What I like to show is how Artificial Intelligence can work on the hidden structures of the world around us. Almost everything is quantifiable and can thus be described in statistical relationships - if X then Y. What we see happening at the moment - the AI boom - is a consequence of more readily available (and reliable) data through the internet, increased computational power and lastly a decade-long development of clever and efficient data-analysis algorithms. Nothing more, nothing less.

January 10, 2024

Artificial Intelligence & Watchmaking - An Introduction  

    Marcus Siems @siemswatches
    Collector, Author, Data Analyst

Artificial Intelligence and Watchmaking:

- Introduction - ChatGPT Interview - Community Outlook -

Artificial Intelligence is a phenomenon of growing interest to all of our lives. It more and more slides from the abstract professional world into our everyday experience and daily routines. And increasingly we find reports on the matter popping up in the horological world as well (for example [here], [here], [here], or [here]). All of these developments are typically met with a lot of buzz and enthusiasm... probably for two main reasons:

First, seeing that such a traditional and long-established industry like watch-making starts to use AI on a larger scale feels surreal after it seemingly moved through time unaltered for Centuries. And second, because 99.9% of us (including me) don't really know what's going on when we type a query into ChatGPT and the likes.

AI-generated (DeepAI) image of an artificial intelligence working in watchmaking.

Even though Artificial Intelligence is one of our most advanced technologies it brings a mythical aspect with it. Probably much like people must have felt when for the first time faced with the findings that our world fully consists of miniature particles or that our brains (and thus we) are functioning through electro-chemical circuits. We first have to believe individual scientists before we can understand it as a society.

However, as long as AI stays in the realm of mythical creatures we won't be able to assess whether a new horological application - be it a website, design, or tool - is truly new and potentially useful or just a toy. We would be like Dorothy following the AI-brick-road to any Wizard that'll promise us what we want. But show us the red shoes! Rather, let's take on this endeavor and try to better understand what Artificial Intelligence is, and what it can be used for.

"Machine Learning Captcha" by Randall Munroe (xkcd).

1) Artificial Intelligence - Meaning

I think a lot of power that strikes us with awe already lies in the name. Intelligence probably is one of the characteristics we as humans most closely relate to ourselves and that we take to distinguish us from the rest of the animal kingdom^[1]. It is a word that has a lot of weight. But what we might attribute to the word and what is actually done in AI research and applications might strongly dissociate. 

Because pragmatically speaking what Artificial Intelligence does is predicting an outcome based on data it has been trained on (more on that below). In other words it's a statistical method to identify and quantify patterns. Nothing more, nothing less. In this way it indeed can work like the human brain. We too seek to find patterns in the world around us and once found we have a hard time not seeing them anymore. Take the following example of a Dalmatian:

Black & White splashes - can you spot the Dalmatian? Photo Courtesy of the Michael Bach Database.

When seeing the image for the first time everything looks like random patches of black and white. At some point structures emerge like a tree in the background, a road and potentially leafs. The dog (a Dalmatian) is quite hard to find but once you've seen it, you can't unsee it anymore. However, that's your brain putting order to the blank space between black patches... strictly speaking there is no dog.

This pattern identification - even on an abstract level - is not intelligence, it is an association. Be it the association between dots and a dog; the sun being out and a warm feeling; the watch market and retail prices increasing; height and weight of a human; etc. Thus, it can be better understood as learning - or in the more common term for the methods used to generate AI as machine learning. And I find this wording grabs the essence better. It doesn't take a highly developed organism (or machine) to establish associations and learn... it is merely a statistical relationship between two things.

Staying with dogs... even completely random stimuli can be paired and associated with one another. We all (hopefully) know from High School about classical conditioning, and Ivan Pavlov reported this form of statistical learning already in 1897. Photo Courtesy of Communication Theory.

2) Artificial Intelligence - Basics

So how can a machine "learn" and what does it take to build your personal Artificial Intelligence^[2]? I argue it all boils down to three main ingredients and steps:

- training data

- a variable to predict

- a model that connects the two

Breaking it down like this it doesn't appear like magic anymore, does it? But let's make it a tad more explicit what I mean with these three ingredients and play a game of... How much is my watch worth?

