Register a SA Forums Account here!
JOINING THE SA FORUMS WILL REMOVE THIS BIG AD, THE ANNOYING UNDERLINED ADS, AND STUPID INTERSTITIAL ADS!!!

You can: log in, read the tech support FAQ, or request your lost password. This dumb message (and those ads) will appear on every screen until you register! Get rid of this crap by registering your own SA Forums Account and joining roughly 150,000 Goons, for the one-time price of $9.95! We charge money because it costs us money per month for bills, and since we don't believe in showing ads to our users, we try to make the money back through forum registrations.
(Thread IKs: OwlFancier)
 
  • Post
  • Reply
The Question IRL
Jun 8, 2013

Only two contestants left! Here is Doom's chance for revenge...

Tesseraction posted:

The judge releases her ruling at 3 so I'm curious if it'll really take him an hour to read it. Maybe this painfully slow reading ability is the real reason he kept losing acting roles.

But nice of him to tell people where to gather with bottles of piss ready for him opening his fat gob.

To me, this screams of his lawyer going "I have prepared a script for what you should say when you lose this, in the vain hope that an apology outside of court might prevent future damages and stop you from doing a Trump."

Adbot
ADBOT LOVES YOU

Tesseraction
Apr 5, 2009

Lozza Fox, red-eyed and railed up, standing in front of the court: "I only lost because the judge is a paedophile"

Ms Adequate
Oct 30, 2011

Baby even when I'm dead and gone
You will always be my only one, my only one
When the night is calling
No matter who I become
You will always be my only one, my only one, my only one
When the night is calling



fuctifino posted:

https://twitter.com/GBNewsSpin/status/1751333623607750893

This was quite a thing too. You can tell she's simply reading a script, and does it in such a weird disconnected way as though she's on a load of tranquilisers. I wonder who actually wrote those words?

e: lmao vvvv

Usual bag of ballix from the execution lovers. Smirking about it with obvious glee, bitching about what an easy ride he'll have in checks notes a secure prison for the criminally insane, and claiming it'd be kinder to kill him than keep him in that prison. All without a hint of shame at the obvious contradiction.

Also the consistent "ah well obviously I only support execution when it's beyond a reasonable doubt" and the argument that it's cheaper at the same time. Of course it's cheap if you assume you can just know every necessary piece of information accurately. In reality killing someone is more expensive than locking them up for life. They take that poo poo fairly seriously in America, insofar as you get your day in court and you get a lot of appeals - and there are still plenty of people who get railroaded or just falsely convicted. An ostensibly fair and open system that at first glance takes a lot of serious measures to ensure they get the right person and that person really deserves to die still falls well short of that. It's pretty much impossible to actually be sure in all cases, or if not, it's prohibitively expensive.

Not that I really need to spell that out here, pretty sure we all know her type are wrong'uns who don't care and just want their baby brained regressive lust for revenge to be the law again.

I did like the "Not!" like a kid attempting sarcasm for the first time at playtime though.

Mano
Jul 11, 2012

mossyfisk posted:

Surely you can just audit the compiling?

The LLM / network isn't really compiled but trained. You could certainly have the training program (which might be compiled but e.g. Python often isn't compiled) log it's input (picture name, size, checksum, description, url, ?). But since you don't necessarily have the pictures themselves afterwards, this does not prove everything.
And obviously even small changes in the picture change most of the logged values.
The description of the picture might be wrong, but then via other things might get linked to the real thing.

also this leaves out that you can't remove a picture afterwards if you should find out that it was copyrighted.


E: this stuff is not my specialty, I learned somethings like 20 years ago and I have friends who use it (on own data).

mossyfisk
Nov 8, 2010

FF0000
Just make them provide the source images and literally have someone from a regulatory body watch to confirm it's what they feed into the big GPU racks or whatever. Keep a copy archived for when you want to check rights etc. I don't think there's actually a technical barrier?

cat botherer
Jan 6, 2022

I am interested in most phases of data processing.

Mano posted:

The LLM / network isn't really compiled but trained. You could certainly have the training program (which might be compiled but e.g. Python often isn't compiled) log it's input (picture name, size, checksum, description, url, ?). But since you don't necessarily have the pictures themselves afterwards, this does not prove everything.
And obviously even small changes in the picture change most of the logged values.
The description of the picture might be wrong, but then via other things might get linked to the real thing.

also this leaves out that you can't remove a picture afterwards if you should find out that it was copyrighted.


E: this stuff is not my specialty, I learned somethings like 20 years ago and I have friends who use it (on own data).
:actually:, the popular DL packages do actually compile the model graphs. But yeah, that won't help with logging the training data (which people would have every incentive to bullshit anyway).

