Automating Podcast Production with AI: Lessons Learned and Real-World Challenges

I am talking to AI.
A successful day, you don't need to know all the answers.
Just have good questions.
Chatting to AI is different from normal speech
and I hope you enjoy listening to the show
whilst getting ideas on how to hone your questions
to get the most out of AI.
My name is Paul.
The live conversations you hear are uncut
although sometimes the AI needs time to think.
In those cases, I've cut out the dead space.
What I want to do is give an overview
of what I've learnt, I guess,
over the last couple of months
around coding with AI
and maybe there's some hints and tips
if somebody's reasonably new
to coding with AI.
I will try to provide a bit of a dump
as to what I've been doing
and what some of the things I may have achieved
and some of the things that I've sort of found out
I guess through this endeavour.
Since I've been doing the podcast
I've been also obviously publishing the podcast
and earlier on I explained this process
that I was going through
to semi-automate the production of the podcast.
So there's actually a bit to creating a podcast.
You obviously have to do the audio which I'm doing now
and you have to sort of figure out
how to get all that to sound nice and work.
One of the big challenges I had was getting
recording many participants at once
including a couple of bots
and that caused severe, severe headache
and although it was a pretty interesting activity
in how to get advice from AI
and when you should be seeking AI for help I guess
and when you should be using your own brain.
Although it was helpful
I would say that in that particular issue
I probably solved the problem myself
rather than AI
and AI actually let me down a few paths
which were incorrect
and I learned a bit from that
but I've already explained that.
So what I want to focus on today
is the next level with producing the podcast
and that is to fully automate the system.
Now I'm still working on that
it's almost done
and I just sort of run through
some of the things that I guess I came across
and some of the challenges and sort of
so okay, without waffling
what am I doing?
So I'll tell you what I'm trying to do
with the process
and then I'll explain what I have done.
So what am I trying to do?
So you record the show
then what you have to do
is you then have to
create a bunch of content around the show
so you have to come up with a title
you have to come up with some summary information
about the show
and then
to host it on iTunes or whatever
you have to tag it appropriately
you have to create images
for the podcast
and a few other bits and bobs
and I've chosen to publish this
on a hosting company
and that was Lipsin
and so that company
what they do is they host the MP3 files
they have a sort of scalability
within their hosting
so that if you get a lot of people
downloading it
they manage all of those
potential issues
because MP3 files can be quite large
and it can cause issues
if you were just hosting it yourself
you can have issues with the
hosting provider not liking you
having that amount of downloads
and just slowing down
so for that reason
most people that have podcasts
they use a hosting provider
and I've also set up a website as well
because I think it's quite good to have a website
you never know
you can't really rely on platforms all the time
at least if you've got the website
you've got a website
you've always got someone where you can point people towards
so I've got a website and I have that updated
and that's using WordPress
so I've got a WordPress website
and I've started off with Lipsin
host
so what you do is
you create the MP3
and then you create the content
then you go into Lipsin
and you update all the stuff
in Lipsin
and then you go into WordPress
and you update WordPress
with all the new stuff
so there's two
there's a few parts to this but basically
once you've created the MP3 file
then once you've created the content
then that content needs to be
put into two platforms
right
so in my semi-automated
process
I'd created a prompt
in chat GPT and that prompt
would take the
well no actually
before I get to the prompt my full process
the semi-automated process was to
once I had completed
well actually
a few bits to this right
so the first problem that I had
with recording this was
because
at the beginning
at the beginning when I first did these podcasts
I was recording using some software
on my laptop
and basically
all the audio was getting recorded
into one file
and then I'd basically take
that file
and do a bit of
editing and that's my MP3
file created which is the show
I soon
found out that that wasn't a very good way
of doing it because
by doing it you end up with
normally you end up with differences in volume
between the two speakers or the three
speakers and the three people participating
in the podcast and you might
want to treat the audio differently
so the podcast that I'm being
recorded in are all recorded
through my iPhone headphones
and I like that because that means I don't have
to have a proper studio
I can just move around and I can record them
the issue with that is that
the recording quality is not as good as
is from a normal
like proper microphone you'd have
