Me, 1996: No, "Goodtimes" is a hoax, an email can't run computer code or give you a virus
Me, 2000, miserable: Microsoft did something weird and now emails can give you a virus. Bubbleboy is real
---
Me, 2010: No, the "Enhance" bit from CSI is not real. There aren't algorithms that can add information to an image.
Me, 2025, miserable: Microsoft did something weird and now the "Enhance" button is real, but the details are fake. Law enforcement IS using it and innocent people will go to prison
@bob @mcc That is absolutely absurd, yeah. I don't get how a scientist looking at something *no one else has seen ever before* can justify *hey, let's just extrapolate everything else we've already seen onto this new thing*.
Why even bother looking to the stars if you just assume there's nothing new in the heavens?
@kevinriggle @xgranade @doephin @mcc
In the 90s we used to think that when people had access to more information it would inevitably lead to a freer world, because truth would drive lies out of circulation in a sort of informational-reverse-Gresham's-Law.
Sadly the opposite seems to be the case: a decent-sized chunk of humanity sees truth in all its complexity and recoils from it, preferring a lie that preserves their simplistic preconceptions. Alex Jones and Rupert Murdoch made their fortunes off this.
And those people get to serve on juries.
@passenger @kevinriggle @xgranade @mcc "fun" fact- once when I did jury duty the judge asked "would you find the defendant guilty right now? If so raise your hand" pre-trial. Like we didn't even know the dude's name.
*Everyone* except me raised their hands.
yeaah. Trial by jury: definitely one of those "best of worst options"things
@passenger @kevinriggle @xgranade @doephin @mcc There's a bit of a group project effect when there's this much knowledge. Whoever is willing to speak up and take control of the group will end up guiding it.
People who are knowledge seeking tend to recognize their own ignorance and not seek to control the group.
People who have enough smarts to have a bit of knowledge alongside ego will think they are brilliant and seek the power trip.
The rest just want the assignment done.
@passenger @kevinriggle @xgranade @doephin @mcc Humility in knowledge seeking isn't seen as an asset so we get stuck with the loudest voice in the room who knows just enough to sound smarter than everyone else and people feel deeply comforted by the confidence of the person who is loud and seems smart.
I've noticed this when people call me "smart" on topics I'm passionate about (because passion can look like confidence). Its deeply uncomfortable realizing that people use this to grift.
That's a really interesting insight, thank you!
I've noticed that this is often something where deference to authority plays a huge part. For example, in corporate events people will assume that management have the most knowledge, in legal events they'll defer to police, et cetera. Even if that person doesn't actually have anything useful to say, their rank will count in the same way as "being loud and seeming smart."
During 2020, when we set up the mutual aid movements, I (and some others) ended up unexpectedly in that position because we had been talking about mutual aid for years. It was weird.
Wow
Want an infuriating quote?
"ATSC had been dealing with doubters for ten years and that the device was merely being criticised because of its "primitive" appearance. He said: "We are working on a new model that has flashing lights" "
@ShadSterling @mcc Remember sitting at the cinema in Man of Steel and hearing a character shout "it's coming in via the RSS feeds!" ?
... more innocent times.
@artcollisions @mcc @psistarpsiii
Hmmmmm......
😀
@artcollisions @mcc @psistarpsiii That is from scanning a photograph, though, so it’s much more plausible, since scanning software often shows a low resolution image first and lets you scan a high resolution image of an area later.
But still not plausible.
@mcc You just reminded me of that meme image that took CSI Miami characters and had them zoom in on a license plate in a grainy 320x240 security camera image, "enhance" and get the license numbers via magic, then they zoomed in on a screw of the license plate, enhanced that, and they had a reflection of the killer in the act that apparently would hold up in court followed by the annoying guy from the show saying "looks like... *puts on sunglasses* the killer is screwed."
