Software that explains itself

The Philosopher Developer

October 14, 2019

My first job as a developer was at a company called MP Capital, an algorithmic trading firm in Philadelphia. Our boss, an experienced trader, would present models to us and we would do our best to code them. Sometimes he would question why our models resulted in certain particularly risky trades. To help him understand cases like these, he asked us to build a feature that would provide a breakdown of every trade: essentially a record of the algorithm that triggered it, every step of the way. At the time I didn't appreciate just how unique this was.

At Google, I remember the search team had something similar: a tool that could provide a rule-based explanation for why a given page was listed for a given profile at a given position in the search results.

More recently I've been working on a side project with a friend that involves scanning documents. The code is fairly complex in some areas; at times it's unclear why it produces certain results. Again, I've found myself wanting the same functionality: not just a result, but a result accompanied by an explanation.

There is no pit of success here. By default, software does not explain itself. The paradigm of every programming language I've worked with requires the programmer to define the inputs and outputs. Anything the programmer left out—e.g., what was the previous value of this local variable, did the loop exit early, was an error raised and handled—is hidden away.

In this paradigm, visibility into the why? of software is a feature that the developer has to build. Responsible engineers add copious logging to their code to capture the processes happening internally as software runs. But the usefulness of logs is limited by the foresight of the developer writing them. It's also limited by how easy they are to access.

One of the biggest fears1 with computers is that one day we will develop advanced artificial intelligence whose behavior we can't understand. I can all but guarantee this will happen if we never take the time to build the capability for AI to explain itself. Hidden-by-default should not be an option.

Imagine a new software paradigm where nothing is hidden. Where every output of the software is accompanied by a token entitling me to a full explanation of how that output was produced. Or better yet, where the explanation is somehow part of the output. What would that look like?

The models we're using today—e.g. machine learning, which requires large volumes of data—almost ensure that we won't understand the decisions made by artificial intelligence in the future. If explanation were baked into the tools we used, how might that change? What would we allow ourselves to build, and how much sooner might we notice when we're headed in the wrong direction?

The proverbial frog allows itself to be slowly boiled2 to death because it doesn't understand what's happening. Give the poor thing a thermometer it can read.

  1. Incidentally, I wrote about this almost a decade ago

  2. My understanding is that this is not actually a real, demonstrable phenomenon. I still think it's a useful metaphor that rings true for the human experience. 

Did you enjoy reading this? Buy Me A Coffee