The Glossary
The terms I use to reason about the economics of intelligence — each defined here at the source, each backed by a full analysis. Start with Joule Wars; the rest of the vocabulary hangs off it.
- Joule WarsCORE CONCEPT
- The transition of the AI industry from competing on model capability toward competing on energy efficiency — who produces the most useful intelligence per joule. Frames AI, semiconductors, cloud infrastructure, power generation and geopolitics as a single economic system.
→ full definition & FAQ - Intelligence per jouleMETRIC
- The ratio of useful cognitive output — tokens, decisions, solved tasks — to the energy consumed producing it. More fundamental than performance-per-watt: FLOPS per watt is dimensionally FLOPs per joule (the seconds cancel).
→ why joules, not watts - Agent-hourUNIT OF WORK
- Work performed by an autonomous AI agent rather than a human. Agent-hours run in parallel and don't consume a human's sequential attention — one person can direct 60+ in a day, breaking the 24-hour ceiling on individual output.
→ The Unit of Work Is the Agent-Hour - Verification costBOTTLENECK
- The cost of confirming machine-generated output is correct. Models collapse authoring cost toward zero; verification cost holds — so it becomes the binding constraint on software work, and the skill that stays scarce.
→ Verification Cost Is the New Bottleneck - The harnessMOAT
- The orchestration layer above the model — the tooling that holds your context, reads your codebase, executes your workflows. Models are swappable; the harness owns the switching costs and the customer.
→ Who Owns Your Harness? - Clearance warsPOLICY
- The phase of AI competition where frontier capability ships not when it is built but when it is permitted — government clearance granted customer by customer. The gate, not the model, becomes the story.
→ The Model Wars Are Over. The Clearance Wars Begin. - Language world modelARCHITECTURE
- A model trained to simulate the environment itself — predicting what the world will do next — rather than to pick the next action. Imagination before action, so agents can test plans against a simulation instead of the real, irreversible world.
→ Language World Models: Predict Before You Act - Social permission to burn tokensLEGITIMACY
- Society's conditional tolerance for AI's energy consumption: if tokens don't visibly improve real outcomes, the license to spend scarce energy generating them gets revoked. Capability doesn't protect you; outcomes do.
→ The Social Permission to Burn Tokens - The liability stackADOPTION
- The layered structure of legal responsibility that decides where AI actually gets adopted: freely in paperwork, stalled at decisions where a wrong answer has no settled owner. Explains healthcare AI's paperwork-only "revolution".
→ The Liability Stack: Why Healthcare AI Stalls
Michał Piszczek
CEO / CTO / FOUNDER
CTO of Archdesk; founder and former CEO of Robotero (acquired), Rejsomat.pl, Inclify and Lextron.ai. This vocabulary comes from the essays on the blog — read the full profile.