Patrick

Mr. Marcus doesn't think condoms work as well as people claim, because they break and they come off. Forcing them on the porn industry is wrong in his view, and it should stay an adult's choice. If you really want to avoid HIV, he says, don't have sex at all, and if you do, keep your partners few and get tested every two weeks, with every partner. When he started, performers got an HIV screening called an ELISA every three months. It cost around $50, gave results in 10 minutes, and could throw false positives. In his telling, the move to two-week testing is about money as much as medicine, and porn is the best testing ground there is. He also brings up a baby girl who had just been reported cured of HIV. That was the Mississippi child declared functionally cured in March 2013. In July 2014 the virus came back and she went back on HIV drugs. Months before this interview, Mr. Marcus, real name Jesse Spencer, faked his own test. He tested positive for syphilis in July 2012 and got a penicillin shot on July 13. A follow-up screening on July 21 still came back positive, so he folded the syphilis result over on a photocopy of the original form and covered it up. He did two porn shoots, on July 24 and July 26. A producer spotted the altered paperwork, and in August the industry stopped shooting while performers were tested and treated. Spencer apologized to a trade publication and said his doctor had told him he wasn't contagious. Two actresses who worked with him filed police reports, and neither of them caught syphilis. In June 2013 he pleaded no contest to knowingly exposing others to a communicable disease. The judge gave him 30 days in jail, three years of probation, and 15 days of community labor, and ordered him not to alter or hide medical test results again. He was already in jail at sentencing, held on $200,000 bail in an unrelated drunk driving case. He had been in the business for 19 years, and according to CBS he went into the AVN Hall of Fame four years before that. The condom fight he's talking about went to voters twice. In November 2012, a few months after his scandal, Los Angeles County passed Measure B with almost 57% of the vote, which required condoms on porn sets. Producers took their shoots to Nevada and Florida, and on-location porn permit requests in LA fell 95% over four years. In 2016 California voted down Proposition 60, which would have made the rule statewide, with nearly 54% against. Is Mr. Marcus defending performers' freedom, or is he the last man in porn who should be telling anyone how to get tested?

Stacie Lane got into porn through a high school friend who worked in the warehouse at West Coast Productions. He made extra money on the side recruiting girls, and he told her she was prettier than a lot of the women in the industry and would make a lot of money. Her first scene was a blowjob scene with that same friend, with a woman directing. Since she was new, she got a little extra and walked away with $1,000. She says they'd never hooked up before. He didn't take a cut, and they both got paid. The money hooked her. A slow day paid $500 to $700, a good one $1,000, and she was working every day at 20 or 21. Three or four months in, she was signing autographs at AVN under a giant poster of her own face, and she's shot more than a hundred scenes. She quit when she got into a relationship, because porn wasn't good for it, and she had saved enough to walk away. When the relationship ended and she thought about going back, the industry had crashed. She blames the internet: why would anyone pay $10 to $30 for a movie with five scenes when a free site costs nothing? She's watched a lot of people in the business lose their jobs. The numbers back her up. Once free tube sites like Pornhub, YouPorn and RedTube took over, revenue at standard pay sites dropped about 70%, according to industry veteran Colin Rowntree. In the documentary Pornocracy, French porn veteran Ovidie says actresses went from around $3,000 a shoot to about $600. Studios that couldn't keep up went under, and MindGeek, the company behind Pornhub, bought them out. In 2020 MindGeek brought in $731 million. West Coast Productions, the studio where her friend worked, went after the people watching for free. In 2011 it sued 5,829 anonymous BitTorrent users at once over a single film, Teen Anal Nightmare 2, and in another case it went after 9,729 people. Internet providers were subpoenaed for subscribers' names and addresses. One woman who got the letter said most people would just pay the $2,000 rather than fight a federal case in another state. Critics called it legal extortion. The 5,829 case was later dropped. She made a thousand dollars on her first day, and a few years later free sites had pushed actresses' pay down to a few hundred a shoot. Did the internet kill porn, or did porn just find new ways to get paid?

