My Big AI Successes
All but invisible upgrades.
Everything in the world is AI enhanced these days. I’ve been poking around with ChatGPT and I actually have some successes to report.
Attention Conservation Notice:
BLUF: AI is great for enhancing familiar tasks or working in areas where one is already well grounded. Completely new stuff will get you lost in LLM hallucinations.
Unix:
This is my 40th year of using Unix. The Internet was just receiving DNS, while both BGP4 and the web were eight years in the future. There are Unix tool chains that just shoot out of my fingers, I don’t even have to think about it. The two big AI values I’ve found are learning of new tools, or improving the manner in which I use existing tools.
Among the new tools I’ve discovered there’s seq, which will print a sequence of numbers. This is super handy for scanning a range of IP addresses, or concocting a way to step through a bunch of web pages for preservation. I have long used tmux, and screen before it, but you’re stuck with serial processing of batch jobs … unless you notice parallel. You can cat a file of commands, pipe it to parallel, and set it to use however many cores you want to dedicate to the work.
The Unix find command will take a file specification, search for it, and either display it for you, or apply a tool chain to it using the -exec option. I used to do an unadorned find followed with greg, awk, sed, etc. Since I’ve had ChatGPT handy to immediately offer a solution to things I can’t do from memory, what I can do from memory has expanded quite a bit.
I used to screw around endlessly hunting for stuff in files in a directory hierarchy. This is a very slick alternative.
find . -name “whatever* -exec grep -niH {} \;
And here’s a thing I made this morning - a command I named 24 that shows all files modified in the last day, ignoring any log files. Since I was looking to visually review I added a linefeed between each one. Simple, but it gets used often during the day when I’m slogging through my PHP duties. I think I’m saving … an hour a day, maybe.
find . -type f -mmin -1440 -exec ls -l {} \; | grep -v "log" |awk {'print $0 "\n"'}
Programming:
ChatGPT wasn’t a very good programmer when I gave it Python tasks, but it’s a stickler for the various formatting rules, which is nice. What I found it to be really good at was offering samples of invocations of unfamiliar APIs and libraries. I’ll do the overall logic, but please show me how to make JRandomFunction() work. It’s also pretty good at coughing up a list of packages that will address a certain problem, comparing/contrasting them in an easily understood fashion. It’s like … fighter plane entries on Wikipedia … everything you want is in the same place each and every time you look.
My programming career involved Pascal (1985-1997), Perl (1997-2012), and Python(2012-2025). Last December I was confronted with a complex problem involving some data PHP code. Since I am good at declarative programming and used to parsing HTML for data scraping it’s been less awful than it could have been.
The recent upgrade of ChatGPT to include a Reason option. I am not sure what the underpinnings are but it’s much more limited in terms of how often you can use it, but it’s a MUCH smarter service. Yesterday I fed it a hundred lines of vexing PHP spaghetti and it offered three things to check. All three made sense upon reading and the second one hit the nail on the head. This sort of learning would have involved a solid weekend with a PHP book when I was starting out, and I’m well past the age of reading once and instantly absorbing things.
Evaluation:
If you want a new tool and you Google, there will somewhere between not one and $uncountable results, depending on what the problem is. If the thing is freemium or paid there will be “listicles”, the top five whatever widgets. Sometimes it’s OK, mostly it’s just miserable. If there’s a product that has a large cash advantage they’ll squelch the competition. Asking ChatGPT for a top five to top ten list summarizes the responses. It’s still skewed but at least it’s all in one place, instead of trying to get details out of a dozen freeform posts.
It helps with hardware evaluation, too. If there are thinly trafficked areas, say driver issues in audio on certain miniPCs, the LLM will offer a full palette of possibilities. Trying to find something using Google is always an exercise in frustration - the people who write on such topics often solve the problem for themselves, not for any other reader. This means a lot of unstated assumptions don’t get memorialized.
Suggestions:
Start using an AI assistance with a topic you know well. My very first time I was trying to ID an inter-war cruiser from a grainy photo. During that process the LLM insisted that Iowa class cruisers (they’re battleships) had a second turret lower than the first (they’re superfiring, navy for above and behind), and that this was done to the guns could be depressed to hit close range targets(!) I also wanted it to help identify the car detective Michener was driving in Penny Dreadful: City of Angels. We knew it was a Packard by sight, Michener said it was a V-12, and we knew that the story was set in 1938. I came at this question a variety of ways, and eventually figured out it was a 1937 model on my own. The LLM never did converge, it seems to be a Pontiac fanatic.
Those are fun examples, but when it’s work things like that are decidedly unfunny. Those systems will hallucinate like crazy at the drop of a hat, if you’re doing something complex and new … fuggedaboutit.
Conclusion:
No LLM can do what I do (yet), but I’ve got one doing what an entry level intern interested in Python would do. That worries me … I got where I am by noticing the tiniest bit of daylight here and there, then slipping through those cracks. That is going to get harder and harder for a new person … this isn’t going to improve our demographics problem, I think it’ll make it worse.
As the super senior people age out … no bench of potential replacements. You want to see what that looks like end stage, check out Russia’s oil industry. They basically stopped training engineers when the Soviet Union collapsed thirty five years ago and now even the “smart young guys” are pushing sixty. Russians drink hard and die young … and even if they can make peace before Ukraine Long Neptunes all their fractional distillation towers … wouldn’t put my money on their production.
