AI is not magic — it is math, history, and human choice. We build from the root and ask: who does this serve?
Course Philosophy
The machine doesn't think. You teach it to.
Now you must decide what.
AI isn't magic — it's math, history, and human choice wrapped in an abstraction. We dismantle the abstraction layer by layer, learning from the root. We choose our own adventure, build things that matter to our communities, and ask the questions that textbooks skip.
18
Weeks of Deep Learning
3
Adventure Tracks
0
Exams. All Projects.
∞
Questions Encouraged
Our Foundation
Why This Course Exists
Most AI courses teach you to use tools. This one teaches you to think — about systems, power, and whose intelligence gets valued. We start at the root: probability, logic, linear algebra. We build upward. We ask critical questions at every layer.
01
First Principles First
We don't start with TensorFlow. We start with a probability, a line, a decision. You'll understand every layer before you stack them. Amy Ko's research shows that reading before writing — understanding before doing — creates more robust, transferable learning.
02
Choose Your Adventure
Three tracks share core concepts and diverge on projects. You pick problems that matter to your community. Jeff Anderson's framework: "You are the world's leading expert on your own learning." Your final project is yours to define.
03
Liberation Through Understanding
Rooted in Paulo Freire and bell hooks: you are not an empty vessel to be filled with AI facts. You are a co-creator of knowledge. Critical consciousness about who built AI systems — and who they harm — is not a "soft" topic. It is the most advanced technical skill in this course.
AK
Rooted in Dr. Amy J. Ko's Research
Dr. Ko at the University of Washington has spent 25+ years researching how to make computing education equitable, joyous, and liberatory. Her discoveries directly shape this course: justice-focused CS requires student trust and agency; understanding ML requires understanding uncertainty; reading programs before writing them builds deeper mastery; scaffolded problem solving beats unguided trial-and-error. CS assessments often aren't fair — so we ditch traditional assessments entirely.
Equitable Computing EdRead Before WriteScaffolded Problem SolvingJustice-Focused CSUnderstanding Uncertainty in MLStudent Agency + Trust
Jeff Anderson's Principles — Applied to CS 180
Rule 1: Health Comes First
Your physical, mental, and emotional health always comes before course content. If you must choose between wellbeing and a deadline, choose wellbeing. Come back when you're ready. This is not a platitude — it is the first rule.
Rule 2: Deep Learning > Shallow
We don't optimize for grades. We optimize for understanding. A neural net you truly understand is worth infinitely more than one you copied. Find your sweet spot: the edge of productive struggle.
Rule 3: Show Up, Show Out
Your presence enriches everyone. Peer instruction is the most powerful learning tool we have — Amy Ko's research confirms this. When you teach, you learn deepest. Bring your full self.
Rule 4: Finish Vegetables First
Complete the core learning before passion projects. Vegetables = foundational AI concepts. Dessert = building something you care about. Both required. The best final projects come from students who did both.
Rule 5: Critical Consciousness = Technical Skill
Understanding who built a system, for whom, and what it encodes is not soft — it's the most advanced skill here. You have agency in this room. Use it.
The 2-Min Question Rule
When stuck, give yourself 2 minutes. Then write a precise question: not "I don't get it" but "I'm on step 4 of backprop and I understand the chain rule but can't see why we multiply the Jacobian here." That specificity is expertise in formation.
Intellectual Lineage
Alan Turing
On computable numbers. The Turing Test. What does it mean for a machine to think?
Claude Shannon
Information theory. Entropy. The mathematical foundation of all communication.
Paulo Freire
Pedagogy of the Oppressed. Students are not empty vessels. Knowledge is liberation.
bell hooks
Teaching to Transgress. The classroom as a site of freedom and transformation.
Joy Buolamwini
Gender Shades. Bias in facial recognition. The coded gaze and its consequences.
Safiya Umoja Noble
Algorithms of Oppression. Search engines and the politics of information retrieval.
Ruha Benjamin
Race After Technology. How "neutral" automation encodes and amplifies racial hierarchy.
Amy J. Ko
Equitable, joyous, liberatory learning about computing for all. The north star.
Choose Your Path
Three Tracks. One Destination.
Every student engages the same core concepts each week — probability, optimization, learning, ethics. Then we diverge into project tracks. Tracks aren't ceilings. They're starting points. You can move up mid-semester. You can also propose your own track.
