My calculator is a transformer
radarsat1 shows an RPN interpreter compiled into Transformer weights; “2 3 + 2 *” returns 10. The residual stream acts as registers, attention weights are compiler-calculated, while nonlinear MLP logic is still trained. The prototype is 1.1 GB; the key point is calculable attention weights, not a practical calculator.
Why it matters: A RPN interpreter compiled into Transformer weights shows attention computed by a compiler instead of trained.