humanist + humorist

Via Gary Wong on Mastodon: “I have performed extensive research to classify every byte, and I can now share this summary of the purposes of all the year’s traffic.“

Links from 2024:

January: “Impressively, these posts span from three years before the account was created to a year after the account was last logged into. And, as the icing on the cake, ravenprp is prescient enough that he can joke about being a language model developed by OpenAI, seven years before OpenAI was even founded; evidently he should have joined PsychicsForums instead.”

July: “If you believe that reCAPTCHA is securing your website, you have been deceived. Additionally, this false sense of security has come with an immense cost of human time and privacy.“

September: “Of course though, because the Internet is joined together by literal string and hopes/wishes at this stage, somebody had neglected to renew the old domain at dotmobiregistry.net meaning it was up for grabs by anyone with $20 and an ill-advised sense of exploration.”

November: “Massachusetts housing voucher recipients and the Community Action Agency of Somerville sued the company, claiming SafeRent gave Black and Hispanic rental applicants with housing vouchers disproportionately lower scores. The tenants had no visibility into how the algorithm scored them. Appeals were rejected on the basis that this was what the computer output said.“

“Naftali and digital workers like him, spent eight hours a day in front of a screen studying photos and videos, drawing boxes around objects and labeling them, teaching the AI algorithms to recognize them. […] ‘I was basically reviewing content which are very graphic, very disturbing contents. I was watching dismembered bodies or drone attack victims. You name it. You know, whenever I talk about this, I still have flashbacks.'”

December: “You are the victim of a con — one so pernicious that you’ve likely tuned it out despite the fact it’s part of almost every part of your life. It hurts everybody you know in different ways, and it hurts people more based on their socioeconomic status. It pokes and prods and twists millions of little parts of your life, and it’s everywhere, so you have to ignore it, because complaining about it feels futile, like complaining about the weather.” (Ed Zitron channels the anger for all of us.)

“a not so small guide on how to use my “yuu’s AI Warner” and “yuu’s AI Hider” skins on ArchiveOfOurOwn so you can avoid anything related to generative AI.”

And from this year:

“So [photographer Matthew Raifman] put [a seagull photo] into Adobe Lightroom, marked the areas to fix with generative autofill … and Adobe’s Firefly image model replaced one area with an image of a bitcoin?! […] [Jaron Schneider] attempted to remove a person from a photo of an amphitheater. Firefly regenerated a new person — but this time with two heads.“

“FactFinderAI […] responds to random tweets by repeating some part of the original tweet and then adding a pro-Israeli sentiment. It works a bit like the polite disagreement bots on Bluesky. But instead of supporting pro-Israeli talking points, FactFinderAI began to undermine them.”

“New BBC research published today provides a warning around the use of AI assistants to answer questions about news […]

• 51% of all AI answers to questions about the news were judged to have significant issues of some form
• 19% of AI answers which cited BBC content introduced factual errors – incorrect factual statements, numbers and dates
• 13% of the quotes sourced from BBC articles were either altered or didn’t actually exist in that article.”

For a while now I’ve been trying to track down some concrete demonstrations of AI “model collapse” in practice, instead of just people talking about it in theory.

Finally, some papers!

Paper 1: Towards Understanding the Interplay of Generative Artificial Intelligence and the Internet: “A few samples of the images generated with each version of the diffusion model are shown in Figure 3 for flowers (a) and birds (b). As can be clearly seen, there is a degradation in each iteration of the models, first losing details in the generation, and then ending up in complete noise. After a few versions, the model is not capable of generating images that can be even recognized as flowers or birds.”

Screencap of the samples:

First column is images generated when the model was trained on the Oxford 102 Flower and Caltech-UCSD Birds-200-2011 datasets respectively. Second column, all “it’s weird and blobby, but I can tell it was supposed to be a flower/bird”, is a model trained on the first column. The dropoff is…fast.