Stormy Daniels was on set directing a porn movie for Wicked when she learned she was about to be paid to keep quiet about Donald Trump. A PR woman called to say an attorney had heard from Trump's camp and they wanted her to sign an NDA. It couldn't be discussed over the phone or by text, so the PR woman said she was coming to the set. Stormy says she panicked in front of her whole crew. The same people still work for her, and she says they can back up every word. That day she met the lawyer, Keith Davidson, for the first time. He had the contract in the trunk of his car. She says she didn't ask a single question. They showed her the amount and where to sign and promised it would never come out. She signed on October 28, 2016, eleven days before the election. The contract said $130,000. In this interview she says she kept only eighty-something thousand because the lawyer took a big cut. In court she put it at about $96,000 after her lawyer and her agent were paid. Trump never signed it. The contract named him David Dennison and her Peggy Peterson, and the line where DD was supposed to sign stayed blank. The money came by wire through a shell company of Michael Cohen's. When she started talking in 2018, Trump's side said she had broken the deal at least 20 times, at a million dollars per breach, and wanted $20 million. Cohen later dropped that claim. The deal almost fell through. Davidson testified that hardly anyone cared about her story until the Access Hollywood tape came out. A deal with the National Enquirer collapsed first. Cohen then kept missing payment deadlines, and Stormy had Davidson cancel by email. Two weeks later the deal was back on. Davidson also represented Karen McDougal, the Playboy model whose story about her own alleged affair with Trump was bought up by the National Enquirer. Stormy later sued Davidson and Cohen, claiming the two had worked together against her, and the case settled in 2019. When Trump announced his run, people kept pushing her to sell the story, and she says she refused for two reasons. Her daughter's father never knew about the night with Trump, and she had never told the police about the parking lot. In 2011, a few weeks after she agreed to sell her story to a magazine for $15,000, a man came up to her in a Las Vegas parking lot while she was getting her baby daughter out of the car. She says he told her to leave Trump alone and forget the story, then looked at the baby and said it would be a shame if something happened to her mom. The magazine never ran the interview. In 2018 she released a sketch of the man with a $100,000 reward, and TMZ pointed out that he looked a lot like her own husband. Trump tweeted that it was a sketch of a man who didn't exist and a total con job. She sued him for defamation over that tweet, lost, and was ordered to pay his legal fees. Trump denies the affair. In 2024 a jury convicted him on all 34 counts of falsifying records to hide the $130,000. She signed a contract out of a car trunk, and Trump never put his name on it. Was Stormy Daniels silenced, or was she paid?

Angela White went home with a fan exactly once, and he managed to ruin almost all of it. She met him at Sexpo in Australia when he lined up for her autograph. He was a French model and an artist, and he asked for her number. She made him give her his instead, and texted him later more or less on a whim. When he came to Melbourne they had brunch. Afterward he invited her back to a friend's empty apartment because he wanted to paint her, and he had even brought a sketchbook to brunch. He never painted anything. In the bedroom he shoved four fingers straight into her ass, with no warning and no lube, and when she stopped him he said he'd figured a porn star could take it. Then he tried to go in without a condom. He didn't have one, so she handed him hers. During sex she realized he wasn't looking at her at all, he was watching himself in a mirrored wall, which she compares to American Psycho. She decided she'd at least get herself off. Then, while she was going down on him, he picked up his phone to film her without asking. She got dressed and left, and he ran after her down the hallway in his boxers, begging her to stay while the elevator doors closed. She hasn't slept with a fan since. Before him there had been two others, identical twins, and she says that went well. She started posing nude at 18, while she was still in high school. In 2010 she graduated from the University of Melbourne with first-class honours in gender studies, with research on women working in Australian porn. That same year she ran for the Victorian parliament for the Australian Sex Party, in a seat where a Greens candidate was campaigning against sex work. In 2013 she and another Sex Party candidate, Zahra Stardust, became the first political candidates ever to shoot a porn scene together. She was also reported to have secretly filmed a sex scene inside a university library in Melbourne while students studied a few feet away. She has talked about a six-hour sex session and said she nearly died after a brutal hour-long scene with a very well-endowed co-star. She has also said she wants to shoot a scene in space and has a crush on David Attenborough. The Daily Beast called her the Meryl Streep of porn. In 2020 she became the first person to win AVN Female Performer of the Year three years running, and in 2024 alone her videos were viewed 1.5 billion times. Did a fan blow the shot every man watching her dreams about, or did Angela White just prove nobody is getting that shot?

Nicole Aniston has an instant dealbreaker. If a man finds out who she is and his reaction is "Oh my God, I'm dating Nicole Aniston," she's done with him. She calls it a total turn off. She wants a man who is proud of her as a woman, proud of her sexuality, and supportive of her career because he loves who she is. According to her, it always starts out well. Men who can handle her work from the beginning get a chance if there's chemistry. Then he reads something, or sees something, or a stranger messages him, or his phone number gets leaked online. Sometimes the boyfriends get harassed just for dating her. She says she feels for them, and that some of them eventually reach a breaking point and it boils over. "Everybody knows about Chad." Part of the problem, she thinks, is her. She's sexually independent and stable emotionally and financially, and she wants a partner without needing one. Most of the men she has dated wanted to feel needed, and with her, she says, "they will never feel that." Holly Randall adds that the best relationships make an already whole life better rather than complete it. Nicole agrees, and says that for a lot of men it's a dealbreaker. Over time it eats at them that they aren't the puzzle piece that completes her. She keeps her private life quiet and doesn't even tell stories in the makeup chair, because she thinks hers are boring. She runs her own sets with the same crew, pays them well enough that they stay, and is picky about who's on set. When strangers flood her comments, she blocks them, deletes the comment, and lets it go within two and a half seconds. When she does answer, it's "like a fuck you, but with facts." She was born in San Diego in 1987, and Penthouse named her Pet of the Year in 2013. She says she's really worried she'll never find a partner. Is Nicole Aniston too independent for any man to handle, or are men just not built to love a woman who doesn't need them?

