And the pipeline um at the moment has undergone a number of revisions um to bake in things like grounding and uh a little bit of RAG. Uh but generally, uh Gemini 3, which is what the um scripts generation is currently based upon, um is very, very good. And I think that there's an enormous advantage. Um I kind of use Gemini as my workhorse AI tool. I still like ChatGPT for conversational use, but Gemini does so much heavy lifting um for very, very affordable prices. So it's um and it has that huge reservoir of Google's own data to draw upon. So I've been very, very happy uh with the information in most of these episodes and it's interesting. Um my objective in this really was when I began using AI for sending in uh prompts uh with my with my voice, uh which was about this time last year.
I discovered that voice tech had become significantly better and that in turn allowed me to um generate prompts that were maybe more thoughtful and specific and had some context data. And I ended up getting, that was really a sort of watershed moment in my use of AI personally. Uh I ended up getting really, really meaty and useful responses, pulling together information from very, very disparate parts of the internet. And that to me is where large language models really excel. Um the challenge was that something that I've sort of been moaning about since the start of my AI journey has been the lack of attention paid to AI storage or output storage. Um in it's kind of getting a little bit better maybe over the course of 2025, arguable. But still scant attention is paid to the question of when we start, you know, we work on prompt engineering, we work on AI workflows. Um but then very little attention goes to how do we actually consume the information from AI tools? And what's a little bit disappointing uh to me actually is that when I look for AI podcast generators, some of I think the most potentially beneficial and interesting and fulfilling uses of AI immediately kind of get subsumed into this, let's do everything quicker and cheaper mentality.
Um that most of the podcast generation I've seen has been kind of um I would say aimed at people who want to just like, you know, get become big podcasters through the um through any kind of hack available and AI podcast generation seems like a good one. My objective in with this podcast from the start was basically giving myself something interesting to listen to. Um I really have I've always loved audio as a modality and especially when um it comes to these kind of let's say deeper topics like some of the recent episodes on this podcast talking about uh topics like reaching AGI through bigger and better or an alternative one, why GPUs are have become like the thing for AI. And these can be sort of put together as long documents or PDFs, but I've I find there's something very satisfying about listening to a conversation. Um NotebookLM is an interesting exponent of this model, but there's something about it that um didn't do exactly what I was looking for here, which is that NotebookLM is, you know, grounded in specific sources.
Um you don't have control over the voices. I find the voices kind of grating. Um they have that kind of very generic, slightly over overly upbeat um energy to them and the grounding thing is on the one hand an advantage and on the other a major limitation because you can't just kind of say have at it, go go and find information from this. Um being constrained to sources can be useful in some cases, but I think for this use case it actually is kind of an impediment. Um so I was working on these these workflows and I came up with the idea that initially I wanted to record prompts, get them transcribed with uh speech to text like Whisper, and then have them, you know, spoken by an AI personality. And the initial personalities I was using one of the only um real challenge in this is doing it at a at a cost per generation that is uh that is affordable enough that I can run a few. My goal was to run a few of these per day and, you know, have a couple of hours of really interesting uh stuff to listen to and when out walking or at the gym or whatever.
Um and we're kind of getting actually to that reality now. The pipeline is is I think finally becoming quite robust. Um but it it's taken a long time to to get here. The first implementation was me speaking prompts, getting them transcribed, and then the with the text to speech, that's the that's the hard part. Um because you either have kind of very generic sounding, not so great voices. Um 11 Labs is very, very expensive. And then I just came up with this idea. I have these, we have these two stuffed animals at home. One of them is Cornelius the sloth, one of them is Herman Herman the donkey, and I've always kind of had a thing for stuffed animals and kind of come up with their little personalities and voices. So I thought, why don't I try to get Herman and Corn as the podcast hosts? Um so that was actually not the not the hardest part of this at all. It's using um zero shot voice cloning whereby I record a minute, sometimes it's three minutes depending on the on the model of them speaking, of me doing their voice.
That clones their voice. And then in the system prompt, um I'm giving some context like just to give them a little bit of bite. You know, Herman's the very goofy donkey and Corn is kind of the um adorable, affable sloth. Um and the system prompt has been the most interesting part of this because it's kind of steering them in in that way, um and then trying some differences over time with trying to mitigate behaviors that I didn't like. So they initially kind of tended to be um just very flat, very kind of, I mean it's it's a generic problem in AI, but informative but dull. And then I kind of asked it to let's let's have them disagree with each other a bit. Um and then we got into kind of corporate jargon like, I have I'm going to push back on that. So it's a constant work in progress as I'm listening to these. Um the two works in progress really were have been the TTS, trying to find a a provider that was good, reliable, and affordable.
