SketchSSM: approximate reads not the state, avoiding compounding compression errors
sehoonkim418 · x · 2026-10-08
The author's principle: keep the state exact and approximate only reads. Once per window, the full state is updated and a compact sketch refreshed in the same kernel; within the window every step reads only the sketch, so no compression error is ever written into the state. In contrast, shrinking the state via pruning or quantization compounds errors each step — even 2–4× less traffic sharply hurts long-reasoning accuracy.
More from Infra
- Chrome's new DecisionModel API reverse-engineered: prompts, limits and engine tests — dejanseo · 2026-10-08
- China's Power Glut Meets Data Centers; Immersion Cooling Traced to Bitcoin Miners — teortaxesTex · 2026-10-08
- omarchy-cluster runs the full 753B-param GLM-5.3 across four old Macs as one endpoint — natesiggard · 2026-10-08
- Only Samsung HBM meets Nvidia Vera Rubin performance requirements, per leak — zephyr_z9 · 2026-10-08
- Box CEO on agent compute: one app serving 100M users would need $2.8B in infra — inductionheads · 2026-10-08
- Firmus, valued near $44bn, may shelve ASX IPO as investors balk — nordicinst · 2026-10-08