Welcome to the Woof Algorithm

This is a podcast about dogs. And it’s completely made by AI, artificial intelligence. It was built to explore the current capabilities of the available tools, to see what an average person could do with minimal effort and cost.

Everything in this show was made by AI, from the tile art to the plan for each episode, to the meta data and the actual episode audio itself, to all the individual posts on this site for each episode (except for this page). No edits have been made to any podcast episodes; they are as the AI tools spat out.

Here’s the overall process used to make this podcast:

I began experimenting with ChatGPT and NotebookLM to fiddle with the best potential outputs. I settled on a podcast about dogs, because there is a ton of readily available information about dogs online, and the current crop of AI tools have been well-trained on that data. It’s also not very timely, so there’s no breaking news and far less chance of hallucinations creating erroneous content.

I asked ChatGPT to come up with a title for a podcast about dogs, made by AI. The top three suggestions were: The Woof Algorithm; Bark and Bytes, and Paws & Protocols. I chose The Woof Algorithm.

Then I asked ChatGPT to come up with 13 episodes, each focused on a popular breed. And to generate an episode title and description, and some marketing copy for each. Those are the episodes you see, and I was surprised when ChatGPT suggested the 13th episode be a special one about the history of humans and dogs.

Once I had the breeds, I turned to ChatGPT. I wanted a detailed prompt that would generate the source material for each breed. I turned to Chat GPT Playground, where you (as a paying subscriber) can input a simple prompt and turn it into detailed machine instructions. I started with “I’d like to make a background document for a podcast episode featuring a specific dog breed. I want to thoroughly document the breed’s history, main traits, suitable owner profile, and other interesting or humorous facts.” I entered that into the generator, and it gave me this detailed prompt:

Now that’s a prompt! I put that into ChatGPT, and specified the breed in the first line only. Out came a beautiful background document. Here’s an example of what was generated for Chihuahuas. Here’s the one for Bernese Mountain Dog. This was exactly what I was hoping to get.

The next step was to make a new notebook in NotebookLM. I uploaded a PDF of the background document, and then simply hit “Generate Audio” in the Notebook Guide. In just a few minutes, the audio was made. The first draft of the audio was quite good; however, the AI hosts kept referring to “the source”, and in some cases even said something like, “according to chihuahua dot pdf,” which obviously wasn’t going to work.

I wanted to add another source, hoping the hosts would say “sources”. I went back to ChatGPT Playground and asked for a prompt that would generate 10 interesting facts about a specific dog breed, with sources. This generated the following machine instruction prompt:

List 10 interesting facts about [Breed]. Each fact should be a few sentences long and include a source citation.

# Steps
1. Research and identify interesting facts about [Breed] using reputable sources.
2. Compose each fact in a few comprehensive sentences.
3. Provide a source citation for each fact. Ensure sources are reliable and credible.
4. Organize the facts into a clear and cohesive document.

# Output Format
- Each fact should be presented in a paragraph format, consisting of a few sentences.
- At the end of each fact, include a citation in the following format: (Source: [Author/Organization], [Publication Year], [Title/Link])

# Examples
Example Fact 1: [Breed] are known for their friendly and tolerant attitudes, making them excellent family pets. They are also intelligent and versatile working dogs. (Source: American Kennel Club, 2021, [Link to source])

(Note: Provide a real citation for completeness in actual implementation)

# Notes
- Make sure that each fact is original and not redundant.
- Double-check the credibility of your sources to ensure accuracy.

Plopping this into ChatGPT and specifying a breed up top gave another great output document. Here’s an example for the Poodle Facts, and here’s the Siberian Husky Facts. So now, inside each Notebook, I had a background document on the breed and a document with 10 interesting facts. When generating the audio around these, the output was much more reliable. There’s the occasional reference to “sources”, but that’s fine.

Once I had these prompts, it was simple to generate the background documents and facts for each breed. It only took a few minutes each. I scanned them to make sure they looked good, and occasionally verified things externally. It all looked quite solid.

Then I set up a Notebook for each breed in NotebookLM, and generated the audio. It was really quite impressive — in just a few minutes, the episode was created. There were some real limitations, though. I tried a few edits in the documents to get the tool understand this was a sequential podcast with multiple episodes, but those worked poorly. The best outcome was with the straight background and fact documents, and to just let it fly.

Midway through the breeds, Google did release an update that allowed for custom instructions when generating audio. This was a great addition. After some experimentation, I landed on these instructions:

• Don’t mention “the sources” specifically
• Spend a little time fleshing out each idea
• Have each host talk for a bit without interrupting each other so much 
• Don’t say “exactly” so much
• Try to be around 10 minutes long

This worked pretty reliably, and the audio you hear in the podcast is the output with those instructions, using the background documents linked. I made no further edits to anything.

I used Spotify for Podcasters to create the podcast. ChatGPT had already made suggested episode titles and descriptions, which I used verbatim. I asked ChatGPT to make a podcast tile art for the show, using some blend of dogs and AI. I was pretty happy with the outcome. Then I asked ChatGPT to make an episode thumbnail image for each breed. Once I got the look I wanted, it was a simple to make them all by prompting, “Now make another one for a Poodle”, and so on. Took just a few minutes.

Uploading everything into Spotify took about an hour, but then I had a fully functional podcast! I launched it, and let my dog-loving kids be the beta testers. They thought it sounded fine.

The final step was to go back into each breed’s notebook in NotebookLM, and use the Notebook Guide to generate an FAQ. Those are verbatim what you see pasted into each post here for the specific breed. And then I put the show in the other major podcast players.

So that’s it. Start to finish, it was probably 10-12 hours of total effort. I bought this domain ($50), and I bought the Easy Podcast Pro plugin for WordPress ($50). And of course I have a ChatGPT subscription. All in all, this was an exercise in seeing what these tools can do, for a layperson. And if you assume this is the worst they will ever be, it points to a future that is somewhat sobering for creators.

However, I remain optimistic. While I’m sure there will be a large wave of AI slop podcasts — arguably one I just contributed to — I also believe these tools will unleash a lot of creativity in people, and make talented people even more talented and efficient. And in the end, make more good than bad.