| December 11, 2019
What happens when you feed a cutting-edge, text-generation AI with two books and over ten years of blog posts? We've been exploring OpenAI's GPT-2 text generation model and it is impressive, but it still needs a lot of coaxing to get the results that we want from it. Sadly, I think it will be some time before it can write my blog posts for me.
A few months ago, Horacio Gonzalez, Senior Data Science Director at Kantar, suggested the idea of developing nAIgel, an artificial intelligence version of yours truly. Since then, as time permitted, Moises Arizpe, Data Science Director, has been responsible for nAIgel’s care and feeding. We are now at the stage where nAIgel can generate responses when prompted with different phrases. The results might best be described as intriguing. Responses often start with a reasonably coherent sentence but then the quality degrades. For instance:
"A brand is"
“A brand is the product of a long-term commitment to providing the best experience possible for our customers.”
Now that sounds pretty good, but after expanding on that statement nAIgel asks,
“How do you make the right decisions to build a sustainable and long-term brand?
And later states,
“We need to understand how our customers and team would feel about a brand and how they would react to that brand. It is a little like using your seal of approval to decide whether to eat a tuna sandwich or a steak.”
I am really not sure how it ends up talking about choosing between a tuna sandwich and steak! What is notable, however, is that the model has adopted specific characteristics from my blog posts, like the rhetorical question and analogy used above. It has even created a new concept not referenced anywhere in my writings:
“Brand Purity is the ability to define a clear vision of what the brand will be in the future.”
However, taken overall, the responses given by the model are far from coherent. The fundamental problem is that we are using a language model to create the content, but it has no inherent understanding of how brands, media and marketing work.
Also, of concern is that the original language model was trained using data from Reddit. Could the reference to “purity” be the product of that data set? Might there be a risk that nAIgel will turn out like Microsoft’s Tay? As Kyle Findlay notes in this article, sampling is not just important to the practice of survey research it applies just as much to training an AI.
Finally, my content dates from as early as 2006, and my thoughts on how brands, media and marketing work has evolved over time. Currently the model is going to ascribe equal weight to a post written in 2006 as one written over a decade later.
None of these challenges are insurmountable. No doubt we could train the model by giving it feedback on what was ‘good’ and ‘bad’ to have a better understanding of brands, but it would be a lengthy and time-consuming task that would require direct intervention from me. What do you think? Worth a try? Please share your thoughts.