Fuck AI

Finding those thousands of tables for a fitness plan is the problem though. The internet has been enshittified so much that it's easier and faster to ask the AI to crawl through the sewage of ad-infected clickbait sites.

The internet is a complete waste of ressources.

We can use it to share information, but there has never been a time where this wasn't accompanied by a massive abuse of bandwidth by someone trying to sell stuff in the same breath.

I'm convinced that 90% of the people who think AI is amazing just never learned how to properly use office software. I saw someone insisting that it was amazing AI could send letters to hundreds officials all across the state to advocate for removing Flock cameras, saving them the time of doing each one individually. You have never heard of form letters? You think every letter you get from Chase Bank has someone typing it?? Did they just give up on teaching this shit to kids?

I had an intern this summer working on a project that involved mapping two data sets based on a description field. I showed him how to write a Python script using a fuzzy matching library to compare the two data sets and find matches. We then talked through the downsides of using a contextless string comparison and the situations that create false positives and false negatives.

We then tried to use a few of the AI tools to match them, hoping to demonstrate how a tool that understands context can do a better job at the expense of far more computation. All they would do was write and run a Python script using the same library, but every request would start with writing a new, slightly different script. Incredibly wasteful.

It turned into a great demo of how even if it returns a useful result, it's rarely better than a person could do in a few minutes and is far less consistent.

nobody has hear of a template

Eggsfuckingactly. Company says use AI as a starting template! Uh, hello, what if we used a fucking template as a template?? And that way we could have a deterministic source?? And we could track and fix bugs in the template?? No? We want to roll the dice each time to create the template? Ok. Ok...

Certainly there’s some open-source version of the app out there somewhere.

... which is what Claude was plagiarizing to build these 'custom' apps.

Certainly there's some open-source version of the app out there somewhere. Or at least enough to lazily glue together into the app.

Yeah, but that code is probably copyleft. They probably want the LLM to launder the license for them so they can get away with using it in their proprietary commercial product.

Yes but, if you do that you don’t get to completely give away your thinking faculties to the magic misinformation machine 🤡

Management types don't understand how things work, so they make bizarre decisions

Well, sloppers were never smart to begin with.

Hello fellow firefox mobile enjoyer

The only reason I even use pacman is because I like seeing it gobble all the little dots. Nomnomnom.

Phew, lucky me. I always type it as two separate lines.

bash: apt: command not found

Just the readme and repo description are very strongly AI-flavoured, and I haven't looked at the code at all

I've been saying that for years: all LLM code output is "tainted" in terms of copyright. At the very least it's all copyleft because copyleft code was used to train it, or more likely it's just impossible to use legally at all because it's conflictingly-licensed stuff smushed together.

Yea, the first one wasn't written by Claude. This Nick person doesn't seem to know what he's talking about.

That’s not the way LLMs work at all. There’s zero weight modification in the transformer model back from prompts. It happened because it was statistically likely to at some point.

Do we really want patents on open source projects that help track the quality of a night sky for photography?

Is there really an issue if such an application just is slopped out there?

No, the other app was published on GitHub just a few weeks before, it can't be inside the training data.

It's more likely this: Claude code can and will do web searches to integrate the prompt if too generic or dubious, found the competing GitHub repo with the same name, got the claude.md with all the description, features and inner details, and rebuilt the copy on that.