We also got you covered if you're not only theoretically interested in this question... "4 Tips for Valuing your Vintage Watch: Beginner's Guide" - A video on the heuristics by Felix Goldammer.

First, the prediction... it can be made on anything unknown that you want to describe: what move to make playing chess, is the person in front of you sick or not, what is the correct response to your question, will this new watch design be a success? Or as in our little example here: How much is my watch worth? The only important thing here is that the semantic question you're asking can have quantifiable answers (here: price in Euro).

Second, to predict the outcome you need relevant information. For example how much a vintage watch is worth depends on its condition, scarcity, the complications, reputation of the brand, etc. etc. Importantly, you can't take any type of information... for example the strap color or the designer's mother's maiden name shouldn’t really be of importance here.

Third, a connection between the input information to the output prediction. This third part is the statistical operation that we classically call machine learning and thus the central aspect of AI. This connection (or model) needs to be trained on the input information of - in our example - several watches and their respective known prices. So we might see that - on average(!) - watches in better condition are valued higher. certain brands are more en vogue and add a premium to prices, complications usually cost a little extra as well, etc., etc.

When you ask Eric Wind of Wind Vintage what's important when valueing a watch: "Condition, condition, condition!". Photo Courtesy of Esquire.

This is a semantic description of a machine-learned quantifiable interaction between market factors on the price of watches. It is a very simplified description to what for example is going on at OpenAI and ChatGPT or image generating AIs but at its core that is what's going on, too*. But we’ll get to these applications in a bit.

3) Artificial Intelligence - Different Models

Now let's take a quick dive into the middle part - the machine-learning model and its assumptions. There are numerous ways how to infer the outcome from a set of information. The most simplistic being the monotonous (sometimes linear) relationship, for example: The better the condition the higher the price of a watch:

Figure 1. (Toy-)Example scenarios where machine learning algorithms can be applied. The colored points depict single watches, the dashed line implies the prediction from the algorithm. A gradual prediction of price based on the condition (left) and a boundary to predict keeping/selling a watch based on emotional attachment and wrist time each piece gets (right). The data depicted is fictional and only meant to aid context.

The dashed line in the figure describes the model prediction. For example it might be a 1-to-1 linear relationship and price = condition * a constant. Maybe there are some effects beyond linearity and our relationship is rather exponential - if you increase the condition on the upper end a bit it dramatically increases the price… This exact relationship between price and condition is what can be learned in the process… from a machine**…

4) Artificial Intelligence - Modern Algorithms

All these simple methods I just introduced are Century old and build the foundation of our modern experience with AI. One of the earliest mass applied machine learning came during World War I… Charles Spearman introduced factor analysis to simplify large amounts of psychological questions and tasks and defined - interestingly - a general parameter for intelligence^[3]. This first quantification of intelligence was then used to assess British soldiers and find suitable officers in their ranks beyond the ancient but still prevalent aristocratic system.

You can find "A Brief History of Artificial Intelligence" covering the last 70 years of major developments by BootstrapLabs here.

In the 1970s & 80s applications using a slight modification of these methods were introduced - artificial neural networks***. Here, an input object with all its features gets weighted (model) to predict a reaction. For example we got a watch that shows several features - size, age, material, condition, brand, etc. - which we combined call a layer. And weighting these features we could (in an optimized world) estimate its best sales price. It is thus an extension on the linear estimation from above.

Figure 2. Schematic of Artificial Neural Network Architecture with one layer. The architecture is only to depict an intuition and the "layer names" on the bottom are memory aids and not constraints or actual modeling results.

This very simple "architecture" can relate the input (a watch) with the output (its price)... Again this relation is learned over thousands and thousands of watches  with known prices before it can be applied to predict the price of a watch that is not already in the database. It will further always include some error, i.e. unknown parameters like for example auction bidders moods.