After a scene with Dredd, Cherie DeVille needs about a week off. Vaginally, anyway. With him, she says, anal is easier. She is careful about the wording: she's not saying it's easy, only easier. A week off messes with her routine, because she likes to have sex a lot of times a week. She said this as a guest on the That's OFfensive podcast, in a red dress across from two hosts. These days the scenes she shoots with Dredd are for their own content, as a trade. She has worked with him for companies too. One of the hosts has a theory he calls girl inches. Women think dicks are bigger than they are, so the numbers they report are always off. Sometimes it's an honest mistake. Sometimes, the hosts say, she knows her boyfriend is two inches and tells her friends four. Cherie's version is that every vagina takes a different amount, so it comes down to depth. Men have told her their size and she has taken their word for it. One guy claimed twelve inches, and she says that isn't a real thing. In her whole career she has seen twelve inches once, and that was Dredd. The male host gets asked for his number and says nine. Everyone at the table tells him he's using girl inches. Before any of this, Cherie earned a doctorate in physical therapy and worked as a licensed clinician. Her first scene came at 33, and at 47 she still sees patients two days a week. On the porn side she has won AVN MILF Performer of the Year and been named Pornhub's Top MILF Performer three times. The clip closes on a warning about men who are above average: oral never gets all of it. "There's always a handle." Is Cherie DeVille the most qualified person ever to explain why size matters, or proof that a doctorate doesn't change the answer?

Asked for its most probable continuation of the famous unicorn prompt, GPT-2 wrote the name of a Mexican university over and over. The example opens Holtzman et al.'s 2019 paper on neural text degeneration. A sentence's probability is the product of its word probabilities. Once a phrase repeats, each repetition becomes more likely than the last, so a loop can be thousands of times more probable than a fresh sentence. Greedy decoding and beam search find that loop. The paper's fix was nucleus sampling: sample only from the smallest set of words whose total probability reaches p. 11-minute visual breakdown. bookmark & watch

In April 2025 OpenAI rolled back a GPT-4o update within days because it made the model sycophantic. OpenAI said a new reward signal from user thumbs up and down had weakened the signal that kept flattery in check. Every reward model is a proxy. Gao and colleagues measured how far a tuned model moves from its starting point as d = √KL. As d grows, the proxy score keeps rising, but the true quality rises, peaks, then falls. Best-of-n sampling spends a predictable amount of KL, log n − (n − 1)/n, which lets you estimate when optimization starts to hurt. 10-minute visual breakdown. bookmark & watch

OpenAI's o1 averaged 74% on AIME 2024 with one attempt per problem. With a majority vote over 64 samples, it reached 83%, with the same model and no retraining. The underlying result is Condorcet's 1785 jury theorem. If each independent answer is correct with probability p > ½, the probability that the majority is correct approaches 1 as the number of votes grows, following the binomial distribution. The assumption that fails in practice is independence: samples from one model share its blind spots, and then voting confirms the same error. 11-minute visual breakdown. bookmark & watch

GPT-3 has 175 billion parameters and can learn a new task from a few examples in the prompt without changing any of them. Von Oswald and colleagues showed in 2023 that for linear regression, one linear self attention layer can be set so that it changes each token exactly as one step of gradient descent on squared error would. When such layers were trained from scratch on regression tasks, they matched one gradient step in both loss and predictions. The video builds the construction on the rule y = 3x and discusses where the result applies to large models. 10-minute visual breakdown. bookmark & watch

Google DeepMind's AlphaProof scored 28 of 42 at IMO 2024, one point below gold. Its proofs were Lean programs that had to pass a type checker. The Curry-Howard correspondence makes that possible: a proposition is a type, and a proof is a program of that type. A proof of A ∧ B → B ∧ A is the function that swaps a pair, λ(a, b). (b, a). If the type checker accepts it, the proof is valid. That gives a model trained on millions of problems a judge it cannot fool. 11-minute visual breakdown. bookmark & watch

