Over the last weeks, I have been on the road a lot for various engagements with customers as well as for our research activities. In virtually every meeting, the whole area of artificial intelligence and especially machine and deep learning come up as a discussion topic. This is great as I think the whole AI/ML/DL area is incredibly exciting and I keep being surprised and impressed by the incredible applications and examples that make the headlines on a very regular basis.
It is clear from all the discusses that AI is on the top of hype cycle and this is concerning in that the expectation on what AI will deliver are perhaps inflated. Especially those not overly well versed in the concept and underlying technology often start to expound on the fabulous opportunities that their products and services have if we just sprinkle a bit of AI dust over them.
Even worse are the cases where individuals start to talk about their expectations in terms that in no uncertain terms would require “General AI” rather than the “Narrow AI” technologies available today. Inflated expectations are not helping anyone and will only lead to disappointment.
The challenge is that we are at a stage in society, as I discussed in last week’s post, where we’re moving to the “post-intelligent design” era. In ML/DL, as humans we are building systems that are building (or rather training) systems that accomplish incredible feats, but we don’t actually know how these systems work. This is a major departure from the engineering approaches over the last decades and even the last centuries.
Over the last weeks, I have been writing about the software engineering challenges associated with building AI systems, but it seems that the key message that I am looking to get across is broader than what I have been communicating so far. The key idea is that AI is not a silver bullet. AI is not going to magically solve problems without any significant investments from our end.
For a machine or deep learning model to work well, we need accurate, clean and typically labelled training and validation data, a well designed model, iterations with several alternative designs to figure out which model performs best, reliable data pipelines to hook the model to, monitoring and logging to track model performance during operation, continuous remodeling and retraining to support continuous deployment, etc. As I’ve been outlining here, here, here and here, building production-quality ML/DL systems requires a solid engineering approach. In that sense, these technologies are tools in our toolbox rather than silver bullets that magically solve global warming, poverty, inequality and everything else that ails the world. This also goes the other way: it doesn’t make sense to blame AI for everything that goes bad in the world either.
Concluding, despite the hopes and expectations of lots of people that I meet, artificial intelligence, machine learning and deep learning are no silver bullets. These are novel technologies in our toolbox that help us solve problems that we were unable to solve earlier or at least solve as well. But, as the saying goes, there is no free lunch. Achieving success using AI/ML/DL requires engineering, discipline and operationalization. Although any advanced technology may seem indistinguishable from magic when looking at it from the outside, as Arthur C. Clarke – the famous science fiction writer once quipped, on the inside it’s typically heavyweight engineering. So, apply AI to your heart’s content, but remember that, as Thomas Edison mentioned, that it comes dressed in overalls.