The Importance of Human +?Machine

At some point last year, it became popular to make the claim that AI was going to take all our jobs. Personally, I feel as if that trend has died down a bit. Though there’s still some trepidation among many people, it seems like the limitations of AI are much better understood right now than they were even a year ago.

Recently, I’ve seen a shift in the conversation around AI away from how machines are going to rule us all and towards where they fall short. Among AI practitioners, this conversation has naturally turned towards what we can do to address some of those short-comings.

As someone who’s working in an AI-first startup, though, I’ve never really expected AI to take over the world within the next couple of decades. Though I have high hopes for the future of AI, I still think we’re a long way off from seeing it perform the types of generalist tasks that we would need it to in order to classify it as intelligent enough to reach human capacity.

Rather, what we know at Apteo is that the future of business and technology, at least in finance, is going to benefit from some combination of man + machine. Right now, ML and AI are really good at finding subtle patterns in data (sometimes you’ll hear the term “latent features” here).

But using the patterns they find requires a certain level of subjective analysis. Patterns that existed once or twice in the past, which may have significantly affected the learning trajectory of an ML algorithm, may not exist in the future. In finance, this is immediately evident. Quant strategies gain and lose efficacy all the time. Things that have worked for a long time suddenly may stop working when there’s a change in the underlying economic state of being.

We’re still living in a world where human subjectivity is better than cold, objective pattern-finding in many areas (not all, since some industries and areas of application are truly numbers-first). Understanding how to apply complicated predictions can be just as important as generating those predictions in the first place.

The way we do this is to analyze where our predictions are good, and where they aren’t. We tend to do better with companies where there’s a copious amount of data and where there aren’t too many hidden or unsurfaced business risks. That means that when we get a highly positive prediction that doesn’t fit into this category, we need to take a closer look at it to understand where it may fall short, or what data from history may have caused the generation of that prediction.

As time goes on, we’ll likely need to do this for fewer predictions and fewer companies. The inclusion of additional data, both in terms of number of instances and different types of data, will help to mitigate this. In fact, I’m actually a big believer in not getting in the way of a functioning model.

But until we do have fully functional models that aren’t biased by bad input data or missing data, it’s important to monitor model performance and use human subjectivity where it’s needed, and that’s going to be the case for the foreseeable future.