I do automation maintenance for a large central fill pharmacy in USA. Automated pill bottling, bottle bagging, and bag boxing. Top of the line equipment, or ar least it was last year when the new machines were built. They recently implemented a ChatGPT-Based agent they call Inspector Gadget.
They trained the Inspector with their machine documents, complete with ID numbers to every single data and power cables.
Let’s say a machine goes down, and you don’t know why. You take a picture of the PLC Computer, with all of the indicator lights, it then tells you what part of the machine to look at based on the lights it sees.You can also take a picture of physical parts and stuff, but I dont need/ haven’t found a use case for that. It cant tell the difference between moving or not moving parts lol. Might be good for less experienced folks I guess. The PLC code-referencing certainly saves time over using the manuals, I’ll give it that. It’s not perfect, but does save time when used for what it’s good at, at least.
My most recent performance review was entirely automated. They made a tool that goes through all your CRs (what we call PRs) writes summaries and enters them into a tool. (You are allowed to edit or alter the summaries). Then another AI took all of your code review summaries and summarized them as projects. Then another agent checked the HR documents outlining what an SDEII vs SDEIII is expected to do and a summary was written saying how you did compared to expectations.
As someone who dreads annual reviews (which somehow now happen 2 to 4 times per year) it was a lot easier but it was also very creepy and big brother.
Yes, we have a fancy tool that does all of this on the backend for us. I’m doing a pretty heavy edit on top of it. It was easily confused by a few things that aren’t really in the artifacts like “was this document important?”
My new problem is how to gracefully tell non devs that I have to throw away their 13,000 line slop monolith and start over whenever I get handed software that someone that doesn’t understand a single line of code made.
I keep having this stuff fall into my lap and I try to be kind but people are still getting offended. Such is the life of a slop repair person.
My new job has a mandate. “Use AI” there is no problem statement, or requirements just a solution AI!
This feels like when everyone said cloud loudly and it was supposed to make stuff better.
In my experience it’s just expensive
In other news other than using AI to draft boring stuff like COI requirements and junk for the state contest and some updates I haven’t had time for. My biggest use case was automating my unemployment search requirements so I didn’t have to think about it when I was between jobs. I don’t think I got any actual call backs from anything AI was used to apply to but hey the unemployment weekly reports where easy.
This is so dehumanizing from my perspective, like they don’t care enough about you to be reviewed by actual people. You’re just a business statistic, not a person.
That means welcome to the club. You are supposed to use AI before anything else. Use AI first and only resort to meat brain if the situation calls for it. (The business will likely say the situation never calls for it)
It is absolutely dehumanizing and it’s one stop closer to getting laid off because an AI bot decided I didn’t ship enough code in November or whatever. Like it’s obvious where this train is heading. AI overlord that tracks your “performance” and decides whether you feed your family or not.
Yeah; I am getting jaded. Sometimes the task I have in front of me is so complex that I get discouraged just trying to break it down into prompts for AI. Then I bail and just do it.
Is anyone using AI for large scale applications? I do not mean websites, scripts, dev-ops or the like. I mean real world applications with business rules and complex, dynamic UX? My largest application is a healthcare member enrollment application for managing our members across multiple states (I work for Kaiser Permanente). It is probably over 1 million lines of code to implement UX for all regions of the country. Is this something that AI can grok? Whenever I try to do anything complex, the AI seems to revert to pure html and independent logic. It does not seem to understand how things work together in an application. It is fine for matching keys in one array vs. another array, but if I ask it to do anything that requires understanding the app; I get code that is more trouble than it is worth.
It’s all about 1. Breaking it down into small chunks. 2. Being incredibly specific. 3. Carefully reviewing the work.
Look into something called Spec Based Coding. You’ll burn a bunch of tokens having the agent examine the existing code base directly, make a plan for how to implement the work and have you review it. Then it’ll make a step by step list of individual tasks and you will review that it. Plus an architecture diagram, example code etc. then the final work gets reviews too.
It’s a structured more rigid way to use AI for programming compared to “vibe” which will not work for big or complex projects.
I’ve been using a tool called codesight to reduce the amount of stuff the agent has to read when examining the code. That, with Rust Token Killer and the Superpowers plugin have been my go-to for coding.
Oh yeah, and the Caveman plugin for terse responses.
