# Recipes in the Wild By Paul Engle June 1, 2017

The Recipes Project blog is, starting today, running a Virtual Conversation on the theme, “What is a Recipe?” I featured this in the editorial of the latest edition of Whewells Gazette the Weekly #histSTM Links List. Inspired by a comparison that I made between algorithms and recipes and a question that I posed, Paul Engle, author of the very excellent Conciatore: The Life and Times of 17th Century Glassmaker Antonio Neri and writer of the Conciatore Blog, sent me the following essay stating, “Feel free to do with it what you will.” So what I will is to post it here as a very welcome guest blog post from an excellent historian of technology who really knows what a recipe is.

It has been suggested at Whewell’s Gazette in a recent editorial that in considering recipes, particularly technical recipes and their relation to algorithms, that, “the two words are in their essence synonyms and there isn’t really a difference.” [1] With all due respect to the author of this passage, I do not think that is quite right.

A recipe is much more than an algorithm, in fact I propose that while algorithms are quite powerful tools, they occupy a rather distinct niche in the universe of recipes. We do agree on some things however,

“For me a recipe is quite simply a set of instructions, which describe how to complete successfully a given task. The task does not necessarily have to have anything to do with cooking, the first thought that pops up when we hear the word recipe.” [2]

I have thirty-odd years of empirical experience writing and following technical recipes in a laboratory setting; I have several shelves full of them that I am looking at right now. I have been programming computers and dealing with algorithms, dare I say it, since the days of punch cards and paper tape. This is a subject particularly dear to me and besides, I sense an irresistible opportunity to make a fool of myself, so here goes.

In the realms of mathematics and computer science, an algorithm is a set of instructions that enjoy several conditions favorable over recipes; a well-defined environment where it does not matter if it is raining or sunny outside and an output or result that is usually unambiguous. For recipes, not so much; even the lowly baker known that on humid days, a prized and tested bread recipes must be adjusted to produce an edible product. These adjustments do not always take a form that can easily be measured or quantified and this starts to get at the heart of the matter.

Any day of the week, rain or shine, a computer running a straightforward algorithm can generate the first million digits of pi, (yes, the millionth digit is 1). While there may be a certain amount of difficulty in verifying a result, it is something that is done quite routinely. While some simple recipes fall into this form, many others do not. Consider that some technical recipes seem to work even if we do not know how. Others require “experienced” practitioners, not because of anything magical going on, but simply because the most reliable results are obtained by one who has done it before. Even with seemingly simple, well-documented tasks like polishing a material, there can be an enormous number of variables involved, some unknown, others that are not practical or possible to control.

An algorithm generally lives in an artificially constructed, tightly controlled environment, recipes, on the other hand, operate in the wild. An aspect of technical recipes often missed by outsiders is the level of attention that must be paid to the interaction of your “product” with its environment. This may mean frequent observation and testing, or, in the kitchen, it may mean tasting the gumbo every few minutes and making appropriate adjustments. No matter if the result is a well-polished sample in a materials laboratory, or a well-seasoned bowl of soup in the French Quarter, what makes the result “good” is not necessarily easy to define. We can calibrate our equipment and take great care with our materials. We can scrutinize the results, and take measurements until the cows come home, but in many instances, this is only a starting point; learning to perform a recipe “well” can be like a mini-education. Writing that down stepwise can be like trying to capture everything you learned at cooking school.

It is in this setting, where there are many variables to keep track of, many unknowns, and even the results may be hard to characterize, that we step into the realm of “art.” A successful outcome depends as much on what you bring to the table as what is written on the page. A recipe becomes like a roadmap for threading your way through a complex maze of decision points. Here is where I get passionate about my subject. Practicing a recipe, in a sense, can be viewed as the purest form of empirical science. And this can take place in a laboratory or in a kitchen. If science is the study of the way the world actually behaves, then going through a series of steps and paying close attention to what is happening, is as good as it gets. It is not a matter of imposing ones will on the world, but of interacting with nature and moving toward a result given the constraints of reality; there is a give and take. A scientific experiment can be viewed as the act of developing a new recipe toward a specific result. Writing that recipe down is an exercise in determining the important variables to pay attention to and capturing a method in a way that is repeatable by others.

