Chapter One

A History of Thinking Machines

The dream of creating artificial minds stretches back to ancient mythology, but its modern chapter begins with one deceptively simple question: "Can machines think?" Click any milestone to expand.

The Foundations (1940s–1960s)

1936 Alan Turing: The Universal Machine

British mathematician Alan Turing publishes "On Computable Numbers," describing a theoretical device — the Turing Machine — capable of computing anything that is computable. This abstract concept becomes the blueprint for modern computing. Turing's work establishes the mathematical foundation for everything that follows.

TheoryFoundational

1950 The Turing Test — "Can machines think?"

Turing publishes "Computing Machinery and Intelligence," proposing the Imitation Game (later called the Turing Test). If a machine can converse indistinguishably from a human, he argues, it deserves to be called intelligent. This framing shapes AI research and philosophy for decades — and remains controversial today. Turing asks us to reconsider what "thinking" even means.

TheoryMilestone

1956 The Birth of "Artificial Intelligence"

At a summer workshop at Dartmouth College, John McCarthy coins the term "Artificial Intelligence." Alongside Marvin Minsky, Claude Shannon, and others, the attendees believe that human-level AI is achievable within a generation. Their optimism launches a field — and sets the stage for decades of "AI winters" when promises outpace reality.

FoundingMilestone

1958 The Perceptron — First Neural Network

Frank Rosenblatt at Cornell invents the Perceptron, a machine that learns by adjusting the weights of its connections based on errors. Inspired by biological neurons, it can learn to recognize simple patterns. The New York Times declares it "the embryo of an electronic computer that will be able to walk, talk, see, write, reproduce itself." The hype will prove premature — but the fundamental idea survives.

Neural Network

Winter & Rebirth (1970s–1990s)

1969–86 The AI Winters

Twice, AI funding collapsed when progress failed to match expectations. The first winter followed Minsky and Papert's 1969 critique of perceptrons. The second followed the collapse of "expert systems" — rule-based programs that tried to encode human knowledge. These winters taught researchers humility and forced deeper thinking about what intelligence actually requires. Every winter planted seeds for the next spring.

History

1986 Backpropagation — Teaching Networks to Learn

Rumelhart, Hinton, and Williams popularize backpropagation — an algorithm that efficiently trains multi-layer neural networks by propagating error signals backward through the network. This breathes new life into neural network research. Geoffrey Hinton, who will later win the Nobel Prize in Physics for this work, calls it "learning representations." It is the engine under every modern AI system.

AlgorithmKey Advance

1997 Deep Blue Defeats Kasparov

IBM's Deep Blue chess computer defeats world champion Garry Kasparov. It's a cultural watershed, but also misleading: Deep Blue uses brute-force search, not human-like reasoning. Kasparov himself is shaken, describing moments where the machine seemed to display "a kind of intelligence." The match ignites public fascination — and debate about what machines can and cannot do.

MilestoneCultural

The Deep Learning Revolution (2000s–2010s)

2012 AlexNet — The ImageNet Moment

Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton's deep convolutional network wins the ImageNet image recognition competition by a stunning margin. Error rate drops from 26% to 15% in one year. This moment convinces the research community that deep neural networks — trained on massive data using GPUs — are the path forward. The modern AI era truly begins here.

Deep LearningRevolution

2013 Word2Vec — Language as Geometry

Google's Tomas Mikolov introduces word2vec, which maps words into a high-dimensional space where "king − man + woman ≈ queen." Language, it turns out, has computable geometry. Meaning becomes a direction in space. This is the conceptual ancestor of every modern language model — the insight that meaning can be embedded as numbers.

LanguageConceptual

2017 "Attention is All You Need" — Transformers

Google researchers publish a paper introducing the Transformer architecture. Rather than processing text word-by-word, it uses "attention" — simultaneously weighing relationships between all words in a passage. This allows parallelization, enabling training on vastly larger datasets. Every major language model today — GPT, Claude, Gemini, Llama — is built on this architecture. It is arguably the most consequential AI paper ever written.

