The essence of Artificial Intelligence is time travel

One of my favorite writers is Matt Levin. He writes for Bloomberg, and has this natural (and funny) ability to explain complex financial concepts to lay-people like me. Check out his (mostly) daily emails and various podcast appearances (here and here) where he explains is writing process. Matt, very famously, exclaimed(?) in his blog post:

The essence of finance is time travel. Saving is about moving resources from the present into the future; financing is about moving resources from the future back into the present. Stock prices reflect cash flows into an infinite future; a long-term interest rate contains within it predictions about a whole series of future short-term interest rates.

The essence of AI (and any data-driven insight or decision) is time travel.

In a sense we time travel when we use data to understand what happened (describe), to explain why it happened (diagnose), to predict what will happen (predict), and to plan against that prediction (prescribe).

The size of data does not matter – it could be very small samples (i.e. from personal experiences) or “Big” data (i.e. a million user click-throughs). As long as we used the data that exists, to inform our knowledge, we’re doing time travel. Isn’t that exciting?

When presenting last quarters sales figures, we often describe the circumstances and results, as if they were unfolding in that very instance. Similarly, when we predict a sales forecast (using a simple regression technique, or an advanced neural net model, we’re using historical data and projecting it out in the future (with some assumptions that are expected to remain valid in the future).

So just to recap 2021, it was a breakthrough year for Tesla and for electric vehicles in general. And while we battled, and everyone did, with supply chain challenges through the year, we managed to grow our volumes by nearly 90% last year. This level of growth didn’t happen by coincidence. It was a result of ingenuity and hard work across multiple teams throughout the company. Additionally, we reached the highest operating margin in the industry in the last widely reported quarter at over 14% GAAP operating margin. Lastly, thanks to $5.5 billion of [Inaudible] small finger by now — $5.5 billion of GAAP net income in 2021, our accumulated profitability since the inception of the company became positive, which I think makes us a real company at this point. This is a critical milestone for the company.

Elon Musk — Tesla CEO from Q4’21 Earnings Call Transcript

As it would be true with time travel, we must remember that moving back and forth in time is not only a little difficult, but we must accommodate for external situations (macro variables) which may have not been captured in the data. See, in the AI modeling world, we like to emphasize on data cleanliness, completeness, and consistency. A model in a computer is much cleaner and simple than the real world because our focus is on building a highly accurate model. But in that endeavor we often lose the details which we believe, at the time, to be less useful. Such as anomalous (very high or low values which seem to skew our results or exaggerate certain results) or missing values. It is very convenient to filter out rows and columns than to dive into another rabbit hole (of why the data is missing… what might have happened to the data – which is another time travel dimension).

In fact, this may be the reason why most data scientists spend less time doing EDA, or exploratory data analysis, than they should. Everyone’s looking for what the model says, so let’s give them an answer… how we arrived at the answer is a conversation for later (which we often don’t get to… until it’s too late).