English· Español· Deutsch· Nederlands· Français· 日本語· ქართული· 繁體中文· 简体中文· Português· Русский· العربية· हिन्दी· Italiano· 한국어· Polski· Svenska· Türkçe· Українська· Tiếng Việt· Bahasa Indonesia

un

技能职业。真实事业。您的道路。

+

创建一节课

登录以创建和分享您自己的课程。
aerospace

Orbital Mechanics: Getting to Space

From Newton's cannonball to Hohmann transfers: the physics that gets rockets into orbit and keeps satellites from falling.
agriculture

Agriculture: Growing Food

Soil science, crop management & sustainable farming: knowledge that feeds the world & the careers that depend on it.
agriculture

Geometry of Agriculture

Center pivot circles, contour farming, row spacing, drainage slope & field layout: geometry that feeds the world.
automotive

Automotive: How Cars Work

From internal combustion to EVs: understand engines, drivetrains, and diagnostics, the foundation of every auto tech career.
aviation

Aviation: Principles of Flight

Lift, drag, thrust & weight: four forces that get aircraft airborne, & the careers that keep them flying.
aviation

Geometry of Aviation

Flight vectors, bank angles, turn radius, wind triangles, and approach geometry: the math pilots use every day.
business

Eight Forms of Capital

Financial capital represents one of eight. Living, material, intellectual, experiential, social, cultural & spiritual capital each compound differently; each can grow, deplete, or exchange. Learn to see all of them.
business

Entrepreneurship: Starting a Business

From idea to revenue: learn business models, market validation, and the fundamentals every founder needs before launch day.
business

ML Triangle: Beyond Unturf

Unturf represents one instance of the machine learning triangle pattern, not the only one. Two solos with complementary specifications plus one machine learning engine plus one gap equals a bridge, across any of the eight forms of capital. Learn to recognize the pattern, find a complementary solo & write a specification that makes the triangle form.
business

What Does Kanban Mean?

From Toyota factory floors to cross-functional work centers: learn a pull system, the board, WIP limits & how to coordinate work across a product studio, bakery, construction site, or any multi-discipline operation without meetings or managers.
business

合作的几何

三角化、Voronoi 域与镶嵌：机器学习三角的几何原理。两个特性互补的单元，一个机器学习桥梁，一个需要填上的差距。
cnc machining

CNC Machining: Precision Manufacturing

From G-code to finished parts: learn how computer-controlled machines shape metal, plastic, and composites with thousandths-of-an-inch precision.
cnc machining

数控加工几何

坐标系、工艺路径、G代码弧、工作范围和GD&T：使精密制造成为可能的几何学。
communication

A Diagram Is Worth 3,000 Words

Visual communication gives you a professional superpower. This lesson covers why diagrams outperform prose, major diagram types, how to read complex visuals systematically & how to design diagrams that actually communicate.
computer science

Build a Stemmer: Harris 1956 and the Roots of NLP

Implement an English stemmer from scratch: learn morphological analysis, suffix stripping, and test-driven development.
computer science

Capacity Planning: Art & Science

John Allspaw's discipline of running services that scale without surprise: measurement, forecasting, ceiling tests, headroom buffers, & the diagonal scaling decision.
computer science

Cybersecurity: Defense in Depth

Think like a defender: learn threat modeling, network security, and incident response from the ground up.
computer science

Data Science: From Spreadsheets to Models

Bridge the gap from Excel to Python: learn how data scientists clean, explore, and model real-world data.
computer science

Geometry in Computer Science and Machine Learning

How geometric thinking powers everything from nearest-neighbor search to neural network training: vectors, distances, manifolds, and loss landscapes.
computer science

Geometry of Cybersecurity

Network topology, attack surface geometry, graph theory in threat modeling, and the geometric structure of encryption.
computer science

Site Reliability Engineering

What Google invented when ops became too important to leave to manual labor. SLOs, error budgets, toil elimination, blameless postmortems, & the four golden signals.
computer science

容量规划几何

容量工作的可视结构: Little's Law 作为面积, 排队曲线的肘部在 80%, 趋势推断作为斜率, & 头room 尺寸作为 2D 安全边距。
computer science

站稳性可靠性工程几何

SRE 可视结构下的几何:延迟分布作为长尾曲线、错误预算作为2D锥体、依赖图作为DAG、仪表板布局作为信息几何。
construction

Construction: Building from the Ground Up

From foundations to framing to finish: understand how buildings come together and the trades that make it happen.
construction

Geometry of Construction

Roof pitch, rafter angles, stair geometry, arch types, and site layout: the geometry carpenters and builders use every day.
cosmetology

Geometry of Haircutting

The head as a sphere, elevation and projection angles, graduation wedges, traveling guides, and the balanced silhouette: precision geometry behind a great cut.
cs algorithms

算法实践：代码审查与复杂性缺陷

专业软件开发需要阅读代码以了解算法复杂性，而不仅仅是正确性。这一课将介绍复杂性意识的代码审查、生产系统中的实际模式、面试式分析以及在1000+开源项目中发现的MOAD-0001缺陷——并指向未汉明课程以进行全面的调查。
culinary

Culinary Arts: Kitchen Fundamentals

Knife skills, flavor profiles & kitchen management: foundation every professional cook needs before the first ticket fires.
culinary

烹饪艺术的几何

切割几何、盘装组成、烘焙比例和锅底尺寸：厨房中的精确几何。
education

Geometry of Facilitation

Vision cones and occlusion, the room as a graph and the sweep as a watchman route, triangulating where help is needed, and the proxemics of intervention distance: the geometry a floating teacher works in.
education

