thekb.eu — Research & Education

Diffusion Language Models Explained: How Google's Diffusion Gemma Works

Fri, 12 Jun 2026 00:00:00 GMT

The **MindStudio Team** publishes an *explainer* (June 12, 2026) on **modèles de langage par diffusion**, drawing on **Diffusion Gemma**, the first **open weights** implementation of this architecture, from **Google**. The starting point is a clash of paradigms. The **autoregressive** models that dominate today (GPT-4, Claude, standard Gemma) generate text **sequentially, one token at a time, left to right**, via **causal attention**. Each output depends on all preceding tokens: generation cannot be **parallelized** across positions, and each token is **fixed** once produced — the model cannot revisit its choices. **Diffusion** models proceed differently: they start from a **noised/masked** sequence and **refine it iteratively** toward a coherent output (**masked diffusion**, or *absorbing diffusion*). The forward pass progressively masks tokens; the model learns to reconstruct them; inference reverses the process over several adjustable **denoising steps**. Attention is **bidirectional**: the model sees the whole sequence in both directions and can **update any position at any step** — "changing its mind" about earlier tokens. The metaphor: writing a **draft and then revising it**, rather than a final copy produced word by word. **Diffusion Gemma**: **2 billion parameters**, **Transformer base derived from Gemma 2**, released in early 2025, weights on **Hugging Face** (Google DeepMind), also on AI Studio and Vertex AI. Key adaptations: removal of causal masking, **noise conditioning**, and simultaneous prediction of distributions over **all masked positions**. Advantages: **parallelism** (a 500-token text would require **50-100 steps** rather than 500 sequential passes, hence potentially much higher speed on long outputs), natural **infilling** and **constrained generation** (templates, code completion with surrounding context, rewriting while preserving the beginning/end), and **built-in revision**. The article states a **clear limitation**: at the 2B scale, Diffusion Gemma **does not match** the best autoregressive models (GPT-4o, Gemini 1.5 Pro) on reasoning, instruction-following, and knowledge — the gap is "closing" without being closed. For conversational use, multi-step reasoning, or token-by-token streaming, autoregressive models remain the better choice. The lineage traces back to **images**: Stable Diffusion and DALL-E left autoregression behind years ago; the open question is whether the same principle holds for text. Diffusion Gemma, through its open weights, makes it a **testing ground** for controllable generation.

A.I. Companies Are Eating Higher Education

Thu, 12 Feb 2026 00:00:00 GMT

Matthew Connelly, vice dean for AI initiatives at Columbia University, writes an alarmist op-ed in the New York Times denouncing the way artificial intelligence companies are taking over higher education, with the unwitting complicity of university administrators. **Aggressive strategies by AI companies**: Connelly describes an arsenal of tactics. Anthropic imposes exorbitant fees for enterprise accounts while paying "campus ambassadors" to promote Claude, creating conflicts of interest when these ambassadors sit on student government. OpenAI developed a ChatGPT text detector that is 99.9% accurate but refused to make it available to educators, fearing that watermarking would push users toward competitors. During final exams, OpenAI offers ChatGPT Plus free to students, Google gives premium access for the entire year, and Perplexity runs sign-up competitions on campuses. **Emblematic case of drift**: a Columbia student, Roy Lee, developed an AI tool to cheat on job interviews. Far from being sanctioned by the industry, Andreessen Horowitz admired his "audacious approach" and raised $15 million for his company Cluely, whose manifesto announces its intent to "cheat on everything." **Infrastructure ambitions**: OpenAI aspires for its bots to become "part of the core infrastructure of higher education," from admissions to academic advising. Google invites students to upload their lecture recordings to NotebookLM, a practice Columbia prohibits without authorization. Universities have no access to the data their students and faculty feed into these systems. **Impact on learning**: research shows that students using AI read less attentively, write with less precision and originality, and do not realize what they are losing. Professors report a notable decline in questions asked in class. The central paradox: the skills needed to harness AI's real potential — critical reading, analytical thinking, argumentative writing — are precisely those that passive AI use erodes. **Call to resistance**: Connelly concludes with a military metaphor: wars can be lost before they are declared if defenders abandon strategic ground without a fight. For universities, that ground is human intelligence itself. He calls on educators to defend and advance human intelligence rather than be seduced by unbalanced partnerships with an industry whose interests fundamentally diverge from the educational mission.

