A good rule of thumb regarding AI is “where there’s rich training data, there will be exceptional AI.” We have already seen this with text and images. In the last month, we have had a taste of a new one: podcasting.
This seems like an odd place for AI to work out, but it really has. Google has a product called NotebookLM. For a while, it was just a research tool. You could load up documents and web links, and then you could chat with the AI and ask questions about them. It could summarise, and it could critique.
Then, a few weeks ago, they added a simple feature. At the push of a button, you could generate a podcast.
The result was a podcast of between 5 to 20 minutes between a man and a woman that could sound like any old NPR podcast. But what was surprising was how interesting and compelling it was. I gave it some papers of mine and found the results outstanding. For instance, here is one about a technical paper of mine on entrepreneurial experimentation.
It has analogies, clear thinking and thoughtful insights. I have done podcasts, and frankly, this is as good, if not better. Occasionally, there are things to quibble with, but this happens when anyone else discusses your research. For this one, you can actually compare the AI version to a podcast I did on the subject.
For good measure, here is one on today’s Nobel prize winners, trained with the Nobel committee’s popular and technical explanations. It is as good as it gets.
I enjoyed providing this, as one of the winners has a reputation for being pretty negative about AI.
That said, the generated podcasts have a notable bias. I tested this by training it on transcripts of Trump speeches from a few weeks ago. Here is the result. I was hoping for an earnest attempt to explain what he was saying, but it tipped towards the craziness. To be sure, it is not inaccurate, I was just hoping for something that was more willing to adhere to the text.
The point here is that this comfortably passes the Turing Test for podcasts — if you didn’t know, you wouldn’t suspect this was AI generated — and it is also interesting — even if you know it is AI generated, it is actually useful. I’m going to generate podcasts for each chapter of my books and include them on my website. It really brings it all to life.
In the meantime, the Internet being the Internet, people have discovered how to go straight to the lowest common denominator.