Твиттерский эксперт №1

Эксперт-аналитик №1 международного экспертно-аналитического конгресса 'Мировая Экспертиза и Аналитика'.

[Твиттерский эксперт №1 @bitexe]

Спор с твиттерским оппонентом.

[Твиттерский эксперт №1 @bitexe]

@BolsheNeVovan @gourev Там мексы. У них есть свои женщины.

[Привет. Взаимная подписка с адекватными @since1795]

Mythos может буквально перевернуть науку. Она уже:
— Ускорила некоторые этапы разработки лекарств примерно в 10 раз;
— Научилась предлагать новые научные гипотезы, а не только пересказывать существующие знания;
— Превзошла недавнюю научную работу, опубликованную в журнале Sciencе

[Твиттерский эксперт №1 @bitexe]

@hellzau Кондиционер поддерживает желаемую разницу температур и влажности воздуха внутри и снаружи, и временно выключается, когда эта разница достигается.

Если окно открыть, то эта разница становится недостижимой, и кондиционер будет работать без перерыва.

"Деньги на ветер" ☝️🧐

[Probability and Statistics @probnstat]

Wasserstein Learning Theory is a rapidly growing area of machine learning that uses ideas from optimal transport to analyze probability distributions, generalization, and learning dynamics. At its core is the Wasserstein distance, which measures the minimum cost of transporting one probability distribution into another. Unlike divergences such as KL divergence, Wasserstein distances incorporate the geometry of the underlying space, making them particularly suitable for comparing complex distributions.

In probability and statistics, Wasserstein metrics are used to study convergence of distributions, concentration inequalities, empirical processes, and distributional robustness. In machine learning, they provide powerful tools for domain adaptation, distribution shift analysis, generative modeling, and robust optimization. The success of Wasserstein GANs demonstrated how transport-based objectives can stabilize training and improve sample quality.

In deep learning, Wasserstein methods help analyze representation learning, neural network dynamics, and generalization under distributional changes. In reinforcement learning, Wasserstein distances are widely used in distributional RL, where agents learn entire return distributions rather than only expected rewards. They also appear in robust RL and exploration under uncertainty.

The deeper insight is that learning often involves comparing distributions rather than individual observations. By incorporating geometry into probability, Wasserstein learning theory provides a principled framework for understanding robustness, generalization, and adaptation in modern AI systems.

https://t.co/wWtARbSLIi

[Rimsha Bhardwaj @heyrimsha]

An Australian mathematician from Perth who spent a decade at Meta building the framework half the AI world runs on, then moved to OpenAI, then co-founded a company with the former CTO of OpenAI, just accepted what is reported to be the largest individual hiring package in tech history.

Mark Zuckerberg paid roughly $1.5 billion over six years to bring one person back to Meta.

His name is Andrew Tulloch.

Here is the story, because almost no one outside the AI infrastructure world knows what one engineer is worth right now.

Andrew grew up in Perth, Australia. He studied at the University of Sydney and graduated with first class honors in mathematics. He went on to the University of Cambridge and earned a master's in mathematical statistics and machine learning. He started his career as a quantitative strategist at Goldman Sachs, applying advanced mathematical models to financial systems.

In 2012 he joined Facebook, before it was Meta. He stayed for more than a decade. During that time he became one of the core technical contributors to PyTorch, the deep learning framework that now runs the majority of AI research on Earth. The framework was a Facebook AI Research project that grew into a community standard. Andrew worked on the underlying systems, the distributed training stack, and the hardware-aware optimization that made it production-ready at the scale Facebook needed.

Then he left for OpenAI, where he contributed to advanced models inside the company that built ChatGPT and GPT-4.

In February 2025 he co-founded Thinking Machines Lab with Mira Murati, John Schulman, Barret Zoph, Lilian Weng, and Luke Metz. Murati, the former OpenAI CTO who had run ChatGPT, GPT-4, and Sora, had walked out of OpenAI in September 2024 with no public explanation. Six months later her startup was real and Andrew was on the founding team.