Sometimes these models can actually output almost exact copies of some of their training data. This paper shows some examples for Stable Diffusion, but I've seen similar results with generative LLMs. However, that's still impossible to do at scale, and models certainly never copy the vast majority of their training examples (it would be information-theoretically impossible).

mossyfisk posted:

Just make them provide the source images and literally have someone from a regulatory body watch to confirm it's what they feed into the big GPU racks or whatever. Keep a copy archived for when you want to check rights etc. I don't think there's actually a technical barrier?
That would be impractical, especially with how much training goes on, and the sheer volume of images or w/e.

sinky
Feb 22, 2011



Slippery Tilde

Mano
Jul 11, 2012

cat botherer posted:

:actually:, the popular DL packages do actually compile the model graphs. But yeah, that won't help with logging the training data (which people would have every incentive to bullshit anyway).
oh, interesting

cat botherer posted:

Sometimes these models can actually output almost exact copies of some of their training data. This paper shows some examples for Stable Diffusion, but I've seen similar results with generative LLMs. However, that's still impossible to do at scale, and models certainly never copy the vast majority of their training examples (it would be information-theoretically impossible).

That would be impractical, especially with how much training goes on, and the sheer volume of images or w/e.

well yeah, storing the entire internet is kinda hard

fuctifino
Jun 11, 2001

https://twitter.com/LozzaFox/status/1751983862240944581

:munch:

Tweeted at one minute past 3. I guess it's quite a brutal and short judgement.

fuctifino fucked around with this message at 16:07 on Jan 29, 2024

Necrothatcher
Mar 26, 2005




https://x.com/PA/status/1751985068833210823?s=20

Failed Imagineer
Sep 22, 2018

Wow he can read so fast

Necrothatcher
Mar 26, 2005





gonna go down and watch this - might grab some popcorn first and live that emoji for real

Pistol_Pete
Sep 15, 2007

Oven Wrangler
Just randomly saying that it's weighing on me how this Palestinian genocide is ongoing and the UK establishment's (and the greater western one's) response to it is to firmly pretend that they can't see anything, while they aid and abet Israel and shriek accusations at anyone who stands up and speaks out about what's happening.

Just so much cognitive dissonance at the moment.

Failed Imagineer
Sep 22, 2018

Necrothatcher posted:

gonna go down and watch this - might grab some popcorn first and live that emoji for real

Pack of monster munch and a Cornetto

smellmycheese
Feb 1, 2016

lol. And indeed lol again

Jedit
Dec 10, 2011

Proudly supporting vanilla legends 1994-2014

Well, there goes another UKMT poster for saying incredibly inappropriate poo poo.

mossyfisk
Nov 8, 2010

FF0000

cat botherer posted:

That would be impractical, especially with how much training goes on, and the sheer volume of images or w/e.

It can be done, the models are built using exactly such archives of image data - it's certainly within the capability of the state. These companies simply want to pretend it's impossible and skimp on accountability. It would be more expensive to comply with these hypothetical regulatory processes, but so what? They couldn't sell a product that doesn't pass, so they'd have to either comply or shut down.

smellmycheese
Feb 1, 2016

Jedit posted:

Well, there goes another UKMT poster for saying incredibly inappropriate poo poo.

I didnt think referring to Cornetto's was probe-worthy?

Edit: Eurrrk. Just seen the offending post lol

fuctifino
Jun 11, 2001

https://twitter.com/LozzaFox/status/1751991214994329942

Necrothatcher
Mar 26, 2005




Say what you will, the man can draw a crowd.



Also Piers Corbyn is here.

Angepain
Jul 13, 2012

what keeps happening to my clothes

he seems to be rambling a bunch about demanding to know exactly what the definition of a racist is. not sure what this has to do with calling someone a pedo

Necrothatcher
Mar 26, 2005




fuctifino
Jun 11, 2001

Angepain posted:

he seems to be rambling a bunch about demanding to know exactly what the definition of a racist is. not sure what this has to do with calling someone a pedo

Cocaine is a terrible drug

The Question IRL
Jun 8, 2013

Only two contestants left! Here is Doom's chance for revenge...

Stopping to light a cigarette during your statement where you explain that you aren't an arsehole for defaming a bunch of people by calling them "paedophiles."

Classy, an in keeping with all the cigarette chat from earlier. Does anyone know what brand he's likely to be smoking?

