My job is so f**king unbelievable.

I'll try to sum it up by first telling you about the folks I work with:

First, there is this supermodel wanna-be chick. Yeah, okay, she is pretty hot, but damn is she completely useless.
The girl is constantly fixing her hair or putting on make-up.
She is extremely self-centred and has never once considered the needs or wants of anyone but herself.
She is as dumb as a box of rocks, and I still find it surprising that she has enough brain power to continue to breathe.

The next chick is completely the opposite. She might even be one of the smartest people on the planet.
Her career opportunities are endless, and yet she is here with us. She is a zero on a scale of 1 to 10.
I'm not sure she even showers, much less shaves her "womanly" parts.
I think she might be a lesbian, because every time we drive by the hardware store she moans like a cat in heat.

But the jewel of the crowd has got to be the fucking stoner. And this guy is more than just your average pothead.
In fact, he is baked before he comes to work, during work, and I'm sure after work.
He probably hasn't been sober any time in the last ten years, and he's only 22.
He dresses like a beatnik throwback from the 1960's, and to make things worse, he brings his big f**king dog to work.
Every f**king day I have to look at this huge Great Dane walk around half-stoned from the second-hand smoke.
Hell, sometimes I even think it's trying to talk with its constant bellowing.
Also, both of them are constantly hungry, requiring multiple stops to McDonald's and Burger King, every single f**king day.

Anyway, I drive these dicks around in my van and we solve mysteries and s**t.

Context: New to typescript. Writing a thing, doing it for work, good opportunity to stretch my dev legs. Using a propriety lib, alternatives not an option.

Rant begin:
SOOOO, who the fuck thought THIS was a good idea:
1. Lib has minified react in dev (because closed source) meaning no downstream errors AND the entire premise of the lib is that a widget is a react component, so I'm writing typescript react the entire time without downstream errors
2. SHIT docs. By that, I mean there's an API reference page that's so sparse there's literally a set of CRUCIAL interfaces that only say the word 'Interface' on them. That's it. that's what i get. It's an interface. NO FUCKING SHIT SHERLOCK, what the fuck is it though? What's its purpose? Is it an interface for a dog? A dog that has a 'shit' property? or a cat? or a cat eating dog shit? Nobody fucking knows - the docs sure as fuck don't care.
3. No syntax highlighting - editors, IDEs (i've tried a few) can't even find the lib inside this environment, so Code and everything else thinks I'm importing shit that doesn't even exist - so no error prediction, code completion based on syntax of the library, none of that.
4. There are some EXTREMELY basic samples - these samples exclusively use React classes - no function components, no hooks, nada - just classes and even perfect replicas of the sample code display erratic behavior like errors about missing props, so that's mostly FUCKING USELESS
5. And this... this is where the straw breaks the fucking camel's back... there's no... there's no hot reloading... Do you know what that (in conjunction with the previous 4 fuckups) means?

When I write anything or I fuck up (which of course I'm doing every time I write half a line because how the fuck?) I have to restart the client and server EVERY FUCKING TIME and manually test to see if the error (THAT ONLY GETS REPORTED IN THE LOCAL UI) is gone or different.
Then, once I see the error, it isn't an error: it's the minified React error-decoder link and guess what? It isn't really clickable a link OR copyable, meaning that every FUCKING time I get a new error, I have to MANUALLY TYPE A FUCKING 50 CHAR URL TO FIND OUT A GENERIC REACT ERROR MESSAGE WITHOUT A LINE NUMBER OR ANY FUCKING CONTEXT. I HAVE TO DO THIS CONSTANTLY TO SEE IF ANYTHING I'M DOING EVEN WORKS.

6. There's no github to complain to the maintainers or search for issues because it's NOT FUCKING OPEN SOURCE so there is literally nothing to be fucking done about it.

This is due in a week and a half, found out about it last Friday. How's your day going?

PS: good to be back after a long respite from dev ranting.

Any night, 1:30am, bedtime: "Yes! I can't WAIT for tomorrow to begin! I'm gonna make SO much progress on that personal project that I just KNOW is gonna change the world and make me a billionaire! My time is now!"

Next day, 9am, first call of the day: "Ugh, waking up SUUUUUCKS! But, fine, just gotta get through the workday, then it's beast mode time!"

5pm: "Ugh, that day SUCKED... meeting after meeting, constant interruptions for the few minutes I got to hack code, SO many emails, and hey, good day, only five new things pushed down from corporate to bang my head against! Feelings pretty mentally exhausted, but it's all good, I fortunately love this programming stuff, so first dinner, then a little exercise, spend some time with the family, and then it's time to COOOODE!"

10pm: "Ok, house is FINALLY quiet (fucking dog), just a little noise from my daughter staying up way too late again... kinda spent, but this project still excites me, and I may not get as much done as I was hoping, but fine, I can still make some tangible progress and that's what matters. Maybe just one last quick check of email, Reddit, make sure there's no new Hot Ones or Honest Trailers I gotta watch, update IDEA plugins and see what's new, then it's work time! Nothing can stop me now!"

Any night, 1:30am, bedtime: "SHIT! I GOT FUCK ALL DONE AGAIN! GO DAAAAAAAMN IIIIIT!!!!"

The next step for improving large language models (if not diffusion) is hot-encoding.

The idea is pretty straightforward:
Generate many prompts, or take many prompts as a training and validation set. Do partial inference, and find the intersection of best overall performance with least computation.

Then save the state of the network during partial inference, and use that for all subsequent inferences. Sort of like LoRa, but for inference, instead of fine-tuning.

Inference, after-all, is what matters. And there has to be some subset of prompt-based initializations of a network, that perform, regardless of the prompt, (generally) as well as a full inference step.

Likewise with diffusion, there likely exists some priors (based on the training data) that speed up reconstruction or lower the network loss, allowing us to substitute a 'snapshot' that has the correct distribution, without necessarily performing a full generation.

Another idea I had was 'semantic centering' instead of regional image labelling. The idea is to find some patch of an object within an image, and ask, for all such patches that belong to an object, what best describes the object? if it were a dog, what patch of the image is "most dog-like" etc. I could see it as being much closer to how the human brain quickly identifies objects by short-cuts. The size of such patches could be adjusted to minimize the cross-entropy of classification relative to the tested size of each patch (pixel-sized patches for example might lead to too high a training loss). Of course it might allow us to do a scattershot 'at a glance' type lookup of potential image contents, even if you get multiple categories for a single pixel, it greatly narrows the total span of categories you need to do subsequent searches for.

In other news I'm starting a new ML blackbook for various ideas. Old one is mostly outdated now, and I think I scanned it (and since buried it somewhere amongst my ten thousand other files like a digital hoarder) and lost it.

I have some other 'low-hanging fruit' type ideas for improving existing and emerging models but I'll save those for another time.