Now imagine you don't have only one layer that directly gives you an output but that this layer feeds its information into another, "deeper" layer. The deeper layer defines relationships between the features, and another deeper layer after that the features of features and another after that and after that... each summarizing and abstracting its input through relations of features from the previous layer...

Figure 3. Schematic of Artificial Neural Network Architecture with several "hidden" layers. The architecture is only to depict an intuition and the "layer names" on the bottom are memory aids and not constraints or actual modeling results.

You can also think of each layer and each abstraction in the sense of language (Large Language Models). The first layer describes letters, the second layer the relation between letters thus "syllables". The third is on the relation between syllables -> words, the fourth layer is words -> phrases or sentences and then on to paragraphs and abstract meaning^*4. These abstractions - that can also be learned from the algorithm without constraining it - are the major advantage of these "Deep Neural Networks" over everything in the field of Artificial Intelligence that came before.

A good analogy for the deep layer structure of artificial intelligence is the plot of the Christopher Nolan movie "Inception" from 2010. Each layer further down is like another dream in a dream... It gets more absurd and more surreal with each dream. Photo a hand-drawn plot structure of the movie from Christopher Nolan himself.

Of course these labels for the different layers are purely schematic. Words, sentences and paragraphs are entities that we can grasp from our daily experience. But a statistical model will build very abstract representation that won't follow our constrained understanding. Nevertheless, these deep layers of abstraction are the reason we can have a conversation with ChatGPT and similar algorithms^[4]. It means that the algorithm doesn't need to have seen a particular paragraph, sentence or question before to understand its meaning because it can find meta-relationships from other text snippets.

However, as you can imagine, to get to this point you need a LOT of processing power and text data for the algorithm to learn and function. For example ChatGPT 3.5 (free version) has been trained for weeks and months on thousands of GPUs with written text, books and websites and finally been integrating feedback from thousands of human operators.

5) Artificial Intelligence - Conclusion on its Essence

So I guess everyone is sighing in relief now that we worked through this rather intense chapter of methodology. However, I hope that I could convey that Artificial Intelligence and Machine Learning are 1) not new, 2) at its core a "simple" prediction models, and 3) not a magical tool gifted to us from the robot overlords to distract us by giving book summaries and generating steampunk images of our pets.

An AI-generated image (DeepAI) of "Several watch gear trains working with computer circuits".

Much rather Artificial Intelligence works on the hidden and latent structures of the world around us. Almost everything is quantifiable and can thus be described in statistical relationships - if X then Y. What we see happening at the moment - the AI boom - is a consequence of more readily available data through the internet, increased computational power and lastly decades of development of clever and efficient data-analysis algorithms. Nothing more, nothing less.

In the next part I want to address in detail what modern AI can and cannot do in our beloved world of watches. To aid this - largely speculative - discussion I also spoke with several experts in the field of horology who all had some overlap with the Tech world in the past. That'll be fun!

* I strongly apologize for the over-simplification.

** By a clever analysis approach but not exactly intelligence^[1]

*** There are several empirical differences between carbon and silicon brains - but let's not get into this. ANNs might (at best) be called a simplified model of biological function.

^*4 You can think of similar layers of abstraction also in the sense of visual/image processing. The first layer are light/dark edges confined to a certain location, the next layer can be angles, the next layers more and more complex shapes that can be found anywhere in the scene and the last layers are object identities, which can be even independent of presentation angle, etc.

References & Recommended Reading

^[1] [Recommendation] On Intelligence; Jeff Hawkins & Sandra Blakeslee, St. Martins Griffin Publishing, New York;

^[2] [Recommended SciFi] Foundation Trilogy (1942-53); Isaac Asimov, Gnome Press;

^[3] The proof and measurement of association between two things (1904); Charles Spearman, Am. J. Psychol.;

^[4] [Recommendation] The Most Human Human - What Artificial Intelligence Teaches Us About Being Alive; Brian Christian (2011), Anchor;

All rights on text and graphics reserved to the Author.