software is not abandoned anymore – part 2

During my MSc, I did a lot of reading on Event Calculus (this book and this book) which lead to software like the decreasoner from the author of the second book. The Discrete Event Calcuklus reasoner is written in C (with support for Cygwin) and Python2. So, once again, armed with my good friend Mistral and GLM-5.3, I asked it to perform the following tasks

  • Modernize the C code and support only Linux (because WSL2)
  • Migrate from Python2 to Python3
  • Work with minisat only
  • And finally reenable support for relsat

This seems to work fine and serves as a good blueprint for anyone wishing to work with this corner of Artificial Intelligence these many years after the books were written. And this is why I mostly did this exercise. It seems I can release the code for the next person to pick it up and I might do it in due time.

Relsat’s licensing on the other hand is unclear to me. Which is fine, since it sent me down a new rabbit hole, reading the initial paper [pdf] and wanting to have a new cleanroom implementation of this. Something like using a not so capable model, where I actually read the paper (instead of the model) and really direct it what to write (codestral I am looking at you).

The above thought about codestral sparked yet another thought to me: Maybe this is how coding should be taught: Using an inferior local model so that the student has the time to develop the capability to understand what is being written before moving forward to the industrial stuff. Kind of as with literature: not everyone can be an author (and certainly not one producing in volumes) but everyone can be taught to read and evaluate what has been written.

We’ll see…

Software is not abandoned anymore

At work we make use of hound code search. It is plain simple and fast. It is also virtually unmaintained for 3 or more years. So when some of our developers came to me with some questions (“Can’t it do X?”) my initial reply was: it is what it is, we cannot do much.

But then, I thought I should make some more use of my (work) Mistral subscription. So I switched to GLM-5.3 and (using pi) I did the following things:

  • I asked it to drop support for any repository that is not git (bye bye svn, etc)
  • I asked it to modernize Go (there have been a number of security releases for Go tooling)
  • I asked it to make it a multistage build
  • I asked it to make it a rootless container
  • I asked it to implement the features our developers wanted

Of the five tasks above I could easily and quickly implement the three. I would want to implement the fifth, but it would always take the least priority possible, so never and I would never deal with dropping non git repository support. It would also require better Go proficiency from me than I currently have.

Yes, within an hour we got our own fork that, for now, nobody else is using. But for this kind of software I am willing to keep working with our version. We could have switched to some fancier software, but that would come with its own operational tax – quite different from the fire and forget setup for hound (plus a Jenkins job that recreates the configuration each time a repository is added).

So that’a a true win for LLM maintained abandonware. The scope is clearly framed by the source code and the documentation. The added feature, if well defined, can be implemented without much hassle.

And life goes on.