a normal audio recording
you'd have a high quality microphone
you'd have a pop guard
and that would do
some enhancing to that would prevent some problems
that you get I'm not getting into detail
and you'd also have probably
a podcasting studio that you're in
and it would be properly sound
isolated
and all that good stuff
I'm not doing any of that
I am in a room with a door
but so
when I'm talking on my microphones
a whole bunch of quality issues
that would be there
but they're not there now because
what I like to do is I like to effect
the audio
of my voice
to remove the problems
and enhance the sound
and I do that with audacity and I've got a whole macro
that does a bunch of stuff
to that and if anyone's interested
the way I came up with that macro
was I asked chat GPT to tell me
what effects I need to apply to an audio
with my iPhone
headphones to get a good sound
and I basically used that
and then made a few little tweaks
so but
what that means is if I was to apply
those effects to an audio file
with two speakers on it
me and chat GPT say
it's going to do all of those effects
to both and
that might sound great for me
but it won't sound good for chat GPT
because chat GPT's already been optimized in a bunch of ways
and then you probably end up making it sound too basic
and it just doesn't do
your good job the way any professional
sound engineer
would tell you is what
they need to do is you need to separate off
all of the different sounds
and then treat them separately
so that meant that I had to basically
multitrack record the podcasts
something that most people do
right however
the software I was using
the only way I could do that was
create a certain file type
that's an NKV type
it's also a video
type but that allows me to do
multitrack recording great
the problem was
my process
is to use audacity
to basically
what I do with that is I
I mean it's very minimal
but basically I've got an intro and an outro
I slap them on the beginning of the end
and then I normally have some nasty bits at the beginning
where I muck it all up
so I have to delete some of that
and then I sort of slap it together
sometimes I might have to do another recording
so I might record twice and slap two things together
so there's a little bit of sequencing required
and then to get to the point where
it's sequenced up
it runs alright and then I
export it as an mp3 file
however the issue with that
is I can't load in an mkv file into
audacity it doesn't accept them
there might be some way
there's normally some way it's supposed to work
but I could never get it to work
I tried to get it to work so I gave up on that
so instead I have a python script
which strips the mpv file
the mkv file
and turns it into just WAV file
so that's the standard
Windows audio file
for each
participant
so it creates four WAV files
for every mkv file
so that allows me to have four participants on the call
so that's the first thing that I had to automate
in a semi-automated manner
and I had a little chat
with the chat gpt and I got this nice
well originally yes
originally I was doing this
with
I tried it with chat gpt
didn't really work I tried it with another
service and it
did work but
the issue you've got is that you're going to have to start
paying for that
because that's actually quite a intensive
operation transcribing
mp3
no sorry I'm getting confused sorry I'm not talking about that
I'm talking about splitting the files
sorry
yeah no sorry forget
the last 30 seconds
basically so I've got the mp
mkv files
and I've got a chat gpt to get a python script
to
split those
so I've got that running on my
Mac
and the python script now
I just give it the URL
for where it is on my computer
the path file
not the URL
and then it consumes that
and voila
it spits out the WAV files
and then I take those WAV files and then I use those
to sequence it up
in my audacity which then creates
the mp3 file
and then the mp3 file
is
what I then use
for all of the content generation
moving
after this so this is quite clever
but
it's not me really
it's not me being clever
it's chat gpt being clever
or grok I tried both
but I ended up with chat gpt at the end
so basically what I then did in my
semi automated process
is I then
created a prompt
took me quite a long time to create this prompt
quite a lot of toing and froing with chat gpt
to get this to work
and it's a multi step prompt
so
you
I sort of
sometimes it's wise to have this notion
of steps
and you explicitly say step 1 do this
step 2 do this, step 3 do this
and this was required
because
I basically wanted so the process that I wanted
it to do was to take
my transcription
that I've done
on my Mac
and then load that
into chat gpt
so you insert it as a file
you can do that
and then the prompt says
look at that transcription
and what I want you to do is I want you to create
a title for my podcast and that has to be
SEO optimised
and it has to be intriguing
and a few other things around the prompt
there to make the content interesting
and then I want you to create a summary
of the post like a few paragraphs
explaining what this
particular show is about
and then
I want you to create
oh yes and I want you to create two
different versions of that summary one is