船长面对船员：如何经营自主学习工作室

一个实际课程针对家长、Pod协调员和职业学校教师：浮动教师如何管理一间自我学习节奏的房间：每日节奏、注意力分配以及评估而无需给分.
electrical

110V运行：插座、开关和灯具

从拉线到连接：正确一次性连接插座、开关环路和灯具。符合NEC电码、箱子填充和线路终端技术。
electrical

Breaker Boxes and Fuse Boxes: Overcurrent Protection

From service entrance to individual circuits: how breaker panels, sub-panels, and fuse boxes protect your wiring, and what every electrician needs to read one correctly.
electrical

Electrical Wiring: Residential Circuits

Understand residential electrical systems: from breaker panels to outlet wiring, US and UK residential wiring compared, the knowledge every electrician starts with.
electrical

Electrical: Industrial Systems and Datacenters

Three-phase power, motor control centers, VFDs, and how datacenters keep the lights on: industrial electrical from switchgear to server rack.
electrical

Geometry of Electrical Work

Conduit bending math, box fill calculations, wire routing geometry, and the electromagnetic fields that follow geometric laws.
entrepreneurship

unturf Partnership Program

unturf holds machine-learning-first ventures as an umbrella organization. This lesson covers the structure, permacomputer economy, extreme leverage, ML Triangle, and how external partners join: no change required to your existing enterprise.
environmental

Renewable Energy: Powering the Future

Solar, wind, hydro & storage: understand the energy transition & booming careers in clean power.
finance

Personal Finance: Money Fundamentals

Budgeting, investing, and avoiding debt traps: the financial literacy every adult needs and nobody teaches in school.
geometry of cs algorithms

算法的几何

每个算法都有一个形状。O(1)画出一条平坦的地平线。O(N)画出一条对角线。O(N²)画出一个向上开口的抛物线。二分搜索将一条线段分成两半。哈希表将一个键映射到一个桶上，像一个函数将一个域映射到一个范围上。排序网络可视化为并行比较线。这个课堂根据算法的几何特征来阅读。
geometry of machine learning

Geometry of PAC Learning

Every chapter of PAC learning carries a shape. ε & δ draw a budget plane. VC shattering draws dichotomies on point clouds. PAC-Bayes draws a posterior cloud over hypothesis space. Double descent draws a curve that classical theory predicted should rise but actually falls. Scaling laws draw a 3D power-law surface. Empirical audit draws a Beta posterior tightening into a needle. Six chapters, six geometries, one activity.
geometry of machine learning

Geometry of ANDREA: Five Spatial Pictures of Language Model Training

Every part of ANDREA carries a shape. Embedding space draws semantic distance as Euclidean distance in R^768. Attention draws a projection of past tokens onto a query subspace. Loss surface draws a high-dimensional terrain that gradient descent walks. Bandit weights draw a discrete simplex where each vertex is one data source & every interior point is a curriculum mix. Curriculum warmup draws a restricted submanifold projection. Five geometries, one model. Walk each picture & lock in the spatial intuition that abstract equations alone cannot deliver.
geometry of unhamming

无汉明几何

汉明看到几何学无处不在：n维空间，误差校正中的球包装，高维直觉的崩溃。无汉明将这一点推广：O(N²)为面积，O(N)为线，工厂模型为有向无环图（DAG），并将复杂性类别视为几何形状。每个算法都有一个形状。学习看到形状可以预测成本。
hamming on hamming

Hamming on Hamming

Richard Hamming's graduate lectures on style of thinking in science & engineering: knowledge explosion, creativity through analogy, expert blindness, data reliability, measurement distortion, & the habits separating people who do important work from those who stay busy.
hamming on hamming

Hamming几何

Hamming工作的几何结构: 二进制n空间中的Hamming距离、通过球packing进行错误校正、√n与n随机行走的论证以及知识翻倍和专长半衰期的指数增长模型。
hamming on hamming

Foundations of the Digital Revolution

Hamming's nine-point argument for why digital beats analog: reliability of discrete values, IC interconnection economics, the shift from material to information economy, robots & manufacturing, the equivalent product insight, & the military information dominance thesis.
hamming on hamming

Geometry of the Digital Revolution

The geometric structures behind Hamming's Chapter 2: logarithmic cost curves & orders of magnitude, signal-to-noise ratio geometry in analog vs digital channels, & exponential models for technology adoption — the S-curve as a logistic equation.
hamming on hamming

History of Computer Hardware

Hamming's survey of hardware evolution: Napier bones through ENIAC & IBM 701, the exponential speed curve from relay to VLSI, three physical limits (molecular size, speed of light, heat dissipation), & the lesson experts who predicted only 3-5 computers would ever be needed failed to learn.
hamming on hamming

Geometry of Computer Hardware

The geometric structures behind Hamming's Chapter 3: exponential growth & physical ceilings on a log-log plot, the speed-of-light constraint as a sphere of influence, heat dissipation as a surface-area-to-volume problem, & parallel speedup via Amdahl's Law.
hamming on hamming

History of Computer Software

Hamming's survey of software evolution: from absolute binary to FORTRAN & ALGOL, spaghetti code & relocatable libraries, the psychological vs logical language design distinction, & four criteria for successful programming languages that predict which survive 60 years.
hamming on hamming

计算机软件几何

汉明第4章的几何结构:解析树与形式文法作为有向无环图、语言设计中的信息论冗余度、复杂性类别(大O)作为几何生长曲线以及语言接受曲线作为网络效应模型。
hamming on hamming

计算机应用的历史

Hamming对计算从数值计算转向经济驱动应用的调查: 技术采用曲线, 将计算机化应用于等价产品见解, 波音共享设计带故事,模式识别成为下一个前沿,以及从给同一讲座30年里学到的职业教训。
hamming on hamming

Geometry of Computer Applications

The geometric structures behind Hamming's Chapter 5: the logistic S-curve as a differential equation, network-effect growth as a quadratic utility function (Metcalfe's Law), & optimization landscapes — gradient descent, local minima, & why shared mutable state produces phantom gradients.
hamming on hamming

