General AI thread

[Dai]

Shouldnt the main concern be where the bosses will look when the next round of money saving redundancies are taking place, especially if your department has already been severely trimmed already?

[tiger_rick]

48 minutes ago, ReturnOfTheMack said:

I may be misreading it, but your work replaced some of your team for AI? 

No, you're not misreading it.

27 minutes ago, Dai said:

Shouldnt the main concern be where the bosses will look when the next round of money saving redundancies are taking place, especially if your department has already been severely trimmed already?

I've no doubt that AI is taking all IT jobs eventually. I'm only still around because someone has to tell Claude what to do and I'm the expert on the product. When Claude can work it out for itself, I'll be emptying bins or something.

Long before AI, I've been in more redundancy rounds than I can remember. It's a turbulent industry. 

[hallicks]

4 hours ago, tiger_rick said:

I've no doubt that AI is taking all IT jobs eventually.

I don’t think so, at least not the LLM generation of AI. Bad things have happened to people that gave a chatbot full write-access to their hard drive. It’s only providing you with statistical probabilistic answers to prompts, it doesn’t “think”, but the way the companies anthropomorphise the outputs and “conversations” you have with it are all designed to make you think they do.

I’ve been a bash script dabbler for many years, I’m not a programmer and never will be, but parts of my job require some kind of automation and I can usually figure something out in bash with prolonged stack exchange searching and equally prolonged periods of staring at the screen, thinking “why doesn’t that work?” The periods when I don’t have to look them (because they’re solid enough) mean that I have to re-learn stuff every time I go back to it, but it always comes back eventually and I’m happy not to cede this cognitive surrender on a skill I would like to maintain.

Where I have used Claude was to build an internal-facing webpage that calls a load of bash scripts that I wrote for machine communication and control. I can do basic bits of html and php but anything beyond a Geocities type thing is beyond me. So I got pretty fucking deep into building this thing with Claude, and I hate it. I can broadly understand what it’s doing with the code it makes, but quite often it throws something in that makes no sense and sometimes just fucking bricks the page, and I’ve spent hours feeding error logs back to it to try and get it to come back. And because it’s beyond me, I cannot fix this myself. I am completely at the mercy of this synthetic text extruder being able to figure out its own mistakes, and it’s shit.

Maybe I’m bad at prompting, maybe what I’m doing is a bit too niche and so it doesn’t have amount of good probabilistic data on what I’m trying to do, but I hate it so fucking much and I don’t have the time to immerse myself in JavaScript to bridge that gap. Though maybe I should make the time.

I should also say I tried to use it for three other ideas and it couldn’t make anything that worked despite many tries. So overall, it hasn’t been reliable for me and I hate myself when I do use it and can’t wait for all these cunts to go out of business so I never have to use it again. 

There’s a great article from Rob Horning about how LLMs are changing our attitude to learning:

Why work at learning anything if it appears that there is no arena in which that effort can be put to use? 

LLMs convey the sense that everything is already known and processed. What might be a tool of actual genesis offers instead an experience of completed synthesis, of closing doors instead of opening possibilities, of preexisting links between things rather than the possibility of searching for new ideas (even if it is very unlikely one will find any). Models present “content” drained of the difficulties of discovery, with a synthetic enthusiasm that susceptible users seem to experience as revelatory.

[Sergio Mendacious]

11 hours ago, tiger_rick said:

Warning: Positive AI Post. (Sorry.)

This could go in the work thread, I guess but, I've been meaning to write about this for a few days. I'm an AI sceptic. I fucking hate any AI "art" or "graphics" or any of that bollocks. 

I've been asked to use AI in our dev team for the last year or so. Originally Copilot and then Claude. I didn't really. I used it a token few times (pun intended) for stuff I could have googled but otherwise, I was happier writing stuff myself and working with my couple of other devs to give them work. I told my boss during a meeting where I got a kick-in for not using it "I don't want to work with a bloody robot".

A couple of months ago, both my devs, along with most others in our department were let go. I was furious. Devastated for them, as people, and fuming that our fucking cheapskate people had cut us to the bare bones leaving just me and fucking Claude. If I was in any position to, I would definitely have quit but I can assure you, I'm not. 

Anyway, given the amount of work we had to deliver and the complexity of some of it, I didn't really have any choice but to try and embrace Claude. I can only admit at this point that it has been incredible. That's probably no surprise to anyone who works with it already, or maybe anyone who doesn't hate the AI world but, it was a surprise to me. The quality of the code, the thoroughness of the conversation, the things it thinks of that even I hadn't. It's really brilliant. What I've acheived in the last couple of months has been incredibly reinvigorating for me. Instead of going to work dreading every conversation and every task being thrown on the pile, we're smashing it. We've simultaenously done all of the customer tasks we have to do and also thrown tons and tons of the internal ideas we've had on some board for years that we'd never have gotten around to in a month of Sundays. There's something incredibly satisfying about being able to do things you want to do as well as those you have to do.

I really don't like having a positive view of it. It doesn't really sit well with me but it's the truth. Fuck knows how long it will last. Sure they'll start quibbling about tokens or something at some point and want something cheaper.

Rick, this is very similar to my positive experience. In publishing, I'm seeing a load of people chuck themselves into the worst kind of slop; generating books over a weekend that are absolute shite, proudly making slop covers in AI and then asking illustrators to copy those covers. It's depressing.

My manager is very robotic, so when they said "we need a citizen developer from your team," he said "that's you now." I'm not exactly in a position to say "I'm out, then," so I went with it.

So I went off and built a massive suite of production-ready puzzle generators — they don't use AI to make the puzzles, but the coding to make the apps for people to make the puzzles was all down in VS Code with a codex agent. Made an app to make the barcode and related stuff on the back cover, and one to make file quality checklists.

So yeah, I don't feel super positive about doing it, but I do love being able to make some time saving stuff.