01

What is Artificial Intelligence?

Try it — type a prompt and hit Generate
Your response will appear here…

"AI is any system that perceives its environment and takes actions to maximize its chance of achieving goals."

Artificial Intelligence refers to machines designed to perform tasks that typically require human intelligence — recognizing speech, making decisions, translating languages, generating text or images. Modern AI learns from data rather than following hand-crafted rules. It finds patterns that humans couldn't write down explicitly.

Definition: Software trained on data to perform cognitive tasks — pattern recognition, prediction, generation — without explicit step-by-step programming for each task.
02

How Large Language Models Work

Watch token-by-token prediction
Loading…
Step 1 of 5

"An LLM doesn't understand — it predicts, with extraordinary precision."

LLMs are trained on billions of text examples to predict the next token (word fragment) given all previous tokens. The Transformer architecture uses attention mechanisms to weigh which previous tokens matter most for each prediction step. Training on internet-scale text creates emergent capabilities — from reasoning to coding — that weren't explicitly programmed.

Definition: An LLM is a neural network trained via next-token prediction on massive text corpora, producing fluent text by sampling from learned probability distributions.
03

AI vs ML vs Deep Learning

Click each ring to learn more
Artificial Intelligence
Machine Learning
Deep
Learning
Click a ring above to explore each layer

"Every deep learning model is ML. Every ML model is AI. But not all AI uses learning."

AI is the broad field of building intelligent systems. Machine Learning is a subset where systems learn from data rather than following explicit rules. Deep Learning is a subset of ML using multi-layer neural networks. LLMs like Claude or GPT-4 are deep learning models, which makes them ML, which makes them AI.

Hierarchy: AI ⊃ Machine Learning ⊃ Deep Learning ⊃ Large Language Models. Each level adds a more specific technical approach.
04

What AI Can and Can't Do

Tap "Reveal Next" to uncover each capability
AI Can
Write & summarize
Generate images & code
Translate languages
Recognize patterns
Answer questions instantly
AI Can't
Truly understand
Have emotions
Be reliably factual
Remember past sessions
Replace human judgment

"The gap between impressive outputs and genuine understanding is where most AI mistakes live."

Current AI excels at pattern-matching and generation at massive scale. But it has no persistent memory across sessions, no grounded understanding of the world, and can confidently produce false information — a phenomenon called hallucination. Knowing these limits makes you a much more effective AI user.

Key limit: AI models are statistical pattern matchers — they don't "know" things the way humans do. Always verify important outputs from any AI system.
05

Key Players in AI

O
tap to flip
GPT-4o
Multimodal model by OpenAI. Powers ChatGPT and DALL·E.
A
tap to flip
Claude 3
Safety-focused AI by Anthropic. Best for long docs & nuance.
G
tap to flip
Gemini
Google DeepMind. Powers Search & Workspace.
M
tap to flip
Llama 3
Meta's open-weight model. Free to run & fine-tune locally.

"A handful of labs are building systems that will reshape how humanity works and creates."

OpenAI leads with GPT-4o and DALL-E. Anthropic (makers of Claude) focuses heavily on AI safety research. Google DeepMind powers Search and Workspace with Gemini. Meta releases open-weight models like Llama that anyone can download and run locally. Smaller players like Mistral and xAI are rapidly closing the gap.

Open vs. Closed: Closed models (GPT, Claude, Gemini) keep their weights private — you access them via API. Open models (Llama, Mistral) release weights for anyone to run, fine-tune, and study.
06

AI as a Tool, Not a Replacement

Drag to see the productivity effect
Human Only Human + AI
Productivity 30%
30%
Skilled human, standard pace.

"The best outcomes happen when human judgment steers AI capability."

AI works best as an amplifier — it multiplies what a skilled person can produce. A writer with AI can draft faster and explore more iterations. A developer with AI can explore more solutions per hour. The human provides direction, taste, context, and judgment; AI provides speed, scale, and tireless generation.

Mental model: Think of AI as a very fast, knowledgeable intern — remarkably capable, but needs clear direction, context, and active supervision to produce quality work.
07

AI Safety & Ethics

Click each topic to expand
Alignment is the challenge of ensuring AI systems pursue the goals we actually intend. Misaligned AI could optimize for proxy metrics in harmful ways. Labs use RLHF (Reinforcement Learning from Human Feedback) and Constitutional AI to steer model behavior toward safe, helpful outcomes.
AI models train on vast datasets that may include personal information without explicit consent. Questions around data ownership, the right to be forgotten, and opt-out mechanisms are central. The EU AI Act and GDPR set early frameworks for AI data governance.
AI systems trained on historical data can encode and amplify existing biases — in hiring, lending, criminal justice, and more. Fairness requires careful dataset curation, bias audits, diverse testing, and ongoing monitoring well after deployment.

"Building powerful AI without safety constraints is like building a rocket without guidance systems."

AI safety covers alignment (ensuring AI pursues intended goals), robustness (resisting manipulation and adversarial inputs), and fairness (not encoding or amplifying harmful biases). Labs like Anthropic, DeepMind, and OpenAI have dedicated safety research teams working on these problems before they become crises.

Alignment: The challenge of ensuring an AI system's goals and behaviors match what its designers and users actually intend — not just the literal instruction, but the spirit behind it.
08

The Future of AI

Click each era to explore what's coming
1
2
3
4
Now
2025–26
2027+?
?
Narrow AI (Now)
AI is superhuman at specific tasks — chess, protein folding, code, image generation — but cannot generalize across domains without human guidance.

"We are building the most transformative technology in human history, and we're doing it in real time."

We're currently in the narrow AI era — systems that are superhuman at specific tasks but can't generalize without guidance. AI agents that autonomously complete multi-step goals are emerging now. AGI (human-level general intelligence) and ASI (superhuman general intelligence) remain the most debated and uncertain frontiers in technology history.

AGI: Artificial General Intelligence — a system that can perform any intellectual task a human can, across any domain, with equivalent capability. No consensus on timeline, definition, or achievability.

You've finished AI Basics!

You now understand what AI is, how LLMs work, the key players, and the limits and future of AI technology.

Continue: AI Tools