Models need tokens to think

A roughly fixed, small amount of compute happens per forward pass, so a model cannot do arbitrary work inside one token — reasoning must be distributed across many tokens. andrej-karpathy's labeling demo: an answer-first solution ("The answer is $3. Here's why…") is training poison because it teaches single-token guessing; the step-by-step version spreads the arithmetic so no token carries too big a leap.

Explains a whole family of behaviors:

• Chain-of-thought works because it is the computation, not a presentation choice; RL-discovered reasoning (emergent-reasoning-via-rl) bloats response length for exactly this reason.
• Counting fails ("how many dots" crams everything into one token); spelling fails because models see tokens, not characters (strawberry's three Rs); both fixed by use code — copy-pasting tokens into Python is easy, and the interpreter does the work.
• 9.11 > 9.9 persists for murkier reasons (activations associated with Bible-verse markers) — a reminder the Swiss cheese has holes theory doesn't predict.

The archive connection: this is why grilling-session-style workflows spend tokens on alignment and why feedback-loop-ceiling tools beat mental arithmetic — same principle, practitioner form.

Source: report