The Foundation · Chapter 1
How We Got Here
From rule-based systems to emergent intelligence — the 8-year arc that changed everything. Understanding this history is not optional. It's your competitive advantage.
🗺️The 60-Year Shortcut

For 60 years, AI researchers tried to encode human intelligence manually — writing explicit rules, decision trees, and logic systems. They made progress on narrow tasks. But the fundamental insight they kept missing: intelligence isn't rules. It's pattern recognition at scale.

The shift happened when researchers stopped trying to tell computers what to think, and started letting computers learn what to think from data. That shift — from symbolic AI to statistical learning — is what made the modern AI era possible.

⏱️The Critical Timeline8 years that changed everything
2012
AlexNet — Deep Learning Proves Itself
A neural network trained on GPUs smashes the ImageNet competition. Computer vision researchers realize that more data + more compute = better performance. The scaling hypothesis is born.
2017
"Attention Is All You Need" — The Transformer Paper
Google Brain publishes the architecture that changes everything. Transformers use "self-attention" to process all parts of a sequence simultaneously — not one word at a time like previous RNNs. Training becomes parallelizable. Scale becomes possible.
2018
BERT & GPT-1 — Pre-training + Fine-tuning
Google's BERT and OpenAI's GPT-1 show that a model pre-trained on massive text can be fine-tuned for specific tasks with very little labeled data. Transfer learning arrives. One big model, many applications.
2020
GPT-3 — 175 Billion Parameters, Emergent Ability
OpenAI releases GPT-3 — 100× larger than its predecessor. Something unexpected happens: at scale, the model starts doing things it wasn't trained to do. It writes code, translates languages, reasons through problems. "Emergent capabilities" become the central mystery of AI.
2021–22
Diffusion Models — The Image Generation Explosion
DALL-E, Stable Diffusion, Midjourney arrive. A completely different architecture (diffusion — learning to remove noise from images) produces photorealistic image generation. AI stops being a text-only phenomenon.
2022
ChatGPT — AI Reaches Everyone
OpenAI wraps GPT-3.5 in a chat interface. 100 million users in 60 days — the fastest product adoption in history. The AI moment crosses from research into everyday life. Every industry starts asking: what does this do to us?
2023–24
GPT-4, Claude, Gemini — The Frontier Race
Multimodal models arrive (text + image + audio). Reasoning improves dramatically. Models pass bar exams, medical licensing tests, write working code. The gap between AI capability and institutional understanding becomes a business opportunity.
2025
Agentic AI — Models That Act
Models stop just answering questions — they plan, use tools, browse the web, write and run code, coordinate with other models. The shift from AI-as-tool to AI-as-colleague begins. This is where you are entering the field.
Why this matters for founders: Every major AI transition created a 12–18 month window where people who understood the new architecture could build businesses that incumbents couldn't. GPT-3 → ChatGPT was one of those windows. Agentic AI + graph architectures is the next one. You're in it right now.
The Foundation · Chapter 2
The Model Zoo
There is no single "AI." There are dozens of architectures, each built for different tasks. Knowing which tool fits which problem is your most valuable technical skill.
🤖Large Language Models (LLMs)Text Generation

What they do: Predict the next token in a sequence. That's it. The magic is that at sufficient scale, "predict the next token" produces language that reasons, explains, argues, codes, and creates.

How they're trained: Pre-training on massive text corpora (books, web, code) → Supervised fine-tuning on example conversations → RLHF (Reinforcement Learning from Human Feedback) to align with human preferences.

Key players: GPT-4o (OpenAI) · Claude Sonnet/Opus (Anthropic) · Gemini Ultra (Google) · Llama 3 (Meta, open source) · Mistral (European, open source)

As a founder: LLMs are your creative partner, code reviewer, customer discovery synthesizer, and pitch draft generator. The skill isn't using them — it's prompting them precisely enough to get reproducible, high-quality output.
🎛️LoRAs — Low-Rank AdaptationFine-Tuning

The problem they solve: Full fine-tuning a 70B parameter model costs hundreds of thousands of dollars and requires a GPU cluster. LoRAs let you customize a model's behavior by training only a small set of additional weights (0.1–1% of the total) layered on top of the frozen base model.

