You love the blog, so subscribe to the Beervana Podcast on iTunes or Soundcloud today!

Friday, December 16, 2016

Won't A.I. Beer Inevitably Result in Bud Light?

I somehow missed the yuuuuge rollout this summer of IntelligentX, a Britain-based project to produce beer that responds via algorithm to customer choice:
After you’ve tried one of our four bottled conditioned beers, you can tell our A.I. what you think of it, via our online feedback system. This data is then used by our algorithm to brew the next batch.
Because our A.I. is constantly reacting to user feedback, we can brew beer that matches what you want, more quickly than anyone else can. That means we get more data and you get a better, fresher beer. 
Paging through Google search results, I see they got a massive amount of press for this, nearly all of it happy to goose the hype machine. (Admittedly, combining "AI" and "beer" is genius clickbait and I kick myself for not thinking of it first.) For some reason, there's a new round of hype, which is how I stumbled on this just now. 

Lab coats make this seem extra sciencey.
Source: 10x

As far as the project goes: whatever. It seems almost inevitable that someone would have invented this in 2016. (In fact, if Elon Musk is right, it was algorithms that led us here in the first place--algorithms within algorthims.) Every start up in the world is trying to figure out how to monetize app-collected data, and fusing that to the jillion-dollar beer industry is a no-brainer. 

I am surprised no one pointed out the three fatal flaws behind the rationale for this project, though. My guess is most brewers would have seen it immediately. First, incorporating feedback into recipes is what breweries do. Markets function as far more effective algorithms than a consequence-free thirteen seconds tapping buttons on a smart phone. There are five thousand US breweries making at a minimum 50,000 beers in the US, and the feedback of preference is available when a customer spends actual cash on a beer. 

Second, people have different tastes. I was out for a beer last night with BeerAdvocate's Ben Keene and Eater Seattle's Adam Callaghan. We discussed how breweries now use IPAs as points of distinction--each one is different, to suit different tastes. And third, collecting the input of hive mind means finding the exact average of their preferences. It will produce a beer most people think is adequate. "In pursuit of the perfectly average beer" was probably not the motto they were shooting for. Brewers know you can either appeal weakly to the largest group (light beer) or powerfully to a small subset (distinctive IPAs), but you cannot simultaneously do both.

And one final thought I throw out to the techies out there. I'm not perfectly versed on my sci-fi futurism, but does any of this actually have anything to do with AI? I mean, find me a HAL who's making his own beer, and then we're talking.


  1. I'm afraid wearing lab coats that aren't done up means they lose science points from me

  2. You're right there Ed. That's how medics [spits] wear them.

  3. Regarding the AI bit, my reading of it tells me that they are using machine learning (ML), which, speaking generally, involves giving computers very large sets of data with minimal instruction and letting it figure out the patterns for itself. For example, giving an algorithm thousands of pictures of animals, and it learning on its own, through many iterations, to tell which contains a dog. A more practical example: it's what gmail uses to determine if a message is spam or not, by observing the thousands of messages people mark "spam", and comparing those patterns against all new incoming messages. It's perhaps the simplest form of what might be called "AI" and indeed I would argue that AI deserves a higher standard than just ML, but that's another discussion.

    "IntelligentX" (ugh) don't specify how exactly they're using this technology, but I would guess that they assembled a repository of many beer recipes (you need quite a large dataset to get useful results), perhaps in combination with favorability data or something else that indicates the "quality" of the result, and are letting the machines find commonalities and develop new twists that fit those patterns.

    You are right though that this already happens, more efficiently, through the market, and Bud Light or something even more terrible would probably be the inevitable result if the algorithm is left to run on its own. See Microsfot's "Tay" twitter bot for an example of how ML can go horribly wrong. It also may require static, reproducible conditions, so I would think anything wild/spontaneous is out by definition. Given that so many environmental factors determine the outcome of a beer, methinks the variables here are too numerous and uncontrollable to get truly great beer.

    Ultimately ML is just a tool. If it is given expert guidance, it might actually produce something genuinely interesting. But who knows what the bonafides of these people are. I'd love to see what a team of Portland's best brewers would be able to guide this to produce. But to me, this simply reads as novelty, one of just a million other startups right now that are trying to make [thing] + AI = profit.