The company raised $2 billion in its first five months. The valuation hit $12 billion. They built a product called Tinker, which lets developers fine-tune frontier models without managing distributed compute. Their public bet was different from everyone else's. While other labs raced to build bigger models, Thinking Machines focused on smarter post-training techniques.

Then Mark Zuckerberg made his move.

Meta had been losing ground in the AI race. Zuckerberg tried to acquire Thinking Machines for a reported $1 billion. Murati refused. He responded with a direct campaign to hire her most valuable people. The primary target was Andrew.

Meta's pitch to him was reportedly a package worth up to $1.5 billion over six years, combining salary, bonuses, and stock awards. The number stunned the industry when it leaked. It would be one of the most expensive individual hires in the history of technology.

Andrew initially refused.

Then in October 2025 he accepted.

He joined Meta Superintelligence Labs, the new division Zuckerberg had created in June 2025 under Alexandr Wang, the 28-year-old former Scale AI CEO whom Meta had installed as its first Chief AI Officer. Meta paid $14.3 billion for Scale AI to bring Wang in. Yann LeCun, who had led Meta AI for 12 years, departed soon after. 600 researchers were cut from FAIR. The first closed-source model from Superintelligence Labs, Muse Spark, launched on April 8, 2026.

Andrew now works on the infrastructure problem at the scale of hundreds of thousands of GPUs. Public reporting describes Meta's target as around 350,000 NVIDIA H100s and roughly 600,000 H100-equivalents of compute. At that scale, even a 10 percent efficiency gain saves Meta hundreds of millions of dollars. That is the kind of impact only a small number of engineers on Earth can deliver, and Meta decided he was worth more than the GDP of small countries to have him doing it.

A mathematician from Perth who once worked on quant trading at Goldman Sachs just became the most expensive engineer in the world.

He spent a decade quietly building the foundations.

Then everyone realized what those foundations were worth.

[Такие Дела @takiedelaru]

82-летняя Галочка с деменцией часами молчала в пустой квартире. Внучка Юля купила умную колонку — теперь Алиса напоминает пить таблетки, читает книги и просто составляет компанию. Этой весной врачи сказали: развитие болезни удалось замедлить.

[Твиттерский эксперт №1 @bitexe]

@mcgospodi 👍👍👍

[Philosophy Of Physics @PhilosophyOfPhy]

In physics, your life’s work is practically guaranteed to be proven incomplete or obsolete by the next generation. Accepting this takes a profound level of existential humility.

"We physicists are in a strange position. We spend our entire lives trying to build a beautiful, perfect house of cards. And the highest honor we can hope for is that, fifty years from now, some young kid will come along, blow our house down, and build a bigger one on top of our ruins."

- Max Born, to his students in Göttingen

[Твиттерский эксперт №1 @bitexe]

@Zusatzzeichen Кто нибудь, спросите у @grokа, верен ли подсчёт или нужно вводить какой коэффициент.

[Борис Трушин @TrushinBV]

Наглядные доказательства некоторых свойств чисел Фибоначчи

[Don Pu @brazzers_don]

Впервые человеку была сделана инъекция препарата, который должен обратить вспять процесс старения

Американский биотехнологический стартап Life Biosciences объявил о том, что первому пациенту была сделана инъекция препарата для клеточного перепрограммирования. Предполагается, что лекарство позволит избавиться от возрастных заболеваний, буквально омолаживая клетки организма.

Ранее подобные эксперименты проводились только на мышах и обезьянах, теперь же ученые получили возможность доказать, что их методика пригодна и для человека. Единственное, что известно о самом эксперименте, это то, что укол был сделан в глаз пациенту с глаукомой.

"Мы в первую очередь изучаем возможность восстановления функций, обращения вспять заболеваний на самом фундаментальном уровне в организме", — заявил гендиректор Life Biosciences Джерри Маклафлин.

[Math Files @Math_files]

Best opening for a math lecture

[Millie Marconi @MillieMarconnni]

A self-taught Irish schoolteacher wrote a book in 1854 that almost nobody read for 80 years, until a 21-year-old MIT student picked it up and realized it could be used to design every computer in human history.

