It’s been awhile, hasn’t it? Long time, no see.
About 3 weeks actually. Normally, I try to get this out every 2-ish, so a little behind.
But for good reason. I’ve been spending most of my free time building.
I don’t know if you noticed or not, but there have been some pretty cool things coming out, AI-wise. ChatGPT seems to be the most widely noticed, but there’s a piece to that most people missed.
And that’s text-davinci-003, the model that ChatGPT is a fine-tune of.
This was a huge jump in quality from text-davinci-002. In fact, it made my life way easier for a few different tasks and that led me to start building out some cool things.
First of all, I was playing around with analyzing Twitter accounts:
Just over 6500 impressions for that tweet, which was pretty nice, considering I normally sit around 200. Just proves out my philosophy of offering functionality for free before the product gets built, because this is a great way to test the process and see what interest is like.
I actually started building that script with text-davinci-002 as the model and by simply switching it to text-davinci-003, the quality improved a ton. I was able to get it perform some extra analysis in addition to just breaking out the summary.
The result: Who Followed Me?
I’m currently onboarding users as beta testers, so if you want to check it out, go ahead and sign up. I’m offering the first 100 users a free year of premium access.
The premise: learn more about the people who follow you and decide whether or not to follow them back. I realized that I was losing a lot of people in the influx of bots/scammers following me and I wanted an easy way to make sure I didn’t miss any. So this tool scans accounts and sends a weekly digest of new followers. I’m planning on adding some additional functionality too, but this is the initial use case I’m covering. During the beta period, reports will come out on a daily basis, but after that, I’ll move the digest to weekly.
And the first day of testing had me hit my email limits because I messed up the cron expression to run my Celery tasks.
Whoops, ended up sending a ton of emails to a couple users before I hit my limit of 100 emails per day. Glad I was on the free plan instead of picking up all sorts of charges before I noticed it. But got that fixed this morning and will try again for real tomorrow. Also planning on putting in some safeguards to only send a given report once. Live and learn!
Buildspace Results
In case you missed the last issue, I ended up joining a Buildspace build for a GPT-3 writing assistant. That was really cool. I ended up not only completing the build in 2 days, but there was also an opportunity to join a community of builders called Nights & Weekends. It’s kinda like an accelerator, with community and events and a meetup in San Fransisco at the end of the 6 weeks. I enjoyed the building of the webapp and Chrome browser extension, so I decided to apply to the program.
And found out on Saturday that I was accepted!
There’s a part of the community that’s my favorite: the push to GTFOL.
That stands for “Get The Fuck Off Localhost”.
This is something that it took me so long to learn. It’s the hardest step to make, but once you do, you become a bit addicted to the rush. That’s by far the fastest way to iterate and get feedback, just launching and seeing what happens.
So let’s do it! Time to ship, ship, ship!
What did I end up building? Not Who Followed Me?, that was a separate project. Instead, I focused on another pain point that I’m facing: Podcast Show Notes.
You can check out the super basic version here. Drop in a transcript and get a summary of the episode. I’ve got a couple of other things I can do with it as well, my goal is to make it possible to put together the show notes for an episode in less than 5 minutes. It doesn’t work for super long episodes at this point, but I’ve got a few ways I plan on dealing with that.
And that’s going to be the starting point for the Ben’s Bites Hackathon that I decided to join!
For those who aren’t familiar, Ben’s Bites is a daily newsletter that sends out links to new AI tools that are hitting the market and all sorts of extra content.
I’ve used it to discover several new tools over the past week or two since I’ve been a member. Feel free to use my referral link if you sign up. If at least one person does, I’ll get access to his database of AI tools. So this week, I’ll be building out the tool and tweeting out some progress. If you know any podcasters who might be interested in trying out the tool, let me know! Also, looking for connections to producers at bigger podcast networks because after some conversations last week, that seems like the best possible approach to making this both profitable and accessible to users.
GPTChat + Midjourney
I’ve been using GPTChat to help generate some better prompts for Midjourney. I wanted to create a couple of scenes that featured time travelers visiting different types of futures, a dystopian one and a more utopian one.
So the first one I tried ended up coming out like a movie poster.
So I told GPTChat that it created movie poster-esque images and I wanted a more realistic scene. It explained the reasoning that happened and suggested that I rewrite the prompt in a certain way and provided an example.
The result:
Much better. Then I asked it to rewrite that prompt but as a more utopian scene. It gave me the movie poster effect again, so I again asked it to make it realistic.
The rewritten version:
Way better results than I get when I use my very simple prompts. Definitely going to keep doing it this way.
Transformations in Midjourney
Another thing I’ve been having fun with is image transformations. Based on an idea of a coworker, I decided to turn my dog into a Pixar character.
The original image:
As a Pixar character:
The process:
I set up a private Discord server that I could upload images to. That made it easy to upload images for free and grab a link to it. Then you can use the /imagine prompt with <link to image> in the style of a Pixar character, cute, adorable --no glasses --v 4
Note: it likes to add glasses to characters if you don’t add the —no glasses. And if you don’t add cute and adorable, it tends to make the faces sad. Probably a bunch of things in there to think about, but not going to worry about that today.
Me as a Pixar character:
Finally, I’ll end with me as a Marvel villain.
Not bad eh?
Thanks for reading! Stay tuned for updates!