Haeuserschlucht

What was the last time you realized that what you did was actually a great accomplishment and you didn't even realize it at that time?

I discussed about it in length and while it was much better following than the average participant at a forum, it was still not really helpful, and it took a long time to be somewhat decent. As LLMs are full of billions of books and articles, does that mean that most books are bad for dating advice, even indirectly?

I propose a theory. My theory is that LLMs do not require a lot of computing power. What they do require is a small computer, but it's blown out of proportion with unnecessary information to artificially increase demand for high power hardware. So the reality why models are the way they are, especially local models, is because they actually suck at optimizing on purpose. As I said, it's even added to complexity to increase sales of hardware.

What are your pros and what are your cons? And do not refer to secondhand information as in "well I was told" or "well I read a paper".

AI studio is pretty much a paywall now. Which alternatives do you use?

Have you ever been able to code in the first place?

It started like this. I programmed in Pascal when I was in school. I learned basic. I learned assembler. I literally assembled code. Sometimes I thought it out, planned it, and created for that time quite elaborate code, controlling a panoramic head, for instance, things like that. So at this time, I was really proud of myself. I was good at this. I was getting better.

And fast forward to today, you have AI and vibe coding. Back in the day, often when I got an error message, I just looked on the internet and tried something out, did trial and error. After a while, it worked either with my own code or with code from a GitHub repository.

Today I don't even do that. All I do is I start the Gemini CLI and have it code for me. And then I watch some video unless it has some issues. I noticed that the code is better and it all comes down to good description of the problem.

I noticed the problems that I fought with for ages, that I spent hours with debugging and shit. It just created in 15 minutes. So I was wondering, was I even ever able to code? Or was it just a huge waste of time?

Because now I can see it from the outside and see the amount of time that I would have used to debug this and to write it in the first place, then copy paste it from other sources, which might be outdated. I would have done something for five days.

Partly I would be excited about this. Yes, sure, it's challenging your brain, but ultimately you have something and you're probably not even that excited about it.

Have you ever been really able to code? Or was it more like trying something, it doesn't really work, you ask in a forum like Stack Overflow, and if that still doesn't work, you just do something else?

I had Pascal in school, later at university Java for the first time. I programmed a little afterwards for Android, given that it was also Java. But often it was Stack Overflow here, Stack Overflow there, trying to match my source code with what I found online, looking at manuals or whatever. It was tedious. Just tedious.

Ultimately I had something that was fun for five minutes at best. Often I just wanted to see if I liked something, but for that I had to create a prototype. And this prototype didn’t work. It didn’t compile.

I saw some repository, wanted it to work, then there was a compilation error. This module was lacking. This dependency wasn’t there. All of this shit.

So I’m coming back to it. Was I ever able to code? Or what is the benchmark for coding? And is it really a disservice to my intellect if I stop doing that and just have AI create it?

There are changes all the time. Repositories are not up to date. Libraries don’t work with each other. One updates, the other doesn’t. Then you have to fix it. It’s depressing and annoying, and I don’t really see the drawback of doing it differently now.

If a repository doesn’t do what I want, I load it, start Gemini CLI, and have it corrected. Reverse engineering protocols or hardware, in my experience, is disgusting. Really hard and frustrating.

So what is your take on the whole situation?

Entering [flagged] in search did not yield results.

I'm asking because it seems that nobody really does that. Yes, there are some projects here and there, but ultimately everybody just jumps over to cloud LLMs. Everything is cloud. People pay for GPU usage somewhere in the middle of nowhere. But nobody really uses local LLMs long term. They say, "Well, it's so great. Local LLMs work on small devices they even work on your mobile phone."

I have to say there's one exception for me and that's Whisper. I actually do use Whisper a lot. But I just don't use local LLMs. They're just really, really bad compared to cloud GPUs.

And I don't know why, because for me it seems that having a speech-to-text model is much more challenging to create than just a model that creates text.

But it seems that they really cannot remove the differences and have it run on consumer computers. And so I also go back to cloud LLMs, all privacy aside.

I'm curious, how did you find people that you can spend time with and do something that is not productive, that is just there to be fun, that is just there, that doesn't demand anything from you?

You're in a group where you are not accepted because you accomplished something big, but just because they like you, for being you, for being authentic, and where you just have your peers where you can go to and you just feel great.

And I'm wondering how to find that because the general advice is like, well, just go to some gym, or just go to a public place. But what they don't take into account is that in a public place, you do not communicate with people. You don't. First of all, you don't in general. So if you initiate contact, it's rather weird for them that you do because they don't know you and they are tremendously suspicious just because they don't know you, not because they have a real reason, just because they are conditioned that way.

And they are quickly asking basically, where's this leading to? What is the purpose of this conversation? When you say, well, I just want to talk a little, they say, okay, I'm not interested.

And they give you a weird look and they run away and they tell their friends, well, there was this weird guy that just wanted to communicate. What a weirdo.

So like all these suggestions, they are all about accomplishing something. So let's say for instance, you just wanted some person, not on the internet, but in real life, where you could just gossip. This basically describes it perfectly. Even though I'm not interested in gossiping itself, but I'm interested in being respected without having to do something for it as in have a certain degree or have written a book or any of that, but just being respected, being appreciated.

When I say something, they enjoy seeing me, they enjoy spending time with me. They just think that I'm a great person and they are respectful.

And this is just not happening. And that's the thing. It's just not happening in those places like in a gym, in my experience. And I don't think anyone has a different experience really. They are just bad places because especially a gym, especially even a FabLab or something like this, they are all places where you accomplish something.

Allegedly, they are there for communicating, for finding new friends, but go fuck yourself. You don't find new friends there. What you find is perfectionists who value you less if you are not a perfectionist yourself and you are only measured by what you accomplished.

So I am a self-taught programmer and I have my way of creating programs. So it's easy to read, to comprehend and then I made the mistake of giving a part of this code to an LLM. And it turned out that it shortened my code to about 10% of the lines that I used and the code is now much more performant. At first I was psyched but then I thought I put so much effort into that. Years of my life went into being able to code like this, I used to be proud of it, and this LLM just does it in a fraction of the time. Just like that. Better, quicker. It now feels like it's not my code anymore. It's not my idea anymore.

Yes, it threw me into a small existential crisis.

Can you relate?

PS: I'm certain that every artist right now and every voice actor can relate.

So you want to have private documents analyzed by a cloud LLM, but not the private data shared that is inside, like names or very private details. What do you do? Is there a github repository that acts as a mim between AI - website and your computer and redacts everything that is being uploaded?