How AI actually works.
The mechanism behind the chat box, the business behind the free tier, and the environmental figures that nobody agrees on. Written from primary sources, with the disagreements shown rather than smoothed over.
New foundation guide · Agentic AI
How AI agents work
Follow the complete loop from goal and context through tool calls, verification, failure handling, guardrails, and human approval.
System design guide · Agent context
What is context engineering?
Learn how retrieval, tools, memory, permissions, and working state shape what an agent can do.
Quality guide · AI systems
How to evaluate AI systems
Build task sets, choose graders, test agent trajectories, and catch regressions before users do.
Data guide · AI foundations
How AI training data works
Follow collection, filtering, licensing, human feedback, synthetic data, retrieval, provenance, and model limits.
Prompt design guide · AI foundations
What is meta-prompting?
Learn when an AI should help design, improve, compare, and test the prompts your team uses.
Operations guide · Responsible AI
AI governance in practice
Inventory use cases, assign ownership, map risk, measure behavior, and manage incidents.
Implementation guide · Agentic AI
How to build an AI agent
Choose a workflow, design tools and permissions, evaluate the trajectory, and launch a reversible pilot.
Architecture guide · Grounded AI
How retrieval-augmented generation works
Follow the path from source documents and search to grounded answers, citations, permissions, and evaluation.
Workflow pattern · Prompt engineering
What is prompt chaining?
Split complex tasks into checked steps, then decide whether the extra calls earn their cost and latency.
Economics guide · AI infrastructure
How AI model pricing works
Understand tokens, caching, batch work, tool calls, quotas, and the cost of a completed workflow.
Architecture decision · AI for business
AI agents vs workflows
Choose control for known steps, flexibility for unknown steps, and a hybrid when you need both.
ChatGPT · Mechanism
How ChatGPT Works
How ChatGPT actually works: tokens, pretraining, RLHF and context windows, and why the mechanism explains every mistake it makes.
Claude · Mechanism
How Claude Works
How Claude works and how it differs from ChatGPT: constitutional AI, long context, and why XML tags change the quality of its answers.
Gemini · Mechanism
How Gemini Works
How Gemini works: natively multimodal from the start, grounded in Google Search, and wired into Workspace. What that changes in practice.
ChatGPT · Business
How ChatGPT Makes Money
How OpenAI makes money from ChatGPT: consumer subscriptions, business seats and API usage, and why the free tier exists at all.
ChatGPT · Impact
How Much Water and Energy ChatGPT Uses
How much water and electricity a ChatGPT query really uses. The official figure, the academic figure, and the accounting difference behind the gap.
Claude · Mechanism
How Claude Code, Skills and Cowork Work
How Claude Code actually works as an agent, what Skills are, and what Claude Cowork does differently. Three products, one underlying idea.
Now put it to work
Knowing the mechanism is what makes the prompting advice make sense. Every technique on the rest of the site is a workaround for something explained here.