Most people know the decimal (Hindu-Arabic) system we use every day. Many also know Roman numerals (I, V, X, L, C, D, M) from clocks, books, and movies. But what if I told you that for centuries, Eastern Europe had its own elegant letter-based numeral system?

🧮 How many different words are needed to name all numbers from 1 to 100?

🧮 Look at this brilliant table from Malinin & Burenin’s Arithmetic (1897) — the thorough Russian classic that built real mathematical mastery. It shows how our decimal system is organized into classes (Units, Thousands, Millions, Billions, Trillions…) and repeating orders within each class:

My new article 'The Well-Kept Secret of Elite Calculus: Why Top Chinese and Russian Students Don’t Use Apostol or Spivak' has been published on Medium #math medium.com/p/3116f99...

scikit-learn has a contradiction hiding in plain sight. LogisticRegression applies L2 regularisation by default (C=1.0). Regularisation shrinks coefficients toward zero, which flattens probabilities toward 0.5. This is a deliberate injection of bias.

Most companies forecast hierarchies wrong. They train 10,000 SKU models. Then they “sum them up.” Or worse — they forecast the total and allocate down with last year’s proportions. This guarantees one thing: Incoherence. Total ≠ Sum of parts. Regions ≠ Country.

📜 From the iconic Malinin & Burenin Arithmetic book — a classic Russian Empire school textbook: “The digits we use are called Arabic, because they were borrowed by Europeans from the Arabs around the mid-10th century.

"The Myth of Calibraiton" based on Chapter of my 'Applied Conformal Prediction' book. If you think that Logistic Regression is a good model I have some news for you. www.youtube.com/watc...

🧮 The Ultimate Russian Arithmetic Mastery Three legendary classics in one unbeatable bundle: • Malinin & Burenin’s Arithmetic — rigorous applied techniques, commercial math, and advanced real-world problems.

🧮 Two legendary Russian arithmetic textbooks. One unbeatable package. Kiselev’s Arithmetic (1884): concise, elegant, and rigorously theoretical — every operation proven from first principles.

The Brier Score (and why it mattered) Calibration in machine learning did not originate in AI. It was first studied seriously in weather forecasting, where incorrect probability estimates immediately translate into bad decisions: aviation planning, agriculture, disaster preparedness.

Forecasting is broken. Not quietly broken. Structurally broken. For four decades the M-competitions cultivated the mantra that “simple methods always win.” Exponential smoothing wasn’t just a tool — it was raised to Golden Calf status.

Why do so many USA schools turn simple subtraction into this mess?

🧮 The ultimate Arithmetic Masters bundle — three legendary Russian classics. Kiselev’s Arithmetic: concise, proof-based foundations (214 sections). Every operation derived from first principles.

🧠 New paper from MIT/Berkeley just dropped: ORCA Most “test-time compute” scaling wastes a ton of samples. ORCA uses conformal prediction to dynamically calibrate exactly how many reasoning samples an LLM actually needs per question. Result on Qwen2.5-32B: → 47% compute savings

What makes Malinin & Burenin different from Kiselev: Kiselev gives you the theory — every operation proved from first principles. Malinin & Burenin gives you the application — commercial arithmetic, currency conversion, multi-step proportional reasoning, the rule of mixtures.

Many people know Kiselev's Arithmetic. Almost nobody in the West has heard of Malinin and Burenin. Yet their textbook went through 19 editions and was the standard arithmetic course in Russian gymnasia for decades.

For years, people kept asking me the same question. "When are you going to write the forecasting book?" I kept saying "soon." That "soon" turned into a year of writing, rewriting, testing every model, every method, every line of code.

European universities — and Berkeley — can only dream of classes like this. Stanford University operates on an entirely different level.

Delighted to share that my new paper is now live on Zenodo: “The Manokhin Probability Matrix: A Diagnostic Framework for Classifier Probability Quality” 🔥🔥🔥 zenodo.org/records/1...

XGBoost — The Kaggle Fairy Tale Is Over XGBoost is miscalibrated 96.7% of the time. Not a typo. Not cherry-picked. 30 real-world datasets. 150 cross-validation folds. Every fold measured with the Spiegelhalter Z-statistic — the gold standard for calibration testing. Mean |Z| = 8.70.

Performing math vs knowing math. That's the difference between modern textbooks and one written in 1884. A.P. Kiselev was a Russian schoolteacher who got tired of watching students struggle with verbose, overcomplicated textbooks that taught procedures without understanding.

My translation of Kiselev Arithmetic is hot 🔥🔥🔥🔥top new release in popular and elementary Arithmetic!

Another paper showing why conformal prediction deserves serious attention in self-driving. The paper is: “Uncertainty-Aware Vision-based Risk Object Identification via Conformal Risk Tube Prediction.” The core idea is exactly the right one. In driving, risk is not a neat binary label.

