Perpetuum Mobile

@perpetuum_mobile
189 Followers
477 Following
258 Posts

Tooting in 🇵🇱 and 🇬🇧, learning 🇪🇸 and 🇨🇵

Father of 2, Husband, #PHP and #Python #developer, #infosec enthusiast, #bughunter, #foss herald, #llm and #ai researcher, #observability, #automation, used-to-be a #gamer #EliteDangerous #EVEOnline #WOTBlitz #NoMansSky #PerpetuumOnline #MSFS2020 #retroGaming

he/him

LocationPoland
@thebluemarble np. odpływ w podłodze łazienki?
@ErikMenzel that's in fact really easy to be done :-) you can actually clone the whole HALs personality and way of saying things just by the means of system prompt
@juliewebgirl if I want to edit one photo and want full creative control I just start the GIMP ;-) and do it manually. It's not that hard really. Just split into layers... Alter colors... Repeat until entire image is to your liking

@juliewebgirl I tried once building an U-net to add colors to B&W photos, but I was getting really washed out hues. So, being an engineer, I started wondering WHY I'm getting just a little effect, instead of color photos. And I quickly realized why: statistics. If you put 100 red shirts and 100 blue shirts in training, the model will average it out... and output greyish purple. It's all math after all!

Next time I will try to add a secondary GAN network to force the model to choose a color.

@juliewebgirl if I make a model myself on data that I own, using GPU that is physically in my own hardware, just to add color to old photos of my family, what's wrong with that?

I'm not trying to defend commercial LLMs here, I'm talking about classical ML

Corvus Catus

@m0bi modelom LLM najlepiej wychodzi uśrednianie. I biorąc pod uwagę że to są w gruncie rzeczy narzędzia oparte o statystykę, nie powinno to dziwić. Dlatego uważam, że używanie LLM do tworzenia treści to nie jest dobry pomysł. Nawet humanizujące prompty, które mają nadać "twórczości" z generatora unikalny styl autora, nadal są uśredniane przez tą samą brutalną statystykę.

"To nie jest gamechanger, to ja mój pies mamy po trzy nogi"

@juliewebgirl it depends on what AI you use to restore colors, if you use GenAI tools like multimodal LLM, then yes, it might add things that weren't there (vide DLSS 5). If you design a custom U-net or GAN in pytorch, which is trained on photos converted to B&W, it should do just fine.

I'll try to make this kind of model if I have some free time. Should be a fun experiment to do!

@grymas1000lecia Claude ma dobrze zrobiony RLHF

I'm a software developer and sysadmin who could really use being #fedihired.

What I'd really like to do is Rust, but once you ignore the dubious crypto and AI stuff, there seems to be nothing out there. Prove me wrong with a counterexample!

I've spent decades fixing Enterprise mudballs mostly written in #Perl. If you've got a crufty legacy system that everybody else is too scared to touch, I'm your man. I love fixing stuff like that.

I've also done commercial #Scala, #Python, #C/#C++, and although I don't usually admit it on my CV but these are now Trying Times when everything is on the table, even #PHP (the longest six months of my life).

Perl naturally leads into Unix system administration and infrastructure. I've built and maintained mail clusters, VoIP systems, network monitoring, DNS management platforms, that sort of thing. If it's non-sexy but something which needs to be done, I'm there.

Available immediately, for contract or permie, onsite in Amsterdam/Randstad or remote to anywhere.

Drop me a private mention or mail [email protected] if you have or know of something.