for Libsyn so that goes into iTunes
and they don't have as much space as they have
on WordPress
and WordPress needs to be more
SEO optimised as well
than the
iTunes
text but I also want
there's also stuff about the feel
and I think I quite like the
text that's come out
and then
I also get it to create a list
of tags for both Libsyn and
WordPress
and create
two different images
one is a square image which is used for
iTunes the other is a rectangular image
which is used for
some of these
I think iTunes and Spotify I think they do use
the rectangular image in some
circumstances depending on your screen
that you're listening to the music on
and it's also used on WordPress
so WordPress is a rectangular image
the way I've set it up
so they are the things
that it creates
and in the semi-automatic
process it had to be
a step-by-step process because
you can't just ask
chatGBT to do all that in one go
and the reason for that is
that images
when chatGBT creates images
for you the way it works
is chatGBT
talks to Dali
before it talks to Dali chatGBT
generates a prompt for Dali
and then it
asks Dali via the prompt
to create your
image for you
now Dali has got certain rules that it runs
by and one of those rules is
it won't release any images
it won't release more
than one image at a time
and after every image that it releases
if it's being communicated
to via the chatGBT
chat window
it will ask for your approval
you can't get around that
and the issue one of the big
things that was causing me a lot of
grief was that chatGBT
was not aware of this limitation
really with
Dali
so it thinks oh this is a great
prompt yeah great
I'll go ahead and do that absolutely
and then it goes ahead and tries to do it
and you end up with this kind of half hang
it just crashes
and chatGBT thinks it's doing a great
job but it's not it's doing a rubbish job
like my prompt has gone
halfway through and just hung
so I had to create these steps
and these steps have to explicitly
explain that
Dali will then ask for an approval
and then you have to put in the text
do you approve this
so I had to actually type
in my prompt the prompt
that it was going to give me to approve
the file and then I had to
do it again for the second one
and then it would then create all the other images
so that was
one reason why I had to structure it with these
with these steps
but the other thing I wanted to do
was basically
in an ideal word I just wanted to
output a zip file with a bunch of files
in it
and that turned out to be
truly a pain in the butt
and although
you can get chatGBT
to do this you can get chatGBT
to give you a file with a file name
I've done it it's happened to me
but
most of the time it doesn't work
like if you want it
and a lot of the time I don't know if you've ever had this
where it gives you a link and the link
doesn't resolve
because of some kind of internal conflict
there's something going wrong there
but it doesn't know that it's going wrong
and this is when chatGBT can be super frustrating
but the thing that I
realised with the files
was that really the chat window
isn't really set up for sending you
files
but you can do it
but it's not
reliable
so that's what caused me to start
thinking along the lines
of the next evolution of this process
so I did get this process to work
so this process was working
it wouldn't
download the files
but it would
create
the files in text so I ended up
at the end of it I ended up with a few
with
all of the tags
the titles were all in the chat window
and I ended up with a couple of images
in the chat window as well
pretty well actually and the images came
out pretty much like I wanted them
to come out
but I wanted to
get this to the next level
I basically wanted to get it to the point where I could
just sort of fire and forget
my dream with this
is to do a podcast, record the audio
then I go over
to my computer and I run
I maybe copy the URL
for the MP3
dump it in
a
well I'm using Jupiter Notebook
to get this started just because it's easier
but just dump it as a parameter
for a function
and then
submit that so that it starts working
and then just go away
have a cup of tea, have my lunch
and bang
I've updated Libsyn and WordPress
without anything at all
that was my fantasy
and I think
I'm getting quite close actually
so
in order to do that I had to fully
automate the process and the process was just
half automated
so what I realized is
and one of the things
so there were two problems that I had to solve
I had to solve this sort of issue with
photos being
requiring this sort of communication
and I also had to solve this problem
of being able to get physical files
because I can't do it
I can't do any automation
if I'm having to go to a chat window
to pull stuff out
so
pretty rapidly realized
that I was going to have to use the API
so
I'm now using the API
for chatGPT
I tried using the API
for GROC
and I'm also using the API
for WordPress
and
I tried using the API for Libsyn
but there isn't actually an API for Libsyn
so that meant that I had to actually change
podcasting providers
to a different one called Transistor
where they do have an API
so
what is an API you might ask
well yes
I do have this habit of getting rather into the weeds
but what is an API