Why this matters for startups: You can take a base model (Llama, Mistral) and fine-tune it on your proprietary data — your customer conversations, your domain documentation, your company's writing style — at a fraction of the cost. A $100 LoRA fine-tune can produce a model that behaves like it was purpose-built for your domain.

Full Fine-Tuning
Update all model weights. Expensive. Requires massive GPU compute. Suitable for large research labs only.
LoRA Fine-Tuning
Add a small adapter matrix. Train only the adapter. $10–$500 on Replicate or RunPod. Accessible to any startup.

Real example: Artiquity's artist twin system uses LoRAs — each artist's style is encoded in a fine-tuned adapter that can be attached to a base image model. The artist's identity travels with the weights.

🎨Diffusion ModelsImage / Audio / Video Generation

The core idea: Start with pure noise. Learn to remove noise in tiny steps, guided by a text description. After enough steps, you have a coherent image that matches the description. It's "sculpture by denoising."

Why transformers ≠ diffusion: Language models predict sequences left-to-right. Diffusion models generate entire 2D (or 3D) outputs simultaneously, refining them iteratively. The two architectures are often combined — a transformer encodes the text prompt; the diffusion model generates the image.

ModelBest ForAccessCost
Stable Diffusion (open)Customization, LoRA, local deployRunPod, Replicate~$0.01/image
FLUX.1 UltraPhotorealistic, high fidelityfal.ai~$0.05/image
DALL-E 3Prompt-following, OpenAI integrationAPI~$0.04/image
MidjourneyArtistic style, aestheticsDiscord$10/mo sub
Sora / VeoVideo generationLimited previewPremium
🔗Embeddings — The Hidden Layer Everything UsesSemantic Search / RAG

What they are: Mathematical representations of meaning. Every word, sentence, or document gets converted to a vector — a list of ~1500 numbers — where similar meanings end up geometrically close to each other in that high-dimensional space.

Why they power everything: Embeddings are what make semantic search, RAG (Retrieval-Augmented Generation), and the Knowledge Graph layer of the Trinity work. When you ask your IAM bot a question, it embeds your question, searches the graph for nearby concepts, and surfaces the most relevant context before generating a response.

Trinity Graph insight: The -ity vocabulary works as embeddings because abstract state words cluster semantically. "Authenticity," "Integrity," and "Reciprocity" land near each other in embedding space — they form a neighborhood the model can navigate.
🤝Multimodal Models — Seeing, Hearing, Reading Together

The frontier has moved to multimodal — single models that process text, images, audio, and video together. GPT-4o, Gemini Ultra, and Claude 3.5 can look at a photo, understand its content, and discuss it in natural language. This collapses entire product categories:

📄
Document Intelligence
Read PDFs, extract structured data, compare contracts, fill databases.
🏗️
Visual Analysis
Read architectural drawings, analyze satellite imagery, inspect photos for defects.
🎤
Voice Interfaces
Real-time speech-to-text + LLM response + text-to-speech. Sub-300ms round trips.
🎬
Video Understanding
Describe what's happening in a video clip. Coming: real-time video analysis.
The Foundation · Chapter 3
How LLMs Actually Work
You don't need to know the math. You need to understand the mental model. This section gives you the intuitions that make you a better prompter, buyer, and builder.
🔤Tokenization — Text Becomes Numbers

LLMs don't read words — they read tokens, which are chunks of text (roughly 4 characters on average). The word "serendipity" might be 3 tokens: "ser" + "end" + "ipity." Every token gets assigned a number (its ID in the vocabulary).

Why it matters: Models have context windows measured in tokens (GPT-4: ~128k, Claude: 200k). Knowing that 1 page ≈ 500 tokens, 1 book ≈ 100k tokens helps you design prompts efficiently.

🎯Attention — What the Model Focuses On

The core innovation of the Transformer is self-attention: for every token it's generating, the model can look back at every other token in the context and decide how much it matters. It's like reading a contract where your eye keeps jumping back to the key clause that changes everything.