Limits of Machine Reasoning: I

Hamming's examination of the fundamental question: can machines think? The Turing test, tacit knowledge, game-playing programs, expert systems, & the boundary between rule-based logic & human judgment.
hamming on hamming

Geometry of Machine Reasoning: I

Search tree geometry: branching factor, depth, exponential node counts, alpha-beta pruning reduction, & the geometric structure that makes rule-based game programs tractable.
hamming on hamming

Limits of Machine Reasoning: II

Hamming on machine creativity: the geometry theorem-proving program & the isosceles triangle proof it discovered, computer music at Bell Labs, the automation-displacement question, & the human-machine collaboration frontier.
hamming on hamming

Geometry of Machine Reasoning: II

Formal proof as a path in axiom space; projective duality & the selfcongruence proof of the isosceles triangle theorem; Nyquist sampling geometry; proof-length minimization as a search problem.
hamming on hamming

Limits of Machine Reasoning: III

Hamming's rebuttals to common objections about machine capability, the question of career relevance, & the transition to n-dimensional design space — bridging machine reasoning limits to the geometry of high-dimensional problems.
hamming on hamming

Geometry of Machine Reasoning: III

Decision boundary geometry: linear separability, the XOR non-separability problem, the perceptron as a hyperplane classifier, feature space transformations that enable separation, & the kernel trick as a geometric embedding.
hamming on hamming

Hamming Ch. 9: n-Dimensional Space

Hamming's treatment of n-dimensional geometry: Stirling's approximation, the Gamma function, the volume of n-dimensional spheres, & the counterintuitive collapse of high-dimensional geometry that governs all engineering design with many free parameters.
hamming on hamming

n维空间的几何

斯托灵近似作为几何工具, n-球体体积公式及其数值验证, 明确计算的角落悖论, 高维几何在优化和机器学习中的意义。
hamming on hamming

Hamming Ch. 10: Coding Theory I

Shannon's communication model, the two-stage encoding architecture (source coding & channel coding), variable-length codes, the Kraft inequality, Shannon entropy, the noiseless coding theorem, & prefix-free decoding trees.
hamming on hamming

编码理论几何I

前缀免费代码的二叉树几何,克夫不等式作为树结构的几何约束,香农熵作为平均编码长度的下界,以及带有具体示例的数值验证。
hamming on hamming

Hamming Ch. 11: Coding Theory II

Huffman coding algorithm in full detail: the merge-and-split procedure, minimum average length L, code non-uniqueness, bushy vs long trees, variance of code lengths, block codes vs variable-length codes, & self-compiling programs.
hamming on hamming

编码理论之几何 II

Huffman 编码作为贪婪树构造:期望编码长度作为加权和,编码长度方差,稀疏树与长树的权衡,以及一个计算最优 Huffman 代码的数值示例及验证无噪声编码定理。
hamming on hamming

Error Correcting Codes

Hamming's Chapter 12: the discovery story of error-correcting codes, from rectangular parity checks to the (7,4) Hamming code, syndrome decoding, SECDED, & the sphere-packing interpretation of error correction.
hamming on hamming

Geometry of Error Correcting Codes

The linear-algebraic geometry behind error-correcting codes: parity check matrices as linear maps, syndrome decoding as a quotient-space operation, coset leaders & standard arrays, & the (7,4) Hamming code as a perfect tiling of binary 7-space.
hamming on hamming

Information Theory

Hamming's Chapter 13: Shannon's definition of information & entropy, Gibbs' inequality & the noiseless coding theorem, binary symmetric channel capacity, & the fundamental noisy coding theorem — what it proves, what it costs, & why error-correcting codes complement rather than replace it.
hamming on hamming

Geometry of Information Theory

The geometric structures underlying Shannon's information theory: entropy as a concave function on the probability simplex, channel capacity as a maximum over input distributions, KL divergence as a geometric distance in probability space, & rate-distortion geometry.
hamming on hamming

Digital Filters I: Eigenfunctions & Sampling

Hamming's introduction to digital filters: why complex exponentials serve as eigenfunctions of linear time-invariant systems, the transfer function as eigenvalue, & the Nyquist sampling theorem as the geometric foundation of all digital signal processing.
hamming on hamming

Geometry of Digital Filters I

Frequency as angle on the unit circle: complex exponentials trace circular paths, sampling maps to rotations, & the Nyquist limit emerges from the geometry of half a revolution. Aliasing as wraparound on the circle.
hamming on hamming

数字滤波器 II：根据条件设计与吉布斯现象

设计一个满足精确频率条件的滤波器：导出 3 样本均值滤波器、传递函数公式以及限制任何截断傅里叶级数设计的吉布斯现象 —— 以汉明对吉布斯发现的账述
hamming on hamming

Geometry of Digital Filters II

The Z-plane as geometric design space: pole-zero diagrams encode frequency response as distances from the unit circle, zeros cancel frequencies, poles boost them, & stability reads directly from pole positions relative to the unit circle boundary.
hamming on hamming

Digital Filters III: Fourier Design & Kaiser's Method

The systematic Fourier series approach to filter design: start with an ideal response, compute Fourier coefficients, truncate, apply a window to suppress Gibbs ripple, & use Kaiser's method to choose both N and the window from ripple & transition-width specifications.
hamming on hamming

Geometry of Digital Filters III

Fourier series as orthogonal projection: basis functions span a function space, Fourier coefficients are inner products, truncation equals projection onto a subspace, & the Gibbs phenomenon emerges from sinc convolution at jump discontinuities.
hamming on hamming

Digital Filters IV: Recursive Filters & Feedback

IIR (recursive) filters: feedback enables sharp frequency response with fewer coefficients, stability requires poles inside the unit circle, & the classic analog filter families (Butterworth, Chebyshev, elliptic) transformed to discrete time — with Hamming's shower story as the concrete feedback instability example.
hamming on hamming

