Inside Grok

nano-Grok · ~33k

Each stack of cubes is one word already in the prompt. You’re watching one guess.

When you hit send in Grok, ChatGPT, or Claude, the model does not retrieve an answer. It predicts the , appends it, and predicts again. The paragraph you read is that loop. This page is — using a finished recipe, not training one.

The cubes are one pass of that loop, shrunk. Each vertical stack is a token already on the tape. Each cube is one number in that token’s working vector. You are running : two layers, 16-wide, four experts. Ask is a different box — the live model, not these cubes.

Adds one word to the reply. Hit it a few times — that’s a chat.

What the cubes read

Each chip is one column. These are the only words this toy knows.

3 tok · < $0.0001 · 14 J

Skip to

Created out of curiosity by Phil Holland, inspired by Brendan Bycroft’s LLM Visualization.

Ask Grok 4.6

Offline

Two machines. This box = Grok 4.6. The cubes = a pocket demo. Hitting Ask sends your question to 4.6 and, if any words match the toy dictionary, those columns light up.

This ask, in pieces

Your question
5 pieces
Extra think (high)
~20
A short reply
~80
About
≈ $0.0006
Energy
0.10 Wh

0.10 Wh is about 38 seconds of a 10 W bulb. You pay for pieces of text. Energy follows how much work the GPUs did — extra think, and running every specialist instead of two, both burn more.

Cubes can show 4 of those words: “why is the sky”. 1 don’t fit this toy dictionary.

Think budget: high

Try a question that uses toy words — “why is the sky blue” — so the cubes and 4.6 move together.

Field notes

0/8

  • Write the next word
  • Load a prompt
  • Pin a cube
  • Look from another word
  • See an expert light up
  • See a real guess
  • Watch one full pass
  • Finish the tour

Each chapter button ticks a note. Hover a also counts.

Chapter 1 of 8. Space continues. P watches a forward pass.