mathieu@laptop:~$ fortune
mathieu@laptop:~/blog/engineering$ cat can-you-think-big-enough.txt
What started as an autocomplete on steroids has evolved to an almost fully autonomous terminal that does most of the work and I can't help but think: are we thinking big enough?
I've been developing almost 30 years now and there has always been a constant: This can be done better, more efficient and faster. And this trend, that most of the developers have experienced, leads to smarter tooling and better standards. Dreamweaver and Frontpage came from the same necessity as better W3C standards and adulting programming languages.
And I do believe that AI (talking LLM here) is following the same constant. And while I do still have (many) issues with it, I can't ignore the fact that after 29 years I haven't wrote as much code anymore than I ever did. But looking statistically, I've done more in the past 2 months than I have in the past 2 years. And it doesn't seems to slow down. "Am I thinking big enough?" and, hopefully not, "are we rushing towards a certain abyss?".
Lets look as some examples.
We had an issue with our, almost-15-year-running-and-still-counting, application. Little background: This application, one of our CMS, is still maintained and holds code from 15 year ago. I know this code by hearth.
Back to the issue: from the description of the bug, I knew this was a difficult one, at least a bug that was deep entangled and might touch many different layers of the application. I know where to look, how to test it, but I know it might take a whole day. So I asked Claude first to help me figure it out. Not only did it find exactly the issue: it fixed it, tested it and added a direct solution to the 'problems' it created on a production database, all in 10 minutes.
But this is not thinking big enough. Is it? I ran the exact same problem but now without my guidance: I just gave it the description of the issue that was reported. And you might guess it: it found and fixed almost exactly the same situation. We all can fill in the gaps here.
Another example.
LLM excels in language. I let Fable run it against 300+ issues in a project, cross-reference all issues and search for overlapping and duplicated issues. Then I merged them and let Fable hold each issue against the current code database, closing and tagging each resolved task. The ones that are left are categorized, summarized (some are results of multiple tasks) and pre-prompted for AI agents.
Agents can than individual pickup these tasks, create mockups (if needed, this may be optional), and implement the tasks. Each tasks gets it's own video/screenshot with a hand-over, ready to be validated.
This is not science fiction, this is real implementable, testable features.
But is this thinking big enough?
mathieu@laptop:~$ fortune
