Learning to add
Explore modular addition (a + b) mod 113. A small linear model learns from products of sine and cosine features in your browser. The original transformer and its recorded activations are unavailable.
━ training pairs right ━ unseen pairs right ━ clock alignment
CLOCK 1 · FREQUENCY 8 · CLEAN 0.00
CLOCK 2 · FREQUENCY 21 · CLEAN 0.00
CLOCK 3 · FREQUENCY 32 · CLEAN 0.00
a sum in use
LOCAL TRAINING
94 + 78 = 59 (mod 113)CURRENT CHECKPOINT
Illustrative activation patterns