As computer algorithms move into the realms of artificial intelligence, driverless cars and the like, they will start to encounter the same difficulties as our baker does on a humid day. Perhaps a true test of machine intelligence will be how well an algorithm negotiates real-world recipes.

[1] Christie, Thony 2017. Whewell’s Ghost blog, “Editorial, Whewell’s Gazette: Year 03, Vol. #41” 31 May 2017.
[2] Op. Cit.

Filed under History of science, History of Technology

### 8 responses to “Recipes in the Wild By Paul Engle June 1, 2017”

1. The difference between a recipe and an algorithm becomes very clear in the context of training people for industrial jobs. Where there really is an algorithm to determine what to do next to accomplish a goal, training new employees isn’t particularly problematic; but it turns out that such algorithms are generally lacking. In fact, where algorithms are available, procedures are usually automated anyhow. If turnover is gradual, the older employees can train the new folks; but turnover is often not gradual. For example, a great many workers in the nuclear power industry were hired at about the same time when the industry got off the ground. That cohort is retiring en masse and the utilities have to retrain a new cohort. Hence the efforts to define what it is that a maintenance man in a power plant actually does so that the new guys will have a recipe for checking valves, monitoring plant noises, etc. That turns out to be very hard to do, in part because the old employees don’t necessarily know how they do their job (“I’ve course I have a method. I just don’t know what it is!”) but also because recipes that sound unambiguous to people who already know how to do the job are likely to be mysterious to people who don’t.

The same problem comes up in hermeneutics because it turns out there is no algorithm for interpreting a text. A book is more like a recipe than a formula and what has to be preserved from generation to generation isn’t just paper or parchment with legible words but a tradition of how to read. Same reason you generally need grandma or cooking school to learn how to cook.

• Thanks for your thoughtful reply! A related point is that recipes occupy a space between written knowledge, and strictly “hands on” knowledge passed between successive generations through apprenticeship etc.
-Paul

• C M Graney

Nice topic, Thony! I have been thinking of this as I hear breathless media coverage of robots, ‘AI’, driverless cars, and the like. Often these reports reference IBM’s Deep Blue and Gary Kasparov. Yet it seems to me that what Deep Blue vs. Kasparov shows is not the power of computers, but the power of people when they create algorithms/recipes. In theory, the steps taken by Deep Blue or any other chess-playing program are simply a recipe, steps carried out very fast by the machine. But a person could follow those steps by hand. What IBM’s developers did was successfully reduce high-level chess to a recipe – as my game-playing son says, to a “monostrat”, a one-strategy game. I think the comment about algorithms in driverless cars and the like starting to encounter the same difficulties as a baker does on a humid day is excellent. Specifically, I imagine driverless cars being quickly co-opted by all those terrorists who currently use cars to deliver car bombs or run people down in the streets. It will be like developing a recipe for high-explosives, and thinking that only people who quarry rock or push highways and railroads through mountains will use it.

• I note that people worry about neural net programs that arrive at favorable results because nobody understands how they do it, even though human beings have also been flying by the seat of their pants from the get go. We often don’t know what we’re doing. This is sometimes called genius, sometimes hubris. To be fair to silicon intelligence we ought to compare our disaster rate to theirs, not to some absolute standard. As the punchline of the joke points our, you don’t have to outrun the bear.

2. Aha. I just recently saw that Lorraine Daston has been making this exact comparison. This seems to relate to her argument about “rationality” (connoting rules) versus “reason” (connoting the judgment that governs how rules are to be applied), with the attendant claim that the latter fell into decline during the Cold War period when faith in algorithmic reasoning supposedly flourished. This is a claim with which I have very strong disagreements.

• Thank you very much for sharing the Lorraine Daston link. I just spent the past hour listening to her (mind blowing) lecture. There is so much to think about here that it will take me a while to absorb. I highly recommend it to anyone interest in this subject, but if you can’t watch it in full, there are interesting bits about algorithms at 33:40 and then 51:30 through the end.

For what it is worth, in the spirit of broad context, there are some interesting resonances (for me anyway) between her survey of rules and exceptions and this overview of Gödel’s incompleteness theorem by Marcus du Sautoy that I happened to be watching recently. https://www.youtube.com/watch?v=O4ndIDcDSGc