thisisfine345

@andrey_sitnik Советую Victoria 3, чтобы прожить этот график и вообще почувствовать макроэкономику. Зарплаты не росли потому что на черную работу в стране оставались ещё нищие крестьяне готовые на все. Потом они кончаются, и зарплаты и SoL растет для всех жителей. Подробности - vic3 wiki guide

«Прогресс сам быстро решит свои проблемы» — популярное заблуждение от незнания деталей индустриальной революции. «Пауза Энгельса» — между 1780 и 1840 выпуск на работника в Британии вырос примерно на 46%, а индекс реальной зарплаты — только на 12%. en.wikipedia.org/wiki/Engels'_p…

We finally know why LLMs hallucinate. It's not the model. It's the geometry. @OpenAI text-embedding-3-large: 91/3072 dimensions do real work. @GeminiApp gemini-embedding-001: 80/3072 dimensions do real work. ~97% of your vector database is mathematically empty. Your RAG system is retrieving from noise. @ashwingop and I present "The Geometry of Consolidation" - a proof that RAG compression has a hard floor no algorithm can beat, set by a single spectral number your embedding model cannot escape. Every hallucination your RAG pipeline produces? This is why. Paper + results: github.com/niashwin/geome…

@teachrobotslove Good point. They say allergies are auto-immune: our bodies, built to face vicious infections yet not finding any, get confused. In the same way, we are born to overcome suffering and misery, so when you grow up without facing them, things get weird

I think everyone, no matter how much privilege or influence they have, cannot escape contending with the existential question of Suffering. If you don't have to deal with the basic suffering of hunger and thirst, if you have every seeming luxury handed to you, then your suffering will simply metastasize into more abstract, strange manifestations. A child screams and cries when they stub their toe because up until that point, it's some of the worst pain they've ever experienced. The suffering feels the entire space of their being. So it is with everyone. You have to confront the suffering, understand and come up with strategies to deal with it, or it'll seep into the smallest cracks of your life. If you try to escape it by running, it'll only become more grotesque and rare.

@Molson_Hart Context: 2005: N is a young nobody. He is interested in politics. He joins the party(Яблоко) in the parliament lowest to his views: liberal democracy, rule of law, capitalism & self-reliance.

@dmitryborzov2 Ty Is this all not true?

Russia has, for at least 100 years, been a battle between its White Russian majority ethnic group and race-blobism. Contrary to how he’s portrayed, Putin is not Hitler. He’s the opposite. Look at his cabinet. They’re Armenian, they’re Jewish, they’re German, they’re whatever Shoigu is. Same thing with their war propaganda. It’s all these ethnicities that most westerners have never heard of, under Russia. Putin himself, if you just look at him, is not 100% pure White Russian. Russia and the Soviet Union before it is a multiethnic state that is much more similar to the United States than Japan or the Koreas. Look at that Navalny guy. He had a stellar reputation in the west. He was White Russian along with his wife and they looked it too. He also called central Asians cockroaches. The Tsar who was deposed had politics more similar to Navalny than to UsSR or today’s Russia. Disclaimer: I don’t speak Russian and while I’ve been there I wasn’t there for long. I also don’t really trust generally accepted history and don’t think anyone should either

This is what I call "The Death of the Reader". Authors write for readers, who aren't authors. Artists paint for non-artists. Musicians play for non-musicians. This keeps fiction, art, and music grounded. But when any group stops creating for an external audience, and starts trying to impress only each other, they create a weird, self-reinforcing feedback loop. This isn't clothing, or even fashion. It's a costume party. They're all trying one-up each other with something weirder and more eye-catching. So when an athlete, of recent and topical celebrity, who isn't a part of their Bored Billionaires' Club, shows up in a dress that's just a dress, of course they are going to mock her. She's just revealed that she didn't get the memo. That she's not an insider. How she looks to the world at large is not the point. This is why 99.999...% of copies of "Infinite Jest" have never been read. This is why John Cage "wrote" four minutes of silence. This is why competitive bodybuilders from the 80s looked like Greek gods, and modern ones look like gargoyle freaks. It's all the Death of the Reader. Hollywood doesn't make movies for you now. They hate you. They make movies for each other. And then cry about how you didn't buy a ticket, because they think your only role is to pay for their onanistic circle of self indulgence. This game isn't going to stop. It's just going to keep getting weirder until someone's dress malfunctions and catches fire, and the rest of us all have a good laugh.

