Ask HN: How are you avoiding LLM anxiety with your side projects? | Hacker News Reader

Ask HN: How are you avoiding LLM anxiety with your side projects?

26 pointscandiddevmike35 comments

I'm having a really hard time working on new, non-LLM related side projects due to anxiety that everything I'm doing is effectively worthless. LLMs have killed any enjoyment I had with programming and solving problems, as I keep jumping ahead and seeing everything being reduced to "folks will eventually just ask a LLM to do X, why bother creating something to do it".

View on HN

> I'm having a really hard time working on new, non-LLM related side projects due to anxiety that everything I'm doing is effectively worthless.

Is there really only joy in programming stuff if you think what you're doing is valuable to others?

I thought it was the opposite: because I don't need to be reliable with a side project, then I can take the time to do things in a slow or impractical way, as tidily as I like, or do it however I want to. -- It's a chance to play with whatever.

I'm finding it quite fun to play around with non-LLM stuff.

I think it's also quite fun to play around with LLM related stuff, and to see what it can do, and what it can't do.

Spend a couple days trying to get an llm to do the project you want to do and you'll see that you have nothing to worry about.

chefs kiss

It honestly really depends on the project...

I wrote a police intake form site for an Indian reservation with multiple pages and local storage and reports in a couple of hours with GPT4. I think someone who is highly experienced with web dev could do the same in similar time but personally it would have taken me hours and hours to get the CSS right but GPT4 just spat out good CSS with minimal prompting (this is with February GPT4 - modern GPT4 does not give you nearly as good of output anymore because they lowered the amount it try-hards on what you ask)

I wrote a python TK app to quickly cut out portions of YouTube videos by time and turn them into GIFs in like 5-10 minutes using GPT4. Likewise with pulling audio from YouTube videos into mp3 files and using FFMPEG to convert. Just minutes for each.

I have an application for World of Warcraft that lets you compare cast timelines and timing and such between two players for a given fight and it lays out the timelines vertically next to each other like a Git Diff with color-coding and everything and it only took a few hours with careful prompting. Maybe again something a highly experienced web dev could do in an hour or two but not me personally.

I tried using it on doing some multiplayer game dev synchronization code and it was basically hopeless, even if I was very specific and detailed in what I was asking for.

I currently have an idea for a competitor to Synesthesia (the game) that has the sheet music scrolling right to left bar by bar as a different visualization and you would hold notes like in DDR where they have a glowing trail. GPT4 is utterly hopeless here even if I completely outline the steps. I personally don't have the expertise with musicXML to do this kind of thing but I feel as though it would be a marketable idea

Synesthesia added the ability to switch to a grand staff sheet music standard notation several years ago, unless you're talking about something else.

Talking about a more dance game like view of the grand staff. Imagine the bars scrolling right to left and the beat line staying put.

LLMs won’t do many things. But most of all, they can’t take away your enjoyment of building things if you do it for the enjoyment of building things. I think a key to building something is not worrying about whether anyone else cares. Make it because you enjoy making it. If you take pride in it it’ll be usable and enjoyable by yourself at least, and possibly others. But I’ve found in my career that how much “use” something finds is largely unrelated to anything I value - the best things last 6 months, the worst things I made still run 20 years later. But I enjoyed the best things, and thus I value them more than the stuff that found a lot more utility by others.

The other thing is LLMs are not a end but a means to an end. For instance I wanted to help my daughter learn typing. I forked a typing tutor and ripped out the boring backend dictionary based challenge generator and wired it up to a llama and stable diffusion and made it generate Minecraft stories with illustrations based off a LLM and SD lora I created for her to practice with (there’s a lot of Minecraft fan fic and she loves Minecraft books). It was a lot fun, she learned typing, and I learned enough about LLM and image generation to not fear them as tools but realize they’re pieces in a more powerful toolbox to be harnessed towards our end. This, btw, is why open source AI is critical.

I hate to undermine your first point and OP... But that Minecraft typing tutor seems incredibly useful.

I know a fanfic LORA is probably not publishable(?) but the rest seems like something many parents would love.