TabPFN lost to CatBoost. Again. Our calibration study — arXiv:2601.19944 — benchmarked 21 classifiers across 30 binary tasks from TabArena-v0.1. We report expected rank across all datasets and folds, so this is not cherry-picking — it’s aggregate, robust, and hard to game.

The only forecasting advice you’ll ever need 🥋 When it comes to forecasting: • R is clearly superior to Python (and Julia is for nerds) • Transformers & LLMs rule supreme — time series is just language but with numbers, duh

Did you know? At China’s elite “985” universities (Tsinghua, PKU, Fudan – where ~80% of top tech/finance talent comes from), core math like calculus & linear algebra don’t use American textbooks as the main text.

New Youtube video 'The Rachinsky Method: Decoding the 19th Century Mental Math' #math

My new article 'Did a MIPT FRTK Grad Secretly Lead the Pentium III? The Untold Soviet Brain-Drain Story That Shaped Your CPU' has been published on Medium

2026, and people still treat “Transformer” and “LLM” as exact synonyms. It’s honestly wild how sticky this idea remains.

For over a decade, large parts of academia—Russia included—were effectively captured by Bayesian dogma. Students at elite places like HSE were fed the same story: priors, posteriors, subjective beliefs dressed up as “rigorous” uncertainty.

There is only ONE technology today that can produce calibrated predictive distributions with guarantees. Not Bayesian. Not deep ensembles. Not quantile regression.

The Rachinsky book is complete, and it is a beauty. 1001 mental arithmetic problems. A real mathematical classic. Get it this weekend: valeman.gumroad.com/...

LLMs are getting very good at text-to-SQL. They are also very good at being confidently wrong. That’s the real problem.

Just use an uninformative prior." There is no such thing. Every prior is informative. You're just pretending yours isn't.

Kiselev’s Arithmetic has shot to the top of the charts on Amazon, claiming the #1 spot in the Hot New Releases #math

I am finalising the release of my translation of Sergei Rachinsky’s 1001 Problems in Mental Arithmetic, and I can now say it plainly: the book is complete, and it is a beauty. Rachinsky was not a classroom theorist.

Bayesians spent 60 years arguing about which prior to use. Conformal prediction spent 60 years proving it works with any distribution. One community debates philosophy. The other ships guarantees.

When I began translating the iconic math textbooks by Kiselev and Rachinsky, I was struck to discover that they were far more than ordinary schoolteachers. Andrei Kiselev was an accomplished mathematician in his own right.

When I began translating the iconic math textbooks by Kiselev and Rachinsky, I was struck to discover that they were far more than ordinary schoolteachers. Andrei Kiselev was an accomplished mathematician in his own right.

My new article “ I Fed a Berkeley Prof’s Anti-CP Article to Grok — Every Single Claim Flagged as Fake News” valeman.medium.com/i...

Why Soviet Math Students Were Strong People often ask: Why were Soviet and Eastern European students so strong in mathematics? Why do Eastern European students dominate math olympics? Part of the answer lies in the textbooks.

Before calculators. Before apps. Before gamified maths. There was hard mental arithmetic, done properly. Sergei Rachinsky’s problem book is complete and nearly ready for release. A beautiful historical work, and still deeply useful today. valeman.gumroad.com/...

A professor of botany at Moscow University gave up academic life to teach peasant children in a village school. Out of that work came one of the most remarkable mental arithmetic books ever written. I’ve completed my translation of it, and it is a beauty. Get it this weekend on Gumroad:

Most companies forecast hierarchies wrong. They train 10,000 SKU models. Then they “sum them up.” Or worse — they forecast the total and allocate down with last year’s proportions. This guarantees one thing: Incoherence. Total ≠ Sum of parts. Regions ≠ Country.

This is not a modern “activity book.” It is 1001 mental arithmetic problems forged in a real school through daily teaching, daily drill, and very high expectations. No fluff. No gimmicks. Just numbers, memory, speed, and thought. The book is complete: valeman.gumroad.com/...

I’ve been saying for years: Kaggle is one of the worst ways to actually learn machine learning.

I’m delighted to announce the full release of my translation of Sergei Rachinsky’s 1001 Problems in Mental Arithmetic. This is a beautiful mathematical classic: 1001 problems in mental arithmetic shaped by years of real teaching, real discipline, and very high expectations.

Just read a groundbreaking new paper that could reshape the AI copyright debate: "Alignment Whack-a-Mole: Finetuning Activates Verbatim Recall of Copyrighted Books in Large Language Models" (arXiv:2603.20957)

Economists have finally stumbled upon conformal prediction, only about three decades after its 1998 debut.