Mauve 👁💜 @mauve@mastodon.mauve.moe

I think my philosophy when making software is that it should work for people with zero money or no bank account / credit card.

I know it's not a popular mindset to be in since money and profit is everything in the tech world.

I think it comes from growing up as a kid with no disposable income or access to anything but my shitty computer.

I'd rather support people with almost nothing than people with latest and greatest tech gizmos and spare cash for subscription services. 😅

Thread of stuff I wanna buy. Will "like" my own toot if I get it.

I'm gonna use this yhread as a way to keep track of links to read and like my own posts when I get around to them.

Feels good to be back in the part of the codebase I can be more productive in.

We should just have Techaro's Anubis be an HTTP standard that browsers implement instead of all this captcha as a service BS.

https://anubis.techaro.lol/

Can we stop the "prove you're human" checks now that vision models can solve them better than some humans and standardize on a hashcash HTTP challenge/response so we can mitigate the scraping asymmetry while ditching the surveillance JS bloat? Please?

http://www.hashcash.org/

LOL. Someone made it so you could invite a person to a project with the "Blocked" role. Imagine.

ADHD meds are more effective when paired with old school gabber mixes. #ProTip

So pissed off I'm switching back to a regular keyboard so I can rage type faster.

Slop devs don't respect api stability

playing MORSE to learn how to write morse code. Seems fun so far.

Been revisiting local models again. The state of the art for it's size is apparently MiniCPM5 right now. It can do very basic code related tasks but it's still way too stupid to do actual tasks. Might be enough for my "speech to code to speech" idea though?

https://huggingface.co/openbmb/MiniCPM5-2B

The hardest part of software development is dependency management. Whatever funky language features you like or hate are secondary to being able to actually run the code.

Seriously, every dependency is a time bomb waiting to screw your builds. Forking a dependency for your project? Good luck keeping anything together after everything shifts in a year or two.

Debugging haskell dependency hell 💀

I think of LLMs as tools that decompress a few words into more words. E.g. instructions into tool calls. My goal is to use as little power and as few words as possible to make it output the words I want faster than I can type it. It's like a keyboard / search tool but with a bunch more shortcut options.

With enough training on speech -> code, it might make it possible to have a model do code -> speech so that you could have n lines fed into the reader to get back a speech description. E.g. "for loop over users, call purge data on each". Maybe a "tranalate to grug" could work too. Again with like a 2b model it should be enough but fast as heck.

Using an LSP can help make it generic between languages, but regular text could be good enough. I just really like being able to "go to definition" and "list all call sites". Rename symbol is pretty convenient too.

Navigating could be the same but with a JEV style model. Trigger navigate action. Say "clear all users", Model gets passed all the var names in the file including clearAllUsers(). LLM choses option you most likely meant and you go to the next occurance.

Slop based audio coding: Describe change in English (select star from users where name equals bob). Editor passes few lines around cursor and file name to LLM. "Turn this into code", LLM outputs diff. With small models and small changes you could probably edit at the speed of speech.