STEP 00 · WRITE
Awaiting the first transition.
Chain of Thought as computation
A prompt is input.
Thought is a Turing tape.
Transformers cannot think in one pass. Chain of Thought gives them a writable tape — each token a cell, each step a transition. Type something ordinary. Watch it become a machine.
δ
⟨ qidle , □ ⟩ → ⟨ qidle , □ , N ⟩
Expanded program
Massive CoT prompt
This is what you would paste into any model. It forces serial writes onto a tape.
Execution
The machine running
qhalt · accept
The mapping
Why Chain of Thought is a Turing machine
A transformer without intermediate tokens is a constant-depth circuit. It reads the prompt once and must emit an answer. That is not enough for anything inherently serial — multi-step arithmetic, planning, counting, search. Those problems need writable memory between steps.
A Turing machine is exactly that: a finite control Q, an unbounded tape Γ, and a transition δ(q, σ) → (q′, σ′, D). Chain of Thought is the same object wearing different clothes.
- Tape Γ The context window. Every thought token is a cell the model can later re-read.
- Head Attention. The next forward pass reads the cells it already wrote.
- State Q The finite control — parse, split, step, check, revise, halt.
- δ transition One generated sentence: read prior cells, write one new result, move right, change state.
- Why it works Serial problems become serial tokens. Depth is no longer bounded by the architecture — it is bounded by the tape.