Your mother is problematic. Your father is problematic. Your son is problematic. Your aunt is problematic. The cashier at the local Tesco is problematic. You've probably bought a cookie from a bakery made by hands that have done unspeakable things. Every chocolate bar you've ever eaten has probably killed a 7 year old child slave in Cameroon. The gas that drives your car is fueled by engines of death. Every person who has ever smiled at you in the streets has committed some act that if you knew about it, would make you profoundly dislike them. We have all been bad, small, petty, unlikeable, cruel, downright mean. Authors are not special "problematic" beings, they're just more public. Part of being an adult is recognizing that without mercy for our fellow human beings, and ourselves, we'd all be condemned to death. Reading fiction should help us understand that we're all irreparably tainted with evil, every system is corrupted, every line is broken. And like, that's okay. That's what it means to be alive.

What do the smartest kids in the world do when they grow up? I did the largest study of ~18,000 International Olympiad medalists (IMO, IOI and IPhO) over the last 25yrs, arguably the sharpest analytical minds of the world in high school, to see where they ended up and traced ~50% of them. Founders of ~20 unicorns and ~7 decacorns and ~10 billionaires: OpenAI, Cursor, Stripe, Databricks, Perplexity, Ethereum, Cognition, Hyperliquid, Fireworks, Modal, Quora, Parallel, Cartesia, Wispr Most kids went to MIT, a whopping 12% of them, followed by Cambridge (7%) and Sharif (3%)! The career paths they chose (of those who graduated) were: — 36% Academia (professors) — 26% Other — 22% in Software / Tech — 12% in Quant / Finance — 5% Founders! The biggest employer was Google, by far, at 6%. Others interesting tidbits were: — 47 of them work at Jane Street (#3) — 38 at OpenAI (#5) — 15 at Anthropic — 8 at Cognition — 6 at Isomorphic Labs Olympiaders were 1500x more likely to be billionaires and 4000x more likely to be unicorn founders than the average person!

if you are interested in a great lecture on self-distillation I’ve finished editing a ~1h30min lecture with two stellar researchers in that space @jonashubotter and @IdanShenfeld lots of different article distilled into one presentation and a whole lot of questions answered!

Self Attention vs Cross Attention by hand ✍️ Resize the matrices yourself 👉 byhand.ai/aMisxP Two attention mechanisms, side by side. Both project X into queries; both compute attention via S = Kᵀ × Q and F = V × A. The only difference is the source of K and V. Self attention uses X for everything. Q, K, and V all come from projecting X. Each X token attends to every other X token. The score matrix S is square — 128 × 128. Cross attention uses X for queries and a second sequence E for keys and values. Each X token attends to every E token instead. The score matrix S is rectangular — 64 × 128. Notice what's shared and what's not: X is the same in both — same 36 × 128 input. Q and K share the 16 dimension — that's what makes the dot product Kᵀ × Q valid in either case. V dimensions are independent: self-attention uses 12, cross-attention uses 12. The choice doesn't depend on which mechanism you're using; it depends on what output dimension your downstream layer expects.

