Overcoming AI Challenges in Podcasting

John Barker discusses his challenges in using AI for podcasting, specifically for repetitive tasks and workflows in the 'Titan of Tech' podcast. John mentions the use of Descript for text-based video editing and transcription. However, he conveys his struggle with the time-consuming nature of the editing process and the development of show notes, YouTube descriptions, and chapter timestamps. Despite improving efficiency with AI tools like ChatGPT and Descript, the John warns against wholly trusting AI-generated results due to issues with repeatability, interpretation of commands by the software, and the AI picking up wrong or minor topics as key points from a conversation. John underlines the importance of verifying the results generated by the AI.

00:00 Introduction to AI Challenges in Repeatable Tasks

00:23 The Journey of Podcasting and AI Tools

01:11 The Struggles of Editing and AI Assistance

02:25 The Evolution of AI Tools in Podcasting

03:04 Building a Standard Operating Procedure with AI

03:45 The Unpredictability of AI Results

06:32 Exploring Descript's AI Tools

09:10 The Misinterpretation of AI Summaries

10:03 The Need for Verification in AI Results

11:06 Conclusion: The Importance of Verifying AI Outputs

I got three things I want to go over with you today regarding some

problems I've run into with AI and doing repeatable tasks or kind of

repeatable workflows if you're trying to build standard operating procedure.

And this has to do with me using AI to help with editing the video and audio

versions of the podcast, Titan of Tech.

So, uh, brief backstory.

This is my second go at a podcast.

I had the Business Samurai about two years ago where I did 45 to 50 episodes.

And I was using a few of the same tools.

I've got one new podcast platform, but I was using Descript essentially

as a text based video editor.

No interest in becoming a video editing person.

Descript was super easy to use because it would come back and give you a

transcription of your entire conversation.

You just threw the video file into it.

It would do the transcription automatically.

It could detect the person that was speaking and you could give it a name.

So if it, It goes, Hey, who is this voice?

And I go, Hey, it's John.

And I put my name in there and it would know it was me the rest of the time.

That was my voice detected.

And I could do that for however many people were a part of the conversation.

Super cool.

So that was obviously a pretty good time saver.

But one of the things that was driving me nuts because as much as I

love prepping for the conversations and having the conversations with

everyone, I find it to be fun.

The editing process is very time consuming.

It's arduous.

It's draining for me.

Um, and unless you're somebody that's got a well, well monetized

podcast where you've got somebody that can help you with the producing

aspect and the posting aspect.

Most people that are doing this are doing this themselves.

And so when it comes to doing things like.

Hey, I need show notes to upload with the podcast.

I need a YouTube description, or I would like to build in the chapters.

You're seeing this a little bit more with, with, um, some YouTube and

with some podcasts where you've got your timestamps that are in there.

And to do that is incredibly time consuming.

I mean, it was great.

I had the transcript in front of me, so it wasn't like I had to keep

listening to a podcast because the script helped me with the transcript,

but it was one of those things of.

Okay, I've got a few of these to do.

What was the ones we talked about?

What were the key highlights?

You know, I've got my notes from the conversation.

I was taking notes during the conversation for follow ups, things

of that nature, but it was still very, very time consuming, boom, jump forward

now to, you know, 2023 and 2024.

As I, as I'm going through this process, you've got chat GPT and also with

the script AI tools baked in to help with this process and while it has

been an incredible time saver and the problems I'm going to tell you about.

You still got to be very careful with the results that you get, and

you still, uh, are going to have to look at it with a critical eye.

Don't just take it verbatim that it's going to come back the

way you need it to come back.

So let's jump into this.

Over the past week, as I've went through the first couple, uh, and

starting to build out how I need to go edit these, what is the information

I need to include for YouTube?

What's the information that I need to include for Captivate, which is

the podcast publishing platform?

I wanted to build a standard operating procedure, no different than I would

have any other IT department or any other business that I work with.

Start creating.

For tasks that are very, very repeatable, you're not necessarily going to do this

on your first go around or something that you're going to do once every five

years or something like that, unless it's very involved process, but for

something that you're going, man, we're going to have to do this, you know,

once a week, twice a week, whatever, you know, that time is by your third

time, you're starting to get a pattern.

You're starting to get a rhythm.

So one of the things that I did was just launch a word document and said,

all right, what are the exact steps?

What is the exact information that I want to generate?

with the podcast and using ChatGPT and Descript.

So, uh, the first problem that I ran into were the prompts and the process

with ChatGPT were not repeatable.

They were just absolutely not repeatable.

So what do I mean by that?

My first step was, give me, export the transcript.

If you look at Descript, it will either, you can copy and

paste the entire transcript.

Uh, after it goes through the process, or you could export it to a text file.

I have the paid version of chat GPT.

So I was tell it to here's, uh, I had maybe a plug in for it

and it was attached the file.

Now give me some good hooks as a YouTube expert like Mr Beast.

And what would happen was.

It would come back and tell me it couldn't read the file, or it couldn't,

uh, the copy and paste that I did was too long, and I found that to be crazy

because I had just done it two other times, uh, with two other, uh, different

videos in a, you know, obviously starting a new, uh, conversation.

And with zero issues with this, but on my third one, as I'm starting

to actually build a workflow with this, it doesn't work.

So that took a little bit of brute force to go through this.

Maybe by the third time it would work.

Maybe by the fifth time it'll work.

I would just have to keep trying over and over and over again or start

a new conversation to keep trying.

And eventually it worked, which I found to be crazy that it was

either coming back and telling me it couldn't read the file or that the.

Text was too long.

If I decided just to try to copy and paste, so that was one of the things

with the prompts not being repeatable.

