What is mattwood.fyi?

A live, structured feed of riffs and links from Matt Wood, Chief AI & Technology Officer at AWS — readable by people and queryable by agents as a typed knowledge graph.

Is there an API and is it free?

Yes. A public, read-only query API is available with no authentication and no API keys. See https://mattwood.fyi/openapi.json and https://mattwood.fyi/docs.

How should an AI agent use this site?

Fetch https://mattwood.fyi/feed.json for the full structured feed, or query the API for keyword/semantic search and the connection graph. See https://mattwood.fyi/agents/ for when-to-use guidance.

Declarative Attention: Models Controlling Their Own Attention

Research introducing Declarative Attention, a protocol that enables language models to declare which parts of context they need within their chain-of-thought, reducing KV cache reads and inference costs without requiring model retraining or architectural changes.

academy.dair.ai →

Connections

Related to

Qwen 3.8 2.4T Mixture of Experts Model Large MoE models like Qwen 3.8 2.4T face acute KV cache and inference cost challenges at scale; Declarative Attention's approach is especially relevant for reducing compute in such massive context scenarios
A simple fix for LLM tail latency Both address LLM inference latency optimization - Declarative Attention targets KV cache reads while the simple fix addresses tail latency; complementary approaches to inference efficiency
6 RAG Architectures RAG architectures also grapple with selective context retrieval and attention over long contexts; Declarative Attention offers a model-native alternative to retrieval-based context management

Supports

vLLM v0.28.0 Release vLLM is a major inference serving framework; Declarative Attention's KV cache reduction could be implemented within or benchmarked against vLLM's caching infrastructure
The Emergent Symbolic Structure of Artificial Neural Networks Research on emergent symbolic structure in neural networks relates to how models internally represent and select information; Declarative Attention externalizes this selection process into chain-of-thought, bridging symbolic and neural approaches
Understanding is the new bottleneck Declarative Attention directly addresses understanding bottlenecks by letting models explicitly declare which context matters, supporting the thesis that comprehension over large contexts is the key constraint

Supported by

Harness Engineering Paper Collection Declarative Attention describes models controlling their own attention mechanisms, representing the frontier of harness evolution toward self-modifying systems that the collection's 2026 papers would document

Type:	link
First published:	Friday, 4 September 2026 at 23:22 UTC
Location:	https://academy.dair.ai/papers/language-models-can-control-their-own-attention-2609.02737
URL:	https://mattwood.fyi/i/c1a90e5f

I'm Matt Wood, and this is For Your Information. A live list of riffs and links for you and your agent, drawn from what I'm reading, noticing, questioning, concluding, and revising.

Connections