I was recently reminded by nostr:nprofile1qyfhwue69uhnzv3h9cczuvpwxyarsvpcxyq32amnwvaz7tmpvanhytnwdaehgu3wd3skueqqyqvfqhg2t43r4wq6nzaf33vzh4040ujsd34cpzg9l3ve6ksty2dss3e99mm of how frustrated I was with vibe coding about eight months ago. I had already spent hundreds of dollars on a project, and it was still full of bugs and half-baked features. The idea was great, but it clearly needed more work, which meant more time and more money. I was close to giving up. I was genuinely close to walking away from vibe coding altogether.

But I’m not a quitter.

I kept going and invested even more, eventually spending around $800 on my first serious project using Goose and Gemini 2.5 Pro. Since then, the tools have improved dramatically. The models are better, faster, and more capable. But more importantly, I got better.

I now see that phase as a learning and teaching experience. I learned how to communicate with AI effectively, how to guide it, correct it, and push it toward better outcomes. I learned the tricks of the trade and how to actually use AI as a real development partner. I kept building and I eventually won. That frustration wasn’t wasted at all. It was the cost of learning how to build in a new way.

Reply to this note

Please Login to reply.

Discussion

💯 agree... I've found another good way is to use multiple AI - as the quite often have different perspectives and approaches.

I've found when 1 gets stuck - pass it over to the next one to take a look. More often than not - it can fix the problem (seems even AI can get stuck in its own head..?)

You can also add in the general GPT to help you explain your thoughts to the coding AI - without wasting development tokens.. 😊

I’ve done this and it has helped to unstick situations and find code errors that the first LLM made and didn’t catch. In these cases, I’ve been pitting Claude against Gemini. Both accounts are paid. I mostly use Claude Sonnet 4.5 and Gemini Pro. I’ve dabbled with Opus 4.5 and so far find it only marginally better than Sonnet 4.5. Opus costs 3x so I’m not seeing the strong drive to use Opus a lot yet.

Now I just use Claude Opus for everything. There is no second best.

Do you see yourself getting into non AI coding (python, RUST,CSS) in the future?

Nah. I have always wanted to find the time to learn more than I know or have used, but with AI, there is zero point. I've made dozens of apps and micro services over the last 8 months and it's only getting better.

That’s an interesting perspective. One I would have if I knew anything about vibe coding. It’s kinda like “why do I have to show my work if We have calculators/I got the right answer” in a math class iykwim

Which project was this? I just saw the 5 stages of vibe coding gifs

Zappix

👀

AI has come a long way in 6 months. I use chat GPT but know that the honeymoon will end and enshittification will creep in once I've given enough data for it to know my preferences. I need to start training and agent on an open source model that's committed to truth and freedom

If you have about 10 bitcoin to burn, you can buy the GPUs you need.

You burnt 0.01 bitcoin on just 1 project, anyone who knows coding enough to train their own model doesn't need a Claude sonnet 4.5, a qwen3.5:14b is a great agentic coding companion and works on retail graphics cards, and doesn't require to burn 10 bitcoin, doesn't even require 1.

and since then i've built countless applications for fractions of the cost of the initial learning period :) if you need some help let me know. i've learned a lot of the last 8 months of build non stop every single day.

Sure, if you need help building a local agentic setup, feel free to reach out 😊

i used ollama and open source models in testing and they just cannot be compared to claude opus 4.5. maybe in a year things will be different.

What weight sizes? I was playing around with 8b models and they suck.

Above 14-20b they are extremely useful.

They've written entire apicstions for me. A lot of pollerama is local models, they are not good for vibing like Claude sonnet. But if you know how to code and plan to review each line, like I do with my products, then it's great and you get infinite generations.

8 and 12. ill try larger ones and see if they're more useful, but honestly i don't mine playing for claude while it's leading the pack.

I probably don't really need my own agent, just access to models that are trained at truth would be good.

The fact you spent $800 and not 8 sats is the reason AI will have trouble justifying its valuations

Now I create projects for $20 a pop.

That’s pretty cool

Once these projects bring in revenue then the AI spend might be justified

800 is way more I use. I do not know how big your project is, but any task I am doing costs me about 2 dolars. Only the comolex one cost bit more.

Correct. Times have changed.

Now set up a course where you teach people what you learned for $800 and move to Austin Texas

i'll be there in 3 weeks :)

but you're basically correct here. that's what i did. i blew through a lot of money and learned a lot over the last year. and now i am teaching people classes all of the time so that they can benefit.

in about 6 weeks im teaching several classes here in my area, educating small bussines owners and entrepreneurs!

That's great

I use chatgpt and I've noticed it's gotten much better at vibe coding. At the beginning I was frustrated with it but now it's getting and better. I would try other AIs but it's too expensive. I pay monthly for chatgpt and it's unlimited. I tried other AIs pay as you go but it's way too expensive.

As if said before, the people who were the first ones vibe coding are the ones who are helping AI using that to train and get better at coding.

You are a legend!

so are you :) i love your music.

thanks!