Geometry of Digital Filters IV

IIR filter geometry: the Z-transform as a map from time to the complex frequency plane, stability as pole confinement inside the unit circle, Butterworth poles on a circle, elliptic poles on an ellipse, & the bilinear transform as a conformal map from the Laplace s-plane.
hamming on hamming

Simulation I: When Experiments Are Impossible

Hamming's first large-scale simulation — Los Alamos atomic bomb calculations — reveals the core principles: expert domain knowledge, repetitive structure, equivalent results, stability vs instability, & the danger of mistaking a simulation for reality.
hamming on hamming

Geometry of Simulation I: Euler's Method & the Step-Size Region

The geometric structures underlying numerical simulation: Euler's method as forward difference in the (x,y) plane, step size h as a geometric parameter with a stability region in the complex plane, discrete dynamical systems, & the simulation feedback loop as iteration.
hamming on hamming

Simulation II: Why Should Anyone Believe It?

Hamming's central question for every simulation: why should anyone believe it? The validity problem, analog vs digital tradeoffs, Simpson's paradox in simulation data, self-reinforcing models, & the double-blind lesson for honest simulation design.
hamming on hamming

几何模拟 II：模型有效性作为几何拟合

模拟有效性作为几何：将表面拟合到数据点上，残差作为距离，系统偏差作为偏移向量，哈沃恩效应作为隐藏的维度，汉明的验证清单重新定义为模型拟合的几何约束。
hamming on hamming

模拟 III: GIGO、稳定性与解决不可解的问题

汉明挑战了 GIGO 假设:垃圾输入并不总是意味着垃圾输出。计算中的反馈可以弥补不准确的输入。数值 ODE 解决方案作为递归数字滤波器、预测-校正方法、尼克弹道导弹的不稳定性以及汉明通过利用其不稳定性解决的电晶体方程。
hamming on hamming

几何模拟 III: 稳定区域、刚性与高维管道

ODE 解算器的完整几何图景：Euler 和 RK4 在复数 hλ-平面的稳定区域、刚性作为特征值几何、来自第九章的高维管道悖论以及利用不稳定性作为导航信号。
hamming on hamming

哈明第21章：光纤

哈明关于光纤技术的描述：全内反射、临界角、小直径优势、电磁无感和哈明在任何人提出之前就认识到的安全见解。
hamming on hamming

Geometry of Fiber Optics

The geometric structures underlying fiber optics: Snell's law as angle transformation, critical angle as a geometric threshold, numerical aperture as acceptance cone geometry, evanescent wave decay, & single-mode diameter as a constraint on propagation angle space.
hamming on hamming

Hamming Ch. 22: Computer-Aided Instruction

Hamming's analysis of computer-aided instruction (CAI): the Hawthorne effect & its corrosive impact on educational experiments, the double-blind problem in pedagogy, branching programs, immediate feedback, & the question of whether machines can genuinely improve learning.
hamming on hamming

Geometry of Computer-Aided Instruction

The geometric structures underlying learning & memory: power law learning curves, exponential forgetting curves, geometric spacing as optimization, & branching programs as directed graphs — the mathematics that makes adaptive instruction tractable.
hamming on hamming

Hamming Ch. 23: Mathematics

Hamming's exploration of what mathematics is: the Platonic vs Formalist schools, Wigner's 'unreasonable effectiveness,' the universality of mathematics across civilizations, & why abstraction increases applicability.
hamming on hamming

Geometry of Mathematics

The geometric structures underlying mathematics itself: axiomatic systems as formal spaces, proof as a path problem, Gödel incompleteness as a topological property, completeness as connectivity, & the Platonic 'space' of mathematical objects.
hamming on hamming

Hamming Ch. 24: Quantum Mechanics

Hamming's account of quantum mechanics: Heisenberg's matrix mechanics, Schrödinger's wave mechanics, their mathematical equivalence via Hilbert space, the Born rule, & Hamming's act-as-if principle for using formalism without requiring philosophical satisfaction.
hamming on hamming

Geometry of Quantum Mechanics

The geometric structures underlying quantum mechanics: Hilbert space as infinite-dimensional inner product space, quantum state as unit vector, measurement as projection, Born rule as squared projection length, Bloch sphere for qubit geometry, & entanglement as non-separability in tensor product space.
hamming on hamming

Hamming Ch 25: Creativity

Hamming's deep analysis of creativity: how novelty differs from originality, why science fails to recognize creative work when it occurs, the role of the prepared mind, the subconscious gestation process, & the 10 important problems technique.
hamming on hamming

Geometry of Creativity

Creativity modeled as search in combinatorial space: analogical reasoning as graph isomorphism, the prepared mind as a richly connected knowledge graph, serendipity as short paths between distant nodes, & the combinatorial explosion of idea space.
hamming on hamming

Hamming Ch 26: Experts

Hamming on expert resistance: how paradigms filter evidence, the two failure modes of expertise, why great innovations come from outside the field, how to handle expert pronouncements of impossibility, & what to do when you face a change in the dominant paradigm.
hamming on hamming

专家之geometry

范式空间作为一个损失地形图的局部最小值;范式转变作为逃离局部最小值;专家知识作为深度梯度投资;知识的拓扑结构具有多个局部最小值;库恩周期作为梯度下降动态。
hamming on hamming

Hamming Ch 27: Unreliable Data

Hamming on measurement: why data is less reliable than advertised, the systematic-vs-random error distinction, calibration chains, the life-testing dilemma, why data fitting theory too well is suspicious, & practical rules for handling data in the field.
hamming on hamming

Geometry of Unreliable Data

Error geometry: measurement uncertainty as probability distributions, systematic error as bias (shift of mean), random error as variance, propagation of uncertainty through calculations via partial derivatives, chi-squared goodness of fit, & suspiciously clean data.
hamming on hamming