The key insight for prompting: Position in the prompt matters. Instructions at the beginning (system prompt) and at the very end of the user message get the most "attention weight." Content buried in the middle of a long prompt is more likely to be under-weighted.

🎲Temperature — Creativity vs. Predictability

At every step, the model produces a probability distribution over all possible next tokens. Temperature controls how that distribution is sampled:

🧊
Temperature 0
Always picks the most likely token. Deterministic. Use for: code, data extraction, factual Q&A.
⚖️
Temperature 0.7
Balanced. Default for most chat. Use for: analysis, explanations, strategic advice.
🔥
Temperature 1.2+
Unpredictable, creative, sometimes incoherent. Use for: brainstorming, creative writing, ideation.
🧠Hallucination — Why Models Confidently Lie

What it is: The model generates plausible-sounding but factually incorrect content — citations that don't exist, statistics that are wrong, events that never happened. It's not "lying" in any intentional sense: it's predicting the next token so confidently that it generates a completion that seems right, even when it isn't.

Why it happens: The model's training objective was to predict text, not to be accurate. It was never explicitly rewarded for saying "I don't know."

How the Trinity Graph solves it: By grounding every response in the Knowledge Graph — facts with sources, confidence scores, and provenance states (CONFIRMED / CITED / INFERRED / ASSUMED) — the generative layer has verified context to work from. Less confabulation. More citation.

Founder rule: Never ship a user-facing product that passes raw LLM output through without verification. The Trinity Knowledge Graph is your verification layer.
🔮Reasoning Models — The New Frontier

The 2024–25 wave introduced models that "think before they answer" — o3, DeepSeek-R1, Claude Sonnet Extended Thinking. They run an internal chain-of-thought reasoning process before producing a final answer. This dramatically improves performance on complex logic, math, and multi-step problems.

The tradeoff: Slower and more expensive. For a simple customer service query: use Haiku or Flash. For complex analysis or strategic decisions: use a reasoning model. Match the model to the task.