Traité d'Architecture Narrative et de Rhétorique de Conférence : Guide Global des Formats et des Structures de Communication

Sun, 01 Feb 2026 00:00:00 GMT

This comprehensive treatise analyzes the public speaking ecosystem through a taxonomy of formats, design methodologies, and cross-cultural narrative structures, culminating in a strategic recommendation matrix. **Format mapping**: Formats are distinguished by three variables: duration, structural constraint, and psychological contract. The Keynote (45-60 min) occupies the summit of the rhetorical hierarchy, aiming to frame a vision and create emotional engagement. The TED Talk (18 min max) imposes a "single idea" as its pivot. Lightning Talks (5-10 min) rapidly disseminate knowledge without saturation. Auto-advancing formats (Pecha Kucha 20x20s, Ignite 20x15s) transform the presentation into a synchronized performance. The Pitch follows standardized structures (Sequoia: 10 key slides) to demonstrate the viability of a business model. **Design philosophies**: Garr Reynolds (Presentation Zen) advocates radical simplicity inspired by Japanese aesthetics, countering the "Curse of Knowledge" through the elimination of the superfluous and the valorization of empty space (Ma). Nancy Duarte (Slidology) positions the audience as hero and the speaker as guide, requiring that a slide be understood in under six seconds. **Narrative frameworks**: Duarte's Sparkline models great speeches as an oscillation between "What is" and "What could be," creating tension until the final resolution. Sinek's Golden Circle (Why→How→What) creates emotional resonance by starting with the reason for being. StoryBrand (Donald Miller) places the customer at the center according to the Hero with a Thousand Faces model. AIDA and PAS remain the pillars of persuasive copywriting. **Cross-cultural structures**: Kishōtenketsu (Japan/China/Korea) generates interest not through conflict but through a four-part shift in perspective. Jo-ha-kyū modulates tempo through progressive acceleration. Indigenous circular narratives return to the starting point with deepened understanding. African Dilemma Tales end on an open question, turning the audience into a co-creator of meaning. **Recommendation matrix**: For a Keynote, use the Sparkline. For a TED Talk, the Golden Circle. For a Lightning Talk, AIDA. For a Pecha Kucha, the Kishōtenketsu. For a Pitch, PAS. For a Panel, the Dilemma Tale. For an Ignite, the Jo-ha-kyū. This format-framework fit is the key to lasting influence in a message-saturated world.

Maîtriser Claude Code — Détail complet des 12 modules et ~60 leçons

Sun, 01 Feb 2026 00:00:00 GMT

This document constitutes a complete pedagogical curriculum for mastering Claude Code, Anthropic's agentic CLI tool. Organized into 12 modules (A1 to A12) totaling approximately 60 lessons, it follows a rigorous progression from foundational concepts to advanced best practices, with an estimated duration of 3 to 5 hours. **Foundations (A1)**: Agentic coding is a paradigm shift in which the developer moves from line-by-line writing to orchestrating AI agents. Claude Code is distinguished by its terminal-first approach and provides access to three models (Haiku, Sonnet, Opus) suited to different levels of complexity. **Getting started (A2-A3)**: Installation is done with a single command (curl or npm). The conversational workflow relies on specific, contextualized prompts. Four golden rules of agentic prompting: be specific, provide context, state constraints, one objective at a time. The Explore → Plan → Code workflow is recommended by Anthropic for complex tasks. **Commands and memory (A4-A5)**: The essential slash commands (/init, /clear, /compact, /cost, /model, /rewind) structure daily work. The 4-level hierarchical memory system (Enterprise Policy, project CLAUDE.md, Project Rules, User CLAUDE.md) allows context to be shared and customized. Auto Memory allows Claude to write its own notes between sessions. **Security and context (A6-A7)**: Four permission modes (Normal, Auto-accept, Plan, Bypass) offer a control-autonomy spectrum. The OS sandbox (Seatbelt on macOS, bubblewrap on Linux) isolates commands at the kernel level. The 200,000-token context window requires active management via /compact and /clear. **Integrations (A8-A10)**: Claude Code excels at Git management (intelligent commits, conflict resolution, automated PRs). MCP (Model Context Protocol) connects Claude Code to external tools (GitHub, Brave Search, databases). Headless mode (-p flag) enables CI/CD integration via GitHub Actions with budget control. **Advanced customization (A11-A12)**: Custom slash commands (.claude/commands/), skills (automatically detected SKILL.md files), specialized subagents, and deterministic hooks (11+ events) enable extensive automation. The five professional patterns (Explore→Plan→Code→Test→Review, Context priming, Iterative refinement, Model surfing, Session hygiene) form the foundation of expert mastery. Each lesson integrates a provocative Socratic hook, a guided exploration, a tutor-learner dialogue, a hands-on exercise, and an FSRS flashcard for spaced repetition.