Chubby Henparty
Aug 13, 2007


Sue me, Linneker, do it

smellmycheese
Feb 1, 2016

He - an upper class white man - is currently explaining what racism is

sinky
Feb 22, 2011



Slippery Tilde

Monster Munch recalled as they may contain traces of Laurence Fox.

StarkingBarfish
Jun 25, 2006

Novus Ordo Seclorum

His complete absence of any form of chin really sets this picture off. I don't even need a transcript to know he's pontificating about how the sovcit definition of racism he ascribes to is akshually ephebophilia or something.

killerwhat
May 13, 2010


Did you offer him one?

Necrothatcher
Mar 26, 2005




killerwhat posted:

Did you offer him one?

gently caress no, I'm not having his filthy nicotine-stained fingers in my monster munch.

fuctifino
Jun 11, 2001

:lol: and :lmao:

https://twitter.com/bat020/status/1751991126993645672

:allears:

cat botherer
Jan 6, 2022

I am interested in most phases of data processing.

mossyfisk posted:

It can be done, the models are built using exactly such archives of image data - it's certainly within the capability of the state. These companies simply want to pretend it's impossible and skimp on accountability. It would be more expensive to comply with these hypothetical regulatory processes, but so what? They couldn't sell a product that doesn't pass, so they'd have to either comply or shut down.
It is possiible, but not practical. Training datasets can be changed over time, and it's easy to lie about what a model is trained with. It's not only big tech companies doing this training either, it's also oodles of small companies and individuals (you can fine-tune a lot of models on a decent laptop). Think of how hard it is to enforce copyright on the internet when it comes to literal copies on places like Pornhub or even Youtube. This is much harder. It's not like you can have a guy watching over the shoulder of everyone training these models.

I'm not saying this is good. It isn't. Large companies with more to lose would probably be easier to bring into a decent degree of compliance, but ultimately the genie is not going back in the bottle.

cat botherer fucked around with this message at 17:10 on Jan 29, 2024

Tesseraction
Apr 5, 2009

Found by a court of law to be a real dogshit person. Legally declared a oval office.

OwlFancier
Aug 22, 2013

You can't technically prove that someone didn't independently invent something but if you come out with a copy of someone else's work you're still gonna get done for copyright infringement.

I don't see why that would be any different because you invented a complicated machine to generate the thing. If it shits out something that is plagiaristically close to someone else's work then rinse the bastards for it.

cat botherer
Jan 6, 2022

I am interested in most phases of data processing.

OwlFancier posted:

You can't technically prove that someone didn't independently invent something but if you come out with a copy of someone else's work you're still gonna get done for copyright infringement.

I don't see why that would be any different because you invented a complicated machine to generate the thing. If it shits out something that is plagiaristically close to someone else's work then rinse the bastards for it.
Realistically, they could get got for things like outputting Mario or the Coke logo. However, that's only going to work for things like logos or copyrighted characters that show up *a lot*. Individual artists or writers would rarely have their work outputted verbatim, so it would be nearly impossible to prove that their work was used in training.

fuctifino
Jun 11, 2001

I guess he enjoys being sued? He's digging in

https://twitter.com/LozzaFox/status/1752001661441049057

Guavanaut
Nov 27, 2009

Looking At Them Tittys
1969 - 1998



Toilet Rascal

Tesseraction posted:

Found by a court of law to be a real dogshit person. Legally declared a oval office.
At least before the creation of The Supreme Court in 2009 the Law Lords in the House of Lords could legally do that.

Which would have been the funniest outcome of Graham Linehan attempting to circumvent all the courts by appealing directly to the Lords that he was being unfairly harassed last year, even though they no longer have said authority.

Appeal denied. Appellant must begin all future statements with "But I Am a Massive oval office" in a style parody of the HMS Pinafore.

Necrothatcher
Mar 26, 2005




fuctifino posted:

I guess he enjoys being sued? He's digging in

https://twitter.com/LozzaFox/status/1752001661441049057

He also repeatedly said "Gary Lineker is a racist" and then told him to sue him during his little spech.

OwlFancier
Aug 22, 2013

The Question IRL posted:

To me, this screams of his lawyer going "I have prepared a script for what you should say when you lose this, in the vain hope that an apology outside of court might prevent future damages and stop you from doing a Trump."

Ah optimism.

Adbot
ADBOT LOVES YOU

sinky
Feb 22, 2011



Slippery Tilde

fuctifino posted:

I guess he enjoys being sued? He's digging in

https://twitter.com/LozzaFox/status/1752001661441049057

It worked for Donald Trump.


Wait, no it didn't.

  • 1
  • 2
  • 3
  • 4
  • 5
  • Post
  • Reply