Track I
The Curious Explorer
"No CS background required. Just bring your questions."
I
You'll use visual tools, Google Colab notebooks, and real community data to understand AI from the outside in. We learn to read AI systems before we write them — a principle from Amy Ko's research: reading before writing creates deeper mastery.
📋Wk 1–2: AI Audit — analyze a real AI system that affects your community. Document what it does, who it serves, and what it misses.
📊Wk 3–5: Spam Detector using Naive Bayes in pure Python dicts — no libraries. Every probability hand-calculated.
🖼️Wk 6–9: Image Classifier using Teachable Machine, trained on photos meaningful to you. Document every failure.
💬Wk 10–13: Chatbot: rule-based first, then ML-powered. You compare the two and write about what the difference reveals.
📰Wk 14–16: Fake News Detector + a bias audit of your own model.
🎯 Final: "AI That Serves My Community" — a working prototype + a public presentation + a written community impact analysis.
Track II
The Builder
"You know Python. Now let's make it think."
II
You implement core algorithms from scratch in NumPy before ever touching scikit-learn. The rule: never use a library function until you've built it yourself first. Understanding means being able to explain every matrix multiplication.
🔢Wk 1–2: Naive Bayes from scratch with Python + NumPy only. Prove it works on real data.
📈Wk 3–5: Linear Regression with gradient descent — no sklearn. Every update rule derived from the loss.
🧠Wk 6–9: 3-layer neural network with backpropagation in NumPy. MNIST from scratch.
🔍Wk 10–12: A* search + minimax for a game you design yourself.
🎯 Final: Full ML pipeline applied to a social justice dataset of your choosing. Technical rigor + critical analysis of what the model encodes.
Track III
The Architect
"You want the derivations. All of them."
III
Calculus, linear algebra, probability theory — all in play. You'll derive backpropagation from the chain rule, understand attention mechanisms mathematically, and engage critically with published AI research papers. Amy Ko: "understanding ML requires understanding uncertainty."
∂Wk 1–2: Derive the MLE estimator for Gaussian distributions from scratch.
📐Wk 3–5: Prove gradient descent convergence under convexity conditions.
🔁Wk 6–9: Derive backprop via chain rule; implement a mini autograd engine.
👁️Wk 10–12: Implement a mini-transformer with self-attention from scratch.
🎯 Final: A novel research question + replication study of a published paper. Submit as a research preprint with critical analysis of its claims.
18-Week Calendar
Root to Branch: Week by Week
We begin with the question of intelligence itself — historical, philosophical, critical — before touching a line of code. This is Amy Ko's principle of reading before writing, and Freire's principle that consciousness precedes action. Every unit builds deliberately on the last.
Wk
Theme & Core Question
Concepts — From the Root
Lab / Project Work
Unit 1
What is Intelligence? — Epistemology, History, Probability
01Unit 1
What Is Intelligence?
Turing Test, Chinese Room, animal cognition. Who decides what counts as intelligent? History from Dartmouth 1956 to ChatGPT 2024.
Foundational
Symbolic AI vs. Connectionism
Two paradigm wars. Expert systems. The perceptron winter. Why neural nets came back. Critical framing: Freire's banking model of education mirrors rule-based AI.
Critical
Lab: Design Your Turing Test
Students design, administer, and analyze a Turing test on a real LLM. 300-word reflection: what did it reveal and what did it conceal?
Project
🧠 Quick Recall
What would it mean for a machine to "truly" think?
Name one thing the Turing Test cannot measure.
What does "intelligence" assume about the judge?
💼 Careers that need this
AI Product Manager
ML Ethics Researcher
Policy Analyst
02Unit 1
The Grammar of AI: Probability
Random variables, events, joint probability. Bayes' theorem from scratch. Why does probability let us reason under uncertainty?
Foundational
Intuition vs. Formula
Monty Hall, base rate fallacy, Bayesian reasoning in medicine. Humans are bad Bayesians. Ko: "understanding ML means understanding uncertainty." Novice: visual tools. Architect: measure theory intro.
Lab: Spam Classifier by Hand
Build Naive Bayes with Python dicts only — zero libraries. Every probability manually computed. No moving forward until you can explain every number.
Project
🧠 Quick Recall
Write Bayes theorem from memory. What does each term mean?
Why do we use log probabilities in Naive Bayes?
What does the "naive" assumption mean — and when does it fail?
