Evidence-based Software Engineering book: the last year
It’s now three years since my book, Evidence-based Software Engineering: based on the publicly available data, was released. What has happened in the last year, since I wrote about the first two years, and what might happen in the next year or so?
There is now a Discord channel for discussing evidence-based software engineering. Blog readers and anyone with an interest in the subject are most welcome.
I keep a copy of software related papers that I think might be worth looking at again, and have been meaning to make this list public. A question by ysch, a Discord channel member, asked after ways of checking whether a software paper was worth reading. This prompted me to create a Github repo containing the titles of these 7,756 saved papers, along with some data related annotations. On the more general question of paper quality, my view is that most papers are not worth reading, with a few being very well worth reading. People have to develop techniques for rapidly filtering out the high volume of drivel; techniques I use, and understanding the publication ecosystem.
This last year saw the sudden arrival of a new tool, LLMs. My experience with using ChatGPT (and other such LLMs) as an evidence-based research tool is that the answers are too generic or just plain wrong (for several months, one LLM reported that I had a degree in Divinity Studies). If I was writing a book, I suspect that they would provide a worthwhile copy-editing service.
I was hoping that the recently released GPT-4 vision model would do high quality text extraction from scanned pdfs, but the quality of output I have received is about the same as traditional OCR-based tools. I expect that the data extraction ability LLM based tools will get a lot better, because they are at the start of the learning curve and there is a commercial incentive for them to be a lot better.
An LLM is driven by the token weights learned during training. Roughly speaking, the more training data on a topic, the larger the trained weights for that topic. There is not a lot of data (i.e., text) relating to evidence-based software engineering, compared to the huge quantities available for some topics, so responses are generic and often parrot established folklore. The following image was generated by DALL-E3:
There is a tale of software product evolution waiting to be told via the data contained in magazine adverts; the magazines are on bitsavers, we just need LLMs to be good enough to reliably extract advert contents (currently, too many hallucinations).
The book contents continue to survive almost completely unscathed, primarily because reader feedback continues to be almost non-existent. Despite the close to 500k downloads (now averaging 4k-5k downloads per month, from the logs I have, with the mobile friendly version around 10%), most people I meet have not heard of the book. The concept of an evidence-based approach to software engineering continues to be met with blank looks, although a commonly cited listener use case for the book’s data is validating a pet theory (my suggestion that the data may show their pet theory to be wrong is not appreciated).
Analysis/data in the following blog posts, from the last 12-months, belongs in the book in some form or other:
Some human biases in conditional reasoning
Unneeded requirements implemented in Waterfall & Agile
Analysis of Cost Performance Index for 338 projects
Evaluating Story point estimation error
Frequency of non-linear relationships in software engineering data
Analysis of when refactoring becomes cost-effective
An evidence-based software engineering book from 2002
Perturbed expressions may ‘recover’
Predicting the size of the Linux kernel binary
Local variable naming: some previously unexplored factors
Optimal function length: an analysis of the cited data
Some data on the size of Cobol programs/paragraphs
Hardware/Software cost ratio folklore
Criteria for increased productivity investment
Likelihood of encountering a given sequence of statements
Recent Comments