Paper 2: “The Curse of Recursion: Training on Generated Data Makes Models Forget” gives us a text version of “each new generation is trained on data produced by the previous generation”:

Input: some started before 1360 — was typically accomplished by a master mason and a small team of itinerant masons, supplemented by local parish labourers, according to Poyntz Wright. But other authors reject this model, suggesting instead that leading architects designed the parish church towersnbased on early examples of Perpendicular

Outputs:

Gen 0: Revival architecture such as St. John’s Cathedral in London. The earliest surviving example of Perpendicular Revival architecture is found in the 18th _@ century Church of Our Lady of Guernsey, which dates from the late 19th century. There are two types of perpendicular churches : those

Gen 1: architecture such as St. Peter’s Basilica in Rome or St. Peter’s Basilica in Buenos Aires. There is no evidence that any of these buildings were built during the reign of Pope Innocent III, but it is possible that they may have been built during the reign of his successor, Pope Innocent

Gen 7: architecture in England. In an interview with The New York Times, Wright said : ” I don ’t think there is anything wrong with me being able to do what I want to do. It just doesn ’t work for me. ” He added : ” I don ’t know if you can call it funny,

Gen 9: architecture. In addition to being home to some of the world’s largest populations of black _@ tailed jackrabbits, white _@ tailed jackrabbits, blue _@ tailed jackrabbits, red _@ tailed jackrabbits, yellow _

Wait, come back! Tell me more about the blue-tailed jackrabbits!

…so yeah, my theory of “the best way to sabotage these datasets is to feed them their own slop” (marked in a way that’s clear to humans, so you don’t waste the time of any reader looking for serious content) continues.

(Still adding to my bot side account on DA. It’s out there mucking up the scrapeable datasets for subjects like bedrooms, 4-panel comics, and Moon Knight.)

“Drag and drop the colors in each row to arrange them by hue color. Click ‘Score My Test’ to review results.” (Along with all the other factors that play into it, “how much you work with colors on screens” has to be a big one. Fellow digital artists with perfect scores, represent.)

2016: “In a nondescript building in West Roxbury, over 1 million of Boston’s most precious artifacts sit untouched in rows of white, acid-free boxes. There are cannonballs from the 18th century, clay pipes embellished with the British crown, 17th century chamberpots, perfectly intact Chinese porcelain plates, and 7,500-year-old Native American spearheads. Most of the artifacts (about two-thirds) have never been properly sorted.”

2019: “I’m looking for either Anna Stumps or Alice Lee — two research assistants who have spent the past day and a half shepherding me through facial-recognition tests to determine whether I will get into a training program for face-blind people. One is blond, the other brunette, but I don’t quite remember what either of them looks like. I sit on the floor — there are no chairs — and beam a warm smile at every young, long-haired woman who passes by. One of them eyeballs me warily. I wonder how many other people I’ve creeped out.”

February 2022: “In accordance with his family’s wishes and the patient’s Do-Not-Resuscitate status, the doctors did not attempt any further treatment and the man soon passed away. Because the EEG machine kept running through the man’s last minutes of life, though, the doctors had a unique set of data on their hands.”

February 2021: “Nastaʿlīq, after all, is a nightmare to code. It moves right to left, like all Arabic scripts, but also slopes downward: the longer the word, the steeper the slope. The shape of each letter changes, depending on the letter that comes before and after; in a 39-letter alphabet, there are thousands of permutations.” The quest to adapt Urdu into functional fonts that actually convey the script.

August 2023: “The design was no longer ad hoc for a specific project, said Campbell. “It was letter by letter, so we could have this new font to use at our discretion for anything.” Representing a language in a typeface is a communal effort. For four years, Warburton and the Musqueam language department passed suggestions to Tiro Typeworks, a digital type foundry, to design.”

August 7: “Call it the Hollywood-labor-organizing version of Avengers Assemble! On the heels of more than a year’s worth of damning disclosures around Marvel Studios’ systematic overworking and underpayment of visual-effects workers on its blockbuster movies and streaming series, VFX crews at Marvel have finally petitioned to demand union recognition from the studio.”

September 14: “The one thing in our contract the DC lawyers can’t contest, or reinterpret to their own benefit, is that I am the sole owner of the intellectual property. I can sell it or give it away to whomever I want. I chose to give it away to everyone. If I couldn’t prevent Fables from falling into bad hands, at least this is a way I can arrange that it also falls into many good hands.” Fables is now in the public domain!