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.

This paper presents a principled two-stage approach for distilling knowledge from Transformer models into Mamba-based State Space Models, using linearized attention as an intermediate step and achieving near-teacher performance while maintaining the efficiency benefits of SSMs.

A comprehensive guide for platform engineers new to inference, covering how models and GPUs work together, production inference deployment, and tools for building an inference engineering career.

Explores using large language models as judges to perform identity resolution tasks while reducing costs by 75% compared to traditional methods.

Explores how Qwen 3.8 27B can interpret and analyze raw neural network weights without running code, replicating findings from a paper showing frontier models can understand model behaviors by reading weight matrices directly.

LongCat-2.5 is a new version of the LongCat model with improved capabilities for processing long-context information and tasks.

Frontier LLMs can directly read and interpret neural network weights to understand what networks do, and can write or edit weights to optimize network behavior without traditional training loops. This capability, termed weight-space programming, enables tasks like identifying bugs in model weights and improving performance on specific tasks with minimal weight modifications.

A tool for bulk text classification and decision inference optimized for single GPU execution, reading content once and scoring all questions in one pass with OpenAI-compatible API support.

A Lua-based tool that generates decorative Delft-blue floral page margins procedurally for use with LuaLaTeX documents.

A project that compiles an NES CPU emulator into the weights of a Qwen3.5 language model, enabling CPU operations to be performed through model inference.

A philosophical debugging tool that returns questions instead of answers, designed to help developers find their own solutions rather than providing direct fixes.

An AI skill that generates hand-inked, isometric drawings of everyday objects from text descriptions, with interactive 3D models you can rotate to see inside the mechanics.

AWS Lambda has fundamentally changed from a simple, stateless event-driven service into three distinct compute primitives with stateful capabilities, dedicated EC2 instances, and long-running workflow support. The article examines AWS's strategic shift in Lambda's architecture and pricing, moving away from traditional serverless constraints.

Introduces Infinity, a new runtime designed to handle high-concurrency agent workloads. The runtime enables efficient execution of numerous agents simultaneously with optimized resource management.

Dessa is a conversational AI model trained exclusively on speech data, enabling natural dialogue interactions based on spoken language rather than text.

BottleCap AI presents ThinkingCap, an optimized version of Google's Gemma-4-26B-A4B model that reduces thinking tokens by 39% on average while maintaining accuracy. The model achieves up to 69% token reduction in some tasks with minimal accuracy loss of only 0.15 percentage points.

Explores how animations help engineers understand system failures, particularly metastability issues and data loss scenarios, with real-world examples from Postgres and etcd demonstrating how production systems can exhibit these problematic behaviors.

Discusses security measures and best practices for protecting the infrastructure and systems used to train advanced AI models, addressing risks and safeguards in large-scale model development.

Kimi K3.1 Pro represents an advancement in open-source AI intelligence, introducing improvements to the K3 model family for enhanced capabilities and performance.

A developer fine-tuned a 360M-parameter language model to simulate a terminal by predicting screen changes from keystrokes, without an actual shell or filesystem behind it. The model can run vim, tmux, and other programs, with the ability to self-correct and maintain consistent state across commands.

I don't even know how to think about this.

A 35B Mixture-of-Experts agentic model that advances long-horizon problem solving through adaptive capability composition, featuring improvements in task orchestration, verification mechanisms, and scientific discovery capabilities.

OpenStick is an open-source HDMI stick designed for AI agents, featuring open APIs, customizable launchers, and community-built apps with planned Model Context Protocol support for agent connectivity.

Alter Vault is a platform that manages privileged access and credentials for AI agents, providing per-agent identity, policy enforcement, audit logging, and integrations with 140+ services to ensure secure and accountable agent operations.

A framework that optimizes language model agent harnesses by evolving a hierarchy of specialized executable classes using object-oriented design principles, achieving up to 15% improvement over baseline methods on long-horizon task benchmarks.