After a recent whisky club where we tried to emulate a ‘bottle naming panel’ – akin to how the Scotch Malt Whisky Society name their bottles but slightly more insane due to playing Scrawl and Pictionary in the process – we settled on a plan. During the evening, we were challenged – somewhat tipsily – to make an AI bot that emulates SMWS bottles names.
So how do you get a bot to create the plethora of bottle names that SMWS uses, from descriptive ones like “Log Cabin in the Black Forest” and “Ice Cream Cones in a Rose Garden” to slightly off-the-wall ones such as “Thou Art a Spicy One!” and “Fatberg Ahead“, all the way to the somewhat free-spirited “Those Aren’t Regulation Tyres” and “Let’s Head for the Lobby“. They all seem connected, but humans are good at seeing patterns – to help out our poor little robot, we need to utilise an LLM…
Like, subscribe etc: @smwsbottles
There were four main steps in creating the bot:
- Collating and processing (in R) a database of randomly-chosen SMWS bottle names, then QAing them to remove errors and duplicates – ending up with ~7000 unique bottle names from the past 40 years.
- Using this list as a training dataset for GPT-2 to start predicting new names. Around 200 iterations were performed, then three sets of results were created using varying ‘hotness’ (1.0, 1.1, 1.2) – differing from very plausible to slightly whacky.
- Working with a compiled list of all SMWS bottle codes to tweet out, with an X before them to avoid confusion.
- Utilising the Twitter API (in Python) to post the results, using a CRON job to push them out automatically.
All that was run though and once we had some output, there was no curation done at all on the results – what it tweets is what you get. It’s quite fun to watch on that basis and there’s always a surprise around the corner!
You can see some example output from the bot below, and for those logged into Twitter, some recent tweets underneath:
