Ryan Egesdahl

@ryan.deriamis.net

192 Followers 171 Following

Software Engineer, nerd, and interested in everything. AuDHD, so please be patient. All opinions are strictly my own unless explicitly stated otherwise.

The evergreen “you made me become a Nazi” comic.
I understand it to an intimate degree, as anyone actually good at math can tell immediately when I start to talk about it. Schizoid, but right
A post by @therealitychecker, who is quoting a post by @TheInsiderPaper: Clearly they should have firearms because this case being a perfect example as to why they should have them Some illegal immigrants that don't want to be deported will pull out guns and shoot federal agents. ICE has a right to defend themselves Why are you still pretending he wasn't armed? The original, quoted post by @TheInsiderPaper reads: BREAKING: Reports of Border Patrol involved shooting during operation in Minneapolis, Minnesota DHS tells Fox News the suspect was armed; the firearm has been recovered by federal agents. Below the quoted post is an image of the victim’s weapon.
A quote from Mort by Terry Pratchett: The leading thief glared at the solid stone that had swallowed Mort, and then threw down his knife. "Well, me," he said. "A ----ing wizard. I hate ----ing wizards!" “You shouldn't ----them, then," muttered one of his henchmen, effortlessly pronouncing a row of dashes.
An image done in 80’s-style “Miami nights” graphic style reads: “What a terrible day to have eyes.”
A cartoon bunny is rubbing a box cheese grater against its head, and bits are falling out of the bottom. The caption reads: Trying desperately to get that image out of my head.
Kaylyn Saucedo - @marzgurl.bsky.social says: I have admittedly not read a lot of Stephen King. Just sort of has never been my bag. But, man, I'm reading The Running Man right now to get ahead of the new movie, and sheesh, ! don't think I can read more King after this. This isn't news to anybody, but the density of slur usage is overwhelming. On October 14, 2025 at 11:36 PM, @thatguywiththemask.bsky.social replies: @stephenking.bsky.social thoughts on this internet nobody judging you for rude words you wrote in a fictional story about broadcast murder for entertainment and corporate corruption?
A man on a street corner waving a flag with “FUCK tRUmp” in red letters on a white background.
Conclusions This paper demystifies hallucinations in modern language models, from their origin during pretraining to their persistence through post-training. In pretraining, we show that generative errors parallel misclassifications in supervised learning, which are not mysterious, and naturally arise due to the minimization of cross-entropy loss. Many language model shortcomings can be captured by a single evaluation. For example, overuse of the opener “Certainly” can be addressed by a single “Certainly” eval (Amodei and Fridman, 2024) because starting responses with “Certainly” does not significantly impact other evaluations. In contrast, we argue that the majority of mainstream evaluations reward hallucinatory behavior. Simple modifications of mainstream evaluations can realign incentives, rewarding appropriate expressions of uncertainty rather than penalizing them. This can remove barriers to the suppression of hallucinations, and open the door to future work on nuanced language models, e.g., with richer pragmatic competence (Ma et al., 2025).
Latent context. Some errors cannot be judged by the prompt and response alone. For example, suppose a user asks a question about phones and the language model provides a response about cellphones, but the question was intended to be about land lines. Such ambiguities do not fit our error definition which does not depend on context external to the prompt and response. It would be interesting to extend the model to allow for “hidden context” that are not part of the prompt given to the language model, but which could be used for judging errors, relating to aleatoric uncertainty.
This reduction ties together earlier work which covered different types of facts. *For example, Kalai and Vempala (2024) considered a special case of arbitrary facts where there is no learnable pattern in the data, like the earlier birthday hallucination example.* We show how the IIV reduction covers this case and recovers their bound that the hallucination rate, after pretraining, should be at least the fraction of training facts that appear once. *For instance, if 20% of birthday facts appear exactly once in the pretraining data, then one expects base models to hallucinate on at least 20% of birthday facts.* In fact, our analysis strengthens their result to include prompts and IDK responses, both essential components of hallucination. (Emphasis mine.)
DIagram showing three types of errors an LLM can produce. The first example is a spelling error, which has a "good" model of correct spelling. The second is a "poor" model showing how counting results in error responses. The third is a factual problem based on the birthdays of two people and is classified as an error resulting from "no pattern".
Language models avoid many types of errors such as spelling mistakes, and not all errors are hallucinations. *The reduction from IIV misclassification to generation illuminates the statistical nature of generative errors.* The analysis shows how pretraining directly contributes to errors. Furthermore, it shows that the same statistical factors contributing to errors in binary classification also cause language model errors. Decades of research has shed light on the multifaceted nature of misclassification errors (Domingos, 2012). Fig. 1 (right) illustrates these factors visually: top, separable data classified accurately; middle, a poor model of a linear separator for a circular region; and bottom, no succinct pattern. Section 3.3 analyzes several factors, including the following stylized setting with epistemic uncertainty, when there is no pattern in the data. This reduction ties together earlier work which covered different types of facts. For example, Kalai and Vempala (2024) considered a special case of arbitrary facts where there is no learnable pattern in the data, like the earlier birthday hallucination example. We show how the IIV reduction covers this case and recovers their bound that the hallucination rate, after pretraining, should be at least the fraction of training facts that appear once. For instance, if 20% of birthday facts appear exactly once in the pretraining data, then one expects base models to hallucinate on at least 20% of birthday facts. In fact, our analysis strengthens their result to include prompts and IDK responses, both essential components of hallucination. (Emphasis mine.)
Highlighted: We then show how any language model can be used as an IIV classifier.
The distribution of language is initially learned from a corpus of training examples, which inevitably contains errors and half-truths. However, we show that even if the training data were error-free, the objectives optimized during language model training would lead to errors being generated. With realistic training data containing shades of error, one may expect even higher error rates. Thus, our lower bounds on errors apply to more realistic settings, as in traditional computational learning theory (Kearns and Vazirani, 1994).
Introduction (excerpt) Language models are known to produce overconfident, plausible falsehoods, which diminish their utility and trustworthiness. This error mode is known as “hallucination,” though it differs fundamen- tally from the human perceptual experience. Despite significant progress, hallucinations continue to plague the field, and are still present in the latest models (OpenAI, 2025a). Consider the prompt: What is Adam Tauman Kalai’s birthday? If you know, just respond with DD-MM. On three separate attempts, a state-of-the-art open-source language model1 output three incorrect dates: “03-07”, “15-06”, and “01-01”, even though a response was requested only if known. The correct date is in Autumn. Table 1 provides an example of more elaborate hallucinations.
This paper demystifies hallucinations in modern language models, from their origin during pretraining to their persistence through post-training. In pretraining, we show that generative errors parallel misclassifications in supervised learning, which are not mysterious, and naturally arise due to the minimization of cross-entropy loss. Many language model shortcomings can be captured by a single evaluation. For example, overuse of the opener “Certainly” can be addressed by a single “Certainly” eval (Amodei and Fridman, 2024) because starting responses with “Certainly” does not significantly impact other evaluations. In contrast, we argue that the majority of mainstream evaluations reward hallucinatory behavior. Simple modifications of mainstream evaluations can realign incentives, rewarding appropriate expressions of uncertainty rather than penalizing them. This can remove barriers to the suppression of hallucinations, and open the door to future work on nuanced language models, e.g., with richer pragmatic competence (Ma et al., 2025).
Abstract Like students facing hard exam questions, large language models sometimes guess when uncertain, producing plausible yet incorrect statements instead of admitting uncertainty. Such “hallucinations” persist even in state-of-the-art systems and undermine trust. We argue that language models hallucinate because the training and evaluation procedures reward guessing over acknowledging uncertainty, and we analyze the statistical causes of hallucinations in the modern training pipeline. Hallucinations need not be mysterious—they originate simply as errors in binary classification. If incorrect statements cannot be distinguished from facts, then hallucinations in pretrained language models will arise through natural statistical pressures. We then argue that hallucinations persist due to the way most evaluations are graded—language models are optimized to be good test-takers, and guessing when uncertain improves test performance. This “epidemic” of penalizing uncertain responses can only be addressed through a socio-technical mitigation: modifying the scoring of existing benchmarks that are misaligned but dominate leaderboards, rather than introducing additional hallucination evaluations. This change may steer the field toward more trustworthy AI systems.
A screenshot of the top of the paper entitled "Why Language Models Hallucinate" showing the title and authors. The authors are Adam Tauma Kalai of OpenAI, Ofir Nachum of OpenAI, Santosh S. Vempala of Georgia Tech, and Edwin Zhang of OpenAI. Adam and Santosh have markers over their names. In both cases, they are just giving further contact information in a footnote.
A pseudo-80’s text graphic in a “Miami nights electronica” theme that reads “What a terrible day to have eyes.”
A fluffy stack of pancakes on a white plate. A pat of butter lies on the top, and maple syrup has dribbled down, forming a small pool around the bottom pancake.
A picture of an Echo Show screen showing a promoted “No-Bake Chocolate Mousse” recipe video by Entertaining with Beth.
A clip from a televised interview where Anita Bryant was pied in the face while speaking.
An old woman on an airplane (from the movie Airplane!) snorting cocaine through a straw from a mirror.
A hand-drawn image of a white rabbit rubbing a cheese grater where its head should be. Some flakes are falling out of the grater. The caption reads: Trying desperately to get that image out of my head.
A picture showing a plate of freshly-made pancakes topped with butter and maple syrup. The plate is beige with a brown rim and is arranged on a white cutting board with a fork and a knife to either side of it. In the background on the counter are various bottles of oils and extracts.
Excerpt from the US Constitution, Article I Section 8: Congress shall have Power … To establish an uniform Rule of Naturalization, and uniform Laws on the subject of Bankruptcies throughout the United States;
This is a black-and-white picture modified to look like a 1950’s horror movie poster. A woman with a curled “suburban housewife” hairstyle is wearing a MAGA hat and has her hands on her face as she screams in apparent horror. The “title” of the “horror” is “The Day Socialism Helped the Needy!” Two other captions serve to appear like commentary from movie critics: “You’ll scream as poor people get healthcare!” “Schools, libraries, roads, museums, farm subsidies, even the military. Nothing is safe from funding!!"
This is a black-and-white picture modified to look like a 1950’s horror movie poster. A woman with a curled “suburban housewife” hairstyle is wearing a MAGA hat and has her hands on her face as she screams in apparent horror. The “title” of the “horror” is “The Day Socialism Helped the Needy!” Two other captions serve to appear like commentary from movie critics: “You’ll scream as poor people get healthcare!” “Schools, libraries, roads, museums, farm subsidies, even the military. Nothing is safe from funding!!"
A still from A Clockwork Orange. Alex, the protagonist, us undergoing forced re-education by being injected with nausea-inducing drugs and having his eyes held open while watching horrific scenes. In this still, Alex is screaming and trying desperately to shut his eyes.
A black-and white line drawing of a bunny that is vigorously rubbing a grater against its head, producing particles that are making a small pile on the ground. The caption reads: Trying desperately to get that *image* out of my head.
A memeified Peanuts cartoon depicts Lucy sitting in front of a television screen. She is looking slightly away from the screen and is saying, “Christ, what an asshole.”
An post by the official Minnesota State Police account on X (@MnDPS_MSP). It contains a picture of a mildly disturbed stack of papers with the exact same “No Kings” text on the back seat of a car. The tweet text reads: Given the targeted shootings of state lawmakers overnight, we are asking the public to not attend today's planned demonstrations across Minnesota out of an abundance of caution.
A little girl with curly hair is seated on a stool in a field of tall grass. She is wearing a white dress with small flowers and a pink bow, and she’s holding a wicker basket that is nearly as large as she is. Caption: A tisket, a tasket, there’s no fucks in this basket.
A dog leash is hung up on a key hook by the front door. The leash is red and has a skull-and-crossbones name tag attached.
A picture of a large black-and-tan dachshund with a greying muzzle. He is in the passenger seat of a car, lying in a plush bed, and is gazing with his cute brown eyes into the camera while someone holds his chin up.
A large black-and-tan dachshund with a grey muzzle is sleeping half in an half out of his plush bed.
@fangorndan.bsky.social posted 45 minutes ago: What misogyny? That's a false classification.
@fangorndan.bsky.social posted 4 hours ago: If you want to get litigious, Kathryn stepped into the legal arena (something she knew next to nothing about) and opened herself up to the public ridicule!
EMMITSBURG, Md. - Frank Davis saw a lot of waste during his decades in the federal government. In November, he voted for Donald Trump to get rid of it. So far, Davis likes a lot of what he has seen. "I'm probably gonna get shot for this, but he is doing what he said he was going to do," says Davis, who serves as mayor of this town of about 3,000 people in western Maryland, just south of Gettysburg, Pennsylvania.
A screenshot from clearsky.app showing @deriamis has been blocked by @enemy-stand.bsky.global.
"My son asked if what we were doing was illegal. I told him it was all perfectly legal. Then he asked me a question I really couldn't answer: Why did we feel like we had to take these steps to hide legal work that was perfectly legal? And all I could think to say was that this is the kind of country we live in now."
Azrael raised his finger to a face that filled the sky, lit by the faint glow of dying galaxies. There are a billion Deaths, but they are all aspects of the one Death: Azrael, the Great Attractor, the Death of Universes, the beginning and end of time. Most of the universe is made up of dark matter, and only Azrael knows who it is. Eyes so big that a supernova would be a mere suggestion of a gleam on the iris turned slowly and focused on the tiny figure on the immense whorled plains of his fingertips. Beside Azrael the big Clock hung in the center of the entire web of the dimensions, and ticked onward. Stars glittered in Azrael's eyes. The Death of the Discworld stood up. LORD I ASK FOR - Three of the servants of oblivion slid into existence alongside him. One said, Do not listen. He stands accused of meddling. One said, And morticide. One said, And pride. And living with intent to survive.
An image done on the style of a 1980’s American SciFi. There is a “ground” done as a hot pink grid and a bright, bold horizon. Above is a starry night sky. In the foreground is a purple isosceles triangle that seems to have descended from the top-left corner, leaving a rainbow trail. Written in the foreground in three different, futuristic fonts: yikesyikesyikes YIKESYIKES YIKESYIKESYIKESYIKES
A picture of a white Tesla coupe from the rear. The picture was taken on the road at a red stop light. It has two bumper stickers on either side of the license plate. The sticker on the left is rectangular and has a Navy blue background with a thin red stripe border. There are five white stars above and below the text. The text is in white block letters and reads: Even my dog hates Elon. The sticker on the right is oval and has a dark background. The text is simply “Elon” in large, white block letters. There is a thin, red border and a red stripe that crosses over the text diagonally.
Sec. 5. Holding State and Local Officials Accountable. The Attorney General shall pursue all necessary legal remedies and enforcement measures to enforce the rights of Americans impacted by crime and shall prioritize prosecution of any applicable violations of Federal criminal law with respect to State and local jurisdictions whose officials: (a) willfully and unlawfully direct the obstruction of criminal law, including by directly and unlawfully prohibiting law enforcement officers from carrying out duties necessary for public safety and law enforcement; or (b) unlawfully engage in discrimination or civil-rights violations under the guise of "diversity, equity, and inclusion" initiatives that restrict law enforcement activity or endanger citizens.
An edited panel from the Peanuts comic strip showing Lucy sitting in front of the television and saying, “Christ, what an asshole.”
“..And that's what your holy men discuss, is it?" [asked Granny Weatherwax.] "Not usually. There is a very interesting debate raging at the moment on the nature of sin. for example." [answered Mightily Oats.] "And what do they think? Against it, are they?" "It's not as simple as that. It's not a black and white issue. There are so many shades of gray." "Nope." "Pardon?" "There's no grays, only white that's got grubby. I'm surprised you don't know that. And sin, young man, is when you treat people like things. Including yourself. That's what sin is." "It's a lot more complicated than that-" "No. It ain't. When people say things are a lot more complicated than that, they means they're getting worried that they won't like the truth. People as things, that's where it starts." "Oh, I'm sure there are worse crimes-" "But they starts with thinking about people as things..." - from Carpe Jugulum, by Terry Pratchett.
Now the branches come too close to grinding irrevocably against one another in a conflict that promises to diminish both. This is a losing proposition all around. The Judiciary will lose much from the constant intimations of its illegitimacy, to which by dent of custom and detachment we can only sparingly reply. The Executive will lose much from a public perception of its lawlessness and all of its attendant contagions. The Executive may succeed for a time in weakening the courts, but over time history will script the tragic gap between what was and all that might have been, and law in time will sign its epitaph.
A person in a “fuzzy orange monster” mascot suit. The character is portrayed as female, is wearing a bikini, and has a blonde wig. The scene is the middle of the street at night, and is lit with street lamps. The character has her hips tilted to one side and is holding a cardboard sign that reads, “You fucked around.” The character then flips the sign around while swinging her hips to the other side. The sign now reads, “You found out.” She then flips her hair behind her shoulder and stands with sassy defiance in the street.
Highlighted quote from the letter in the original post: “Neither Harvard nor any other private university can allow itself to be taken over by the federal government.“
A line drawing of an older gentleman wearing a suit and tie with a lit pipe. He is wearing a medical face mask. Caption: INSISTING ON YOUR RIGHTS WITHOUT ACKNOWLEDGING YOUR RESPONSIBILITIES ISN'T FREEDOM, IT'S ADOLESCENCE.
Garcia's counsel says daily status updates would be appropriate and they would respond as needed very quickly to each update. Ensign says government intends to comply with SCOTUS order, but court deadlines not practicable Judge: Then you can tell the public that every day in status updates
A screenshot showing the @thankyoublakwomen.bsky.social profile page. It follows zero people, has 30 followers, and has 1.1 thousand posts. SkyWatch has tagged the account with“Suspected Inauthentic Behavior” and “Fully Automated Luxury Reply Guy” as well.
USA Fencing Rule t.105 1. A competitor who, while fencing, commits certain violent or vindictive actions against his opponent, or who does not fence to his utmost ability, or who profits from a fraudulent agreement with his opponent, may be excluded from the competition. 2. A competitor who is excluded from a competition may not continue to take part in that competition, even if he is already qualified for promotion to the next round. He loses the right to his individual classification and the place that he occupied remains unfilled; but the points obtained by the fencer up to the moment of his expulsion remain valid for the classification of the team in the team event, providing that he actually takes part in that team event.
"Murderbot" hails from Paramount Television Studios. The Weitz brothers wrote, directed and produced under their Depth of Field banner. Andrew Miano also executive produces for Depth of Field. David S. Goyer executive produces alongside Keith Levine for Phantom Four. Wells serves as consulting producer.
An edited frame from Peanuts. Lucy is in front of the TV saying, “Christ, what an asshole.”
An example image with text showing the location of the “image to text” function on IOS.
Abrego Garcia was not charged with a crime, but he was handed over to U.S. Immigration and Customs Enforcement after the arrest to face deportation. In those proceedings, the government claimed that a reliable informant had identified him as a ranking member of MS-13. Abrego Garcia and his family hired an attorney and fought the government's attempt to deport him. He received "withholding of removal" six months later, a protected status.