[abroszka33]

I agree, I have over 20 years of software engineering experience and after vibe coding/engineering/architecting (or whatever you want to call it) for a couple months, I also don't see the technology progressing further. LLMs are more or less the same as 6 months ago, incremental improvements, but no meaningful progress. And what we have is just bad. I can use it because I know the exact code I want generated and I will just re-prompt if I don't get what I want, but I'm unconvinced that this is faster than a good search engine and writing code myself.

I think I will keep using it while it's cheap, but once I have to pay the real costs of training/running a flagship modell I think I will quit. It's too expensive as it is for what it does.

[esafak]

Yet, the real cost of inference for a given model complexity is dropping.

[benjiro]

> LLMs are more or less the same as 6 months ago, incremental improvements, but no meaningful progress.

Go back to a older version of a LLM and then say the same. You will notice that older LLM versions do less, have more issues, write worse code etc...

There have been large jumps in the last 2 years but because its not like we go from a LLM to AGI, that people underestimate the gains.

Trust me, try it, try Claude 3.7 > 4.0 > 4.5 > Opus 4.5 ...

> I'm unconvinced that this is faster than a good search engine and writing code myself.

What i see is mostly somebody who is standoffish on LLMs ... Just like i was. I tried to shoehorn their code generation into "my" code, and while it works reasonably, you tend to see the LLM working on "your code", as a invasion. So you never really use its full capabilities.

LLMs really work the best, if you plan, plan, plan, and then have them create the code for you. The moment you try to get the LLMs work inside existing code, that is especially structured how YOU like it, people tend to be more standoffish.

> It's too expensive as it is for what it does.

CoPilot is like 27 Euro/month(year payment + dollar/euro) for 1500 requests here. We pay just for basic 100Mbit internet 45 Euro per month. I mean ... Will it get more expensive in the future, o yes, for sure. But we may also have alternatives by then. Open Source/Open Weight models are getting better and better, especially with MoE.

Pricing is how you look at it... If i do the work what takes me a months in a few days, what is the price then? If i do work that i normally need to outsource. Or code that is some monotoon repeating end me %@#$ ... in a short time by paying a few cents to a LLM, ...

Reality is, thing change, and just like the farmers that complained about the tractor, while their neighbor now did much more work thanks to that thing, LLMs are the same.

I often see the most resistance from us older programmer folks, who are set in our ways, when its us who actually are the best at wrangling LLMs the best. As we have the experience to fast spot where a LLM goes wrong, and guide it down the right path. Tell it, the direction is debugging is totally wrong and where the bug more likely is ...

For the last 2 years i paid just the basic cheap $10 subscription, and use it but never strongly. It helped with those monotoon tasks etc. Until ... a months ago, i needed a specific new large project and decided to just agent / vibe code it, just to try at first. And THEN i realized, how much i was missing out off. Yes, that was not "my" code, and when the click came in my head that "i am the manager, not the programmer", you suddenly gain a lot.

Its that click that is hard for most seasoned veterans. And its ironically often the seasoned guys that complain the most about AI ... when its the same folks that can get the most out of LLMs.

[abroszka33]

> Trust me, try it, try Claude 3.7 > 4.0 > 4.5 > Opus 4.5 ...

I started with Sonnet 4 and now using Opus 4.5. I don't see a meaningful difference. I think I'm a bit more confident to one prompt some issues but that's it. I think the main issue is that I always knew what/how to prompt (same skill as googling) so I can adjust once I learn what a modell can do. Sonnet is kinda the same for me as Opus.

> LLMs really work the best, if you plan, plan, plan, and then have them create the code for you. The moment you try to get the LLMs work inside existing code, that is especially structured how YOU like it, people tend to be more standoffish.

My project is AI code only. Around 30k lines, I never wrote a single line. I know I cannot just let it vibe code because I lost a month getting rid of AI spaghetti that it created in the beginning. It just got stuck rewriting and making new bugs as soon as it fixed one. Since then I'm doing a lot more handholding and a crazy amount of unit/e2e testing. Which btw. is a huge limiting factor. Now I want powerful dev machine again because if unit + e2e testing takes more than a couple seconds it slows down LLMs.

> Pricing is how you look at it... If i do the work what takes me a months in a few days, what is the price then?

I spent around 200 USD on subscriptions so far. I wanted to try out Opus 4.5 so I splurged on a Claude Max subscription this month. It's definitely a very expensive hobby.

> And its ironically often the seasoned guys that complain the most about AI

I think because we understand what it can do.