LLMs... Do you use them?

[broccoli]

LLM chatbots are basically instant gratification machines, with all the benefits and detriments that entails.

I use it to decipher the abbreviations in doctor's notes that are difficult to google. For image generation, it's just a novelty, something that I might say "How funny that a computer could make this" to but nothing that I would want to look at seriously. With fiction, it's useful for making me feel confident in my own ideas because it just repeats my ideas back to me, but then when it offers to write those ideas I find the prose severely lacking and prefer to do it myself. For code, I prefer to give it something I already wrote and ask "What went wrong here?" because it is good at saying something like "You accidentally deleted a comma on line 3 million" that I would never have caught. I tried to get it to write code for me out of nothing and ended up having to fix it so extensively that the end product was entirely my own work...

That being said I think LLMs are pretty good at clustering documents based on context in the text so at work it's very useful for determining in 600,000 doctor's notes which ones had the provider actually discuss the screening questionnaire and which ones the provider performed the screening questionnaire and then didn't follow up on the results, because everyone writes it a little differently and it's nightmarish to regex it and NLP performs slightly worse.

I'm really into algorithmic composition and generative art but the current state of AI-generated stuff is such that I actually wish it had a little more of the machine's touch instead of falling into that weird uncanny valley where it tries to look and sound realistic.

It's just a (inconsistently useful) tool and novelty for me. I couldn't imagine using an LLM as a friend or therapist or lover the way some people online do. Guess that's part of the instant gratification machine thing though.

[Capybara]

Update on my job's AI assistant: It started out useful, but is now shitting itself and nothing comes up for basic searches. (To keep it vague, I work at a store and use it for inventory.) This is making me mad, because I've been on the boat of "AI is flawed but can be good if regulated camp"... until now. Are models hallucinating en masse now? (More on that later)

I really like your reply, and the issues regarding mental health services and how they can make AI chatbots tempting are something I hadn't thought about. I will say that «AI therapy» is something I completely disagree with, as most chatbots tend to just confirm what the user already believes- I don't think therapy is for everyone, but AI can be especially dangerous to those who are prone to magical thinking.

I should've mentioned the AI-induced psychosis that comes from people using it for emotional support, but I figured everyone knew that since there's a new story constantly  I stay quiet about this because I do think OpenAI (etc.) should be sued and change their models for this, but I've also had human therapists do these exact same things with enabling unhealthy behavior. Well, I guess it had to be trained off something...
Besides using AI to "talk to", AI therapy is dangerous for medical advice. (!?!??!) On a Dreamwidth community I reluctantly browse, someone mentioned using ChatGPT for therapy and medication advice... No. NO!!!!

Side note, but I decided to enable Visual Studio Code's built-in Copilot and it's... Really really bad! I've been going back to edit every page for grammar issues, and it literally hallucinated (?) an episode of a TV show that doesn't exist. It's also made up historical events that never happened. My experience using ChatGPT and Grok are similar, they'll claim random bullshit and then go "Oops! Sorry!" when I correct them... I can't believe this... I was hoping VSC Copilot could be an accessibility tool for making coding easier, but it's creepy to see it suggest edits pretending to be me.

I might disable it, but has anyone had decent experiences with VSC Copilot as a tool? I tried to make it fix a broken layout, and it broke it further..... 

[fablefound]

well, folks, i regret to announce i had to use an ai chatbot to make something at work. and it was spectacularly infuriating and incompetent, but i can also see why people think they're so convenient and get straight-up hooked. you can ask these things, over and over again, to explain and zoom in and zoom out on problems. even if it has no idea what it's talking about, it'll hype you up and praise you for asking questions and try again and again.
didn't like it one bit :/

[devils]

Update on my job's AI assistant: It started out useful, but is now shitting itself and nothing comes up for basic searches. (To keep it vague, I work at a store and use it for inventory.) This is making me mad, because I've been on the boat of "AI is flawed but can be good if regulated camp"... until now. Are models hallucinating en masse now? (More on that later)

They do seem to be hallucinating more frequently now, I've noticed that too! Or maybe people just got better at manipulating the machines. Regardless, genAI does need to be regulated at the very least. It's burning the planet, and each day I'm turning more and more against genAI and LLMs.

i can also see why people think they're so convenient and get straight-up hooked. you can ask these things, over and over again, to explain and zoom in and zoom out on problems. even if it has no idea what it's talking about, it'll hype you up and praise you for asking questions and try again and again.

I think the reason they are programmed to hype people up is precisely so people get hooked on it. The more people get hooked on it, the more they're willing to give them data and buy their subscriptions. It's sad, really. OpenAI et al are using technology that could make humanity better (yes, even in creative fields) to just get as much money possible while damaging our species and the world at large. [elongated sigh]

[st3phvee]

Side note, but I decided to enable Visual Studio Code's built-in Copilot and it's... Really really bad! I've been going back to edit every page for grammar issues, and it literally hallucinated (?) an episode of a TV show that doesn't exist. It's also made up historical events that never happened. My experience using ChatGPT and Grok are similar, they'll claim random bullshit and then go "Oops! Sorry!" when I correct them... I can't believe this... I was hoping VSC Copilot could be an accessibility tool for making coding easier, but it's creepy to see it suggest edits pretending to be me.

I might disable it, but has anyone had decent experiences with VSC Copilot as a tool? I tried to make it fix a broken layout, and it broke it further..... 

This is one of the main reasons why I discourage new / hobby web developers from using LLMs to help with coding solutions. LLMs are incapable of actual reasoning, and their algorithms make incorrect predictions on a regular basis (I hate calling this phenomenon "hallucinating," as that's an anthropomorphism that implies these systems have actual intelligence -- they don't). I would never trust Copilot, Claude, or any other LLM to fix something that I don't already know how to fix. I'd rather learn how to do it myself, by reaching out to human coders who actually know what they're talking about.

[fablefound]

I think the reason they are programmed to hype people up is precisely so people get hooked on it. The more people get hooked on it, the more they're willing to give them data and buy their subscriptions. It's sad, really. OpenAI et al are using technology that could make humanity better (yes, even in creative fields) to just get as much money possible while damaging our species and the world at large. [elongated sigh]

EXACTLYYYYYYY they're going straight for our dopamine and trying to get us in the same nasty feedback loop as every other algo fine-tuned to maximize doomscrolling. more data, more engagement, more money. also more servers and more environmental damage but they don't care.