The tech can't do it, but it won't stop people from believing. So called "AI" still can't add details that aren't there without making them up, but it's not going to stop people from thinking it's correct anyway. *Big sigh*
CSI has nothing on Red Dwarf, they're the real experts in enhancing images:
@wurzelmann @mcc Yeah, I remember that one. I'm 99.999% certain it was parodying the same thing.
I rounded down.
@mcc I've heard some histories of dousing rods in police forces, and it's brutal because not only are people/critters getting hurt, but you realize that there are no safety rails and that every single time a technology comes up like this that offers the illusion of magic, cops are just going to jump on top of it and judges will shrug and say, "I guess that's fine."
You see bite mark analysis and all this shit and realize, "oh, this will *always* happen.
@mcc Decades to push back on ideas that are even stupider than "enhance", literally "we have a magic branch that tells us where bodies are."
And then in the space of a year or two, another bullshit grift pops up and you think - even once public opinion shifts, it's going to be 30 years from now and activists will still be arguing to stop this from showing up in court. Because they believed in magic fucking sticks, and nothing has changed.
@mcc
> Microsoft did something weird and now emails can give you a virus.
This whiplash struck me in the late 1990s, yeah. Had to revise the advice I was giving to loved ones.
What an obviously stupid decision by MS, surely everyone who says they want security will abandon this single program that makes them uniquely vulnerable, and use any of the others that don't have this flaw.
The growing horror as I observed exactly the opposite trend, has never left me.
“Law enforcement IS using it and innocent people will go to prison”
I suspect that’s exactly what happened in the UnitedHealthcare CEO shooting case.
They released a bunch of surprisingly clear pictures of the suspect, but the men in the pictures didn’t look like each other.
That would also explain why the men in the photos were abnormally good looking; models and actors are over represented in the AI training sets.
@feich ah yes, let me just do a tax evasion. i'm sure that will work out wonderfully for me as i get sent to prison.
oh right, sorry. strikes only work if everyone does them, don't they? all i have to do is convince everyone in the entire country not to pay their taxes. seems simple enough. once that's done, they won't be able to arrest us all, they won't have the prison space or the money. of course, they also won't have the money for any of the social programs my tax dollars currently support, and the people who depend on those programs will likely die, but hey, at least we're not funding the largest military on earth anymore! that's good, right?
Je länger man über die Aufnahme von Alina Lipp und Thomas Röper auf die EU-Sanktionsliste nachdenkt, desto gravierender wird der damit vollzogene Rechtsbruch. Man muss sich die konkreten Folgen in Deutschland vorstellen, um die wahre Bedeutung zu erfassen.
Je länger man über die Aufnahme von Alina Lipp und Thomas Röper auf die EU-Sanktionsliste nachdenkt, desto gravierender wird der damit vollzogene Rechtsbruch. Man muss sich die konkreten Folgen in Deutschland vorstellen, um die wahre Bedeutung zu erfassen.
@mcc Just the other day I was annoyed at the "AI enhancement" in cameras, remembered that "check behind corners" tech from Bladerunner and yes, people are apparently working on it: https://www.nature.com/articles/s41467-024-45397-7
Looking forward to when someone combines it with gen AI to just make stuff up and get a conviction.
Ordinary cameras cannot directly see objects hidden around corners. Here, from an ordinary indirect 2D photograph, the authors compute a 3D image of a scene hidden behind a doorway by exploiting two perpendicular edges of the doorway.
@mcc This was possible before 2010. There were face enhancement things for CCTV in the ‘90s that were trained on a set of faces and would fill in the gaps.
The key idea is that the image is not the only information you have. You know, a priori, that this blob is a face. Or you know that it’s a number plate and so the shapes must be letters and digits. And maybe it’s a UK number plate and so must be letters and numbers in a particular sequence. And maybe it’s a number plate for a specific type of vehicle so (assuming it’s a valid number plate) must be one of this known set.
