Temperature

Prompt 1 Samples: 0

Drag from the token to aim. Tap Sample to launch.

Sampling Lab

Temp 0: correct and boring. Temp 2: unhinged but occasionally a genius. Thread the token from start to goal.

temperature 0.00

0 · greedy 1 · honest 2 · unhinged

How Temperature Works

You are the decoder. There is exactly one knob: sampling temperature. The token is your next predicted move, and it has to reach the goal across a room full of walls and gaps. Drag from the token to aim a launch vector, set the temperature, then Sample to commit.

Drag from the token to aim direction and power (longer drag = harder launch)
Set the temperature slider — your only real control
Tap Sample to launch the token along the sampled trajectory
Land it on the goal. Bounce off walls. Avoid falling into the void
Fewer samples is a better run. Clear all eight prompts

Low Temp vs. High Temp

At temperature 0 the decode is deterministic: the dotted prediction line is exactly where the token goes — surgical through a tight corridor, but it takes no risks and can get wedged behind a wall. Crank toward 2 and every launch gains random angular spread and velocity variance: now the token can scatter across a gap or rattle out of a trap, but you are gambling on the distribution.

Slop Fact: Temperature literally rescales the logits before the softmax. At 0 it always argmaxes the same boring token; crank it up and the tail of the distribution wakes up, starts hallucinating, and occasionally lands a stroke of genius nobody can reproduce. The gap, meanwhile, will not cross itself.