Thursday 26 July 2018

Test Software Like AI

It seems to me that in the future the role of an application developer in many markets will revolve much less around describing the functionality of an application in code and much more in training AI systems. This splits the developer's role into two main areas, creating the training data and testing that the AI is performing tasks correctly. As often the solutions that the AI may come up with are hard to understand and reason about we may not even bother. If we can test it well enough we just test that it achieves all of its goals.

So why don't we start testing applications in this way now, it's just an extension of where we are naturally going with Test Driven Development. If we write the tests from a black box point of view that confirm the application or significant subsystems function as desired. If we test well enough then the tests become much more valuable than the application its self in many ways. The application can be completely rewritten and we can guarantee that it works in as many cases as possible.

It seems very likely that as AI allows us to develop systems much faster and with much less understanding of how the system actually functions that the role of dev-testers will become the main job in a development team. These engineers will still code, building tools to help them test the applications as I believe it is very hard to make a generic test framework for all applications that can function as well as specifically designed tooling that takes into account the domain.

The machines are coming, let's make sure they do what we want :)

terminator 2 thumbs up GIF

No comments:

Post a Comment