Synthèse : Architectures Narratives & Formats de Conférence - Guide Pratique et Prompts

Sun, 01 Feb 2026 00:00:00 GMT

This synthesis document merges data on narrative engineering and cognitive attention management, starting from the premise that stories are 22 times more memorable than facts alone, but that their impact depends on a strict match between format and narrative framework. **Visionary and long-form formats**: The Keynote (30-60 min) aims to inspire and guide using the Hero's Journey (audience = hero, speaker = mentor) or the extended Sparkline. The TED Talk (max 18 min) requires the standard Sparkline or Sinek's Golden Circle to establish a rapid emotional connection. The Masterclass (2-5h+) uses a "Modular Hero" with embedded mini-journeys for each module. **Short and effective formats**: The Lightning Talk (5-10 min) uses AIDA or a simplified Story Spine to spark curiosity. The Elevator Pitch (30-60 sec) follows AIDA with the 3C rule (Clear, Concise, Convincing). The Startup Pitch (3-10 min) combines PAS to amplify market pain with the Sequoia/YC 10-slide structure. **Constrained and visual formats**: The Pecha Kucha (6m40, 20 automatic slides) relies on Kishōtenketsu to surprise without conflict, with visual storytelling and no bullet points. The Ignite (5 min, 15 seconds per slide) uses Jo-ha-kyū to manage the imposed rhythmic acceleration. **Dialogic formats**: The Panel (45-90 min) structures individual answers with PREP (Point, Reason, Example, Point) and uses the African Dilemma Tale to end on an ambiguity that draws the audience into debate. **Framework dictionary**: Western frameworks (Sparkline, Hero's Journey, Golden Circle, AIDA, PAS, STAR) rely on conflict and linear logic. Cross-cultural frameworks (Kishōtenketsu, circular narratives, Dilemma Tale) favor harmony, cycles, or communal participation. **Principles for designing AI prompts**: Four golden rules emerge. First, position the audience as the hero and the speaker as the guide (Yoda/Luke relationship). Second, let time dictate the structure — never use a full Hero's Journey for a 3-minute pitch. Third, adapt culturally: use Kishōtenketsu for Asian or international audiences sensitive to direct conflict. Fourth, apply cognitive design to slides: one image per idea, no bullet points, following the Presentation Zen philosophy.

Playing Pretend: Expert Personas Don't Improve Factual Accuracy

Sun, 07 Dec 2025 00:00:00 GMT

This study from Wharton's Generative AI Labs examines whether assigning expert personas to AI models improves their performance on difficult objective multiple-choice questions. The researchers tested six models (GPT-4o, GPT-4o-mini, o3-mini, o4-mini, Gemini 2.0 Flash, Gemini 2.5 Flash) on two demanding benchmarks: GPQA Diamond (198 doctoral-level questions) and MMLU-Pro (300 professional-level questions). The protocol compares three conditions: a baseline with no persona, expert personas (expert in physics, mathematics, economics, biology, chemistry, engineering, law, history), and "low-knowledge" personas (Layperson, Young Child, Toddler — "a 4-year-old who believes the moon is made of cheese"). Each model-prompt pair is evaluated over 25 independent responses per question (4,950 runs per pair on GPQA, 7,500 on MMLU-Pro), with 95% confidence intervals. The results are essentially null: most persona conditions produce performance statistically indistinguishable from the baseline. On GPQA Diamond, no expert or low-knowledge persona reliably improves performance; the sole exception is a small gain from the "Young Child" prompt on Gemini 2.5 Flash (RD = 0.098). On MMLU-Pro, no expert persona delivers a statistically significant improvement for 5 of the 6 models, and nine significant negative differences are observed. Low-knowledge personas often degrade accuracy: the "Toddler" persona reduces performance in 4 of 6 models and proves significantly worse than "Layperson" in 5 of 6 models. The notable exception is Gemini 2.0 Flash, which shows modest positive differences with all five expert personas on MMLU-Pro, particularly in engineering and chemistry. Additionally, aligning the expert persona with the question's domain provides no consistent benefit. The researchers identify failure modes: the Gemini Flash models sometimes refuse to answer when assigned an out-of-domain expert persona, and overly narrow role instructions lead the models to underuse their actual knowledge. The practical implications are significant: the widespread practice of persona prompting is likely ineffective for improving factual accuracy. Organizations will derive more value from task-specific instructions, and should test multiple prompt variants for their concrete problems. Personas may nonetheless retain other uses, such as modulating tone or presentation style. The study's limitations (a limited number of models and personas, academic benchmarks) open avenues for future research.

