There is a law in software engineering called "ninety-ninety rule"

The first 90% of the code accounts for the first 90% of development time; the remaining 10% accounts for the other 90%.

which can be interpreted as "final touches are the most difficult".

I think it is also true for vibecoding. As seen in a video I've posted, although the app works, there are some small faults yet to be fixed. And as seen in commit history of PhotoCraft - the work isn't finished, it currently intensifies

I'm looking at those issues... here a guy opened an issue and literally 15 minutes later there is PR with a feature he wanted

https://github.com/storytold/photocraft/issues/2034

I mean: I'll probably get downvoted because we are on Lemmy and the title contains "vibecoding" in it, but holy s**t, are we just getting a viable Photoshop alternative for linux 🤣

There is no bubble bursting in the near future.

We are just entering a stage when AI is becoming a cyber security threat. It can become a weapon. What that means is:

1. There is a large military budget that is going to be invested in it
2. There is an order of magnitude larger cybersecurity budget to be spent by everyone not wanting to become a target of those mentioned in point 1

AI is a weapon and the only way of defending against it is AI. Do not expect budget cut.

Excellent article. Not only shown how to use it, but also what are alternatives and why use unions instead of them.

Guys, please do not downvote this post just because you do not agree who actually solved this problem. The title of the post is correct, unbiased. It says that it's OpenAI that claim to solve this problem. It doesn't say that OpenAI has indeed solved this.

I haven't even noticed how useful it was to me until they disabled it. Glad they came back

your post is not related to the topic of this community

Note from OP: I posted this because I find a topic of "agent swarms" super interesting as seen as an engineering challenge. It's just a novel kind of problem yet to be solved. And this article describes what worked well and what not, what gains did they get making certain changes.

But being honest I see some gaps in their reasoning. SQLite is an open source project which most probably means its code already is part of a training set of frontier LLMs. This calls into question whether language models actually invented or just recalled. Also, as noted in the article, they instructed models to build something based on the docs. That means that the goal was to recreate, not to invent. One of the hardest things in software engineering is to figure what to build and this agent swarm has already been given that information, what made the job much easier.

If the benchmark is solved, then why official arc prize page on X https://xcancel.com/arcprize does not mention it? And instead they post results of models that are not even close to 100%?