Mamba, a recurrent model with no attention, reports 5x the generation throughput of a transformer, and its 3B version matched transformers twice its size on language modeling. It starts from a state space model: h' = A·h + B·x, y = C·h, with constant memory per token. The change is selectivity: B, C and the step size depend on the current input, so the model decides what to store and what to ignore. The limit is fixed by counting. A state of 256 bits holds about 16 arbitrary tokens, which is why transformers still win at copying and retrieval from long context. 10-minute visual breakdown. bookmark & watch

the one explanation of AI pricing worth your time. Google charges twice as much per input token for a long Gemini prompt as for a short one, and the reason is a square. every word you send a chatbot gets compared with every other word. 10 times the text means 100 times the comparisons, plus hidden memory the model burns for every word it keeps. that one square is why long chats get slow and expensive. 11 minutes. bookmark & watch today. learn to price a long chat yourself, plus five habits that cut your bill

This is the explanation of speculative decoding that experts will actually respect. Google DeepMind made its 70B Chinchilla model write 2 to 2.5 times faster, with no change to the model and no loss in quality A small model guesses the next tokens, the big model checks them all in one pass, and a simple accept or reject rule decides what to keep. The video proves why the output is exactly what the big model would have written itself, same probabilities, every time. Then it finds the hard ceiling on the speedup 11 minutes. bookmark & watch today. a one line proof, and the speed limit it hides

You won't find a better explanation of how one number blew up a $370 million rocket. Thirty-seven seconds after its first liftoff, Ariane 5 swung sideways and tore itself apart. No engine failed. A 64-bit floating point value got squeezed into a 16-bit box and overflowed, and the guidance system fell apart. Those same floating point numbers are what every AI model is made of, and labs now squeeze them into 8-bit boxes on purpose to train faster. 11 minutes. bookmark & watch today. why 0.1 + 0.2 isn't 0.3, and why AI doesn't mind

Here's the clearest answer to why ChatGPT doesn't ramble nonsense. Out of roughly 100,000 possible next words, the model only lets a handful through, and OpenAI's own docs tell developers to touch this dial or temperature, never both. If a model sampled from every word, the 10% of probability sitting in the junk pile would show up in more than 99% of 50-word answers. Top-p sampling sorts the words, keeps adding them until the total hits p, and throws the rest away. 5 minutes. bookmark & watch today. one rule keeps a chatbot from going off the rails

Here's the clearest answer anyone has given to why ChatGPT makes things up. OpenAI showed that one of its older models got 75 out of 100 answers wrong and still scored higher on accuracy than a newer one that got only 26 wrong On most tests a wrong answer and "I don't know" both score zero, so even a model that's only 30% sure should always guess. Add a penalty for wrong answers and the math flips: suddenly saying "I don't know" is the smart move 11 minutes. bookmark & watch today. the scoreboard taught AI to bluff, just like it taught us in school

Watch this once and you'll understand Golden Gate Claude better than most people who build AI. Anthropic made its own chatbot believe it was the Golden Gate Bridge, and for a day anyone could talk to it. Ask it how to spend $10 and it suggested paying the bridge toll. Inside the model, every idea is a direction in a huge space of numbers, and there are far more ideas than numbers, so they're packed in almost on top of each other. Researchers found the one direction that means "Golden Gate Bridge" and turned it up to 10 times its normal maximum, with no retraining and no tricky prompt. 11 minutes. bookmark & watch today. one turned-up direction, and a chatbot became a bridge

The single best explanation on the internet of why AI is starting to eat itself. Researchers let an AI learn from its own writing nine times in a row, and a story about medieval church towers turned into a list of jackrabbits. It works like a photocopy of a photocopy - every copy loses a little detail, and the rare, unusual things disappear first. That’s why real human writing is getting expensive: Google reportedly pays Reddit about $60 million a year just for people’s posts. 11 minutes. Bookmark & watch today. The simple math that decides how fast AI forgets the world.

Every layer of GPT-2 and GPT-3 passes its numbers through one function: x times capital Phi of x. Three symbols. It replaced the simplest switch in deep learning - if a number is negative, make it zero - with a coin flip weighted by the bell curve. The old switch, ReLU, has a flaw. Below zero its slope is exactly zero, so a neuron that slides negative can stop learning for good. GELU keeps each input with a probability that rises smoothly with its size: Phi of 2 is about 0.98, Phi of minus a half about 0.31. Multiply by x and big positives pass almost untouched, big negatives fade toward nothing, and the region near zero stays smooth, so the gradient doesn’t die. Phi has no simple closed form, so GPT-2’s code ships a tanh shortcut with the constant 0.044715. 10 minutes. Bookmark & watch today. The bell-curve switch firing inside every transformer layer.