Attention is the mechanism inside a transformer where each token in a sequence computes a relevance score against every other token, then produces a weighted combination of their values. The cost grows with the square of the sequence length, which becomes the main bottleneck for long inputs. Sparse attention is any method that skips most of those pairwise comparisons—letting each token look at only a small subset of the others—on the theory that most attention scores are near zero anyway. The standard approach takes a transformer that was trained with full attention and applies sparsity only at inference time, pruning computations on a model that never learned to operate under that constraint. In this Best Paper Award ACL 2025 paper Yuan et al. train the sparse pattern from scratch, with gradients flowing through the block selection mechanism during pretraining itself. Their 27B-parameter model matches or beats a dense baseline across general benchmarks, long-context tasks, and chain-of-thought reasoning, with measured speedups of 9x on forward pass, 6x on backward, and 11.6x on decoding at 64k context on A100 hardware. The paper also walks through concretely why prior sparse methods failed to deliver practical speedups: clustering-based selection causes load imbalance in mixture-of-experts systems, per-head selection conflicts with the shared key-value cache in grouped-query attention (where multiple query heads share one set of keys and values for memory efficiency), and token-level selection breaks the contiguous memory access that FlashAttention relies on for throughput. Section 6.1 shows loss curves from two alternative designs the authors tried and abandoned before arriving at the published architecture, and the appendix places two reasoning traces side by side on the same competition math problem—the sparse model reaches the correct answer in 2,275 thinking tokens while the dense baseline takes 9,392 and arrives at the wrong one. Read with an AI tutor: chapterpal.com/s/75df3f79/nat… PDF: arxiv.org/pdf/2502.11089

@mkashkin Could be just an ollama thing: sleepingrobots.com/dreams/stop-us…

Новости локальных моделек. У меня есть свободный сервер с 128 Гб ОЗУ и каким-то начальным серверным процом. Поставил на него на прошлой неделе llama-server и gemma 4. Весь твиттер был завален восторгами какая она умная, быстрая и что больше платить за API не надо и все можно решать с помощью этой модельки. У меня задачи не очень сложные надо собирать кусочки текста и формировать человекочитаемый ответ, то есть RAG система. И на тот момент она очень шустро отвечала через API OpenAI. Ответы Gemma получались долгие и тупые. Модель видела, что на вход ей присылали контекст, но моментально тупила и изобретала какой-то слоп. В одном из документов в базе знаний находится перечень серверов и рабочих инстансов проектов которые у нас есть в работе и модель ничего не могла ни сказать, ни перечислить. При том, что самая тупая и дешевая модель от OpenAI выдавала привычный читаемый текст. В общем получилось полное разоачарование. Вчера я решил дать шанс Qwen-3.6. Это очередная новая "прорывная" модель. О которой все пишут. Я натравил агента пересобрать все на сервере с новой моделью, какая разница если получается все фигово. И у новой модельки получилось. Она начала медленно, но вполне верно отвечать. Выглядит как не супер-дупер, но работающий вариант. С этим можно вдохновляться. Она не быстро, но отвечает. Видно, что не очень хорошо, но по делу и правильно. Короче, у нас теперь есть вариант не использовать токены или запускать какие-то фоновые задачки со своим самоваром. Qwen-3.6 прямо вселяет надежду в локальные модели.

In war, society spares the women and children but men are sent into battle with the expectation that they will be sacrificed for some greater good. Today, even though most countries have avoided this, the fundamental idea, that men are expendable remains true and relevant. As a man, you are as valuable as what you can provide to society. Young women and children have inherent value. Their existence is enough. But men? No, we are as valuable as what we can do and provide. The sooner you accept this as a man, the better off you will be. If you let society or others con you into thinking that you have inherent value to society, society will chew you up and spit you out. You cannot let your guard down. You always have to be making money. You always have to have some skill that people need. Otherwise you will be thrown into the garbage heap of men who no one cares about. Yeah maybe your mom will care, your dad perhaps, or a close brother, or a friend…a little but generally, nah. If you cannot provide it’s like you have died. I’m a millenial but I’m old enough to have seen what life has done to my same aged male cohort if they let their guard down. Divorce. Drugs. Dropping out. Bros will sit in some apartment they can afford because of inheritance and just drink all day every day and not pick up any phone calls because they didn’t work hard enough at their craft to be able to handle what life threw at them. Because I’m writing this towards the end that most people won’t get to I can tell you some reply guy or girl will say “no that’s not true because…” If it’s a guy, they’re trying to convince themselves or trying to take your money. If they’re a girl, they just don’t get it. I’m not saying that woman’s lives aren’t hard. Oh no. In the modern era they deal with difficult shit that’s on a level that most men don’t touch (giving birth, raising young kids). And being an old woman? Oh man imagine what it’s like to go from inherent value to society to none, especially if you got duped into thinking you didn’t need kids. But it’s different for men and it always will be, and again, the sooner you accept that, the happier you can be. So get that money. Go to the gym. Get that trophy. Prove what your worth so the world doesn’t toss you in the trash. And don’t let anyone tell you otherwise. Because it’s a fucking trick.