Hamming Ch 28: Systems Engineering

Hamming's systems engineering principles: why optimizing components ruins system performance, the importance of interfaces over components, bounding conditions, the education system as a failed systems engineering case, & how to think as a systems engineer rather than a component optimizer.
hamming on hamming

系统工程几何

生产可能性边界与帕累托边界; 为什么组件最优点位于系统最优点之外; 约束几何和可行区域; 最大原理; 参数空间中接口作为共享边界; 为什么系统最优点位于约束边界上。
hamming on hamming

Hamming Ch 29: You Get What You Measure

Hamming on measurement distortion: how IQ calibration creates the distribution it claims to discover, Goodhart's law before Goodhart named it, the corruption of metrics by targeting, the dynamic range problem in rating scales, & practical defenses against measurement distortion.
hamming on hamming

几何学视角下的你得到你测量的内容

度量扭曲作为从值空间映射到测量空间；Goodhart法则作为梯度上升偏离；多目标优化（帕累托）作为对单指标污染的防御；测量不变性及其重要性。
hamming on hamming

Hamming Ch 30: You and Your Research

Hamming's direct advice on doing great work: the 10 important problems technique, courage to work on hard problems, the open door policy, compound knowledge growth, the role of luck vs preparation, & the central question every researcher must answer.
hamming on hamming

Geometry of You and Your Research

Research portfolio as a point in problem-space; the 10 important problems technique as maintaining frontier coverage; compound interest of knowledge as convex growth; opportunity cost geometry; why avoiding hard problems means missing high-leverage regions of problem-space.
health

EMT Basics: Emergency Medicine Fundamentals

First link in the chain of survival: learn patient assessment, airway management & emergency protocols.
hvac

HVAC: Heating and Cooling Systems

From refrigeration cycles to ductwork: understand systems that keep buildings comfortable & the career that keeps them running.
hvac

HVAC几何

烟道尺寸、通风几何、冷凝器表面积、心理温度图表以及加热和制冷的几何原理。
logistics

Logistics: Supply Chain Fundamentals

How goods move from factory to doorstep: warehousing, routing & the invisible network that keeps the world running.
logistics

物流几何

仓库布局优化,托盘堆叠几何,路径优化,和旅行推销员问题。
machine learning

PAC Learning: From Theory to Production Audit

Probably Approximately Correct (Valiant, 1984) gave a first rigorous answer to 'what does it mean for a machine to learn?' This lesson walks all six chapters: classical PAC, VC dimension, agnostic + PAC-Bayes variants, why deep learning broke classical bounds, scaling laws & data-wall estimates for automated general intelligence, & a production audit pattern that makes empirical PAC operational. Six chapters, one activity.
machine learning

Grow a Language Model: One GPU, One Model

ANDREA grew on a single RTX 4090. This first activity walks through what a language model predicts, why a decoder-only transformer fits 12M to 480M parameters on one consumer GPU, & maps 24 follow-up activities covering every algorithm in a recipe (Autonomous Neural Data Recipe for Education & Agency). Start here.
machine learning

基于Harris分词器的语言模型训练

ANDREA不使用BPE。它根据分布分析发现的词缀边界将词语分割，构建一个包含(256 + N + 1)个token的词汇表。这一部分将介绍分词器的作用、Harris如何找到词缀以及为什么256+N+1结构很重要。
machine learning

Grow a Language Model: Tokenizer Diet

ANDREA-12M wasted 63.6% of its vocabulary on data a model never saw. ANDREA-120M fixed that. This activity covers tokenizer-diet alignment, vocabulary saturation, & a vocab-science discovery that 8192 segments saturate a 1.25B-character corpus while 16,384 produces only 13,106 actual segments.
machine learning

学习语言模型：嵌入与位置

Token IDs 通过两个查找表流入转换器：token嵌入(V × d_model)和学习位置嵌入(T × d_model)。 本活动将介绍向量空间、为什么点积用于测量相似性以及为什么ANDREA使用学习位置嵌入而不是RoPE或正弦波。
machine learning

基于缩放点积注意力的语言模型训练

注意力计算 Q · K^T / sqrt(d_k)，对未来位置进行掩码，应用softmax，并将 V 加权。这个活动涵盖了每个部分，解释了为什么 sqrt(d_k) 因子很重要，为什么引入因果掩码使自回归生成成为可能，以及如何将 768 维输入分成 12 个头，每个头有 64 维。
machine learning

基于多头注意力的语言模型训练

一个注意力头看到一种关系模式。十二个头看到十二种。多头注意力将 d_model 分割为 n_head 平行投影，分别在每个头上进行注意力计算，然后将结果拼接回一起。ANDREA-120M 将768分成12个头，每个64个。这个活动通过分割、平行计算、拼接以及通过训练压力独立出现头部专业化的过程来进行。
machine learning

Grow a Language Model: The Transformer Block

A transformer block stacks two sublayers: multi-head attention & a feed-forward MLP, each wrapped in a residual connection & layer normalization. ANDREA-120M stacks twelve such blocks. This activity walks through the pre-norm structure, why residuals matter, the MLP's 4× expansion ratio, the GELU activation, & how a 12-layer stack composes into a working language model.
machine learning

Grow a Language Model: Loss, Perplexity & The SMMA Curve

Cross-entropy loss measures how surprised the model gets when it sees the correct next token. Perplexity = exp(loss) measures the same surprise on a friendlier scale. SMMA (smoothed moving average) loss tracks training progress in real time. ANDREA-12M reached SMMA 2.0; ANDREA-120M descended to 3.43 at v1 minimum then polished to coherent multi-paragraph chat. This activity walks through the math, the random-chance baseline, & how to read a real loss curve.
machine learning

