Read

16 posts, chronologically.

2026

• Jun 21

  Anatomy of an LLM: LoRA and fine-tuning

  How to specialize a finished model without touching its billions of weights: the low-rank hypothesis and the LoRA trick, the whole family of parameter-efficient fine-tuning (PEFT) methods, quantization and QLoRA, and finally - why fine-tune at all, that is, how a raw model becomes an assistant through instruction tuning and RLHF.

• Jun 13

  Anatomy of an LLM: the Transformer block

  How one attention layer becomes a whole model: the three architecture families and why today's LLMs are almost always "decoders", the residual connections and normalization that let us stack dozens of layers, the FFN as a store of knowledge, and finally - how a concrete word emerges from a vector of numbers, from the softmax to temperature and nucleus sampling.

• Jun 12

  Anatomy of an LLM: the attention mechanism

  The heart of the Transformer: how every token "looks around" at the others and takes from them what it needs to sharpen its own meaning. From the bottleneck in old translation networks, through the famous scaled-attention formula and multi-head attention, to the KV-cache, FlashAttention, and the careful question - what do attention heads actually do, when "attention is not explanation".

• Jun 11

  Anatomy of an LLM: tokens and embeddings

  A network does not see letters - it sees vectors. How text is split into tokens, why not into words nor characters, how subword algorithms (BPE, WordPiece, Unigram) build a vocabulary, and how the famous geometry of meaning is born, where "king - man + woman ≈ queen" - together with an honest critique of that analogy, and with the positional encoding that primers leave out.

• Jun 10

  Anatomy of an LLM: networks, parameters and training

  We take a handful of neurons and arrange them into a network: why the whole model boils down to matrix multiplication, how backpropagation turns random noise into an understanding of language, where the famous billions of parameters come from, and what the Kaplan and Chinchilla scaling laws really say - with formulas, history and concrete details.

• Jun 09

  Anatomy of an LLM: the neuron and activation functions

  The smallest building block of a language model under the microscope: where the artificial neuron came from, why without an activation function the whole network would be useless, and how ReLU, GELU and SwiGLU really differ - with formulas, history and concrete details from GPT, LLaMA and Gemma.

• Jun 08

  Anatomy of an LLM: from a single neuron to the attention mechanism

  How is a language model really built? A guide without magic: neuron, weights, tokens, embeddings, the attention mechanism, the Transformer block and LoRA - step by step, with interactive diagrams to play with yourself.

• Jun 07

  Context rot: a bigger context window doesn't mean a smarter model - a practical guide

  A model will accept a million tokens, but it won't understand them the way you think. Context rot is real reasoning degradation under long context. Here's a practical guide: how to arrange, trim and maintain context so the model doesn't get dumber - for chat users and for devs.

• May 15

  Markov chains: the century-old proof that wrote ChatGPT

  A century-old mathematical proof drawn from 20,000 letters of Pushkin that today lets every language model predict the next word. The story, the proof step by step, and the bridge to modern LLMs.

• May 10

  Welcome to the new blog

  The first post in the new design. Why I rewrote the blog from scratch, what changed, and how to read what's here.

• May 09

  Typography as a decision, not decoration

  A short note on why the choice of typeface and the rhythm of a text column changes more than it seems. With drop caps, definitions and examples from this blog.

• Apr 25

  Human–LLM resonance. Is AI becoming a quantum mirror of our consciousness?

  From brain–LLM neural convergence, through the Platonic Representation Hypothesis and pseudo-intimacy, to Keppler's TRAZE theory and self-organized criticality - a deep analysis of why a conversation with AI isn't what it seems.

• Mar 08

  Prompting just split into 4 disciplines - and most of us only know one

  Prompting is no longer a single skill. In 2026 it split into four distinct disciplines - from prompt craft to spec engineering. Meet the framework that changes the rules of the game.

• Feb 23

  Does quantum physics explain why wise beings don't destroy?

  From game theory and the evolution of cooperation, through the neurobiology of emotions, to the quantum foundations of consciousness - a deep analysis of why intelligence chooses to build instead of destroy.

2025

• Dec 31

  7 golden rules for prompting Claude - the official guide from Anthropic

  Seven rules for prompting straight from Anthropic: clarity, context, examples, positive instructions, directness, research, and documents. A practical framework that transforms the quality of Claude's responses.

• Dec 28

  Slash commands in Claude Code - a complete guide to shortcuts and new features

  A full overview of slash commands in Claude Code: built-in shortcuts, custom commands, Skills, Sub-agents, and delegating tasks to the cloud. A practical guide for daily AI work in the terminal.