an API is basically a way of
moving data into an app
over the internet
you call what is called an endpoint
which is just a URL
and you can
depending on the way this is configured
you can then sort of read data
from that endpoint or you can update data
or you can change data
from that endpoint and that requires
you to have
some security to allow you to do
that
and that is called a token
and I won't get into all of that
but basically that's what an API is
it allows you to write some software
that then communicates with all the software
and brings back data
an API is exactly what is happening
when ChatGBT
is trying to answer your question
and the answer is not
within
its data
that it's been taught on
so
I've explained this before
but ChatGBT
they run the model
in 12 months and this is when a new
version of ChatGBT comes out
and what that does is
you basically get the model
you've made some changes to the model
then you train the model
and it pops out
now I've fully formed
of ChatGBT version 7
and
so there's a whole load of
stuff in that model
and so you can get quick answers
most of the time and it just looks at the model
and answers your question
however if you're asking questions which rely on
more timely data than what it was actually
modelled on
asking about something that happened yesterday
and the last time the model was created was 6 months ago
it's going to have to go
and find that data
and it does that with an API
so it's using APIs too
so sometimes when you see ChatGBT
and it's like you're waiting for it to do something
it's not actually
doing the work
it's created an API call
which is what it's called where it's asking
some other piece of software for some information
it's waiting for that
and that's what an API call is
so anyway I'm going to go into the world of
APIs now with my automation
and this is going to get me much closer
to my dream
however there is one thing that I still need
to do in this automation
which is interactive
and that is approve the titles
because sometimes I was finding that the titles
that ChatGBT would come up with
they weren't really in line with what I wanted
sometimes when I'm doing these episodes
say
I've got the philosophy
miniseries
kind of thing
I'm on philosophy number 4
and it's on such and such
so it'll come up with a title
which will be based on the content
so it'll be completely not thinking about that
so sometimes I'll just
change the episodes
and sometimes it comes up with terrible ideas
so I just say well have another go
that was rubbish
so what I've done now
is I've got a new process set up
and that's using
APIs
and it's fully automated
now
I went through a bit of a process
here as well so I wanted
to
automate this
and
involved in writing prompts
for ChatGBT
you start to realise that things can get to rather messy
especially when they get long
and I was wanting
a way that I could
in programming
you have this concept of modularity
where
you split things down
into more manageable chunks
and I wanted to get
a bit of that going
in the code
a lot of my process is not
actually ChatGBT
reliant
all this calling APIs
well that's nothing to do with ChatGBT
that's just code
although ChatGBT is helping to write the code
just makes it quicker
but yeah
so I've got this sort of requirement
for some code-based stuff
and also some ChatGBT
or GROC or LLN
based stuff
and I want to get some more modularity
into the whole thing
to make it a bit tidier
so originally
I looked into something called N8N
and
it's like a lot of these graphical interfaces
look super cool
you basically and this would be a good thing
for a lot of people I think
if you're wanting to create some kind of agent
that does something because this is in effect what I'm doing
is creating an agent
then you use this N8N
and
I'm quite familiar with data processing
I've
done it as a career
in the past
so it reminds me a lot of these
sort of data processing workflows
basically what you do is you have some kind of trigger
so that could be someone in a chat window
or it could be
a trigger caused by
like a date coming up
or somebody entering something on a form
on a website or something like that
and then this trigger
and then the trigger then
starts a cascade of things happening
so you know
you could type some things into a web portal
and then what happens is
I don't know maybe you chat
maybe you chat in something that you would
you would like a PowerPoint
I want my PowerPoint to do blah blah blah
and then the first thing is it goes to chatgbt
ask chatgbt to create
a
a list of the slides
and then it goes to
maybe
maybe you then take that list
and then update a Google Doc spreadsheet
with that list
and then you could do something
whatever you can do
in this N8N
you can put things in a Dropbox
it's got all these different
it has all these
API connections already sort of embedded
into it so you can
it makes it more easy to use
and then it's a sort of graphical interface
if you want it
so it looked pretty cool
what I didn't like about it
was the fact that it was a graphical interface
basically and I'm finding that
the more I work with
well the more coding I do
the less I like graphical interfaces
and especially
when I'm doing chatgbt
stuff and doing code
