Working Knowledge
Working Knowledge · 2026 Edition Field Guide

What AI Actually Is (and Isn't)

Everything you need to understand AI — and start building with it — explained from zero. Eight short readings, then a complete lexicon. No prior knowledge assumed.

Terms decoded
8
Readings
~33 min
End to end

Eight readings in order, or pick a path below. Gold links open lexicon definitions. Search covers all 225 terms.

Begin
Start Here

Eight readings, in order.

Read top to bottom and the whole field assembles itself — what AI is, how it works, what you can build with it, and where the money goes. Or pick a path, then dive into the lexicon below.

How to use this
  • Readings — eight essays, top to bottom, or follow a path for your situation.
  • Paths — curated shortcuts if you are non-technical, building, or just scanning headlines.
  • Lexicon — search or filter 225 terms; tap any card for the full definition.
The Board

Who's who — and who pays whom.

Most AI infrastructure headlines come down to a few dozen names in three roles. Learn the roles and the deals stop being a blur. The specific names shift; the shape of the board doesn't. As of 2026.

The labs

They build the models
  • OpenAI — GPT; backed by Microsoft
  • Google DeepMind — Gemini; in-house
  • Anthropic — Claude; backed by Amazon & Google
  • Meta — Llama, open-weight
  • xAI — Grok; merged into SpaceX (Feb 2026)
  • Mistral — open-weight, EU
  • Alibaba (Qwen) · DeepSeek · Zhipu (GLM) — China now leads open weights

The clouds

They own the compute
  • Microsoft Azure — backs OpenAI
  • Amazon AWS — backs Anthropic
  • Google Cloud — owns DeepMind
  • SpaceX · Colossus — the Memphis supercomputer; rents to Anthropic & Google
  • Oracle — data-centre capacity
  • CoreWeave — GPU "neocloud"

The silicon

They make the chips
  • Nvidia — GPUs + CUDA; ~75% of accelerators, and slipping
  • AMD — the main challenger
  • Google TPU · Amazon Trainium · Microsoft Maia — in-house silicon
  • Broadcom — custom accelerators
  • TSMC — fabricates nearly all of it
1The money moves in a circle. Nvidia sells chips to the clouds.
2The clouds rent that compute back to the labs.
3The clouds are also the labs' largest investors — Microsoft in OpenAI, Amazon and Google in Anthropic.
4And Nvidia, in turn, takes stakes in the very labs and clouds buying its chips — it floated a non-binding plan to put up to $100 billion into OpenAI, since reported to have stalled and shrunk to a fraction of that.
Follow the arrows Stranger still, Anthropic now rents an entire supercomputer from Elon Musk's SpaceX — which owns its rival xAI. Capital, compute, and equity keep circling the same handful of names, which is exactly why skeptics ask how much of the boom is real demand and how much is money changing pockets.
The Lexicon

The complete field guide.

Every term, organized not alphabetically but by how the pieces fit together. Tap any card to open it — each holds a plain definition, a concrete example, and where it connects.

Start with these 15
No match in the lexicon.
Try a broader term, or tap “All” to see everything.

Where to actually begin.

You don't need to install anything or write a line of code. Six moves turn everything above into something you can use this week.

Pick one assistant and live in it.Open ChatGPT, Claude, or Gemini and use it daily for a week on real tasks — email, summaries, planning. Fluency comes from reps, not reading.
Treat it like a sharp new colleague who forgets every conversation.Give it context, one example of what “good” looks like, and the format you want back. The vaguer the ask, the vaguer the answer.
Verify anything that matters.It writes what sounds right, which isn't the same as right. Check every fact, figure, quote, and citation before you lean on it.
Feed it your own material.Paste in the document, the thread, the spreadsheet. The moment it answers from your stuff instead of the open internet, it gets genuinely useful.
Start free; pay when it buys back time.The free tiers are enough to learn on. Upgrade only once a tool is clearly saving you hours of your week.
Be careful the moment it can act.Answering is low-stakes; doing — sending, buying, deleting — is not. Hand an agent real access slowly, and watch what it does.

How to read the whole field.

Researchers live in the architecture, training, and prompting layers. Builders and operators care about inference, tasks, and evaluation. Safety and policy people orbit alignment. Investors track the hardware — with scaling laws as the thesis tying it together.

And the one thing to never forget: these systems are built to sound right, which isn't the same as being right. Anything that matters gets an independent check.

On the numbers.

The figures here — capital spending, training scale, market positions — move fast. They were verified against current sources as of July 2026 and are framed to age gracefully, but treat any specific number as a snapshot, not a fixed point.

Where a claim was time-sensitive, it’s dated in the text. Everything conceptual — how a model works, what an API is — stays true regardless.

Decoded & designed · Working Knowledge · 2026 Edition

Compiled by AtlasScope.

Figures current as of July 2026.