This lets you treat it as a lossy compression problem with an external dictionary, rather than an image enhancement problem. The data you have is the lossily compressed data and the a priori dataset is the dictionary. Now you ‘just’ have to write the decoder.
The interesting experiment that drove some of this was a project to put biometrics on a magnetic strip for MoD systems that found that 50 bits was sufficient to uniquely identify facial features (which makes intuitive sense: there were around 232 people when they did this work and even if they all have unique faces then 50 bits is ample to have a sparse space containing all of them). From there, it follows that you can extrapolate a face from a few pixels as long as they are the right pixels (spoiler: you usually don’t).
EDIT: I remember a fractal image compression program from the ‘90s that claimed to be able to produce more detail than was in the original. It came on a magazine cover disk (floppy, not even a CD). You fed it an image and it would, effectively, produce a small program that generated something very close to that image. It took several minutes to run on my 386 but then you got an image that was resolution independent. You could make it generate arbitrary levels of detail, zooming into things that were single pixels on the original. Just zooming in a bit gave quite plausible things. They were all extrapolated from the rest of the image and nearby pixels, plus whatever biases were present in the generator. After a bit more zooming, you got obvious nonsense, but by then the entire screen was a single pixel from the original image.
@brezelradar @mcc Even without too much magic, bilinear filtering is a simple function that says 'assume these four pixels form a linear function over the colour space, predict values for places between the pixels'. Bicubic filtering takes the two pixels on each side and fits them to a cubic function. You can always use more pixels and pick higher-order polynomials.
If you pick sufficiently high-order polynomials to the point where you cannot make any claims about the relationship between the polynomial and anything in the real world, you are qualified to call yourself a 'machine learning engineer'.
@david_chisnall @mcc The advantage of the diamond method was that it could be done in real-time when implemented in hardware (cheap ASIC) and the results looked arguably better for SD TV up-scaling to HD / full HD.
And you could test patent violations using a simple VHS cassette with a certain test pattern. 😈
@david_chisnall @mcc If I remember correctly, polynomial interpolation tends to overshoot/deteriorate quickly (unless you only use splines, I guess).
You'd have to do a direct comparison with a bunch of people and certain static and moving pictures.
The test video with a car driving behind a moving grating comes to mind. Not sure how that one is called, its been a while: "Lenna" is only sample I still know by name. Today you'd probably use Bad Apple or Doom for that.😄
@mcc I think Samsung did it first with the S23 'enhancing' ie. faking shots of the moon.
I hope Microsoft are paying Samsung royalties.
https://www.theverge.com/2023/3/13/23637401/samsung-fake-moon-photos-ai-galaxy-s21-s23-ultra
@mcc Maybe the problem is (obsolete) naming?
Email used to be a blob of text with a protocol header. Now it's a web page, complete with embedded scripts, links, and links to other scripts. And the "mail reader" is a full browser (like almost every other app). And the browser has shit security.
"Enhance" is straight-up marketing deception. The word should be "embellish".
mcc
Cassandra Granade 🏳️⚧️
bob
doephin
Kevin Riggle
Passenger
ClaraBlackInk
Excessive Minimalism
mmby
The Cookie Witch
climbertobby
TeflonTrout 
he/him
Kevin Karhan 
Alice 
ShadSterling
Uli Kusterer (Not a kitteh)
aburka 🫣
The Psychotic Network Ferret
Granny Art (Shrimp) (Joni)
Fork_Merge
Magnus Ahltorp
Nazo
Wurzelmann
normy foxyoreos
Bitslingers-R-Us
Sean Kleefeld
Frank J
Geoff Coffey
bignose
Omar Jarjur
feich
Dr Andrew A. Adams #FBPE 🔶
L'égrégore André ꕭꕬ
David Chisnall (*Now with 50% more sarcasm!*)
embix
¡Abbie!
Philippa Cowderoy
Gabe
MarjorieR
Poligofsky 🇨🇦
Robert Petersen 
Dataline