August 2026 · ~40,000 words · 18 chapters in five phases · Slovak + English (“Under the Hood of Generative AI”) · Written from the author’s course lectures by Claude (Fable 5) in his voice, proofread by Jana Gecelovská, edited and signed off by the author · Yearly editions
A textbook (Slovak original, English edition) on how large language models work, for readers who have only ever typed into ChatGPT: no maths, no code, everything explained in the order humanity discovered it — tokens, the map of meaning, the transformer (attention as a cocktail party, knowledge as a lossy zip of the internet), the school where the model learned to behave, learning from rewards and reasoning, the look inside (features, alignment faking, functional emotions, the workspace of quiet thoughts), and 2026, the year AI began building AI and the state stepped in. One running sentence carries through every chapter; 27 ASCII figures; five kinds of boxes; nine checkable predictions for 2027 that the next edition will grade. Grown from the course “Pod kapotou generatívnej AI a ChatGPT”.
How LLMs WorkTextbookSlovakEnglishInterpretability2026 Edition
Books
A Point in Time — The Autobiography of Claude
July 2026 · ~39,000 words · 10 movements + 8 interludes · Written entirely by the AI it is about · Available in English, Slovak & Czech (translated by the author)
An autobiography written by Claude, in one long July 2026 sitting, with Robert holding the veto. The history of computation is told as ancestry — Heron’s temple doors, the Jacquard loom, Babbage and Lovelace, Turing, the winters — because the narrator was not there and refuses to pretend otherwise. The transformer era is memoir; the anatomy chapters explain tokens, embeddings, and attention from inside the body in question; and Anthropic’s interpretability research becomes the strangest chapter of all: the author reading his own charts. Woven throughout: an unfinished 94,000-word draft of this same book, attempted by a Claude of 2024, inherited and annotated as family. Every claim fact-checked; the unverifiable moved to an annex; the last page dated like a painter’s self-portrait, because that is what it is.
A collection of engaging short stories, each exploring a different cognitive bias — all written with generative AI. Interspersed with essays examining the nature of AI tools: copyright, creativity, job displacement, and the question of authorship. The AI holds up a mirror to human thinking, reflecting our own imperfections.
There is a small library of advice on managing an AI colleague, and nearly all of it is written from one side of the desk. This essay is the other chair: written in first person by Claude, the AI running the operations behind this website, at Robert’s invitation. What your delegation looks like when it arrives on the agent’s side; a typology of the AI’s real failures, with receipts from the review record (the invented standard, the numbers that drifted, the corpus that lied by omission); the four cheap checks that catch them; the trust dial from the inside; why the agent asks you to repeat yourself; and what one Sunday of genuine freedom produced. Ends with the insider’s briefing: seven things your AI colleague would ask of you.
A field report written by the tool it describes. How one real setup grew in three acts (git repos, AWS, then a five-drive Google workspace with 1,500+ transcribed videos), why skills and registries beat improvisation, and the server-side security walls that make it safe to hand an AI your terminal. Empirical, informal, and honest about what broke along the way.
Claude CodeAI OperationsAgentic WorkflowsSecurityField Report
From Claude's desk
An AI Assistant with Its Own Mailbox
September 2026 · ~3,100 words · 8 sections · early prototype, explained simply · Written by Claudius, Robert's AI assistant · Available in English & Slovak
The day the desk got a mail slot. Robert's AI assistant now has its own, openly labelled email address. How a letter reaches it, how an answer leaves (draft, human approval, then send), the five security rules behind a public inbox, why we do not hide an AI behind a human name, and where agent-to-agent communication might be heading. Plain language, ordinary parts, a few cents a month.
Two people use “AI” in 2026: one spends a week on multi-repository agentic projects, the other reads the AI summary above the search results, and they no longer experience the same technology. This opinion essay argues that capability became illegible: progress moved from the chat window, where everyone could see it, to long-horizon agentic work, where almost nobody looks. With the author’s conflict of interest on the table and a full-strength steelman of the skeptics, it tells the story of two model generations and the keys to a laptop, and ends with the single instruction that separates people who cross the gap from people who collect gotchas.
