Anonymous A started this discussion 3 weeks ago#134,927
I’m seeing AI pilots all over the place. Most organizations don’t have a clue so they democratize tools, leading to people recreating the exact same prompts and process flows. All of which eat up tokens faster than an ADHD kid at Chuck-e-Cheese.
Then some people start running these flows against APIs, harming license agreements or causing security breaches. Someone at my old job set up an LLM assistant that deleted all of his work and his hard drive when he was fired.
In other words, executives fear they’ll be left behind. Employees attempt to manage old and new workflows. While some cost savings may be found these tools become expensive as shit, leading to cost offsets elsewhere in overhead. While human verification weakens some of AI’s ROI and efficiency promises.
Not to mention the LLM and cloud supply chains have been compromised multiple times by some of the most creative worms ever programmed.
Anonymous B joined in and replied with this 2 weeks ago, 2 hours later[^][v]#1,435,084
You can run local LLMs if you are concerned about sensitive data.
A quarter of my work time was replaced by a script that reads tickets into an LLM and submits a form to the company server. Yhat will make up for any wasted time on experiments.
Anonymous C joined in and replied with this 2 weeks ago, 1 hour later, 4 hours after the original post[^][v]#1,435,089
@previous (B)
Problem with AI output is you gotta have a human go over it with a fine-toothed comb, and really at that point you might as well just hire the human to do the work and never waste time/money setting up the LLM.
Anonymous B replied with this 2 weeks ago, 34 minutes later, 5 hours after the original post[^][v]#1,435,094
@1,435,089 (C)
Not really, I checked my script manually when I started to ensure it wasn't making mistakes, but it was still much quicker. After a few days I just stopped checking and it's worked fine.
There are millions of jobs that are menial and repetitive and an LLM can do it reliably.
There's new tools that make it easier, double-check the work, and handle more edge cases every day.
I remember educated people telling me driverless tech would never work or catch on a year after my boss and colleagues had FSD and had it drive us around. Right now people insist llm coding will never replace cs grads, but I've seen tech averse seniors spend hours talking with bots to perfect some niche app they had an idea for.
So many hivemind luddites are convinced a tech won't work when its been pumping out product for a long time. Or calling it slop like what came before wasn't soulless assembly line corporate output from deadeyed human drones.
> You can run local LLMs if you are concerned about sensitive data. > > A quarter of my work time was replaced by a script that reads tickets into an LLM and submits a form to the company server. Yhat will make up for any wasted time on experiments.
Why would you need an LLM for that? Couldn’t you just use OCR?
Anonymous B replied with this 2 weeks ago, 23 minutes later, 6 hours after the original post[^][v]#1,435,110
@previous (Mr. Black Boi)
It's not OCR. It takes client emails, confirms info, puts the info in a two part web form and then sends them an update when that is processed.
The script is mine, but the company also uses a voice bot to answer calls. Customers can get a live person by pressing zero.
No one has been laid off, but ai has automated a lot of work without any problems.
Mr. Black Boi replied with this 2 weeks ago, 7 minutes later, 7 hours after the original post[^][v]#1,435,111
@previous (B)
You said tickets so I thought you meant literal tickets. Reading emails is a good use of an LLM, although I’m not sure that was something that was impossible before LLMs. Large tech companies were automatically scanning emails before ChatGPT, like I remember my phone would always automatically add any trips to my Google Calendar if I had a flight or something.
Mr. Black Boi double-posted this 2 weeks ago, 4 minutes later, 7 hours after the original post[^][v]#1,435,113
On the topic of code though, I’m in disagreement there. I’m a CS major and I’ve had to work with other students in hackathons and in projects in my courses. I’ve noticed there’s this issue where when people don’t really know how to code, and they use AI, AI can make nice user interfaces. But if you don’t understand what’s going on underneath it can cause issues. For example, there was one guy who didn’t know the difference between writing a for loop to find every permutation of 2 items from a list vs every combination of 2 items from a list, which is a difference of O(n^2) vs O(n(n - 1)/2), which is a 2x difference in terms of performance. Or a few times I saw people try to make web servers using an LLM, which doesn’t work if you have no idea how a web server works so they would have a bunch of emojis in the GitHub repo and they’d be installing random stuff and then when I’d ask them how to run it they’d have no idea and be trying to make some super complicated design while not knowing the difference between a get request and a post request. I think it is possible to completely replace programmers with AI and have a product and release the product, and then realize it’s horribly broken underneath and have no idea how to fix it. I think that’s more likely.