Grow a Language Model: Backprop In Custom CUDA

ANDREA's training engine `microgpt_cuda.cu` ships hand-written forward & backward CUDA kernels for every operation: attention, MLP, layer norm, embeddings. No PyTorch, no JAX, no autograd library. The chain rule lives in C-style code that gets compiled with nvcc & linked into a single binary. This activity walks through forward kernels, backward kernels, gradient accumulation buffers, & why writing your own framework gave ANDREA full control of memory & numerical precision on a single RTX 4090.
machine learning

Grow a Language Model: AdamW & Decoupled Weight Decay

Activity 10 of 25. AdamW separates weight decay from a gradient. Vanilla Adam collapsed ANDREA-120M v1 into repetition loops; weight_decay=0.01 in v2 produced stable training & factual recall. Walk through one optimizer step: first moment m, second moment v, bias correction, decoupled decay, parameter update.
machine learning

Grow a Language Model: LR Warmup & Cosine Decay

Activity 11 of 25. Learning rate schedule decides whether ANDREA reaches coherence or stalls in a bad basin. v1 used instant peak learning rate on fresh weights & collapsed; v2 ramps linearly over 2000 steps then decays cosine to zero. ANDREA-12M used a warm restart at step 25K. Walk through every regime.
machine learning

梯度裁剪：增长语言模型

活动 12/25。梯度裁剪将每个梯度的 L2 范数限制在 max_norm=1.0 之前，AdamW 才会触及一个参数。ANDREA-120M v1 没有裁剪；源切换每 7-42 步产生梯度脉冲，驱动模型进入退化吸引子。三个 CUDA 内核并行计算全局 L2 范数。
machine learning

Grow a Language Model: FP32, FP16, FP8 Precision

Activity 13 of 25. Numerical precision sets the memory footprint, training speed, & numerical stability of a language model. FP32 baselines safety; FP16 doubles throughput on tensor cores; FP8 E4M3 doubles again with NaN risk. ANDREA-120M uses FP16 cuBLAS at 6 steps/min on RTX 4090.
machine learning

Grow a Language Model: Multi-Armed Bandits & UCB1

ANDREA picks training data using a multi-armed bandit. Each data source becomes an arm; each step pulls one arm & gets reward. UCB1 (Auer, Cesa-Bianchi & Fischer, 2002) balances exploration against exploitation through a confidence-bound formula. This activity walks UCB1 from first principles, computes scores by hand, & explains why ANDREA tunes C=0.5 instead of a textbook 1.4.
machine learning

Grow a Language Model: Phase-Based Dice Control

Vanilla UCB1 picks one arm at a time. ANDREA wraps UCB1 inside phase-based dice control: variable phase lengths (7, 14, 21, 28, 42 steps), random arms first, UCB fills remaining focus slots, & 25-33% of phases run fully random. The 42-step ceiling caps damage from bad focus picks. This activity walks the dice math, the phase ladder, & why this structure prevents lock-in.
machine learning

Grow a Language Model: Reward Attribution & EMA

UCB1 needs a reward signal. Where does ANDREA get one? Each CUDA step reports (source, doc_index, loss). Proxy maintains a per-source exponential moving average of loss. Reward = max(0, EMA(t-1) - loss(t)) * 1000. The 1000x scaling matters: without it, per-step loss improvements (~0.001) get drowned by UCB exploration bonus (~0.5). This activity walks the reward pipeline & shows the math behind the scaling factor.
machine learning

Grow a Language Model: Source Floors & Epoch Penalty

UCB1 ranks arms; phase-based dice control assembles focus sets. But raw bandit output can starve priority sources or memorize tiny ones. ANDREA layers two protective rules on top: source floors guarantee minimum sampling weight per source, & epoch penalties downweight sources that have been pulled more than once per document (1/(1+epochs)). Lifetime pulls persist across restarts. This activity walks both rules with worked examples.
machine learning

Grow a Language Model: Coverage Bonus

Doc-level coverage tracking & a 1.3x freshness bonus push a bandit toward unexplored data. CUDA reports which document index sampled each step; a proxy tracks unique doc indices per source. Sources below 50% coverage earn a multiplier scaling toward 1.3x. This activity walks the math, contrasts coverage bonus against epoch penalty, & shows how doc-level freshness keeps a 500K-document gutenberg corpus from collapsing into a 2K-document subset during 200K training steps.
machine learning

Grow a Language Model: Curriculum Warmup

Restrict a firehose to 7 chat & prose sources for the first 20K steps, then activate the full mix. v1 (16 sources from step 0) collapsed into repetition; v2 added warmup & survived; v3 polish (after step 112K) sets curriculum_warmup_steps=0 because the model already learned coherent language. This activity explains why early diversity hurts a freshly initialized 120M model & how a two-phase curriculum balances coherence with breadth.
machine learning

通过形状而不是字符过滤语言模型

v2 受到污染的训练数据：将 session JSONL 中的泄露代理提示放入用户回合 slot。模型学习到用户以多部分 Markdown 语法说话并复制了代理装饰。v2.5 补丁：`has_system_prompt_shape()` 通过形状（头部计数、长度、指纹短语）而不是字符匹配检测泄露提示。三个信号组合成一个丢弃决策。通过重新过滤 hermes3 数据源后的丢弃率：87.7%、90.4%、93.0%。这项活动让检测器学会了教训：装饰演变，结构保持不变。
machine learning

Grow a Language Model: Coherence-Gated Early Stopping

v1 produced `region region region` for 10+ days because eval_chat_quality() wired only to a legacy multi-phase runner & never ran during the firehose curriculum. v2 ships a coherence gate: every sample scores on four metrics (bigram diversity 0-35, trigram diversity 0-35, English word presence 0-20, character diversity 0-10). Auto-halt after 5 consecutive samples score below 30. Back-tested on v1, the gate would have triggered at step 132K, saving 3.8 days of compute. This activity walks the four scorers, the consecutive-counter design, & why composite metrics catch failures a single signal misses.
machine learning

