Home
⏭️

What comes next?

I made a piece of hashart called element. It allows you to generate realistic sounding chemical elements from a seed phrase - the same phrase always gives you the same element.
values
The algorithm that generates the name of the element uses a Markov chain.

A Markov chain or Markov process is a stochastic model describing a sequence of possible events in which the probability of each event depends only on the state attained in the previous event. Informally, this may be thought of as, "What happens next depends only on the state of affairs now.”

These are really fun data structures, relevant in the agent of LLMs, and surprisingly easy to implement. We’ll do so with a few simple code examples, but first let’s build an intuition for what these things are and how they relate to our chemical “name generation.”

🔗 Jordan’s daily routine

I’m a creature of habit. My commute to work can roughly be summarized with the following:
  1. Wake up in my apartment
  2. Walk to the Hoboken PATH terminal
  3. Take the PATH to Manhattan
  4. Walk to the office
We can draw a diagram showing where I currently am.
My commute home, however, has a fork in the road. 80% of the time I walk to PATH and head home. 20% of the time I take the boat ⛴️ 😎. We’ll add these states and, importantly, the probabilities of each route.
Here’s the graph with my PATH commute home added. Notice a new 🟣 end state. Once I’m at Hoboken PATH my commute is pretty much over.
We can also introduce my boat ⛴️ 😎 commute.
Many of our arrows are unlabeled. For simplicity’s sake, let’s split the probabilities evenly. If two arrows come out of a box, they each get 0.5.
Congrats, we’ve created a Markov chain! We can look at various points in the graph and make the following observations:
More interesting observations also appear:
Markov chains do not have any memory, and merely describe “What comes next.” If we start and keep asking “What comes next?” we can trace out some viable routes.
  1. The typical commute
    🎬 → Hoboken terminal → Manhattan PATH → Office → Manhattan PATH → Hoboken terminal → 🏁
  2. The boat commute ⛴️ 😎
    🎬 → Hoboken terminal → Manhattan PATH → Office → Ferry → Hoboken terminal → 🏁
  3. The double shift
    🎬 → Hoboken terminal → Manhattan PATH → Office → Manhattan PATH → Office → Manhattan PATH → Hoboken terminal → 🏁
  4. The “I love trains”
    🎬 → Hoboken terminal → Manhattan PATH → Hoboken terminal → Manhattan PATH → Hoboken terminal → …
  5. The “Nevermind I’ll work from home”
    🎬 → Hoboken terminal → 🏁
There are a countably infinite number of paths my simple commute devolves too, using nothing but the power of probability and forgetting where you came from. Let’s code up some examples.

🔗 Commuting in TypeScript

Here’s my daily commute, represented as a graph on Val Town.
Given an object like { ManhattanPath: 0.8, FerryTerminal: 0.2 }, we can write a small program to choose one of their options, keeping their weights in mind.
Then, repeatedly call this program while traversing our commute graph, starting from START, and stopping once we reach END.

🔗 Lists of characters

Consider the word “keyboard.” We can divide this into a list of letters.
Then, we can represent these as steps just as we did for our commute. Rather than “after the ferry we arrive at Hoboken terminal” we can say “After e we arrive at y”
More interesting things happen if we choose a word like “onomatopoeia,” where from a given letter there are multiple next letters (just as from the Office we could go to the PATH or Ferry terminal).
To build this graph we’ll go letter by letter, but scan the rest of the string to see which letters appear next. We can see how o goes to n, m, p, and e with equal probability.
Let’s write a TypeScript program to do this for us. For simplicity’s sake we won’t keep track of percentages that add up to 1.0, but instead increment a number for each connected letter. (We could do some post-processing if we wanted to)
And we can use this to generate the markov chain for our string “onomatopoeia”
Just as we did with @jdan.randomCommutes, we can generate random strings from this chain.

🔗 Some more words

While these words might be fun to try and pronounce and share some of the same characters as the source word, generating “words that look like onomatopoeia” is not particularly useful. After all, our random element above doesn’t always look like Hydrogen - it looks any of the elements on the periodic table!
To do better, we’ll need more words. We can modify @jdan.markovOfString to operate on a list of strings, registering them on the chain one by one.
Now let’s generate a chain from several different words.
These graphs become harder and harder to visualize as the nodes (letters) and edges (transitions) grow.
Fortunately, our @jdan.traverse function has no problem with graphs of this size.
But the output is… questionable. When we look at a word like mipeipoescrdamaratypescrin, we can find a sequence of letters crd. This is valid on our chain because there is an r which comes after a c, and a d which comes after an r, but the trigram crd is… nonsense. 
$ cat /usr/share/dict/words | grep crd
$
It would be useful to tell our markov chain about these groups-of-three. In particular, for “programming” we can signal that gra is valid because a gr (bigram) is followed by an ra (bigram).

🔗 Letters with memory

Rather than scanning one letter at a time, we can for subsequent pairs of letters. We can make slight tweaks to @jdan.markovOfStrings to accomplish this.
Which, for the same input words, produces a markov chain with a much… “tamer” feel.
Traversing this chain requires some tweaks as well, since for the transition START → pr we want to keep the entire pr, but for pr → ro we only want the o. (A classic off-by-one error!)
And now when we generate 10 words…
🦗🦗🦗
We certainly have more realistic looking words, but nearly all of them are in the training set! By changing letter-by-letter to pair-by-pair, we drastically cut down the noise but left only a few valid paths in the graph.
How can we fix this? We can add more words to the set!

🔗 The Elements

We can ask our dear friend GPT-3 for the chemical elements.
From this array of 109 elements (quite outdated! At the time of writing this in September 2023 there are 118 known elements), we can generate a Markov chain of letter pairs (or bigrams).
Which produces a graph so gnarly it’s quite difficult to visualize.
…But easy to probe. Let’s generate some elements from it.
Giving us elements such as Flutom, Boridium, and Antimonioberbium.

🔗 The order of things

Just as we plotted “Where is Jordan in his commute?” as a series of states and probabilities between them, we can outline the transitions between letters in a set of words. This turns out to be quite chaotic – producing truly nonsense strings just as we produce truly nonsense commutes (who gets off the boat and gets right back on? ⛴️ 😎).
We can tame the madness by mapping out relationship between pairs of letters, or bigrams, sacrificing uniqueness. We can then balance things out by adding many more words to the training set (such as the 109 chemical elements known to GPT-3).
The result are words that are completely made up, but sound reasonable, what more could you ask for?

🔗 Future considerations / exercises

I hope this post was a gentle introduction to state diagrams and the magic behind Markov chains. Here are few questions to ask yourself with your newfound knowledge: