Some users share examples of AI succeeding at coding tasks like adding buttons or building web pages, while others dismiss the builder's criticism of poor AI programming skills as a skill issue or mock the discussion.
No Digg Deeper questions have been answered for this story yet.
Most Activity
Every time I do any work at all I make this tweet, so I guess it's kind of apparent how little work I've been doing recently
These models are so god damned awful at programming it's shocking
@yacineMTB @dillon_mulroy is currently streaming and one of the foremost experts (imo) at browbeating reasonable code out of models. Def recommend tuning in / ripping his dotfiles if you're looking for alpha
@yacineMTB I have an idea: let’s just loop them so they can evaluate their awful code with even worse reasoning. Forever.
@yacineMTB I wonder almost everyday if the people on this app talking about loops/prompts are purposefuly lying or jesters. Everytime I let AI write code for me, I regret it personally.
@yacineMTB claude just oneshotted adding a button to a next.js app
@yacineMTB It's not the programming so much as decision making IMO. On their own they make terrible decisions.
You just need to define what "proper programming" means. One thing at a time. E.g. these are code quality metrics. These are git-level problems. These are product-level agreements. See my cases.
You make you can measure what actually matters to your codebase. I will show examples, because I wanna solve this problem too.
But if you mean "programming" as in "solving problems", then yeah models can't connect facts / intent in a meaningful way yet.
That has to stay human for now🤷
@yacineMTB did you give it the definition of good?
@yacineMTB Yes. We want the models trained on Google's repos, not the ones trained on mine and other retards public repos lol
@bossmuntime @yacineMTB They have but they use multiple agents to catch bad decisions. They talk between each other and you get a better result. That's the only improvement I've seen for agentic workflows. Single agents still make stupid decisions.
@yacineMTB Sounds like a skill issue
@yacineMTB Fingers and toes everywhere
@yacineMTB @grok how many times in the last month has @yacineMTB complained about how terrible LLMs are at programming? Then please pick the funniest one from the month
@BeatGreatFilter @yacineMTB Oh right, adversarial review especially with decorrelated models works well.
But that’s not really an LLM improvement that’s just tooling. It still feels limited. Looping them hits diminishing returns after like 1-2 rounds in my experience.
@yacineMTB the occasional sparks of genius i see from some crazy run make the whole process worth it. far more enjoyable than coding by hand
@BeatGreatFilter @yacineMTB I don’t think LLMs have even really improved on this since the beginning. It’s probably a limitation.
The Cursor guy was in an interview talking about how he sees a future where you’re writing some form of pseudo-code for an agent to translate. I think that makes sense.
@elvisrepo @yacineMTB Google’s repos aren’t good either.
@yacineMTB you forgot to write "make no mistakes" again didn't you?
@yacineMTB Yea I was trying to fix something going into the file fixing it by hand had the LLM fix it then finding out the shit was in two places when it could've been in one fuck you're right
@yacineMTB ofc, they were trained to be mean