His name was George Boole. The book is called An Investigation of the Laws of Thought.

Boole was born in 1815 in Lincoln, England. His family was poor. He left school at 16 to support them. He taught himself Latin, Greek, French, German, and Italian.

Then he taught himself mathematics. By 19 he had opened his own school. By 24 he was publishing original papers in the Cambridge Mathematical Journal, competing with men who had spent decades inside the best universities in Britain.

He never had a degree. He never had a mentor. In 1849, Queen's College in Cork hired him as a professor anyway.

In 1854, he published his masterwork. What he built inside it was something nobody had attempted before at this scale. He turned logic into algebra.

Before Boole, logic was philosophy. You argued in sentences. You reasoned in paragraphs. It was powerful and completely impossible to automate, because there was no formal system underneath it, just language.

Boole stripped it down to arithmetic. He showed that every act of human reasoning could be reduced to operations on two values. True or false. One or zero. AND, OR, NOT. If both conditions are true, the result is true. If neither is, the result is false. Every judgment a human mind makes, every decision, every deduction, could be written as an equation following those rules.

Logicians read it. They found it interesting. Engineers building machines had never heard of it.

For 83 years, the book sat there.

Then in 1937, a 21-year-old MIT master's student named Claude Shannon was working on a thesis about electrical relay circuits. Switches that could be open or closed. Current that either flowed or didn't.

He read Boole and understood something nobody had connected before.

An open switch is a zero. A closed switch is a one. A circuit with two switches in series only carries current when both are closed. That is AND. A circuit with two switches in parallel carries current when either is closed. That is OR. Shannon proved that every possible logical relationship Boole had described could be physically built using wire and switches.

That single insight is the foundation of every computer ever made.

After Shannon, chip designers stopped thinking about electricity and started thinking about logic. Every transistor on every processor running right now is implementing a Boolean operation. Every if-statement in every codebase is Boolean logic. Every database query using AND or OR. Every neural network threshold that fires or doesn't fire. All of it is running the algebra of a self-taught schoolteacher from Lincoln who died 160 years ago.

The strangest part is what happened to Boole at the end.

He was walking to class in November 1864 when he got caught in a rainstorm. He lectured for hours in wet clothes. He went home sick. His wife, Mary, believed in homeopathic medicine and thought the cure should mirror the cause. She wrapped him in wet sheets and poured cold water over him repeatedly.

He died a few days later. He was 49.

He never saw a transistor. He never saw a circuit. He never saw a single physical machine run a single one of his rules.

His book is in the public domain. Free to download. Most engineers use the word Boolean dozens of times a week. Almost none of them know who they are saying.

The man whose logic runs inside every phone, every server, and every AI model on Earth died soaking wet in a small Irish town, 83 years before anyone figured out what he had actually built.

[Твиттерский эксперт №1 @bitexe]

@adagamov + NYC наводнён припаркованными иногородними машинами уже сейчас,

+ собственники жилья рассматривают съехать в другой штат на время чемпионата и сдать своё жильё как Airbnb.

[Ihtesham Ali @ihtesham2005]

A neuroscientist at UC Irvine spent 60 years proving that your brain has a separate memory system for emotional events, and you can hack it to remember almost anything you want.

His name is James McGaugh.

He founded the Center for the Neurobiology of Learning and Memory at the University of California, Irvine, and he has been running experiments on how the brain decides what to keep and what to throw away since 1959. The finding he spent his entire career building, refining, and defending should change how every person on Earth thinks about learning.

The discovery started with a question almost nobody else was asking in the early 1960s. Why is it that you can read an entire textbook chapter and remember almost nothing three days later, but you can recall in vivid detail what you were doing the moment you heard a piece of devastating news from years ago.

The two memories were stored by the same brain in the same skull, and yet one was almost completely erased while the other was preserved frame by frame. McGaugh was convinced that something specific had to be happening in the brain to explain the gap, and he spent the next six decades chasing it.

