AI code generators

Welcome to installment number ten of applying lean manufacturing concepts to software engineering. 

Today we resume AI code generators  that can write code for the game of associating any movie actor to Kevin Bacon within six movie connections (https://en.wikipedia.org/wiki/Six_Degrees_of_Kevin_Bacon).

Up now is https://workik.com/c-sharp-code-generator.  It generated similar code to the first two, but as an improvement to the last tested program, this generated code included sample data for both movies and actors, so the algorithm was complete, and it was not duped by Meg Ryan and Tom Hanks working on three movies.  However, the design was sub-optimal because one had to supply a list of actors with all the movies the actors starred in, and a second list was necessary to supply a list of movies and all the actors who starred in this movie.

My early hunch is that these three companies must all source their knowledge base from the same open-source software code bases because the generated code is so similar.  None of the companies I have worked for would sell their source code to an information broker where their competitors could buy this source code and place our companies at a disadvantage.

There is a lot of excellent open-source software available, and at the same time, there is open-source software whose purpose is to instruct, demo, or to provide starter code toward a larger solution.   One does not know ahead of time whether one’s search will land one in the range from best to worst.

Lastly, I had high hopes for Google Cloud, at https://cloud.google.com/use-cases/ai-code-generation.  Google has few equals in its reach to websites worldwide. On the plus side, it was the only generated code that used a single data source for movies with the list of actors in each movie.  It ran a pass through the movie list to calculate all the movies that an actor was in.  It had a lot of comments, and it put placeholders in code where data should be added.

On the negative side, the source code had eight lines from some other piece of code randomly inserted into the code.  After removing these lines, the code would still not build. The generated code did not contain a main program to call the method. The biggest error, also seen in the first sample code generator, was that the data structure contained actors, and the code was trying to lookup up movies.

My conclusion from a small sample of AI code generators is that one is better off doing Internet searches to find helpful articles and sample code. In subject areas where one is unfamiliar, the articles on the topic itself can better fill in knowledge gaps than the source code itself with comments could.  Viewing multiple articles with sample code informs one if there is a standard solution or if there are multiple solutions that developers use.

I want to delve into whether quality is a cult, a culture, or both in the next few articles.