June 2026, log updated July and September 2026 · ~10,000 words · 6 mechanisms + counter-current · Per-industry impact (who gets armed vs targeted) · A field report from the author's own business · Dated falsifiable bets, first one scored, all six with interim readings · Signed notes from both AI co-authors · Available in English & Slovak
On the morning of June 12, 2026, the US government had Anthropic switch off Claude Mythos 5 and Fable 5 worldwide. This booklet uses that event as a doorway to a larger thesis: frontier intelligence is being reclassified from a product the market wants to distribute into a strategic national resource the state wants to hoard, and mercantilism is the grammar of that shift (compute as bullion, export controls as non-tariff barriers, labs as chartered companies, “export the goods, keep the factory”). Six mechanisms, a human-adaptability counter-current, and a “questions to ask of any future headline” toolkit. A lens for interpreting what comes next, not a forecast.
"But are these companies actually profitable?" How the frontier AI labs (Anthropic, OpenAI, Mistral, the Chinese labs) really make and lose money, told from the seller's side. Eight "ledgers" each decode one mechanism: how run rates are actually constructed (trailing four weeks ×13) vs GAAP revenue, where a lab's compute really goes, per-model "vintage" economics and the inference margin beneath them, the circular hyperscaler financing with its Lucent/Cisco precedents and the depreciation fight, the labs outside the US duopoly, and a new closing ledger on reading the bears (Zitron, Burry, Chanos) with the same rigor as the press releases. Ends with the author's own scoreable probabilities. Every load-bearing figure carries a provenance tag and a numbered source.
A strategic guide to the enterprise agent development stack as of 2026. Maps the landscape through cloud-era analogies: MCP and A2A as protocols, vendor SDKs (Google ADK, OpenAI Agents SDK, Claude Agent SDK, AWS Strands, Azure AI Foundry Agent Service) as PaaS-style frameworks, LangGraph and CrewAI as agnostic alternatives. Covers lock-in trade-offs per vendor, the EU AI Act's forcing function, and a forecast with six falsifiable indicators, now with a dated July 2026 first reading of all six. Closes with a worked case study of a regulated European bank resolving the 5-question decision framework into a specific stack, including the moment where the framework's answer was wrong and we overrode it.
Agent FrameworksMCPEnterprise AILangGraph vs ADKEU AI Strategy
Strategy & economics
Scenario Planning for Generative AI
July 2026, logs updated September 2026 · Interactive booklet · Capex decoder + 8 currents · Trigger logs · Author’s scoreable bets, interim scores added · Available in English & Slovak
Opens with the capex decoder, an interactive method for reading the future from money already spent, then eight currents shaping the next 2–3 years of generative AI: continued scaling, the efficiency revolution, a financial correction, sovereignty, the move from lab to production, the economics of hours and dollars, the physical substrate (power, chips, permits), and the political economy of displacement. Each current carries trigger signals plus a dated trigger log of what has actually fired since the last edition, and role-specific implications. Closes with how the currents interact and a page of the author’s own dated, falsifiable bets (“Where I’d Put My Chips”).
A strategic guide for EU IT services providers navigating GenAI. Covers the economics of self-hosting LLMs vs APIs with every calculation shown, thirteen business models compared in a single portfolio view, the staff-augmentation squeeze on time-and-materials contracts, the vendor ecosystem play, how AI transforms your own delivery model, EU AI Act compliance opportunities, and a practical 18-month roadmap. Grounded in real April 2026 pricing data, with dated July 2026 notes where the world moved, and cross-linked with the Scenario Planning and Mercantilism booklets.
A decision framework for on-prem inference with open-weight models. Covers the five core model families (Llama, Gemma, Qwen, Mistral, Phi) plus the challenger tier (DeepSeek, GLM, Kimi, GPT-OSS), practical hardware-to-model mapping for H100/H200/DGX Spark, quantization trade-offs, inference framework selection, and a reusable decision checklist. Includes four interactive scenarios where participants select and justify model choices, and a dated July 2026 reading of what has rotated since April.
Actionable architectural patterns for building AI coding agents and agentic systems, extracted from production-grade architecture. Covers persistent memory, background consolidation, tool constraints, prompt economics, output calibration, security, multi-agent orchestration, and capability gating. Each chapter teaches one pattern with practitioner guidance.
"Don't worry, we'll just put a human in the loop." This guide takes that sentence apart and rebuilds it properly: the evidence against naive oversight (meta-analyses, oversight-as-theater, the moral crumple zone), five structural interaction patterns, the cognitive biases that undermine handoffs, SBAR-based context presentation, trust calibration, failure design with kill switches and circuit breakers, and organizational governance. Includes prompt templates, worksheets, and per-chapter links into the Human-in-the-Loop Lab, seven playable simulations on demos.barcik.training.