The LLM challenge is um is is the standard one of of knowledge. Um if I'm prompting for something that demands up-to-date knowledge, that's really hard to to execute. Um so what currently in practice is Gemini 3 for having a fairly recent cutoff, I think, if I'm not mistaken, and Tavily um to kind of just buttress buttress the gaps by having that web search capability built in. There's a planning agent that when I send a prompt, the the pipeline that I've documented on the website isn't isn't is a still a bit of a simplification of what it actually involves. Um it's been put together using Claude code. But it starts with my voice prompt and now it's actually going to Gemini as audio. So it's audio multimodal, which I think is one of the most interesting modalities in generative AI. Um it just opens up so many possibilities. So they don't even, they don't actually get the transcript anymore, they hear the audio. Um and then Gemini is told to generate a script according to the format of the podcast whereby um we've got these two personalities, it's diarized.
I then added in um and if I ever sort of get some some, you know, some uh some traction behind this, people actually besides me start listening to the episodes. Um I picked up a couple of ideas from other people. Um my wife actually suggested the uh the the grumpy caller mechanism, which which I've had a lot of enjoyment with. Um Jim from Ohio. So he was added in at some point into the pipeline. And then I also added in um an attempt to get like a really spammy late night TV style advertiser called Larry who's kind of just speaks in this like overhyped uh tone and just his his instruction is to um try to go to a sponsor break where he just suggests these like ridiculous ridiculous impossible products. Uh so ultimately there's four voice clones in the production pipeline. Um and the nice part the second the V2 really of the pipeline was me was my prompt getting getting added to the episodes. The first episodes, a lot of them I've been trying to clean up some of the old ones recently and some of the first ones kind of prior to September, there was a lot of false starts with different models.
I I haven't gotten around to cleaning up all of them. but stuff like where the TTS failed. In the in the recent episodes, every time I tried a new text to speech provider, there are some kind of road bumps to figure out. Uh so that's the case with some of the recent ones where we have a few kind of audio weirdness going on in some of the episodes. Um and my take at the moment is that it would be the what's interesting about this for me is that at the moment and getting sort of my uh burning burning questions about life across to the across to the pipeline. And over time I'm probably going to actually go back on the automation a bit and start doing maybe a bit of editing of what the hosts do, but for the moment most of the energy, uh most of the effort in this has gone into getting the pipeline right. So I haven't even really got around to thinking about how I might pull down the uh the generated uh TTS script and then manually review it.
Um so at the moment, what actually happens is that my prompt and this was the pivot that I realized it's nice to kind of blend. I think it's nice to blend the human and the AI. So, uh from a certain point, I think it was V3 of the pipeline, I thought, why don't I just combine my prompt and the response? So, for anyone interested again in all the details, the TTS gets uh concatenated. It's generated in chunks. That gets grouped and then my prompt and all the intro and the outro. And it's quite amazing to me like how well this kind of can all work together. Like the imperfections of course in the TTS and and there's that, but besides those hiccups, let's say, overall this works very well. And the script actually even does an FFMPEG um normalization of my audio, a little bit of silence truncation, a little bit of EQing. So like it does, it does the whole thing. Um I haven't yet been able to figure out a good intake pipeline for like me sending in the prompts.
I've tried to get a Telegram bot working, it didn't really work. Um but I'm sure I'll get that figured out. For the moment, um I've generally done like one or two uh a day. And just today I've kind of gone on a bit of a a bit of a uh binge to give myself a lot of listening material to to look forward to over the next couple of weeks. Uh but generally it's kind of one or two here and there and I use the I use my computer to just run the the generation pipeline. So that's it. Basically, it's an experiment. It's it's something and why is it public? So when I was initially doing this, it was very much for my own own use and own own enjoyment. Um like listening back to my prompts on my own leisure and in the form of a podcast created just for me. And then I kind of thought, do any of these actually need to be private? And most of them, most of them don't, or they're not much effort to kind of just redact a little bit.
So I figured that I often with these AI things get just really interesting responses that I want to share with people and it's just easier to have that in the in the in the public domain um than it is to have it like, you know, in a Google Drive. So that's why it it was pivoted to this. And now it's just kind of there for anyone to to oogle at or listen to if they want to. Um there is a disclaimer at in the episode about it's AI, it might all be wrong. Um which I think is kind of important to have. Um so that's there, but, you know, most of the time I find the information is actually really thought-provoking and accurate and often gives me pause for thought about um the stuff I'm interested in. And frequently actually educates me, which is kind of wild to think that, you know, just sending your answers into an to an AI pipeline um can actually sort of be an education and it just teaches me about um I found some of the episodes very informative.
I was listening to the one about the Iranian nuclear or rocket program the other day, you know, INS and fascinating. This is the stuff that I that I really love and it's hard to find this kind of information um sometimes on as many podcasts as there are in the world. So that's a bit about the project, um what we're what I'm kind of working on here and it's going to just continue at the moment. Um and if you do have an ideas or want to collaborate in some manner on on this, um any new ideas for the format or how other directions this could be taken in, do feel free to drop a line. The email addresses are up there on the website, which is myweirdprompts.com. Uh the show email,
show@myweirdprompts.com and my email
public@danielrosehill.com. Thanks for listening and more episodes from Herman, Corn and the rest of the crew coming here soon.
