Before LLMs were cool

Back in 2015 I wrote a Tumblr to try to earn commissions on wine sales on Amazon. That was a thing you could do then, in 2015. I guess I should say “wrote” because the only thing I wrote was code. I certainly didn’t write the wine reviews.

What I did, before anyone had ever said “Language Learning Model” (ok maybe some nerds had said it, whatever), was download the text of about 800 Wine Spectator “Daily Picks”. This was pretty easy because they used integers as keys in their URL so you could just go to /page/1 then page/2, etc etc, and Beautiful Soup just dumped the words into a text file.

I fed this download into markovify, a Markov chain generator written in Python. Someone may come “well, actually” me on this but Markov chains are essentially early LLMs that could run on your basic laptop in 2015.

Then I wrote a script that took a referral link to an Amazon wine and spat out a review and (and this might be my favorite part) a rating that was a random number between 77 and 94.

It then posted these reviews to Tumblr. They’re still there, you can go see them in all their glory. The name, Andrey Wines, sounds like a fancy winery, right? I would totally buy a nice cab from them. It’s named for Andrey Markov, who invented the Markov chain. I thought that was SO clever.

I mean look at that (click the image for the post, I’m not sure how to do alt-text on WordPress. I should figure that out). It had no idea the difference between a red and a white and I didn’t try to explain.

I giggle to myself knowing that OpenAI paid Tumblr to let them train their AI on Tumblr posts, and some of those posts are going to be this.