💼 Careers that need this
Data Scientist
Security Engineer
NLP Engineer
Unit 2
AI as Search — State Space, Heuristics, Adversarial Play
03Unit 2
AI as Search
BFS, DFS, A*. The state space. How is intelligence framed as finding a path through a space of possibilities? The frame problem introduced.
Foundational
Heuristics: Art or Science?
Admissibility, consistency, completeness. Novice: interactive visual grid. Builder: implement all three. Architect: prove admissibility guarantee from scratch.
Lab: Maze Solver
Build, visualize, and compare BFS vs. A*. Then ask: what does "optimal" actually mean in a real-world search problem?
Project
🧠 Quick Recall
What does A* guarantee that BFS doesn't? When?
What makes a heuristic "admissible"?
Name a real problem where BFS is the right choice over A*.
💼 Careers that need this
Robotics Engineer
Game Developer
Logistics/Routing
04Unit 2
Games & Adversarial AI
Minimax, alpha-beta pruning. What assumptions does "optimal play" make about your opponent — and about the world?
Game Theory + Ethics
Zero-sum vs. cooperative games. Prisoner's dilemma. Arms races. When AI agents compete, who gets hurt? Critical discussion: autonomous weapons, recommendation algorithms.
Critical
Project: Unbeatable Tic-Tac-Toe
Minimax from scratch. Novice: scaffolded. Builder: with alpha-beta. Architect: prove the speedup with a time-complexity analysis.
Project
🧠 Quick Recall
What does alpha-beta pruning guarantee about the result vs. minimax?
What assumption does minimax make about your opponent?
Name a real domain where minimax would fail — and why.
Predict housing prices in your own neighborhood. Then: whose neighborhoods get valued more, and what does the model encode about society?
Critical
🧠 Quick Recall
Write the MSE loss function. What does each term mean?
Why does the learning rate matter? What happens if it's too high?
What is the difference between the loss and the gradient?
💼 Careers that need this
ML Engineer
Quantitative Analyst
Data Scientist
06Unit 3
Classification + Decision Boundaries
Logistic regression, sigmoid activation, binary cross-entropy loss. From predicting "how much" to deciding "which category."
What Does "Accurate" Mean?
Precision, recall, F1, confusion matrix. Accuracy is a lie when classes are imbalanced. Medical diagnosis, recidivism prediction, hiring algorithms. What should we optimize for?
Critical
Project: Admissions Classifier
Build a college admission classifier. Then audit it for bias. Write a 400-word ethics brief: what does this model assume, and who does it harm?
Project
🧠 Quick Recall
What is the difference between precision and recall?
When would you choose high recall over high precision?
How can a model be 95% accurate and still deeply harmful?
💼 Careers that need this
AI Fairness Researcher
Classification Engineer
Responsible AI Lead
Unit 4
Neural Networks — Perceptron to Backpropagation
07Unit 4
The Neuron: Biology to Math
The biological neuron → the perceptron. Weights, bias, activation functions. The XOR problem: why one layer can't learn everything.
Foundational
Multi-Layer Networks
Hidden layers, ReLU, tanh. Universal approximation theorem (intuitively). Novice: trace the network on paper. Architect: prove the UAT sketch.
Lab: XOR Neural Net
NumPy only. Trace every weight update by hand at least once. You must explain what happens at each node before you run the code.
Project
🧠 Quick Recall
Why can't a single-layer network solve XOR?
What does a ReLU activation do? Draw it.
What does the universal approximation theorem say — and what does it NOT say?
💼 Careers that need this
Deep Learning Engineer
Research Scientist
AI Software Engineer
08Unit 4
Backpropagation: The Chain Rule
The most important algorithm in modern AI — fully demystified. Credit assignment across layers. Computational graphs. The engine of deep learning, exposed.
Foundational
Forward + Backward Pass
Gradient accumulation. Automatic differentiation. Novice: visual diagram. Builder: NumPy full implementation. Architect: build mini-autograd from scratch (like Karpathy's micrograd).
Lab: Handwritten Digit Recognizer
MNIST from scratch vs. PyTorch. Same results, different effort. Celebrate what the abstraction hides — and understand why it hides it.
Project
🧠 Quick Recall
Explain the chain rule applied to a 3-layer network without formulas.
What is a "vanishing gradient" — and which activation functions cause it?