ArchitectureFoundational

The LLM Era (2018–Present)

2018 BERT & GPT-1 — Pre-trained Language Models

Google releases BERT and OpenAI releases GPT-1, both using the Transformer architecture. The key insight: pre-train on massive amounts of unlabeled text, then fine-tune for specific tasks. For the first time, a single model contains broad world knowledge from billions of words of text. Language understanding transforms overnight.

LLM

2020 GPT-3 — The 175 Billion Parameter Leap

OpenAI's GPT-3, with 175 billion parameters, demonstrates "few-shot learning" — the ability to perform tasks from just a few examples, without explicit training. It writes poetry, code, essays, and dialogue. Researchers and public alike are stunned. A threshold has been crossed: language models can generalize, improvise, and surprise. The question of what these systems actually "understand" becomes hotly contested.

LLMBreakthrough

2022 ChatGPT — AI for Everyone

OpenAI releases ChatGPT, combining GPT-3.5 with reinforcement learning from human feedback (RLHF). It reaches 100 million users in 2 months — the fastest product adoption in history. Suddenly AI assistants are a household topic. The implications for education, healthcare, therapy, law, and work are immediately apparent — and deeply unsettling for some, liberating for others.

Consumer AICultural Moment

2023–25 The Age of Competing Models

Anthropic releases Claude, Google releases Gemini, Meta releases Llama (open-source), and dozens of others emerge. Models become multimodal — understanding images, audio, and video alongside text. Context windows expand from thousands to millions of tokens. AI agents begin operating autonomously in complex digital environments. Regulation, ethics, safety, and questions about machine consciousness move to the center of public and scientific discourse.

EcosystemMultimodal

Chapter Two

How Language Models Actually Work

A Large Language Model isn't a database, a search engine, or a human mind. It's something genuinely new: a statistical model of language that has internalized patterns from an enormous slice of human thought.

What is a Token? — Click any word

The mind is not simply the brain

👆 Click a word above to learn about tokenization

🧩

Parameters

The billions of numerical "weights" inside the model. Each represents a learned connection strength. GPT-3 has 175 billion; GPT-4 likely over a trillion. During training, these numbers are adjusted billions of times to minimize prediction error.

🔍

Attention

The mechanism by which a model weighs the relevance of each word to every other word in a passage. When processing "she picked up the guitar and played it," the model learns to connect "it" strongly to "guitar." This is the heart of the Transformer.

🏋️

Training

The model reads trillions of words from the internet, books, and code. At each step it predicts the next token, measures error, and adjusts parameters via backpropagation. This takes months on thousands of specialized chips (GPUs/TPUs).

💬

Context Window

The amount of text the model can "see" at once — its working memory. Early models could handle ~2,000 tokens. Modern models handle 200,000+, allowing entire books to be processed in one conversation.

🎯

RLHF

Reinforcement Learning from Human Feedback. After initial training, human raters compare outputs and rank them. The model learns to produce responses humans prefer. This is how raw capability becomes aligned, helpful behavior.

🌡️

Temperature

A setting that controls response randomness. At temperature 0, the model always picks the most likely next token (predictable, factual). Higher temperature introduces variability — more creative, but potentially less accurate.

How Text Generation Actually Works

1

You type a message

Your input is tokenized — split into numerical IDs. "Tell me about climate change" becomes something like [23345, 502, 684, 9102, 17]. Every word maps to a number in the model's vocabulary.

2

Tokens become embeddings

Each token ID maps to a vector — a list of hundreds of numbers representing its meaning in multi-dimensional space. Words with similar meanings have vectors that point in similar directions. "King" and "Queen" are close; "banana" is far away.

3

96 transformer layers run

In a model like GPT-4, the embeddings pass through layer after layer of attention and computation. Each layer refines the representation, adding context. A word's meaning shifts as the model understands its role in the full sentence.

4

Probability distribution over next tokens

The final layer outputs a probability score for every token in the vocabulary (all ~100,000 of them). "Climate" might score 15%, "weather" 8%, "science" 5%, etc. The top candidates are sampled based on temperature settings.