Also, the results that you got were not repeatable.

So I had the first few times I did this, you know, obviously you've got

your, your conversations are saved on the left hand column and chat

GPT that you can go back and review.

So the first, the first couple of times I did this, the results were fantastic.

Okay.

The summaries were absolutely on point.

The time stamps look good.

Uh, the key, the key highlights from each of the, the, the main sections

where it was like, Hey man, this was a key, you know, key highlight one, key

highlight two, key highlight three.

Um, for those conversations, I knew we're on point.

Um, and when I would use that exact same prompt, I copy and pasted.

That same prompt that I used with a new workflow, it would come back with garbage.

Uh, it, it just wouldn't, it wouldn't work.

It would come back with different formats for some of the stuff.

And it was just like, this is, this is kind of crazy, um, that I can't just

use the same prompt over and over and over again and get repeated results.

Even though the first two came back.

And we're pretty similar, similar enough where I didn't

know too much of the difference.

Why now is the third one doing something different?

So I found that to be a little crazy.

So after I started running into a little bit of those headaches with ChatGPT itself

on the, uh, I think the third and fourth podcast, uh, and video that I was editing,

I said, let me try Descript's tools.

So no surprise now.

We've got a situation where every software tool under the sun is doing some sort

of AI into it, whether they're building their own and turn to the platform or

they're calling back using an API with a chat GPT and kind of baking it in.

So it looks kind of seamless, you know, they're trying to build their own.

Chat, quote unquote versions with inside their tools that are very specific

to the workflow of that software.

So if you go into the script and they've got an action bar, and it's the 1 of the

drop downs is actually called ask a I.

And so under ask a I.

You can tell it to add chapters, you can say, summarize, suggest

titles, social posts, blog posts.

So I went through and I said, Hey, give me social posts.

Like I want to use this and I have used this already for like LinkedIn or for

Instagram or something along that line.

And I'm like, Hey, these are, these are pretty good, you know, summaries.

Maybe instead of using chat GPT, I'll, I'll start using Descripts.

So then I went and said, all right, give me a summary.

It was either give me a summary, give me show notes, doesn't matter, like similar

in vain instead of actually giving me the results of of the summary of what

the conversation was about, it would come back and give me a summary of what the A.

I.

Is supposed to do on the back end to actually create a summary and it did

this multiple times in a row and it was just one of those things like it didn't

know how to interpret its own command.

And these weren't things that I modified.

I know you can modify the command.

You can type your own.

You don't have to use their pre baked ones, um, that they've built in there.

But I said, let me use what they've created themselves.

I know it's in a beta version, but still.

So I would just click the button.

Whatever it came back with is their preferred prompt.

I would use it.

And even it was doing something that was a little bit different each time.

And that was from the software makers themselves.

So, after I get through all this, I start force feeding, uh, force feeding,

uh, forcing my way through, some of these issues and basically just doing

it over and over and over again until I got a result that I was looking for.

Unless it was a situation where it, you know, it was just, uh.

You know, it's pretty close.

You can't take the results exactly.

You've got to go tweak some of the stuff, depending on what

you're asking for it to get back.

, and then it became a situation of getting a prompt and getting

what looked to be good results.

And this is the third problem with this.

But it would pick the wrong topic is like the key.

The primary driver of a conversation I had this happen on two different occasions

for the same conversation that I had ironically enough about project management

AI, where we had talked about something that was maybe 30 or 60 seconds toward

the very, very end of the conversation.

And when I said, go give me a summary, and this was in a chat, I

believe it was picking that little 30 or 60 second clip out of an hour

long conversation and making it.

The priority it was making it.

If you looked at the summary, you would have thought that's the whole thing

that we talked about for an hour versus.

Quite frankly, something that shouldn't have even made the summary, maybe

it would have made the chapter as being its own little chapter, but

nothing that would have made a summary of anything that we talked about.

So, again, you've got to verify the results of these things again.

This has been way, way faster, but not as smooth as I hoped it would be, um,

not as something where I don't know if I need to go take the good results.

I want it that I have gotten in some of the other ones.

Instead of just using that same prompt.

Then start building that prompt out even more detailed and say, I want you

to give me the, the, the show notes, but I want it in this format with these

key highlights to make it more match.

Maybe that will help.

I have to spend some more time doing that, but it definitely wasn't just because

I got good results this one time that I'll get a good results the next time.

So, again, the 3 problems I had where the prompts were not repeatable,

sometimes the, the built in things within the software themselves

that were supposed to be specific.

To video editing were giving me weird like their instructions were

just being repeated and I have to do that over and over again and

others were if you had long files.

Uh, long conversations, it would just pick up a thread of something very minor

and blow it up into something super big and make the whole summary about that.

And so the reason I bring this up is I know there are a lot of people that

are using, , AI, whether it's chat GPT, you're building your own, , language

models internally to sit there and help with proposal writing, writing grants.

, maybe building your own workflows that you still have to verify

the results that you get.

Um, I do know when it comes to, the prompt writing when it comes to creating

images such as mid journey or dolly that just because I used a prompt one

time and got pretty cool results by copying pasted the same one, it's going

to give me something totally different.

I think what shocked me or not necessarily shocked me, but I surprised a little bit.

Was I was feeding chat GPT a lot of the information for it to work itself

off of that, you know, I'm giving it this gigantic transcript of information

already to base the results off of, and it would just give me many, many

different results, um, in different formats or even just the wrong stuff.

So always verify use this stuff.

Be cautious to make sure that you're not posting inaccurate information and.

Okay.

Enjoy have the have the time savings that are out there with these tools

that were not there from before, but you may have to do it 456 times.