Notices by :trebuchet: Kale :trebuchet: (darkestkale@mastodon.social), page 12

I know 'hey the kid who doesn't like reading is Dyslexic! Surprise!' is a bit of a trope, but... yeah, I've actually heard some folks find out they're dyslexic way later that it would seem obvious, so... I guess this'd be a PSA

https://mastodon.social/@thefunnypages/113050756705098665

Listen, kids.

I can understand someone saying 'we've run out of *human written* novels to train AI on'

but...

Not video, when (almost) everyone has a camera in their pocket and they're uploading shit to the cloud

'We have already run out of video to train AI on'

**reads**

**looks at camera**

**reads again**

What the fuck is this foolishness?

My new model is on... mescaline?

> Kara (aka Kara): Hi there everyone waves hello again from time travel adventures into new realms where everything seems normal yet it turns out completely different when things go wrong....

**Triumphant shouting** YESSSSSSSSSSSSSSSSSSSSSSSSSS

Dunno if I'll even get anything useful out of this, but it has been an interesting project.

Still no idea what my endgame is.

An LLM with... a more ethical dataset?

Huh.

Ok, so parameter size is determined by (base) model architecture (which makes sense).

So... my model, currently on gpt-2 small is about 124M

Could go higher, definitely, and I definitely probably will... just not sure if what I need is moar dataset or moar training or MOAR PARAMETERS

So many little levers and buttons and things to toy with.

Thing is - now that I can do it /quickly/, I can already feel my brain going 'yeah, ok, solved problem. Let's go do something else entirely'.

**long shrug**

if anyone wants to know what the fucking deal with #CUDA for #PyTorch is, it's like this:
* You can run multiple versions of CUDA and swap between them using Environment variables
* If you're using Pycharm, you want to manually pull in the Pytorch PLUS Cuda variant of torch via pip, not the built in package manager
* 12.x version of CUDA can suck knob, use 11.x instead.

I fucking trained an epoch (which used to take about 1.5 hrs) in... four fucking minutes.

Gypsy told me the other day that I don't make things anymore, but I do work out how to make things and how to solve problems, and I guess that's right?

I've turned into 'ok, I've proved that X can be done, let's...'

Another day, another half dozen fucking machines^ recommended to me by #aliexpress

^ - yes, literal fucking machines

We have a late entry.

(This tub holds rags. How the fuck did fhey break it?)

Today's entry for Crimes Against Equipment:

Good morning folks

@HauntedOwlbear NUP, YOU GO OWLZ

https://www.youtube.com/watch?v=SREZ-ggSDjM

Faces might change, but there's groups of folks here who mirror people I've seen or known, over and over again

So, it's over 20 years since I first went goth clubbing, and lemme tell you: people repeat through time.

Goth club fashion is a bit themeless tonight, bummer.

Always hoping for a cybergoth resurgence.