The Stakes · Chapter 1
The Dangers
Every powerful technology is a double-edged sword. Understanding the risks isn't pessimism — it's competitive advantage. The founders who understand failure modes design better systems.
🌀
Hallucination & Confabulation
Models generate confident, plausible, wrong answers. Not occasionally — systematically, in domains where they lack grounded training data.
⚠️ Air Canada's chatbot promised a bereavement discount that didn't exist. A court ruled the company liable for its AI's mistake.
⚖️
Bias Amplification
Models trained on human-generated text absorb and amplify historical biases — racial, gender, socioeconomic. At scale, these biases affect hiring, lending, healthcare triage, and policing.
⚠️ Amazon's AI hiring tool systematically downranked résumés from women's colleges. It was trained on 10 years of historically male-dominated hiring decisions.
🎭
Synthetic Media & Disinformation
Realistic AI-generated images, audio, and video of real people saying things they never said. At scale: election interference, financial fraud, reputational destruction.
⚠️ A deepfake audio of a CEO's voice was used to authorize a $25M wire transfer in 2024. No code was written. No technical skills required.
🏭
Economic Displacement
AI automates cognitive tasks that previously required expensive professionals. The disruption affects knowledge workers — the sector that escaped previous waves of automation.
⚠️ Goldman Sachs (2023): AI could automate 25% of current work tasks, affecting 300 million jobs globally. White-collar roles face greater disruption than blue-collar.
🏰
Concentration of Power
Training frontier models costs $100M–$1B+. This concentrates AI capability in 3–5 companies globally. Those who control the infrastructure control the intelligence layer of the economy.
⚠️ OpenAI, Google, Anthropic, Meta, and Microsoft collectively account for >90% of frontier AI capacity. The "open source" movement (Llama, Mistral) is the only structural counterweight.
©️
IP & Attribution Crisis
Models trained on copyrighted content can reproduce that content, mimic specific artists' styles, and generate works that compete with the original creators — with no compensation, no credit, no consent.
⚠️ The New York Times v. OpenAI (2023): Times claims ChatGPT can reproduce verbatim articles, destroying the value of their paywalled content. $1B+ damages sought.
🔒
Privacy & Surveillance Expansion
AI dramatically lowers the cost of surveillance — facial recognition at scale, behavioral prediction from data exhaust, real-time monitoring of communications. What was expensive and impractical becomes cheap and routine.
⚠️ Clearview AI scraped 30 billion faces from the internet without consent. Courts in multiple countries ruled this illegal — after the data had already been sold.
The founder's responsibility: You are not a passive observer of these dangers. Every product you build is either part of the problem or part of the solution. The question to ask before shipping: "Who is harmed by this, and did they consent?"
The Stakes · Chapter 2
The Possibilities
The same capabilities that enable the dangers enable something extraordinary: the compression of expertise, the democratization of creativity, and the first real tools for solving problems that have defeated humanity for decades.
🏥
Healthcare: Closing the Expertise Gap
AI can now pass the USMLE (medical licensing) exam at physician level. For the 5 billion people without access to a doctor, an LLM with a good medical knowledge graph is a genuine intervention. AlphaFold has predicted the structure of 200 million proteins — more than all experimental science combined had discovered in 70 years.
🎓
Education: Personalized at Scale
Khanmigo (Khan Academy's AI tutor) provides Socratic dialogue — asking questions rather than giving answers — personalized to each student's level and pace. For the first time, personalized tutoring (historically only available to wealthy families) is accessible to anyone.
🎨
Creative Industries: Artiquity & Attribution
The challenge isn't that AI will replace artists — it's that AI trained on artists' work generates derivative content without attribution or compensation. Artiquity is the infrastructure layer that makes AI generation traceable to its creative origins. Chain of title for the generative era.
🏗️
Construction & Infrastructure: Enginuity
Complex construction projects fail at massive scale — 85% run over budget, 75% run over schedule. Enginuity's approach: ground planning in real physics, real costs, real constraints — not Excel estimates and gut feel. Near-symbolic reasoning over built environments.
🔬
Scientific Research: Acceleration at Scale
AI agents can now run experiments, analyze results, generate hypotheses, and iterate — autonomously. In drug discovery, what took 12 years of human research can now be explored in months. The bottleneck is shifting from compute to ethical review and deployment.
🌐
Access to Expertise: Democratization
Legal advice, financial planning, medical second opinions, business strategy — currently gated by wealth and geography. AI doesn't eliminate the need for human experts, but it can give everyone a baseline level of guidance that was previously unavailable. This is the most profound social implication of the technology.
The Trinity Graph: Aware AI
Current LLMs are powerful but amnesiac — no persistent memory, no verified knowledge, no sense of who they're talking to. The Trinity Graph architecture changes this: Social Graph grounds the human context, Knowledge Graph grounds the facts, Generative Graph synthesizes with awareness. This is the architecture this course is teaching you to build.
Applied Topics · Ethics
The Ethics of AI
Ethics isn't a lecture topic — it's a design discipline. Every architecture decision is a values decision. This section gives you the frameworks to make those decisions consciously.
⚖️Three Frameworks Every AI Builder Needs

Core principle: The right action maximizes overall well-being across all affected parties.

Applied to AI: Does this product produce more good than harm, summed across all users, affected communities, and society? Not just for the paying customer, but for everyone the system touches.

The test: Run the product through the utilitarian calculator: (Expected benefit × probability of benefit) − (Expected harm × probability of harm). If the expected harm is high-probability and catastrophic, the math changes regardless of the upside.

Case Study
TikTok's Recommendation Algorithm & Eating Disorders
The algorithm maximizes engagement (a utilitarian metric) — but for vulnerable adolescents, maximized engagement means maximized exposure to eating disorder content. The aggregate engagement metric is positive. The harm to the tail population is severe. Utilitarian ethics demands you count both.
📚 Source: WSJ investigation (2021), MIT Media Lab (2022), Arturo Béjar congressional testimony

Core principle (Kant): Act only according to maxims you could will to be universal laws. Treat people as ends in themselves, never merely as means.

Applied to AI: If every AI company used your data practices, your model training approach, your content moderation policy — what world would that produce? The "universalizability test" is the most useful practical tool in AI ethics.