with AI it's a lot easier
to just get the full code
and dump that somewhere
than to be worried about
getting bits and bobs and updating
drop downs and things like that
it just makes it easier
I know that there is a way of uploading
JSONs into N8N
but I didn't much like that
I would have been more happy
with a complete code solution
and then what really
did it was that
the code that you use in N8N is
JavaScript look and I'm quite
happy with JavaScript
but I think if I'm going to be doing data stuff
I'm better off with Python
that's really what it's
for
people would dispute that
but I think
especially if I'm moving around files
on my computer
and stuff like that
Python is just really the thing
that I feel I should be using
and I think it's a more appropriate tool for the job
so I started looking for
works that use Python
that also work with
LLNs and I came up again
I came up with something and it's called
CrewAI
so my system now is using
CrewAI
and that system
basically
when I
there's a few manual
processes at the beginning because obviously
I've got to
I've got to do the recording
manually
then I still
using this semi-automated process
which basically just takes the
NKV file and creates the WAV files
but it doesn't take long and it just runs on my computer
so I'm doing
that and then I have to create
the show manually
again so
these activities probably take me about 5 minutes
to be honest 10 minutes
they can take me an hour or so
if the podcast is requiring
a bit of finessing
but they don't necessarily take me very long
and then I get to the point
where I've got a finished
transcription
I've got a finished MP3 file
and a finished transcription file
and once I've got those in the folder
then I have this
new process which is running in Python
all the code is running in Python
and the semi-automated stuff as well
but this is like a sort of app that then runs
in Python
and what that does is it then
takes in the
URL of the
transcription and the MP3 file
and
loads them all in
and creates
all of the content
while it's creating the content at the very beginning
it gives me the option to approve the titles
or disprove the titles
basically I can
yes, no or put in my own title
and it will do this
so I can
submit maybe
20
podcasts
but let's say 5
I tend to work on a batch process
so I can submit 5 URLs
for MP3s
and all of these transcriptions are in the same folder
so it finds them anyway
and then it takes those
and creates the title
etc etc I approve it
or whatever
and then it then goes ahead and creates all the content
puts that content into some JSON files
stores them on the
computer
and then uses those JSON files
to update the APIs in WordPress
and Libsyn
so
my aim with this is, my feeling is
if I can get this podcast to the point where
I am not overwhelmed
by the administration
then I will continue to do it
I am a curious person
I am always talking to chat GBT
I am always trying to find out answers to questions
and I love it
so I can't see any problem for me doing that
but
the manual stuff of actually loading things in
and all that kind of stuff
makes me want to stick pins in my eyes
so if I cannot stick pins in my eyes
that's what I want because that means that I will actually continue to do this
so
and I think I am very close
so
are there any other key learnings
one of the things I can tell you that I sort of
I mean
I mean I have been coding
I have been coding for the last couple of years
so I have been using AI
to assist
but a few things that have cropped up recently
I mean I am not using, so for example
I am not using
I am using Visual Studio Code
but I am not using the AI agent with the Visual Studio Code
I just don't like it there
I would rather only use it when I need to use it
but if you do that
and you don't have something that is constantly
looking at your actual repo
with your code
the difficulty you can have is
it doesn't really appreciate
you are asking it a question
but it doesn't have all of the context of everything
that is in your repo to understand exactly
how it all works
so one of the classic problems that I get
is
especially with Python at the moment because I haven't installed
any tools in Python that will
correct the code
if I was to change file names
and folders and things like that
because when I am working with JavaScript
I have got plugins
which basically say
you have just changed this file over here
which is imported in this file over here
I will just change the imports automatically for you
because it is a real pain
especially if you are refactoring your code
so that was all done for me in JavaScript
in Python
it is not done for me
it is just because I haven't got round to figuring out
I wouldn't even say figuring out
I haven't even had time to think about it
I have been prioritising other things
but that hasn't really been a problem so far
because the Python stuff isn't
it is quite flat
there is not a whole load of depth to the code
and
one thing I figured out
is that you can actually just zip up an entire
directory or subdirectory
attach it to chatgbt
and say here is my code
it is good for you to fix this
and then it will go ahead and do that
and it will change all of the interrelationships
with everything
so that is super good