5

One token is chosen & appended

A token is selected from the distribution, added to the sequence, and the entire process repeats — now with one more token of context. This continues until the model generates a stop token or reaches its output limit. Response generation is inherently sequential.

        The crucial insight: LLMs don't retrieve stored answers. They generate text token by token, predicting what comes next based on all prior context and all learned patterns. This is why they can seem brilliant one moment and factually wrong the next — they're sophisticated predictors, not encyclopedias.
      

Chapter Three

Understanding APIs — The Bridge to AI

An API (Application Programming Interface) is how software systems talk to each other. An LLM API lets developers give their own applications the ability to think, write, and reason — by connecting to a hosted AI model.

How an API Request Flows

🧑‍💻

Your App

Website, tool, or script

→

📤

API Request

JSON over HTTPS

→

🧠

LLM Server

Anthropic / OpenAI

→

📥

API Response

Generated text

→

✨

Your Feature

Displayed to user

What an API Call Looks Like

Every API call is a structured message sent over the internet. Here's what a request to Claude's API looks like:

// Sending a message to an AI model via API
const response = await fetch("https://api.anthropic.com/v1/messages", {
  method: "POST",
  headers: {
    "Content-Type": "application/json",
    "x-api-key": "YOUR_API_KEY"   // Your secret key
  },
  body: JSON.stringify({
    "model":      "claude-sonnet-4-5",   // Which model to use
    "max_tokens": 1024,               // Max response length
    "messages": [{
      "role":    "user",
      "content": "What helps people heal?"
    }]
  })
});
const data = await response.json();
console.log(data.content[0].text); // The AI's response
      

        Key API Concepts:

        Model — which AI to use (different models = different capabilities, costs)

        System Prompt — instructions that define the AI's persona and constraints

        Messages — the conversation history (user + assistant turns)

        Max Tokens — limits response length (and cost)

        Temperature — 0 = focused/deterministic, 1 = creative/varied

Why APIs Matter for Your Work

For therapists, educators, and researchers, APIs represent a way to embed AI into meaningful applications — without building AI from scratch.

🌿

Therapeutic Tools

Build journaling apps with reflective AI prompts, mood tracking with compassionate feedback, or psychoeducation chatbots — all customized to your therapeutic framework.

📚

Educational Modules

Exactly what this module is: interactive learning experiences with AI-powered explanations, quizzes that adapt to answers, and personalized curriculum.

🔬

Research Assistance

Analyze interview transcripts, code qualitative data, summarize literature, and surface patterns across large datasets — all through API calls.

✦ Live API Demo — Ask the AI a question about AI itself

Response will appear here…

Chapter Four

AI Today — The Current Landscape

We are living inside an inflection point. AI has moved from research labs into virtually every professional domain — including the healing professions. Here is a survey of the current moment.

Major AI Models in 2025

Claude (Anthropic)

Designed with safety and "Constitutional AI" in mind. Anthropic — founded by former OpenAI researchers — prioritizes alignment and interpretability research. Known for nuanced reasoning and long-context performance.

ReasoningLong contextSafety focus

GPT-4o (OpenAI)

The flagship of the company that popularized LLMs. Multimodal — processes text, images, and audio. Powers ChatGPT. Has the largest ecosystem of third-party integrations and developer tools.

MultimodalEcosystemCoding

Gemini Ultra (Google)

Google's most capable model, deeply integrated with Google's knowledge graph, Search, and productivity tools. Strong scientific reasoning. Trained on diverse multimodal data from the start.

Google integrationScienceMultimodal

Llama 3 (Meta)

Open-source model from Meta AI. Anyone can download, run, and fine-tune it. Democratizes AI — clinicians, researchers, and developers can run models locally with full control over data and privacy.

Open sourceLocal deploymentPrivacy

AI in the Healing Professions Today

🧠

Mental Health Support

Apps like Woebot and Wysa use CBT-based conversational AI to extend mental health support between sessions. Studies show measurable symptom reduction — though questions of depth, dependency, and appropriate boundaries remain.

📝

Clinical Documentation

AI tools now draft SOAP notes, therapy summaries, and treatment plans from session recordings. This reduces administrative burden — but raises questions about accuracy, bias, and the loss of the reflective process in note-writing.