What does PyTorch's autograd actually do for you?
💼 Careers that need this
Computer Vision Engineer
ML Platform Engineer
Research Scientist
09Midpoint
Midterm Portfolio Exhibition
Not a test. A community showcase. Students present their first major project. Structured peer feedback. Written learning self-evaluation due. Celebrate progress.
Exhibition
Critical AI Panel Discussion
Guest practitioner (healthcare, education, or criminal justice tech). Class debates: "What should AI never be allowed to decide?" Structured Socratic discussion, not lecture.
Critical
Final Project Design Sprint
Students pitch final project ideas. Structured peer + instructor feedback using the 2-minute question protocol. Learning partnerships form around shared interests.
Project
Unit 5
Unsupervised Learning + Natural Language Processing
10Unit 5
Learning Without Labels
K-means clustering, PCA. What patterns exist in data that no human named? The power and danger of unsupervised discovery.
Dimensionality & Meaning
The curse of dimensionality. Embeddings as compressed meaning. PCA as listening to the data's own story. Word vectors: the geometry of language.
Lab: Cluster Your Playlist
K-means on Spotify audio features. Does the algorithm's grouping match your intuition? When it doesn't — why not? What does the model think "similarity" means?
Project
11Unit 5
Natural Language Processing
Tokenization, TF-IDF, word2vec intuition. From bag-of-words to semantic embeddings. How machines read — and what they miss.
Language as Power
Whose language is "standard"? AAVE, dialects, slang in NLP systems. Safiya Noble on bias in training data. This is not a soft topic — it is a precision engineering problem.
Critical
Project: Sentiment Analyzer
Train on Yelp reviews. Test on text from your community's language. Document every failure. Propose a technically grounded fix.
Project
Unit 6
Computer Vision + Generative AI
12Unit 6
Computer Vision
Convolutional neural networks. Filters, pooling, feature maps. How spatial pattern recognition works — from edge detection to object recognition.
What Machines See
Facial recognition + racial bias. Gender classification failures. Saliency maps: what is the model actually looking at? Buolamwini's "Gender Shades" as required reading.
Critical
Lab: Community Image Classifier
Classify local flora, community murals, or culturally meaningful objects. Test on edge cases you design. Document failures and hypothesize causes.
Project
13Unit 6
Generative AI
How AI creates. GANs, VAEs, diffusion models (conceptual). The latent space as creative medium. Where does generative AI actually come from?
Authorship, Consent, Labor
Who owns AI-generated art? Training data and consent. Displacement of creative workers. Class debates: should this technology exist in its current form?
Critical
Lab: Generate With Purpose
Create a generative tool with declared intent. Write a 500-word artist + engineer statement explaining your technical and ethical choices.
Project
Unit 7
Reinforcement Learning + LLMs + Alignment
14Unit 7
Reinforcement Learning
Agents, environments, rewards, policies. Q-learning. The math of learning through consequence — like humans, but without empathy or self-preservation.
Who Defines the Reward?
Reward hacking. Goodhart's Law. The alignment problem as a political problem. Every reward function is a value judgment. Who made that judgment?
Critical
Lab: Q-Learning Gridworld
Design your own reward function. Notice how your choices shape the agent's behavior. Reflect: what did you accidentally teach it to optimize for?
Project
15Unit 7
Large Language Models
Transformers, self-attention, pre-training + fine-tuning. How ChatGPT actually works — at the architecture level, not the PR level.
Hallucination, Truth, Power
Why LLMs confabulate. Epistemic authority without accountability. Concentrated AI power. Required: Bender et al., "On the Dangers of Stochastic Parrots."
Critical
Lab: Red-Team an LLM
Systematically probe for biases and failure modes. Document findings. Builder/Architect: fine-tune a small model. Write a technical safety audit report.
Project
16Unit 8
AI Ethics & Governance
Algorithmic fairness definitions — and why they conflict mathematically. Explainability, accountability. EU AI Act, California proposals.
Critical
Navigate vs. Transform
Jeff Anderson's framework: how do you navigate harmful systems without internalizing oppression? As a future AI practitioner, what will you uphold — and what will you transform?
Final Project Studio Time
Dedicated studio time. Learning conference check-ins. Peer feedback sessions. Document your process, not just your product.
Project
17Final
Final Exhibition — Day 1
Community showcase. Each student presents: what they built, what they learned, what they'd change. Not a performance — a conversation with peers and guests.