The first major piece of evidence came from a simple animal experiment that almost ended his career when he proposed it. He trained rats to perform a task, and then immediately after the training he injected them with a stimulant drug. The rats who received the injection remembered the task far better than the rats who did not, even though both groups had performed identically during the actual learning.

The drug had not made the rats smarter or faster. It had been administered after the learning was over. Something about the chemical state of the brain in the minutes following a learning event was determining whether the memory survived or vanished.

This was the moment the entire field of memory consolidation was born. McGaugh had proven, against decades of consensus, that memories are not formed at the moment of the experience. They are formed in the hours that follow, and the chemical environment of the brain during that consolidation window decides what gets kept.

The next 50 years of his lab's work mapped out exactly which chemicals were doing the work. The answer turned out to be the stress hormones your body releases when something emotionally significant happens. Epinephrine. Cortisol. And most importantly, a neurotransmitter called norepinephrine, which floods a specific part of your brain called the basolateral amygdala the moment you feel anything strongly.

The amygdala is the small almond-shaped structure deep inside your brain that processes emotion.

McGaugh and his colleagues proved that when this region is activated by emotion, it sends a signal to the hippocampus, which is the brain region responsible for forming new memories, and that signal physically strengthens the consolidation of whatever you were experiencing in that moment.

Emotionally charged events get stamped into the brain with a flood of hormones that say keep this. Neutral events get filed without the stamp and are quietly thrown away over the next few days.

The experiment that made the mechanism unmistakable was published in 1994 by McGaugh's collaborator Larry Cahill, who had trained at the same lab. He showed participants a series of slides that told a story. Half of them saw a neutral version of the story where a boy and his mother visited a hospital.

Half of them saw an emotional version where the boy was hit by a car and rushed to the same hospital. The slides were almost identical. The narration was different.

Two weeks later, the participants were brought back and tested on how much they remembered. The emotional group recalled the middle of the story, where the trauma happened, with significantly higher accuracy than the neutral group recalled the same middle slides. The story was the same. The images were the same. The only thing that had changed was whether the brain was emotionally activated while encoding the information.

Then Cahill ran the experiment again. This time he gave half the emotional group a drug called propranolol, which blocks the action of norepinephrine in the amygdala. The drug did not interfere with their thinking. It did not make them sleepy. It just shut down the chemical pathway McGaugh had spent decades mapping. And the emotional memory advantage disappeared completely.

The group on propranolol remembered the emotional story no better than they remembered the neutral one. The hormone was the difference. Block the hormone, and the brain stopped stamping the memory.

This is the framework McGaugh built over 60 years. The brain has a two-track memory system. The default track is for neutral information, and it is leaky on purpose because most of what your brain processes in a given day is not worth keeping.

The emotional track is for information that arrives with a chemical signal that says this matters, and it preserves the experience with stunning detail because evolution decided that anything emotionally significant was probably important for survival.

The implication is the part almost nobody talks about, and it is the reason this research should be on the wall of every classroom and study room.

If you want to remember something, you have to give your brain a reason to flip the emotional switch on while you are learning it. Information delivered in a flat, neutral, low-stakes environment will be processed through the leaky default system regardless of how many times you re-read it. Information delivered with curiosity, surprise, stakes, embarrassment, awe, or even mild stress will be processed through the emotional system, and the same brain will hold onto it for years.

This is why the lecture you were forced to sit through evaporated by the end of the week, while the question you got humiliated by in front of the class is still perfectly preserved 15 years later. The humiliation was the chemical stamp. The lecture had none.

People who remember enormous amounts of what they read are not gifted. They are emotionally engaged with the material in a way most learners never become. They argue with the author in the margins. They feel actual frustration when something does not make sense. They get genuinely excited when a concept clicks. The frustration and the excitement are not side effects of learning.

They are the mechanism of learning. Every emotion you feel while reading is a small dose of norepinephrine being released into the amygdala, telling the hippocampus to stamp this page into long-term storage.

The fix is almost embarrassing in its simplicity.