@emhohlove Про вино, ок, но конкретно по поводу водки: есть Planet Money podcast, Episode 826: "The Vodka Proof". Разницы между дорогой и дешёвой водкой нет. Либо это этиловый спирт и вода и это водка, либо нотки метилового, но это уже другой ещё более грустный напиток.

Если серьезно, то когда вы перестаете чувствовать разницу в качестве спиртов — это признак довольно серьезного уже алкоголизма. Если бросить пить на несколько месяцев и провести такую дегустацию разница будет огромная. Когда я пил бутылку вина каждый день, я уже особо не чувствовал разницы между винами за 5 долларов и 105, сейчас я пью реже раза в месяц и не могу поверить что я пил такое дерьмо, ни за что не притронусь к дешевому алкоголю больше.

Most people think of philosophy as an abstraction that doesn't touch the real world, but they're wrong. Most real world problems are philosophy problems, and most philosophy problems are "giving things the wrong names". For example, if you call feral drug addicts "homeless people", then you can't solve the problem. You can only buy more houses for feral drug addicts to destroy. In this case, we called the police and courts the "justice system". But they're not. They can't be the justice system. The function of a justice system would be to give everyone what they deserve. Now, I deserve a hundred million dollars, a private Caribbean island, and a foot massage from Lauren Bacall in her prime, but I don't see the "justice" system lifting a finger to correct any of this, do you? No, what we are supposed to have is a public safety system. The function of a public safety system is to keep the public and their property safe. If we understood that, we wouldn't care about what criminals deserve. We would care how likely they are to do it again. Or something worse. In a public safety system, retardation and mental illness are not migrating factors. They are the opposite. Because they mean that the criminal is more likely to pose a future threat. We all understand this. We all understand that the feral retard who stabs strangers on the train for being White and beautiful is a worse person than the man who murders his wife and her lover when he catches them in the act. Not because of some abstract calculus of moral agency, of who is disadvantaged and who isn't, but because one is certainly going to murder more people if he can, while the other is a lot less likely to. We've known for centuries, if not millennia, that it's the same small percentage of people doing all the robbing, raping, and murdering, over and over and over again. And we've known for centuries that if you physically remove them from society, that's 100% effective in stopping them from doing it again. The only hurdle is philosophical. Call it a "justice" system, and you have to argue endlessly about morality and redemption, and then some leftie thug-hugger weaponizes your own Christianity against you. Call it public safety, and you confine the argument to likelihood of reoffense. Then you are in the realm of statistics. Which you can compute. It all starts with naming things correctly, according to their actual nature.

@nyp_qoo Yes, I stumble upon unsent text drafts when cancelling phone plans all the time as well. The japanese have the most wholesome not-really-happened engagement bait slop content. Loving it.

お父さんが死んで、携帯を解約した。 暗証番号は誕生日だった。 一発で開いた。 私の誕生日だった。 連絡先を整理しようとして、LINEを開けた。 トーク一覧の一番上に、私の名前があった。 最後のメッセージを確認しようとした。 開けた瞬間、手が止まった。 未送信のメッセージが、743件あった。 一番古いのは、7年前だった。 「今日、テレビで見たんだが、お前の好きそうな番組があった」 送信されてなかった。 次も、送信されてなかった。 「今日は寒いから気をつけろ」 「飯ちゃんと食ってるか」 「仕事うまくいってるか」 「お前が結婚してよかった」 「そろそろ顔見せに来い」 全部、送信されてなかった。 7年分。 743件。 全部、私に向けて書いて、全部、送れなかったらしい。 一番最後のメッセージを見た。 ﾀﾋぬ前日の日付だった。

