Percepta unveils Spotlight: decoupling intelligence from infinitely growing memory
Recoil42 · reddit · 2026-10-03
Startup Percepta announced Spotlight, a new architecture that separates computation ("intelligence") from memory: the intelligence module's weights stay fixed while knowledge, skills and working state live in a writable, unboundedly growing memory.
Key claims:
- Memory replaces attention: every token reads from and writes to unbounded memory, but learned indexing means each token touches only a few cells, so access cost doesn't grow with memory size.
- Unlike MoE-style sparsity, Spotlight is arbitrarily sparse — the fraction of memory used per step can shrink toward zero as memory grows.
- Since memory can hold skills as well as facts, the model could gain capabilities without retraining.
It's a claim-stage announcement with two companion blog posts, no third-party evals yet.
Related event: Percepta Unveils Spotlight Architecture Decoupling Intelligence from Memory(2 posts)→
More from Research
- ezyang's DeepSeek-V3 series part 3: roofline analysis for training DSv3 on Hopper — ezyang · 2026-10-03
- 36B Open-Source Robot Model Isaac 0.5 Runs Real-Time Under 50ms Without Local GPU — AkshatS07 · 2026-10-03
- Harmfulness eval pipeline: probes on BeaverTails, training on JBB, OOD eval on ClearHarm — maksym_andr · 2026-10-03
- A robot demonstration can lose its most useful half-second: tracking gaps sabotage manipulation learning — Klutzy_Cap8492 · 2026-10-03
- Hinton: leading researchers now think an intelligence explosion may happen soon — geoffreyhinton · 2026-10-03
- Spotlight architecture scrutiny: attention-style memory may carry a large constant factor — teortaxesTex · 2026-10-03