machine learning brewery




At the height of the AlphaGo news cycle, we trained a machine learning model on beer recipes which were highly rated on internet recipe sites. We then took this recipe to an independent Toronto brewery who agreed to help us brew and bottle the beer. This resulted in a series of AI-generated beers with names such as "Frontiers Pale Ale", "Alpha Go A.I.P.A.", and "Lee Sedol A.I.P.A.", which we imbibed while taking bets on the winner of the now-historic Go match between Lee Sedol and DeepMind's reinforcement learning agent.




We also launched a stealth mission to hand-deliver a bottle of beer to the prime minister of Canada during
an announcement ceremony for Pan-Canadian Artificial Strategy funding, along with a "Totally Unsupervised" t-shirt.
Our co-founder Ragavan Thurairatnam was tackled by the prime minister's security detail in the process.

It is unclear as to whether the prime minister imbibed the beer or wore the t-shirt.
We're going to assume that he did, because who wouldn't?



Recipe: Alpha Go A.I.P.A.

We trained a neural network to predict the best grains/hops ratio by mass.

• 12.75 lbs of base malt
• 0.75 lbs of Munich malt
• 4 different hops (1 oz each, totalling 0.58 lbs)
• 7 gallons of water (pre-boil)

Add bittering hops at 65°C and boil for 90 minutes.
A baseline Bayesian Ridge model predicting the rating out of 5 for each grains/hops ratio.