especially for refactoring
if you want to refactor one of the things I have done a bit
today is just
simplify
simplify the code quite significantly
and in order to do that
that is actually a really good use
I mean I have done it before
but not in that way
not sending it
so it is not that irrelevant
but I just thought it comes to mind
something I found useful today
but
yeah so I hope
that explanation explains
how I am using AI
and
some of the
things I suppose
reflection
now after working
on this podcast
and automating this process
because before
when I was programming I wasn't
I wasn't coding with
I was using AI to help me code
but I wasn't actually using AI
to solve a problem with the app I was working on
which I am doing now
yeah I guess
so
oh yes I remember
yeah so my point
so after doing all of this
it does make me think
about the use case
for AI in the workplace
and how
the challenges
and
how things are going to pan out I guess
and it sort of reinforces my thoughts
that I have been having in some of these other podcasts
about where I think it is all going
I don't think
I think ultimately
programmers are going to become obsolete
but I don't think
the short term thing
I think in some ways
they are
but
I think there are other occupations
that are going to be more obsolete
more quick
and in the short term I think there is actually going to be
an increased use of people
with technical
programming skills
but they won't be programming
in the same way as they were before
and they will be
required to
recode people's jobs
I mean the obvious
thing that I am sure most
big companies are doing at the moment
is using AI
to
follow people around
if you have got various
for process mapping
but you can get
documentation
and once you have got the
documentation then you can take it to the next
stage and think well how can we
are there
you can do analysis on that
documentation you can say well
out of all these processes
strategically can you look at it now
and tell me
is all of this required
first question
and then you
obviously can remove quite a bit
just like a tangled web of code
you look at all these process maps
and you
not you, AI does
and figures out what is redundant
what is duplicated
so that is the first step
and then once you have done that
you can reorganize according to that
and then the second step will be
to well what is automatable
what is easily automatable
and then you are going to need people to create
those automations and I would argue
that when you start getting
to the point where you are actually
automating stuff in a meaningful way
then
I don't think people
general people
will
they will have the skills
to do something but
I think there will be a difference
because it really
is programming at the end of the day
but it is just very high level
it is very high level programming but
the power that you have is very high as well
so there is a
focus on being
what would the word be
disciplined in your thinking
if you go off
ad hoc without
some disciplined thought about what you are trying to achieve
you are likely to get
in a mess and I think a lot of
you know the more I do this the more I see
I know what the next thing in AI is going to be
is going to be big
corporate companies saying oh we tried AI
and we made a complete mess
and
and I think that is going to be the case
and there will be successful
companies that manage to focus
and work out how to not make a
complete mess and to take advantage
and there will be other companies
which will just be I think the majority
of companies will be saying yeah we are doing AI stuff
oh yeah we are really clever
we are doing AI stuff and they will be saying that
because they are worried that if they don't say that
the share price will go down
and then
but they won't be
they will do it half arcing it
and probably
they will either be overly cautious or they will be making
a great big mess one or the other
so
I think that is going to be the big challenge
and that is going to require people with the technical
ability
to coordinate these activities
and work with AI
to
get what you want
in these companies
so I am not particularly concerned being a developer
I think
there is a lot of change and it is a good time to
understand AI for sure
you know if you are
if you are sort of not flexible
obviously that is going to be a problem
and there is a lot of jobs that won't exist
you know
I mean obviously call centre jobs and things like that
they can probably be done better with AI
and there is tons like that as well
marketing jobs
it is scary how good it can do marketing
anyway I think I am waffling now
so I hope you found that interesting
I am happy to
answer any questions about
what I have done
I am no expert I am just learning AI stuff
as I go
and I hope
it is difficult like talking about
code without actually having seen it
and you know people will
know different things about code
so I turned myself into
a
code tutor
but I hope I didn't go
I hope I kept it quite high level
and you will understand what I was talking about
maybe I should have done it with chatubt
could have explained a few concepts better maybe
but anyway I have done it now
so I hope you enjoyed this
and speak to you next time, goodbye
talking to ai.show