The beauty is the idea of fanfic derived models doesn’t require distributing the LORA directly, just the trainer and a scraper with URLs. Perhaps today it’s unreasonable to fine tune on an average persons machine but I have a feeling a) this is the way of the future, more so than closed models in closed gardens, and it’ll drive an impetus for more better cheaper consumer kit, b) efficiency of these processes are absurdly bad. They were designed by ML engineers. You get the hard core game engine nerds optimizing these things and you’ll see stuff like vLLM news constantly.

I realized pretty quickly this could be used for anything - paw patrol, Star Trek, or all sorts of fan fiction bases. I think even more commercial, for better or for worse, is making a LORA based off alt.sex.stories.

Anyway I have cleaning up to do but it was written in rust and used diffusers-rs and llm-rs, and can do fine with base OSS models for almost anything. My Lora models do well for making specific characters and situations more prevalent and a better tone, but the base 13B models out there do fine. Stable diffusion 1.5/2.1 also do fine, but do even better with anime or rpg Lora. I plan to release it at some point as OSS (it’s a fork of ttyper-rs)

> . I think even more commercial, for better or for worse, is making a LORA based off alt.sex.stories.

Oh, that was being done even before LLaMA.

And the optimized frameworks are already there, the dots just havent been connected yet.

Yeah I know they’re there - but I think they’re just the beginning of the optimizations. Just working through the stable diffusion code it looked like there were eye watering improvements to make.

You should check out the AITemplate SD demo. I don't think its even SOTA, but its the fastest I've tried (and you can use a LORA by baking it into a diffusers model), though its still using the relatively inefficient diffusers implementation.

There are some Vulkan implementations too, like SHARK, and MLC's Apache TVM demo.

LORA based off alt.sex.stories

There's a ton of adult story generative language models in kobold AI.

I am a mediocre dev as a technologist generalist. LLMs have given me superpowers other people needed to develop or were born with. They’re perfect for my side projects. “Slap it together, ‘good enough’, ship it.”

For me the existence of LLMs means I can avoid the boring parts, so I'm more happy and confident to start side projects :)

Projects with good UX will always be needed, start hacking!

I think a lot of the mega hype around LLMs relies on extrapolating their (very impressive) performance today and assuming that the error rate will decrease significantly over time. I am very skeptical that this will actually happen, since in order to ensure correctness I believe you basically need to build a separate system that can determine correctness, which reduces to what you had to do pre-LLMs to build such a system.

I think they are going to be a really useful tool alongside many other really useful tools which already exist. I think that they will have an impact on many areas where correctness is not very important. I don’t think they will make everything else obsolete, and this line of thinking seems like some kind of weird utopian thought pattern bug in humans.

So if you want to help your anxiety, just become extremely skeptical of LLMs! Practice rolling yours eyes, etc.

I have been working on LLM projects for the past 2 years and I still share a similar anxiety to you. Will what I am building just be a small added feature inside a future ChatGPT or will an AI just be able to do it out of the box in the future?

However, I don't think that is a reason to not build - if anything it is a reason to build faster and just believe that you will be able to adapt to whatever happens in the future. And on a personal note, a part of me envies people who aren't actively working on projects using LLMs -- you guys are the ones who are more likely to stumble across unique problems to solve with AI that most of us just didn't see.

I started to avoid LLM news and I feel much more relaxed. Most of it started to feel like bs hype. It will take a year or so until the hype settles and everyone has a reasonably idea of how it can be used reliably in a product.

By that time all our jobs are automated...

Yes. When I build something, I envisage what the end product is. That is my personal goal. It doesn't influence my programming.

LLM's make me question if the goal I am moving 'to' will even be mine by the time I get there.

Nobody likes working on a sandcastle and discovering Steve next door built the exact same sandcastle as you, halfway through your build.

It has nothing to do with money or praise. It remains to be seen if LLM's can build ALL sandcastles before I even finish my one... that is the anxiety.

The upside of LLM's is that they should make software blackboxes much easier to make.