Exhibition
Learning Conferences
Individual meetings: you present your portfolio and assign your own grade with evidence. Instructor reviews for completeness and provides final feedback.
Portfolio Due
All projects, reflections, concept notes, and final learning self-evaluation. You assign your grade. Evidence required.
18Finale
Final Exhibition — Day 2
Remaining presentations. Community celebration. Reflection on what this class leaves with you — and what you leave behind in it.
Exhibition
Where Do You Go From Here?
Transfer pathways. Careers in ethical AI. Organizations doing justice-focused tech work. You are now a practitioner. Act like one.
Course Co-Evaluation
You evaluate the course, instructor, and your own growth. Your feedback co-creates the next version of this class. This is Freire's dialogic education made real.
Assessment
You Assign Your Grade. Here's Why.
Adapted from Jeff Anderson's ungrading practice, grounded in research showing traditional grading harms learning. We don't optimize for grades. We optimize for real mastery you can use in your career and community.
Your Learning Portfolio
Your portfolio is evidence of your learning journey — not a finished showcase, but a process document. First attempts, failures, revised understanding, growth moments. Include everything.
Concept notes for every major idea, in your language + technical language
All project code with documented thinking, not just clean code
Critical essays: who built this, for whom, and what should change?
Bi-weekly learning reflections
Mid-term and final learning self-evaluation
Evidence of peer teaching + peer feedback given
Your "Conquering College" activities: meta-learning about how you learn
Three Feedback Sources
You receive feedback from three sources, not one. The instructor is the smallest. This mirrors professional practice in tech.
Self-Directed: Check your implementations against first principles. Use the 2-minute question rule. Document your own errors and corrections.
Peer-to-Peer: Learning partnerships (2 people) and learning groups (4–5) meet weekly. You teach each other. Teaching is the deepest learning.
Instructor Conferences: Every 2 weeks, your group presents portfolio progress. 15–20 min. Targeted, specific feedback. Not a grade — a coaching session.
You Assign Your Grade: Final self-evaluation with evidence. If you've done the work deeply and honestly, you earn an A. You present the case for it.
Learning Strategy
How to Succeed in This Course
These aren't productivity hacks. They're how people actually learn hard technical material — especially first-generation students navigating a field that wasn't designed with them in mind.
01
Trace Before You Run
Before running any code, trace through it by hand on paper with 3 small inputs. Predict the output. Amy Ko's research: this is the single highest-leverage learning habit in CS.
02
Teach One Thing Weekly
Explain one concept to a classmate, family member, or rubber duck every week. If you can't explain it plainly, you don't understand it yet. Teaching reveals the gaps that practice hides.
03
Reduce to 3 Elements
When stuck on an algorithm, shrink the problem. n=3 instead of n=100. Draw it. The bug is almost always visible at small scale. Don't reach for debugging tools before reaching for paper.
04
Write Before You Search
Before using ChatGPT or Stack Overflow, write your precise confusion: what you understand, what step breaks, what you've tried. That writing is the learning. The answer is secondary.
05
Read the Error, Don't Flee It
Error messages are AI's way of pointing exactly at the problem. Read the full error. Find the line number. Read the surrounding code. 80% of bugs announce themselves clearly — if you look.
06
Discomfort Is Data
Confusion doesn't mean you're behind — it means you're at the edge of your current model. Write down exactly what's confusing. That specificity is expertise in formation. Then bring it to class.
How It All Connects
The Concept Map
AI isn't a collection of isolated tools — it's a web of ideas that build on each other. Here's how every major concept in this course connects.
Foundation Layer
Probability + Statistics ↓ Bayes Theorem ↓ Maximum Likelihood ↓ Everything Else
Search + Optimization
State Space Search ↓ BFS / DFS / A* ↓ Minimax ↓ Gradient Descent
Learning
Loss Function ↓ Linear → Logistic ↓ Perceptron → MLP ↓ Backpropagation
Who builds it? ↓ Who benefits? ↓ Who is harmed? ↓ What should change?
Real Data
Practice with Real Datasets
Every project in this course can be applied to real data that matters — not iris flowers and toy examples. These datasets are curated to connect technical learning to real-world consequence.
ProPublica COMPAS
Recidivism prediction scores used in US courts. Foundational for bias auditing — used in Machine Bias investigation.