# PhaseSAE Executive One-Pager & Media Brief
**Author:** Conner Kupferberg (conner@connerkupferberg.com)  
**Publication Date:** September 2026  
**Canonical Repo:** https://github.com/connerkup/phase-sae  
**Whitepaper:** https://connerkupferberg.com/blog/phase-sae-time-division-multiplexed-dictionary-learning/  

### Summary
Sparse Autoencoders (SAEs) resolve polysemanticity in LLMs, but evaluating monolithic 64k–262k dictionaries incurs a catastrophic memory-bandwidth bottleneck during autoregressive inference (B=1): an arithmetic intensity of just 0.5 FLOP/byte, leaving GPU cores >99% idle.

PhaseSAE introduces Time-Division Multiplexing (TDM) to dictionary learning. By streaming P parameter slices against sequential residual errors:
1. Peak active parameter bandwidth drops to 1/P (4.0× reduction).
2. Sequential residual orthogonalization eliminates greedy feature competition.
3. Downstream Cross-Entropy recovery jumps +65.9% (40.51% vs 24.42%) on TinyStories-1M at identical L0=16.0.
4. Explained Variance (R²) jumps +65.7% (0.2913 vs 0.1758).
5. Custom Fused Triton GPU kernel pins residuals in SRAM registers, matching monolithic wall-clock latency while preserving 75% memory savings.
6. Anytime Early-Exit provides dynamic 7–23% latency speedup on low-entropy tokens.