YouTube's AI Tutorial Explosion: Democratizing Technical Education at Scale

Wed, 01 Oct 2025 00:00:00 GMT

YouTube is experiencing an **unprecedented explosion** of AI-centered educational content, with **thousands of new tutorials, courses, and explainer videos** published every month. The phenomenon represents a **democratization of technical education** at very large scale — world-class AI learning content freely accessible to anyone with an internet connection, disrupting traditional educational gatekeeping while raising challenges around quality assurance, content discovery, and creator sustainability. **Growth metrics and scale** Platform statistics reveal explosive growth: **more than 500,000 AI/ML tutorial videos** published in 2024–2025, **more than 2.5 billion views** annually on AI educational content, **leading channels** (Andrej Karpathy, StatQuest, 3Blue1Brown, Sentdex) accumulating millions of subscribers, **more than 100 hours** of new AI content uploaded every hour, search volume for "AI tutorial" or "machine learning course" up 800% year-over-year. The scale exceeds traditional academic output — **YouTube has become the primary learning resource** for many aspiring AI practitioners. **Content diversity** The spectrum covers: beginner introductions (AI fundamentals, minimal prerequisites), technical deep dives (master's/PhD level, research paper explanations), tool tutorials (TensorFlow, PyTorch, LangChain), project-based learning (complete end-to-end applications), career guidance (entering the AI field, interview preparation), research summaries, and philosophical discussions (AI ethics, societal implications). **The quality variance challenge** Quality is **highly variable**: excellence (Stanford/MIT professors, industry researchers sharing their expertise for free, production quality rivaling commercial courses, solid pedagogy) sits alongside problematic content (outdated tutorials on deprecated libraries, conceptually wrong explanations, clickbait promising unrealistic results, copy-pasted code without understanding, misleading career advice). **Learners must develop discernment** — views and likes are imperfect quality indicators. **Creator motivations and educational benefits** Motivations blend monetization (advertising, sponsorship, course sales, Patreon), reputation building, recruitment, teaching passion, community building, and marketing. YouTube offers advantages traditional education struggles to match: immediate updates, global accessibility without geographic or financial barriers, self-paced learning, practical implementation focus, visual explanations, community discussion, fast search, and being free of charge. **Discovery challenges and impact on traditional education** The explosion creates findability problems: saturated search results, blurred quality signals, persistent outdated content, absence of guided curricula. YouTube is disrupting institutions: challenging the value of costly degrees, being used as a supplementary resource, prompting teacher adaptation (flipped classrooms), and pushing toward democratization. Communities self-organize: curated playlists, Discord/Reddit communities, GitHub "awesome" lists. Sustainability questions remain: algorithm dependency, creator burnout, platform risk. The explosion represents an **unprecedented democratization of knowledge** while raising urgent questions of quality assurance and educational transformation.