Stop trying to absorb information neutrally. Pick a question you actually care about answering before you open the book. Argue with the material as you read it. Get angry at the parts that feel wrong. Get curious about the parts that surprise you. Try to explain what you learned to someone who would push back on it. Care about the outcome.

Your brain was never designed to remember neutral information. It was designed to remember anything that made you feel something. McGaugh spent 60 years proving that the rest is almost a rounding error.

The voice in your head that tells you to study harder is wrong.

The one that tells you to study warmer is the one your brain actually listens to.

[Рустем Адагамов @adagamov]

🇺🇸 Reuters опубликовало большое расследование о криптовалютных проектах, связанных с семьёй Дональда Трампа. Главный вывод журналистов сформулирован в заголовке: "По криптовалютным правилам Трампа семья всегда выигрывает. Инвесторы — нет".

TLDR здесь такой: Трамп и его семейка обдурили доверчивых трампофилов, вложившихся в их мусорные криптовалюты. Люди потеряли свои вложения, а Трамп разбогател за их счет на пару миллиардов.

По подсчетам агентства, с середины 2024 года семья Трампа заработала около 2,3 млрд долларов на криптовалютных проектах. При этом многие инвесторы, вложившиеся в те же активы, понесли убытки в сопоставимом размере. Reuters утверждает, что во всех случаях использовалась схожая модель: Трампы практически не вкладывали собственных средств, но получали доли в проектах, права на выручку и другие финансовые преимущества благодаря своему имени, политическому влиянию и способности привлекать внимание рынка.

Наибольшую прибыль семье принёс проект World Liberty Financial, который рекламировался как инструмент для "демократизации финансов". По данным Reuters, продажи токенов обеспечили связанным с Трампами структурам более 1,4 млрд долларов дохода. Однако стоимость самих токенов впоследствии упала примерно на 87%, а инвесторы столкнулись ещё и с тем, что их активы оказались фактически заблокированы до 2030 года.

Не менее показательной оказалась история мемкоина $TRUMP, запущенного незадолго до инаугурации Трампа на второй президентский срок. На волне ажиотажа цена токена взлетела, однако затем рухнула примерно на 97% от достигнутого максимума. Несмотря на это, структуры, связанные с семьей президента, заработали более 600 млн долларов на выпуске и продаже монет, тогда как многие частные инвесторы остались с практически обесценившимися активами.

Reuters также анализирует другие проекты, включая криптокомпанию ALT5 Sigma и майнинговую фирму American Bitcoin. В обоих случаях после первоначального всплеска интереса инвесторов последовало резкое снижение стоимости активов. Инвесторы потеряли сотни миллионов долларов, тогда как семья Трампа получила доли и финансовые выгоды на выгодных для себя условиях.

По мнению Reuters, ключевая особенность этой бизнес-модели заключается в том, что доход семьи формируется не за счёт долгосрочного успеха проектов, а на этапе привлечения средств и продажи активов. Как отмечает агентство, Трампы получают прибыль независимо от того, что происходит с проектами дальше. В результате риск в значительной степени переносится на покупателей токенов и акций.

Опрошенные Reuters эксперты называют такую модель крайне спорной с этической точки зрения, однако признают, что доказательств прямого обмена политических решений на инвестиции нет, а значит большинство подобных схем остаются в рамках закона.

Именно поэтому расследование приходит к выводу, вынесенному в заголовок: в криптовалютной империи Трампа семья практически всегда оказывается в выигрыше, тогда как инвесторы слишком часто остаются проигравшей стороной.

[Rimsha Bhardwaj @heyrimsha]

A Stanford assistant professor and a small lab of graduate students sat down in March 2023 and reproduced the behavior of ChatGPT for under $600.

The model they released became the template that every open-source instruction-tuned model on the planet now copies. The evaluation system they built became how the entire field measures AI alignment. Most people scrolling through AI Twitter cannot name him.

His name is Tatsunori Hashimoto. His lab is called the Tatsu Lab.

Here is the story, because almost nobody outside the language model research world knows what one Stanford lab has quietly shipped.