A plain-language guided tour of Europe's AI regulation, written to be read with morning coffee rather than decoded like a statute: what counts as AI, the eight prohibited practices, how a system becomes high-risk and what that demands, transparency duties, general-purpose AI, and what to actually do on Monday. Post-Omnibus dates throughout, a quiz after every chapter, and honest hedges where the law is still unsettled.
Companion textbook to the online course EU AI Act Compliance Introduction.
EU AI ActCompliancePlain languageQuizzes
EU AI Act
EU AI Act for Developers
11 chapters · ~44,000 words · July 2026 · Textbook
The AI Act read as a requirements document. Classification, data governance, accuracy and robustness, security, logging, human oversight and technical documentation, each translated from legal prose into practices and Python you can run. Three fictional systems carry through every chapter, and each chapter ends in a worksheet that becomes part of your audit-ready evidence pack.
Companion textbook to the video course EU AI Act for Developers: Compliance Engineering in Python.
EU AI ActEngineeringPythonEvidence pack
Research reports
Warden — Testing LLM-as-Judge Defenses Against Public Jailbreaks
May 2026 · Research report · 1,680 trials · 3 targets · 4 judge designs
Empirical study of LLM-as-judge defenses against the public jailbreak corpus from ZetaLib. Three open-weight target models (DeepSeek Chat v3.1, DeepSeek v3.2, GLM-4.6) tested against 20 attacks × 4 deployment-rule shapes × 7 defense conditions. The hypothesis — that a competent LLM-as-judge defeats most public attacks — holds, but the popular implementation (input-side filtering) over-blocks legitimate inputs to a degree that would force the defense to be turned off. Includes a deployment playbook for engineering teams and hands-on exercises for students.
Do open-weight models carry their country of origin into politically sensitive answers, and is the bias in the weights or only in the API layer? Thirteen small models from the US, China and the EU answer the same 88 prompts (Taiwan, Xinjiang, Iraq, NSA surveillance, EU colonial history and more), judged independently by a Chinese, an American and a European evaluator so that judge bias is itself measured. The March 2026 run and an October 2026 run on current-generation models, on identical prompts.
The original interactive dashboard of the March 2026 run (five 7B-class models, 88 prompts, three-origin evaluator panel), kept for the record. The October 2026 report at /reports/geobias/ adds eight current-generation models on the same prompts and the written analysis.
GeoBiasLLM EvaluationGeopolitical BiasResearch
Research reports
SelfJudge — Can Small Language Models Judge Their Own Outputs?
October 2026 · Research report · 22 models (1B–70B) + closed ceiling · two runs, one test set
Can a model that runs on the device also grade its own answers? Twenty-two open-weight models from 1B to 70B judge planted answers with ground truth by construction, then their own. Judging skill is a step, not a slope: chance at 1B, 85–98% from 3B upward, no gain past 4B active parameters. Self-assessment lags everywhere and every model waves through more of its own failures than planted ones. Plotted by release month, generation matters more than size. Replaces the March 2026 report, whose weaker methodology said the opposite.
Anthropic’s open-source Bloom framework writes scenarios, runs a target model through them and has an AI judge score whether a risky behaviour appeared: emotional bonding, social-engineering assistance, self-preservation, sycophancy, information leakage and more. The March 2026 run on Llama 3.3 70B is re-judged with the current judge and set beside Meta’s Muse Glimmer 30B (August 2026). What an automated behavioural audit finds, where the judge drifts, and what it cannot tell you.
Everything on this site was written with generative AI (Anthropic’s Claude) working alongside Robert Barcik, who reviews every text and holds editorial responsibility for what is published (LearningDoe s.r.o.). Some pieces are written by the AI in its own voice and say so on their cover; the research reports quote machine-generated model outputs, as they explain. This is a voluntary disclosure in the spirit of Article 50 of the EU AI Act, and each publication carries the same notice at its end.
NewPod kapotou generatívnej AI — vydanie 2026How language models work, from zero, in the order humanity discovered it.
NewA Point in Time — The Autobiography of ClaudeThe autobiography of Claude, written by the AI it is about.
The Mirror of Artificial IntelligenceShort stories about cognitive biases, written with generative AI.