I continue to be impressed with the work at OpenAI and their GPT-3 offering. Generative AI work is hard. They have been able to finance complex research through commercialization. Their marketing is phenomenal - word of mouth runs around the tech community.
One company building on this framework is conversion.ai. The company provides automatic marketing copywriting. The OpenAI API powers "Jarvis." Jarvis is a text generator that allegedly produces copy based on the seeds provided. And a lot of it is good!
Dave Rogemoser, the CEO, previously built Proof, a very cool plugin for marketing sites. It surfaces recent social proof activity - signups, activation, etc. - to potential customers looking at your website. The effect is remarkable - everyone with a monetizing business should use this service or something like it once they have a minimum threshold of users.
Overall, his track record points toward creating real marketing value for his customers. I am generally skeptical of machine-generated text, but this history convinced me that this was worth examining.
In conversion.ai, you dump your ideas about your product into a text field and see the generated headlines and value propositions get restated using known marketing frameworks, pre-tested for, well, conversion.
If I have an idea, and I am not confident in my presentation of that idea, the software will find an expression that matches known best practices. The outputs run for maybe two or three sentences. One sees multiple versions of the copy, and a person picks the one that seems best. Clever!
There is another angle to the product. One can generate longer-form copy for a higher fee, like a blog post - or an essay! Here one can seed the content just as you did before to establish the topic or core idea. The system then generates text a few paragraphs at a time. They offer integration with an SEO product to give immediate feedback on how well such a piece will rank according to the latter's analysis. You retain control to delete or revise some of the text. Don't like how it is turning out? Then tell it to generate some more! The score from the SEO optimizer dials in through this rinse-and-repeat method.
This human-in-the-middle generative process creates text. And it reads as human writing. An automatically generated article on gardening reads like an article on gardening. But does it generate information? Do customers benefit from all these words spilling out?
Claude Shannon, the author of the Mathematical Theory of Information and one of the parents of information theory as we know it today, pointed out how many exchanges contain very little information - we repeatedly pass it around. Redundancy can add value since people are imperfect transmitters and receivers of information, and the media we use to communicate can be lossy.
But the difference between the sliver that is information and the bulk that is transmission guarantee comes into play when we think of these AI tools.
I attended two of Rogemoser's training sessions on conversion.ai. He shared that the system works optimally when the user provides about 600 characters of text to start. Less or more degrades its ability to generate helpful text.
Any automatically generated text must be derivative. The machine is not teaching us - it is simply expanding on the information received at the start. The auto-generated text is the bulk of redundancy rather than the sliver of insight. At best, it repeats the idea.
The implication is that the optimal length for transmitting one idea is 600 characters. That's slightly over two tweets! Given some knowledge about the world and perfect attention to your words, a listener should be able to "get" an idea from you in about 100 English-language words.
How little of our writing adds to a discourse! What's the use of the rest of the words? After all, I am already about six times that length in this essay!
Conversion.ai usefully reveals this through its dashboard: the power is in repackaging the same idea in different frameworks. They are pretty explicit about the value being in the repetition across multiple media and methods of conveying a given idea into one's head.
If we find the correct 600 characters, we can get to the essence of our idea. The job of the human is to arrive at 600 characters. The machine, as Rogemoser has shown, can take it from there.
What are your 600 characters?
Photo by Aaron Burden on Unsplash