The way I have been dealing with this is brushing up on my calculus, and probability theory, as well as matrix math, and filling in any holes in linear algebra I have, and then learning more about neural networks.

LLMs burst into more people's consciousness last year, but so did Stable Diffusion, so I have been finding neural networks in general more interesting. Once I get a better grasp of backpropagation etc. I'll see if I want to move on to learning about transformers and LLMs, or latent diffusion models, or both, or maybe some other type of neural network.

My side project is a model of computation (the bitgrid[1]) that could possibly make LLMs much faster, as a side effect. I'm interested in real time signal processing as a long term goal. For now, the model is so foreign to even me that its a bit brain breaking to think about.

I just got a binary counter working... it's that basic right now.

[1] https://esolangs.org/wiki/Bitgrid

> I'm having a really hard time working on new, non-LLM related side projects due to anxiety that everything I'm doing is effectively worthless.

Well for me, all of my side projects are "worthless". Fun, but worthless. I guess that makes me worry less, although burnout usually gets to me before AI anxiety gets a chance.

The existence of LLMs has made me feel more alive and better able to live in the moment.

So what if a machine can write a novel? I don't even want to read it.

Human intention is what gives meaning and value to my life.

The things that I want to build need me to make them. Why would I wait for a machine to do it later without my imperfect human perspective?

Why not use the LLM to work more quickly? I do.

> folks will eventually just ask a LLM to do X, why bother creating something to do it

Dive into more exotic languages. Maybe LLMs will make simple Python scripting obsolete, but they will be awful at writing, say, OpenCL/Vulkan kernels, engine specific scripts, SIMD intrinsics, triton code and such for some time.

This exactly. If an LLM can do your task, then maybe your task was not that exciting to begin with. Let the LLMs to do the mundane coding.

I do bunch of this "exotic" stuff, OpenCL kernels, matrix multiplications in Rust, Lua-written game development, Bezier curves, I write a lot of Haskell, other of random stuff. I feel like if you are doing something that's a bit off the beaten path suddenly the LLMs are not as useful anymore (I use GitHub CoPilot, and occasionally throw something in ChatGPT). And I don't know how fixable this is with the currently known techniques if it's more of a dataset problem.

LLMs are going to need a major leap forward in their abilities before I get afraid of becoming obsolete. Time will tell.

> matrix multiplications in Rust,

1. not exotic

2. LLMs can do this easily

3. it's already abstracted away in efficient libraries so largely an obsolete skill even before LLMs (maybe it's good for educational purposes)

All of this is fine if you want to tinker with various tools for its own sake but don't lie to yourself that you're operating in some area of programming that's out of reach for LLMs or that other people/employers will find especially useful.

Prompt engineering is the last great frontier.

I worked on Tanimoto similarity today. It was cool: https://www.featurebase.com/blog/tanimoto-similarity-in-feat...

LLMs routinely give me incorrect answers when I ask it questions related to coding. I was nervous about them for a couple weeks, then I started to use them and realized that they're really not that useful for generating anything other than boilerplate

This is what I am working on

https://tildes.net/~comp/16ab/what_programming_technical_pro...

I'm going to have to do a real write up on it thanks to the increasing interest. Also I am making things like

https://mastodon.social/@UP8/110542853660777197

and am thinking now about how to make 3-d worlds around them. To get ready for Apple Vision development I went on Ebay and got myself a Hololens 1! I've also got people trying to get me to start another project I was talking about six months ago. So I'd better quit playing Elden Ring and get to work.

I simply do not understand this entire line of thinking.

LLMs aren't useful. They don't solve problems. They don't substitute for real software. They don't threaten our jobs. They are absolutely irrelevant to my life. Why do you see them as relevant to yours?

i have found concrete uses for them in automating work that otherwise would be tedious. you just need to find problems where errors are acceptable and more than 1 second of thinking is needed. not all problems meet those standards but they exist.

> LLMs aren't useful.

to you?

> They don't solve problems.

for you?

Care to elaborate?

Are you curious why other people find them useful? To see how they solve problems with them?

my side projects target folks that dont want anything to do with LLMs

Reply on news.ycombinator.com