The test: "Would I consent to this if I didn't know which side of the system I'd be on?" (The veil of ignorance, from Rawls.) If you wouldn't accept being an unwitting training data source, you shouldn't build systems that rely on it.

Case Study
Scraping Without Consent — The Clearview Model
Clearview AI scraped 30B+ photos without consent, built a facial recognition database, sold it to law enforcement. Deontologically: users did not consent to being enrolled in a law enforcement database. Their faces were used as means, not ends. Courts in EU, UK, Canada, and Australia ruled it illegal.
⚖️ Kant would call this a categorical violation of the formula of humanity.

Core principle (Aristotle): What would a person of excellent character do? Ethics is not about rules but about cultivating the virtues — honesty, courage, justice, practical wisdom (phronesis) — that produce good action naturally.

Applied to AI: Does building this product make you more or less honest? More or less courageous? Would you be proud to explain how it works to someone it affects? Virtue ethics focuses on the character of the builder, not just the consequences of the product.

Framework
Anthropic's Constitutional AI
Anthropic trained Claude using a set of principles it calls a "constitution" — rules the AI uses to critique and revise its own outputs. This is virtue ethics applied to training: instead of just rewarding outputs that don't break rules, they cultivate a kind of judgment in the model itself.

Core principle (Noddings, Held): Ethics is grounded in relationships and the responsibility to care for the most vulnerable parties. Power asymmetries matter. The person harmed first when systems fail is the person who should drive the design.

Applied to AI: Who is most vulnerable to harm from this system? Design first for them, not for the median user. In AI: children, elderly, immigrants, the economically precarious, and historically marginalized groups are disproportionately harmed by algorithmic systems.

Framework
The SAVESUCCESS Reciprocity Principle
Reciprocity as the load-bearing axiom: "Value must be traceable. Contributions must be attributable. Systems must enable independence." Applied to AI: the people whose data, creativity, and labor train the models must receive a proportional share of the value those models generate.
🔬Ethics Checklist — Before You Ship
  1. Consent: Do the people whose data, content, or likeness you're using know about it and agree?
  2. Harm visibility: Who is most likely to be harmed by this system failing, and did you design for that case?
  3. Transparency: Can users tell when they're interacting with AI? Can they opt out?
  4. Accountability: When the system makes a mistake, who is responsible? Is there a human in the loop?
  5. Reversibility: If this turns out to be harmful, can you turn it off? Or is it already embedded in systems you don't control?
  6. Proportionality: Is the level of AI capability proportional to the stakes of the decision?
  7. Attribution: Does value flow back to the people and data sources that made the system possible?
The founder's shortcut: If you'd be embarrassed to explain your data practices to the people your data came from, you're probably doing something wrong.
Applied Topics · Workshop
Prompt Engineering & Best Practices
Prompting is not a hack — it's a discipline. The difference between a mediocre AI product and a great one often comes down to the quality of the prompts its builders can write.
🔬Anatomy of a High-Quality Prompt

Every effective prompt has five layers. Most people write only two or three. The best prompts include all five:

// SYSTEM PROMPT (sets the identity and operating rules) You are an expert financial analyst specializing in early-stage SaaS metrics. You think rigorously, cite data sources, and acknowledge uncertainty explicitly. // CONTEXT (gives the model what it needs to know) We are analyzing a B2B SaaS company: $2M ARR, 120% NRR, 18-month CAC payback, Series A closing at $12M at 6x ARR. Comparable: Notion at same stage = 8x ARR. // INSTRUCTION (what you want it to do) Evaluate whether this valuation is defensible to a Series B investor in 2026. // FORMAT (how you want the output structured) Structure: (1) Bull case (3 points) (2) Bear case (3 points) (3) Verdict with confidence score 0–100 and the one question that changes everything. // CONSTRAINTS (what to avoid) Do not hedge every sentence. Do not use the word "synergy." If you don't know, say the specific thing you're uncertain about.
🟡 System
Who is the model? What are its operating principles?
🔵 Context
What does it need to know to answer well?
🟢 Instruction
What exactly do you want it to do?
🟠 Format
How should the output be structured?
🔴 Constraints
What should it explicitly avoid?
🎯Core Prompting Patterns
Zero-Shot Basic
Just ask. No examples. Works well for clearly defined tasks with common patterns.
"Classify this customer feedback as Positive, Negative, or Neutral: [text]"
Few-Shot Basic
Show 2–5 examples before the actual question. Dramatically improves consistency on formatting, tone, and edge cases.
"Review: 'Slow delivery.' → Negative
Review: 'Amazing quality!' → Positive
Review: 'It arrived.' → [classify]"
Chain-of-Thought Advanced
Ask the model to "think step by step" before answering. Dramatically improves multi-step reasoning and reduces errors.
"Before answering, work through your reasoning step by step. Then give your final answer."
Role Prompting Basic
Give the model a specific identity and operating context. The model will behave more consistently when it has a clear role to inhabit.
"You are a senior VC partner reviewing this pitch. Be direct, skeptical, and specific."
Constraint Injection Advanced
Explicitly tell the model what NOT to do. Often more effective than telling it what to do. Negative constraints tighten outputs.
"Do not recommend tools you're uncertain about. Do not hedge every sentence. If you don't know, say exactly what you don't know."
Output Scaffolding Advanced
Provide the skeleton of the output format. The model fills it in. Forces structured, comparable outputs across many runs.
"Respond in this format exactly:
PROBLEM: [one sentence]
SOLUTION: [one sentence]
EVIDENCE: [one citation]
RISK: [one sentence]"
Omega Protocol Expert
10-rune sequential prompting engine that forces LLMs through 7 phases of structured reasoning — from grounding through convergence to return. Full spiral as a meta-prompt.
Materiality → Vitality → Interiority → Criticality → Connectivity → Lucidity → Necessity → Reciprocity → Totality → Continuity
Self-Critique Loop Expert
Ask the model to critique its own output, then improve it. Two-pass generation consistently outperforms single-pass for complex analytical tasks.
"First answer the question. Then critique your own answer — what's wrong or incomplete? Then write a revised, improved answer."
The 10 Most Common Prompting Mistakes
  1. Vague instructions: "Write something good about our product." → Good at what? For whom? In what format? What makes it good?
  2. No format specification: Getting a wall of prose when you needed a table. Always specify the output format.
  3. No context: Asking a question without the background the model needs. It will hallucinate the context it doesn't have.
  4. Over-constraining creativity: So many rules the model can't do anything interesting. Constraints should narrow, not paralyze.
  5. Prompt bloat: 2,000-word prompts where 200 words would do. Longer ≠ better. Clarity beats length.
  6. Accepting the first output: Always iterate. Ask it to improve, critique, or approach differently. The first response is rarely the best.
  7. No system prompt: Without a system prompt, the model is operating as a generic assistant. Give it a role and operating principles.
  8. Asking for opinions without context: "Is this a good idea?" → The model has no stakes. Ask it to evaluate against specific criteria.
  9. Trusting citations without verification: Models invent citations. If a source matters, verify it independently.
  10. One-shot for complex tasks: Complex multi-step tasks should be broken into chains — each step's output feeds the next prompt as context.
🔧In-Class Workshop: Prompt Surgery30 minutes

Each pod takes one of the following "broken" prompts and rewrites it into a production-quality version using the five-layer anatomy. Then we compare outputs from the original and improved prompts side-by-side.

❌ Broken Prompt
"Write a pitch for my startup. It does AI for restaurants."
✅ Your task
Add all five layers. Specify the audience, the format, the angle, the constraints. Measure the output quality difference.
❌ Broken Prompt
"Summarize our interviews."
✅ Your task
Add role, context, instruction, format (e.g., "Extract: top 3 pains, key quotes, one sentence synthesis"), constraints (no inference, only direct quotes).
The exercise goal: You should be able to take any vague business request — "analyze our competitors," "draft a cold email," "review this contract" — and turn it into a reproducible, high-quality prompt in under 3 minutes. That skill is worth more than knowing how to train a model.