Grow a Language Model: Checkpoints, Resume, Signals

ANDREA writes a checkpoint every 100 steps. Each one packs a 4-byte step header, a 4-byte parameter count, then three float32 arrays back-to-back: weights, Adam first moment m, Adam second moment v. SIGTERM triggers an immediate checkpoint then exit; SIGUSR1 triggers an on-demand checkpoint without stopping. .loss.json never gets archived (cumulative training history per the never-archive rule). Walk the byte layout, sketch a resume, & lock in why one rule (do not archive .loss.json) outranks every other archival decision.
machine learning

Grow a Language Model: Sample Audit & External Grading

Loss is a number. Reading samples is how we know what that number means. Every 100 steps, ANDREA generates 420 tokens of free-form output, & loss alongside each sample lands in five tiers: 0.30 = factual recall (ppl 1, memorized), 0.74 = textbook one-liner (ppl 2), 1.05 = constraint following (ppl 3), 1.94 = knowledgeable but imprecise (ppl 7), 0.13 = anomalous memorization (zombie arm signal). Six unrelated knowledge domains in 700 steps prove the bandit working. External chat-quality grading rated samples 9.5/10. Walk a real audit, learn the loss tiers, & catch a zombie arm in the act.
machine learning

Grow a Language Model: From microGPT to ANDREA-120M

ANDREA-120M did not converge on its first try. v1 collapsed into 'region region region' & ate 165K of 200K planned steps. Five compounding failures. v2 added five fixes & ran two weeks before data contamination surfaced at step 15K. v2.5 patched the filter. v3 ran clean for 112K steps until a zombie bandit arm reproduced repo-docs verbatim, triggering the polish pivot at step 112,619. Within hours, samples rated 9.5/10. Walk the engineering record from collapse to coherence; learn what each fix solved & what the next failure taught.
math

Geometry of Capital

Radar charts, network graphs & flow diagrams: geometric tools for visualizing capital portfolios, measuring balance across eight dimensions & mapping flows between forms.
math

Geometry of Kanban

Queuing theory, flow diagrams, and cumulative flow curves: the geometric thinking behind WIP limits, Little's Law, and bottleneck analysis in cross-functional work systems.
math

几何在现实世界中的应用：从建筑到机器学习

几何如何影响建筑、计算机图形学、机器学习、导航和工程。
music

Geometry of Woodwinds

Pipe acoustics and why length sets pitch, hole placement as a geometric sequence, why a clarinet overblows a twelfth while a flute overblows an octave, and the circle of fifths as a literal circle: the geometry inside every wind instrument.
music

Recorder to the Woodwind Family

Why the recorder is the cheapest, most durable on-ramp into a whole instrument program: what transfers to the flute, clarinet, oboe, and saxophone, what is new, teaching beginning recorder, and the parallel on-ramps of percussion, piano, and guitar.
nursing

Nursing: Patient Care Fundamentals

Science & art of nursing: from vital signs to patient advocacy, bedrock of healthcare careers.
physics

Nuclear Engineering 401 -- Reactor Design Capstone

A full semester capstone project: design a safe nuclear reactor from first principles. Choose reactor type, fuel, coolant, triple-redundant safety systems, human oversight, siting, and licensing strategy: and defend every decision like a professional engineer.
physics

Nuclear Physics 101: The Atomic Nucleus

Community college-level nuclear physics. Strong force, shell model, radioactive decay, binding energy, fission, fusion, and E=mc²: with real calculations.
physics

Nuclear Physics 201: Reactor Principles

Community college-level nuclear engineering: neutron physics, criticality, four-factor formula, fuel cycles, control rod physics, thermal hydraulics, reactor types & reactor accidents that rewrote safety culture.
physics

Nuclear Physics 301 -- Safety Systems and Defense in Depth

A full community college safety engineering module: defense in depth, triple redundancy, ECCS, containment, reactivity control, electrical systems, I&C, human factors, ALARA, accident lessons, PRA, and waste management.
physics

核物理学几何

反应堆核心格子几何、中子流动分布、屏蔽计算以及倒数平方法。
physics

轨道力学的几何

椭圆轨道、霍曼转移、轨道平面、倾角变化和拉格朗日点：太空飞行的几何。
plumbing

Geometry of Plumbing

Pipe grade calculations, fitting angles, offset geometry, drainage slope, and the physics of water flow in geometric systems.
plumbing

Plumbing: Water In, Waste Out

Understand water supply and drainage systems: the invisible infrastructure that makes modern life possible.
robotics

Geometry of Robotics

Forward and inverse kinematics, workspace envelopes, degrees of freedom, and path planning: the geometry that makes robots move.
robotics

机器人学：移动的机器人

传感器、执行器和控制循环：机器人是如何感知、决策和行动的，从工厂机械臂到火星漫游器。
unhamming

Unhamming: What Hamming Gave & What He Missed

Richard Hamming built a dragon of knowledge from Bell Labs & the cold war era. Unhamming grows from his back: same fire, different flight. No spy versus spy. No predator & prey. A mapping of what Hamming got right, what his era blinded him to, & what a permacomputer extends.
unhamming

The Missing Chapter: Algorithmic Complexity

Hamming covered digital filters, coding theory, simulation, & n-dimensional geometry. He never named Big O. This lesson fills the gap: O(1) through O(N²) with concrete programming examples, the cost of silence, & why MOAD-0001 grew in Hamming's era unnoticed.
unhamming

MOAD-0001: The Sedimentary Defect

A list used where a set belongs. CWE-407: O(N²) membership tests in 1,000+ sites across 60+ software ecosystems. Hamming's era seeded it; our era inherits it. This lesson teaches pattern recognition, code archaeology, & the one-line fix that makes infrastructure 1,000× faster.
unhamming

