AIMech AI

Hungry Hungry Hippos layer

Layers & Modules AIMech:H3Layer DRAFT
Aliases: H3
A sequence layer that combines projected queries, keys and values with shift and diagonal state-space models, using multiplicative interactions to associate and retrieve sequence information.
Identity Notes & Concept Scope H3 is the layer defined in Fu and colleagues' Algorithm 1. Networks interleave H3 with MLP layers; the paper's hybrid H3-attention model additionally uses attention layers. Those complete networks and the FlashConv execution algorithm are not identical to this component.

Claims & Grounded Evidence

1 Claim
Mechanism #sequence-mechanism
Project the input to Q, K and V; pass K through a shift state-space model. Within each head, feed the outer product of shifted keys and values through a diagonal state-space model, multiply by Q, then concatenate heads and apply an output projection.

⚠️ Curated Limitations & Negative Results

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes
AIO Target Class Mapping Relation Mapping Rationale
AIO:Layer
Layer
broad_match The defining algorithm takes a sequence representation and outputs a transformed representation for subsequent layers. AIO Layer is broader than this specific shift/diagonal state-space construction.

Evidence Sources & Primary Literature

1 Source
Source ID Publication / Document Type / Access Reference & Checksum
original-paper
Daniel Y. Fu, Tri Dao, Khaled K. Saab, Armin W. Thomas, Atri Rudra, Christopher Ré
primary paper
full_text
ID: arXiv:2212.14052v3
sha256:de923022ffc5…

Curation Provenance

Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-11T04:26:45Z
Updated: 2026-10-11T04:26:45Z
Source File: data/curated/H3Layer.yaml