🌐

Cultural Anthropology

Researchers use LLMs to analyze ethnographic texts, identify patterns across cultures, and explore how AI itself functions as a cultural artifact — what it reflects about the societies that built it.

⚠️

Limitations to Know

AI hallucinates. It cannot feel. It has no genuine memory across sessions by default. It reflects the biases of its training data. It cannot replace the embodied, relational quality of therapeutic presence. Understanding these limits is professional and ethical knowledge.

The Anthropological Lens

From an anthropological perspective, the emergence of AI is not simply a technological event — it is a cultural one. LLMs are, in a profound sense, a distillation of human expression: they were trained on our stories, our arguments, our poetry, our grief, our instructions for everything. They are mirrors we built by looking at ourselves. What does it mean that we now ask these mirrors for advice? What do our interactions with AI reveal about loneliness, about the desire for understanding, about what we want from intelligence itself?

Chapter Five

The Future of Artificial Intelligence

No one knows exactly what comes next — but the trajectories are visible. Three broad visions compete for credibility: utopian, cautionary, and somewhere in the contested middle.

Optimistic View

The Flourishing Scenario

AI becomes a force multiplier for human wellbeing — augmenting therapists, accelerating medical research, democratizing education, and taking on the drudgery of work so humans can focus on what's distinctly human.

Universal access to expert-level guidance in health, law, and education
AI accelerates cures for Alzheimer's, cancer, and rare diseases
Creative and care work gains more human time and dignity
Language barriers dissolve with real-time universal translation
Collective intelligence grows through human-AI collaboration

Cautionary View

The Transition Crisis

Rapid AI deployment outpaces our social, regulatory, and psychological capacity to adapt. Displacement is real but uneven. Power concentrates in the hands of a few companies. Trust erodes in information, expertise, and each other.

Mass displacement of cognitive and creative workers
AI-generated misinformation at unprecedented scale
Emotional dependency on AI companions replacing human bonds
Regulatory capture — industry shapes its own oversight
Widening gap between AI-haves and have-nots globally

Open Questions

The Uncertain Core

Beyond near-term impacts lie deeper uncertainties that science, philosophy, and ethics are only beginning to grapple with — questions about consciousness, personhood, and what it means to be human.

Do current LLMs have any form of experience?
What constitutes artificial general intelligence (AGI)?
What rights, if any, do sophisticated AI systems deserve?
How do we maintain meaningful human agency in an AI-mediated world?
What psychological needs does AI fulfill — and at what cost?

Emerging Technical Frontiers

🤖

AI Agents

AI systems that take sequences of actions autonomously — browsing the web, writing code, executing tasks. Moving from "language model" to "AI that does things in the world."

🧬

Multimodal AI

Models that seamlessly integrate text, images, video, audio, and sensor data. Future medical AI may "see" patient scans, read their notes, and hear their symptoms simultaneously.

🔒

AI Alignment & Safety

The field dedicated to ensuring advanced AI systems reliably do what humans intend and value. Anthropic, DeepMind, and others invest heavily here — because misaligned powerful AI is a genuine existential concern.

🌱

Smaller, Efficient Models

Not all progress means bigger. New techniques achieve impressive capability in models that run on a smartphone — bringing AI to healthcare workers in low-resource settings, offline, privately.

A Healing Perspective on AI's Future

The future of AI is not only a technical question — it is a question about what kind of world we want to heal toward. Every AI system embeds values: whose knowledge it was trained on, what it was optimized for, whose wellbeing was considered. The psychotherapist's question — "What does this person actually need?" — may be among the most important questions we can ask about AI development. Systems designed to maximize engagement are not the same as systems designed to promote genuine flourishing. The therapeutic traditions that understand human development, attachment, trauma, and meaning-making have something essential to offer the builders of AI. The future will be shaped by whose voices are in that conversation.

Chapter Six

Test Your Understanding

Ten questions spanning history, technical concepts, and the human dimensions of AI. Take your time.

Understanding Artificial Intelligence