Undefect: 在野外发现 MOADs

知道 MOAD 存在与找到它是两回事。这节课将教您检测方法论:如何在代码中读取沉积式模式,如何使用 grep 和静态分析搜索生态系统,以及如何通过 MOAD 管道从确认缺陷到上游修复:扫描、工单、修复、测试、披露、PR、合并。
unhamming

Eight Forms of Capital: What Hamming Never Measured

Hamming taught 'You Get What You Measure' — one of his best insights. He measured financial capital & intellectual capital. Eight forms of capital exist. This lesson extends his measurement framework to all eight, then asks: what does your work grow & what does it drain?
unhamming

永久计算机：超越建造者基础设施的 Infrastructure

Russell Ballestrini 写了一篇永久计算机宣言：基础设施不抽取租金地增长，代码超越作者，免费计算属于每个人。这堂课将 Hamming 的系统工程（第 28 章）扩展到永久计算机设计原则：基于能力的展示、渐进增强和合作基础设施。
unhamming

You & Your Research: Open-Source Edition

Hamming's 'You and Your Research' assumed institutional backing: a Bell Labs, a university, a national lab. Open-source removes that assumption. This lesson extends Hamming's best advice — work on important problems, keep 10 important problems in mind, compound your knowledge — to independent researchers with nothing but commit access & a public repo.
unhamming

无间谍对谍

汉明的时代在冷战二元制中运行:我们 versus 他们, Classified versus 开放,贝尔实验室 versus 其他人。Unhamming在没有这个框架中运行。这堂课探讨了二元制在每个资本队列中的成本以及当你培养一个不需要游戏的棋盘时会发生什么。
unhamming

永久农业与编程：生长与提取

永久农业原则：观察与互动；捕获与存储能量；生产无废弃物。应用于软件：编写能够为周围系统注入能量而非从它们中提取的代码。这堂课将永久农业设计原则映射到永久计算基础设施与MOAD工厂模型。
unhamming

Balance All Workstations: MOAD-0001 & MOAD-0005 Are Coupled

Hamming's systems engineering (Ch. 28): optimizing one component in isolation degrades the whole system. A MOAD factory model applies this at infrastructure scale: fix O(N²) at a high-throughput node & every downstream queue floods. MOAD-0001 & MOAD-0005 couple. Solve both before disclosing either.
unhamming

The Quadrivium: Truth, Freedom, Harmony, Love

Hamming ended 'You and Your Research' with the question every researcher must answer: what are your values, & what will you sacrifice for them? The permacomputer quadrivium gives a concrete answer: Truth (open source), Freedom (voluntary participation), Harmony (self-renewal), Love (wu wei). This lesson applies each to software & research practice.
unhamming

MOAD-0002: An Intertangled Defect

Subsystems coupled through shared mutable global state, mutated synchronously in a single event loop or god-object. A structural coupling, not a timing problem. This lesson teaches why adding a mutex makes it worse, & how phase-snapshot decouples subsystems at the architectural level.
unhamming

MOAD-0003: 泄露的上下文

ThreadLocal持有请求范围的身份。线程返回到池子。ThreadLocal.remove()没有被调用。下一个请求在同一个线程上读取上一个请求的上下文。不是崩溃;一个静默的安全边界违规。这堂课教会你上下文的不匹配以及替代它的生命周期正确的基本原理。
unhamming

MOAD-0004: A Logged Secret

Web servers log HTTP headers verbatim. Authorization, Cookie, & X-API-Key get written to disk in cleartext. Not in memory; on disk, in log shippers, in aggregators, retained for 30-90 days. This lesson teaches the exposure window calculation & the credential denylist fix at the serialization layer.
unhamming

MOAD-0005: A Thundering Herd

cache.get(key) + null check + expensiveCompute(key) + cache.put(key): four non-atomic steps. N threads all miss, all compute simultaneously. The idempotency trap: 'it is OK if two threads write the same result' gets correctness right but ignores cost. This lesson teaches the atomic fix & its coupling to MOAD-0001.
unhamming

MOAD-0006: A Glass Safe

Credentials stored in recoverable form: plaintext, reversible encryption, base64. A single database dump exposes all credentials simultaneously. Unlike MOAD-0004 (accidental log write), a Glass Safe requires credential recovery by architectural design. This lesson teaches why strong encryption does not fix it, & what one-way hash replaces.
unhamming

MOAD-0007: A Flatland Defect

Spatial data queried with O(N) linear scan. Geometry structure discarded. BVH, octree, & k-d tree never built. Named after Edwin Abbott's Flatland (1884): the structure exists; it gets discarded. Three.js Raycaster at 60Hz with 10,000 objects: 600,000 geometry checks per second from a 3-line fix. This lesson teaches the spatial index & why a hash set cannot replace it.
unhamming

MOAD-0009: A Metered Heart

A scheduled job fires on a timer regardless of whether work needs doing. Two forms: a state-repair job that fixes broken state after a delay (users see glitches during the window), & a wasteful report that recomputes from scratch on a fixed interval. This lesson teaches event-driven design as the replacement & the MOAD amplification pattern.
unhamming

Why MOAD Festers: The History of Known Defects

Five defects — O(N²) membership, global state coupling, ThreadLocal identity leak, credential logging, thundering herd — were knowable decades before systematic detection in 2026. This lesson traces the intellectual history, identifies the five structural conditions that allowed each to persist, & builds a diagnosis vocabulary for recognizing unfixed knowable defects.
welding

Geometry of Welding

Joint angles, bevel geometry, weld profiles, distortion control, and fit-up tolerances: geometric precision in metal joining.
welding

Welding Fundamentals: Joining Metal

Learn science & craft behind welding: from arc physics to joint design, foundation of every trade career in metal fabrication.
Unlimited access — no payment required

Complete 2 lessons, write a paragraph about your enterprise & earn solo-tier access through entrepreneur contribution track.

Entrepreneur track →