If you see me unfollow you here and simultaneously follow you on my main account, it's just because I'm trying to keep my feed on here as development-oriented as possible and some of you are just too social ;)

Tiny Farm releasing tomorrow, public domain and completely free! Fits all previous Tiny releases: kenney.nl/assets/serie...

I have almost entirely abandoned Claude code in favor of this. Probably will completely abandon CC soon.

polytoken dot dev@polytoken.dev · 2mo ago

Polytoken 0.2.0 is publicly released! Even though there isn't all of a website yet. An entire thirteen of people have been using it all day and nobody has yelled at me yet. You can get it at docs.polytoken.dev/installation (forgive the mess; WIP) or you just `brew install polytoken/tap/polytoken`!

And people say AI can't produce art.

austin@aparker.io · 2mo ago

gonna use this just to code up monad impls github.com/cyberpapiii/...

Introducing the new Gobbledygook™

Introducing Googlebook, designed for Gemini Intelligence

We’re introducing Googlebook, a new category of laptops designed for Gemini Intelligence and perfectly in sync with your Android phone.

blog.google

Claude you can't just put those words in that order without consequences.

Sorry @dollspace.gay I have incredibly terrible news for you.

Matthew Sheffield@matthew.flux.community · 3mo ago

In totally unsurprising news, Richard Dawkins is developing AI psychosis. Paywall bypass: archive.is/6RdK9

Every single time, the argument is the same. "this tool makes dangerous knowledge too accessible to people who might misuse it." And every single time, the actual function of the restriction is to consolidate control over who gets to know things, not to prevent harm.

Doll@dollspace.gay · 3mo ago

It is categorically impossible to make AI perfectly safe and also useful. Pick one. The alignment question is "Alignment to who?"

oh no no no no, search engines are different because I, personally, find them useful. Sure, everyone knows it's scraping every copyrighted article possible without prior permission and building a model that's exploited for commercial purposes, but what my job presupposes is... what if it wasn't?

*shoots a huge bong toke* Coding agents are GANs where the discriminator is regression tests.

meanwhile, OpenAI claims that GPT-6 can write an accurate regex for HTML parsing

In the UIColor implementation in iOS in order to get RGBA values you call “getRed”, because the signature is getRed(_, green: _, blue: _, alpha: _) This is because it follows objective C coding standards, which were designed by criminals who should all be in jail.

Ok, TODAY I will get actual output from the app ...

Gonna set this as my system prompt.

Ok, today the goal is to get the app to generate something other than progress bars. (Although I do think my progress bar is pretty neat, I don't think end users would be satisfied with that.) Currently implementing the naïve CPU search algorithm (again) as a baseline for "correct".

Rewrote this to just call the GPU process once per sheet with all regions instead of once per region within the sheet, and then collate the results. Now it's down to 3 seconds, from 3.5 minutes. ᕙ(⇀‸↼‶)ᕗ

XaiaX (💻)@developer.xaiax.net · 7mo ago

Using Metal shaders to perform transformation task reduces time from 3.5 minutes to about 25 seconds. Friendship ended with CPU. GPU is now my best friend.

On second thought, perhaps this is a poor choice of parameter name.

Using Metal shaders to perform transformation task reduces time from 3.5 minutes to about 25 seconds. Friendship ended with CPU. GPU is now my best friend.

Is there a good / canonical set of image training data these days? I have the old CIFAR 32x32 images available but I vaguely recall there being some issues with that. (This is not for image recognition, just want something to use to develop hyperplanes for a locality sensitive hash idea I have.)

I feel like I may have found this before, but just discovered in Swift you can do return if someCondition { case_when_true() } else { case_when_false() } Which seems clearer than the traditional ternary format.

The autocomplete code generation in Xcode seems to be getting really good. Seems to be trained on Swift and somewhat context and maybe stylistically aware. I wrote the doc string and the function signature and it completed the rest of the code on its own. This runs entirely offline, too.

Back at it. Been thinking about the screenshots / Live Photos / etc stuff and realized that I can make a more dynamic experience where the user can turn those on and off without needing a full parse / index. The way that it works now is that I parse through local library, which is all thumbnails.

It captures shapes much better in monochrome. I might also make a version of this that uses the luma data from the tiles and the chroma data from the target. (It’s in LAB format already so that should be easy)

Ok time to figure out WTF "Capture of 'someVar' with non-sendable type 'SomeClass<String, Int>' in a '@Sendable' closure" means. (Well, more specifically, figure out how to get it to stop warning me about this.)

This account is a sub account of my main at @xaiax.net and will just be for dev stuff / maybe software in general, so I'm not going to just mirror my following list here. Fave this one if you want me to follow you from here, maybe if I see dev stuff I can help. (Mostly I just do Swift and Python)

Getting back to working on this the only way I know how: turning it into a shitpost.

I'm going to attempt to use this for dev logging, probably in an extremely verbose manner so that I can get all my thoughts down in text in a way I can get to from anywhere and also that other people might be able to read if they care to.