Tatsu grew up between two continents and studied at MIT, where he eventually earned his PhD. After finishing he moved to Stanford as a postdoctoral researcher in 2019, co-advised by Percy Liang and John Duchi, two of the most respected names in machine learning. The combination is unusual. Most postdocs work with one advisor. Tatsu sat at the intersection of statistical machine learning, robustness, and natural language processing, which meant he could draw from both camps.

By 2020 he was hired as an Assistant Professor in the Stanford Computer Science Department. He joined the statistical machine learning and NLP groups. His research focused on something most of the field was ignoring at the time. How do you actually evaluate language models in a way that is rigorous, reproducible, and not gameable?

Then ChatGPT launched in November 2022.

Within four months Tatsu and his students did something nobody else in the open-source world had figured out. They took Meta's just-released Llama model, fine-tuned it on instructions generated by GPT-3.5, and released Stanford Alpaca on March 13, 2023. The training cost less than $600. The resulting model behaved like ChatGPT on most everyday tasks.

The release went nuclear. Within days every open-source AI project on Earth was running variants of the Alpaca recipe. The technique he and his students used became the standard. Every "fine-tune your own ChatGPT" tutorial that exists traces back to this lab in Stanford.

Then he built the evaluation system.

In 2023 his group released AlpacaEval, an automatic evaluator for instruction-following language models. The idea was simple and powerful. Instead of paying humans hundreds of thousands of dollars to evaluate model outputs, you use a strong language model as the judge against a reference model. The results were highly correlated with human expert annotations. Suddenly the entire open-source community had a fast, cheap, reproducible way to compare instruction-tuned models against each other.

Over 100 models have been added to the AlpacaEval leaderboard. Every major open-source release from Mistral to Llama to DeepSeek runs against it. The repository lives at github .com/tatsu-lab/alpaca_eval. It is one of the most cited language model evaluation systems in the field.

In 2024 his group released Length-Controlled AlpacaEval, a debiased version that strips out the trick of making outputs longer to win evaluations. The community had been gaming the original. He fixed it and released the patch.

The Tatsu Lab also released AlpacaFarm, a simulation framework for studying how language models learn from human feedback, with collaborators including Yann Dubois, Xuechen Li, Rohan Taori, Tianyi Zhang, and Ishaan Gulrajani. Several of these collaborators are now at OpenAI, Anthropic, and other frontier labs. His postdocs Niladri Chatterji and Shibani Santurkar both ended up at Meta and OpenAI doing core research.

Tatsu still publishes constantly. His Google Scholar reads like a map of the modern alignment field. He keeps a low public profile. He gives almost no media interviews. His Stanford homepage is a flat list of papers with no styling and no marketing copy.

A Stanford lab that most people outside academic AI cannot name built the open-source ChatGPT recipe, the evaluation system the field now runs on, and trained the researchers who went on to power frontier labs.

He did it from a small group of graduate students.

[Холод @holodmedia]

Онкологи аплодировали стоя докладу о препарате, который справляется с одним из самых опасных видов рака почти вдвое лучше стандартной химиотерапии. Его представили на ежегодном конгрессе Американского общества клинической онкологии.

Как пишет The Insider со ссылкой на эксперта, побочные эффекты от него «относительно мягкие». Лекарству еще предстоят дальнейшие исследования и получение государственной сертификации, но он уже доступен американским гражданам с терминальными стадиями заболевания

[Твиттерский эксперт №1 @bitexe]

@nett00n А результаты выборов публиковать в блокчейне поимённо (по "индексу избирателя", привязанному к снилс и паспорту), с разбивкой по избирательным учаскам и регионам. И сделать эту базу доступной любому для анализа.

[Lesbian (But it is a gender) 🏳️‍🌈🏳️‍⚧️ ⚲✨️ @nett00n]

Неиронично считаю, что все законодательные документы стран должны быть опубликованы в виде git-репозитория с wiki-style кросс-сылками. История изменений с версионированием, авторством, дватами изменений, pull request'ы, удобное сравнение версий, все дела.