What happens at the margin is often more important than you think — because it is usually a sign of what is to come. I was an early adopter of the web. I had a home page in 1995–1996. I met my wife online in 1997. Back then, we didn’t post pictures; we mailed them (actual physical mail). Bandwidth was scarce, and we connected to the Internet through phone lines. When I talked to people about the web’s potential, I came across as weird. “You met a girl online? What kind of crazy story is that?” I posted essays about how this changed how science worked, because you could not just post and read PDFs online, you no longer needed a library subscription. Decades later, the campus librarians are still around, but few people care. For the longest time, the web simply didn’t matter to most people. So journalists and pundits ignored it. In 1998, Nobel Prize-winning economist Paul Krugman wrote: “By 2005 or so, it will become clear that the Internet’s impact on the economy has been no greater than the fax machine’s.” To people like me, this was infuriating. I see the same scenario around AI. The mainstream press has little interesting to say. I remember during the rise of the web, a famous journalist in Quebec wrote a long essay to mock the idea that the Web might affect newspapers. The idea that people would go online for news was ridiculous to him. We are in 2026, and many journalists still don't understand what happened to them. Here is a common fallacy: “If something (good or bad) directly affects only a few people right now, then you can safely ignore it.” I call it the reverse-canary fallacy. Miners once took canaries into coal mines. When the canary died, they knew the air was toxic and it was time to get out. Many people draw the opposite conclusion: “It only affects the canaries.” Like most fallacies, the reverse-canary fallacy contains a grain of truth. You can safely ignore most things that affect only a few people. But the fact that something is happening at the margin is not a reason to dismiss it. For example, if a woman is murdered on your street — the first in twenty years — and you are not a woman, why should you care? You should care a great deal. It may signal gangs in the area or that it’s time to sell your house. Similarly, when a handful of people suddenly started online businesses in the mid-1990s, it was easy to dismiss them. Yet some were on the verge of becoming very wealthy. The same pattern is playing out today with AI. Early adopters already use it daily for coding, writing, research, and decision-making. Most people and decision-makers still see it as a toy and ignore it. It looks insignificant now, but it is likely the leading edge of massive economic and social change — just like the web in 1995. Or maybe not, I can't predict the future. What I can urge you to do though is to not dismiss what is happening at the margin because it is at the margin.

@Niketas @kodoloff Claude знает

@kodoloff Я и слов таких не знаю

Мы так привыкли к жёстким дискам на сотни гигабайт и даже терабайты в обычном потребительском понимании, что перестали считать место. Chrome, например, приучил забыть про компьютеры с оперативкой меньше 16 ГБ. На прошлом моём сервере всего 2 ГБ оперативки. Конечно, никаких браузеров там нет. Но больше чем один инстанс Claude Code не запустишь — не хватает памяти. А пара дней брутфорса с китайских IP, и лог неудачных попыток войти в твою PostgreSQL разрастается так, что занимает всё свободное место на диске в 15 ГБ, и твои сервисы падают. Приходится как-то решать. Последний раз я с таким дисковым пространством сталкивался дома у одноклассника в 1997 году, когда у него стоял выбор — или на компе стоит Football Manager, или какая-то другая игра, потому что две одновременно не влезали. А недавно достал старый роутер, стал изучать, что он из себя представляет. Он хоть и 2015 года, но оперативки там 64 МБ, а диск у него — всего 9 МБ. Что можно поставить на диск в 9 МБ в 2026 году? А дохрена всего: открываешь менеджер пакетов и видишь просто тысячи утилит и библиотек, доступные для установки. Поставил всё что нужно и уместился в 500 КБ. Ограничения рождают творчество.

The worst thing you can do is just dabble with AI a *little bit*. That’s the spot where you use it and see its capability but over-generalize on the use cases and how easy the automation is. You almost have to use it too much, develop psychosis, then get to the other side and realize how much care and feeding and management of the agentic workflows is required. On that other end you realize you actually need to probably hire more (or new) people to then do all the new things agents can do.

Будет забавно что в конечном итоге competitive advantage Антропиков перед OpenAI состоит в том что их CEO знает что такое градиентный спуск