MCP for Beginners - YouTube

Mon, 28 Jul 2025 00:00:00 GMT

The YouTube playlist "MCP for Beginners" created by Microsoft Developer offers a comprehensive introduction to the Model Context Protocol (MCP), a robust framework designed to develop secure, scalable, and practical AI applications. Comprising **12 videos** and totaling **more than 43,713 views**, this playlist serves as an educational resource for anyone seeking to understand and implement AI agents using MCP. **Curriculum structure** The curriculum is structured to guide learners through the different stages of MCP mastery. It begins by exploring the foundational concepts of the Model Context Protocol, establishing a base understanding for beginners. The playlist then dives into practical development, providing concrete guidance on effectively building and deploying AI agents. A significant portion of the content is dedicated to advanced implementation techniques, enabling developers to create sophisticated, multi-modal AI agents. **Key topics covered** Key topics include an introduction to MCP, its foundational concepts, and the security best practices essential to developing robust applications. Learners will also discover how to build their first MCP server and gain insight into the complete workflow of building, testing, and deploying MCP applications with real tools. The playlist further explores advanced aspects such as creating secure, scalable, and multi-modal AI agents, and provides instructions for contributing to the MCP community through tools, documentation, and code. **Practice and real-world applications** To bridge the gap between theory and practice, the playlist incorporates lessons from MCP early adopters, offering valuable insights and development best practices drawn from real-world scenarios. It also shows MCP in action through various case studies, illustrating its application across different industries. A notable practical highlight: **four labs** focused on building AI agents in Visual Studio Code with MCP and AI Toolkit, offering direct application of the concepts learned. The complete course and code samples are available via the provided link, enriching the learning experience. **Accessibility and community** The playlist is designed to be versatile and useful at every organizational scale. Individuals can use it to track their progress and identify skills to strengthen. Teams, even as small as two members, leverage the resource to optimize their current workflows and processes. For large enterprises, the content serves as a critical tool for optimizing overall processes, managing team efficiency, and making informed investment decisions. **Engagement and deployment** Security and data privacy are paramount within the MCP framework. The playlist ensures that data is handled to the highest security standards, reflected in the thorough coverage of security best practices throughout the videos. The pedagogical approach emphasizes real-world applicability, ensuring learners can immediately apply their knowledge to their own AI development projects. This resource from Microsoft Developer represents a significant investment in community education around MCP, positioning the protocol as an essential skill for the future of AI development.

Machine Learning Fundamentals: A Hands-On Guide

Mon, 01 Apr 2024 00:00:00 GMT

Sebastian Raschka, **renowned machine learning educator** and researcher at Lightning AI, publishes the comprehensive guide **"Machine Learning Fundamentals: A Hands-On Guide"**, an accessible yet rigorous introduction to modern ML practices. The book represents the culmination of Raschka's long teaching experience, **combining mathematical foundations with practical implementation** using contemporary tools such as PyTorch. **Author credibility** Raschka brings unique qualifications: a **PhD in computational biology**, author of several influential ML books including "Python Machine Learning", core contributor to **Lightning AI** (a framework simplifying PyTorch training), active educator to thousands of students, and prolific open source contributor. This combination of academic rigor, industry experience, and pedagogical expertise shapes the book's approach. **Structure and content progression** The book follows a **carefully designed learning path**: ML basics (supervised vs. unsupervised, bias-variance tradeoff, model evaluation), classic algorithms (linear models, decision trees, ensembles), neural networks (architectures, backpropagation, optimization), deep learning (CNNs, RNNs, Transformers), practical concerns (overfitting, regularization, hyperparameter tuning), and modern techniques (transfer learning, fine-tuning, deployment). **Hands-on philosophy** A distinctive trait: **every concept is accompanied by working code**. Rather than a purely theoretical treatment, readers implement algorithms from scratch to understand how they work, then use modern frameworks for practical applications. The examples favor **PyTorch** as the primary framework, reflecting the industry's shift toward this platform. **Mathematical foundations and modern landscape** The book balances **rigor and accessibility**: linear algebra, calculus, and probability are introduced as needed rather than in an initial block. The content reflects the **current ML landscape**: transformer architectures, attention mechanisms, self-supervised learning, few-shot learning, and model scaling considerations. Each chapter includes **real-world applications** (image classification, NLP, time-series forecasting, recommendation systems). **Open source ecosystem and target audience** Code repositories are public on GitHub, with accessible Jupyter notebooks and datasets (MNIST, CIFAR, etc.). The book serves several audiences: students entering ML, software engineers branching out into it, data scientists consolidating their fundamentals, and researchers seeking a comprehensive reference. Minimal prerequisites: basic Python and high-school mathematics. **Differentiation and impact** What sets this book apart in a crowded field: Raschka's pedagogical experience, the focus on a modern framework, the theory-practice balance that avoids a purely "recipe" approach, and comprehensive coverage from fundamentals to advanced topics. The book is positioned to become a **standard reference** for ML education, both in university curricula and self-study.