I've been teaching myself Machine Learning to see if I can get a computer to learn to write stuff. The area I'm interested in is generative AI, which is algorithms that create new content entirely on their own, rather than optimizing or automating existing stuff.





Baldwin& isn't very active on social, so my first project was training an AI social media copywriter that could write posts as if it were our agency. We named it TERRY, sort of like Watson's dumb, weird brother than he never talks to except at reunions. It worked well in some ways, but couldn't create long sentences that held together.

I trained TERRY on the social media accounts of most of the ad agencies in the USA. It made for a pretty big data set, but with high variation and complex sentence structures. If I had more data, TERRY could have eventually figured it out, but there are only so many Ad Agency social media accounts. For my next project, I'm gonna try to come up with a dataset of less complex/variable writing. I hope that helps the algorithm figure out basic syntax more completely.

Some of my favorite posts from TERRY, below:

Screen Shot 2017-06-21 at 2.24.21 PM.png


TERRY figured out most of the main talking points for advertising, and the Super Bowl came up often.


When the length stayed short, TERRY could seem pretty profound.



Portly artists are the best artists.



I also trained TERRY on March Madness social posts.





TERRY worked pretty well in short posts, but when he tried to string phrases together they did't make sense. Here, words seem like they're vaguely in the right place, but it's gibberish.