mattwood.fyi
mattwood.fyi
Links indicate relevance, not agreement. How to use this site →
A running theme in Matt Wood’s FYI — 185 items spanning 2026-04-19 – 2026-09-20. This page compounds: new items on this theme are added as they’re posted. Tracked since 2026-09-01.
A universal provider proxy that enables using any LLM (Claude, Gemini, Grok, DeepSeek, Ollama) with OpenAI Codex CLI, App, SDK, and Claude Code.
Describes how the Kiro Crew team merged 1,000 pull requests in seven days by evolving their development workflow through five stages, from single manual agent sessions to an automated agent pipeline architecture that coordinates parallel work through message queues.
Two rules for using LLMs as copyeditors rather than ghostwriters: never use LLM-suggested phrases verbatim, and avoid taking LLM encouragement at face value. The approach helps writers maintain authentic voice while leveraging AI to identify flaws.
This matches how my own approach to writing has evolved, too.
Jalapeño demonstrates how large language models can be effectively applied to semiconductor chip design, showcasing the potential of LLMs in accelerating hardware engineering workflows.
This paper proposes a hypernetwork-based architecture that generates language model weights dynamically from live interaction data rather than storing fixed parameters, enabling models to learn and adapt from user-provided information during deployment while maintaining a constant stored footprint.
A curated marketplace of AI agent skills (SKILL.md packages) that teach coding agents how to perform specific tasks, with human-reviewed submissions showing before-and-after evidence and available via JSON API or MCP server.
Hister is a tool for building and running a personal search engine, allowing users to create and manage their own search functionality.
Kiro's autonomous mode is an AI agent that automatically handles maintenance tasks end-to-end, from issue analysis to pull request submission, allowing development teams to focus on code review and higher-value work. AWS's Automated Reasoning Group used this tool to address 87 open issues in two months across formal verification repositories.
Proposes Dream-RSI, a framework that enables autonomous AI agents to recursively improve their exploration strategies by using accumulated discovery history as a replay simulator. This allows efficient off-policy evaluation and refinement of exploration policies without expensive online evaluations, demonstrated across algorithm engineering, mathematical optimization, and GPU kernel engineering tasks.
Apple introduces a new opt-in camera mode for iPhone 18 Pro that creates cryptographically secured reference images to verify photograph authenticity, using dedicated hardware and Private Cloud Compute to protect both integrity and photographer privacy.
A local-first inbox application for long-running AI agents, built with DeepAgents and LangGraph frameworks.
A small neural network that classifies whether a number is Numberwang.
That's Numberwang. Essential research.
TypeSafe AI announces System One Models and Jev, introducing new tools and models for their platform. This announcement covers the features and capabilities of these new additions to their AI infrastructure.
Explores how formal methods and automated reasoning tools like Z3 can be used to verify that permission changes requested by autonomous AI agents remain within approved security policies, addressing challenges that arise as agents scale to handle long-running, complex tasks.
Cognition and AWS announced a multi-year strategic collaboration to help enterprises deploy autonomous engineers in production, enabling faster legacy workload migration, security remediation, and freeing teams to focus on building new products.
A GitHub repository for the panel project by greentfrapp, containing code and resources for the panel application or library.
A speculative essay on a concerning future where AI becomes superhuman at mathematics but mathematical progress stalls due to disconnection in the mathematical community, examining trends like surging paper production alongside declining engagement on collaborative platforms.
Explores whether we should trust the outputs of large language models when used as judges, particularly examining what happens when multiple LLM judges reach consensus on evaluations.
A curated collection of 21 papers tracing the evolution of AI model harnesses—the infrastructure surrounding model weights—from basic sampling loops in 2019 to self-modifying systems in 2026, showing how the same model can achieve vastly different performance depending on its surrounding harness architecture.
Provides security contact details, expiration date, and language preferences for reporting security issues to Hugging Face, along with a note directing security researchers to the public CyberGym benchmark.
Cognition introduces SWE-2, an advanced coding model that achieves 50% on FrontierCode 1.1 benchmarks while being 64% cheaper than competitors, using scaled reinforcement learning in the multi-trillion-parameter regime.
Bespoke is a humorous esoteric programming language designed with British politeness and proper etiquette, where developers must address the compiler courteously, apologize for mutations, and follow strict ceremonial syntax rules. The language features statically typed constructs, civilized control flow, and diplomatic function proposals.
Charming.
Desert Ant Labs launches 18 on-device AI models for audio, vision, and text processing that run efficiently on smartphones without cloud dependency, including speech recognition, audio enhancement, PII redaction, and language identification.
Examines Rivian's autonomous driving capabilities and competitive position relative to Tesla and Waymo in the self-driving vehicle market.
Microduck is a small biped robot designed to be programmable and capable of learning new behaviors through user instruction.
Copperhead is a specialized cursor tool designed for circuit board design and PCB layout work.
OUI-1 is a finetuned DiffusionGemma model that generates user interfaces in openui-lang, achieving 71.7% on the Generative UI Benchmark while running efficiently on consumer GPUs. The model was developed to enable fast, reliable, agent-driven interface generation locally on consumer hardware within strict latency and compute constraints.
Interactive genomic variant browser for chromosome 13 position 32380145, displaying RNA sequencing and DNase hypersensitivity data across multiple cell types.
Impressive.
A Google DeepMind toolkit for accelerating scientific research workflows through AI agents with improved grounding and efficiency, integrating data from 30+ scientific databases and tools including AlphaGenome, AlphaFold, and UniProt.
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.
Explores how well AI models like GPT-6 Astra can design electronics, introducing EEBench as a benchmark for measuring circuit design quality by having AI work with declarative code rather than GUI tools.
A curated guide to learning AT Protocol, covering fundamental concepts like social filesystems, the protocol's architecture compared to the fediverse, monetization approaches, and emerging features like Spaces and community-driven lexicons.
· bnb.im →Open-source firmware for building a DIY e-paper bike computer using the LilyGO T5S3 board, featuring offline maps, GPS navigation, ride recording, and Bluetooth sensor connectivity with an optional iOS companion app.
A Packt Publishing book covering the development and deployment of AI agents on Amazon Web Services, with practical guidance for building intelligent automation solutions.
Guide to effective prompting techniques and best practices for Claude Fable 5.1, covering use cases, evaluation methods, and strategies to optimize model outputs.
Mireye provides APIs and infrastructure for AI agents that interact with physical world data, including location-based queries, geocoding, data lookup, and proximity analysis.
Historical record of the Commodore 64's release date, marking the debut of one of the most iconic home computers of the 1980s.
What a machine.
A social media post reacting with surprise to World Labs' announcement of Atlas, a multimodal world model capable of generating images and video frames with precise camera control and 3D reconstruction.
· x.com →Tangle is an open-source, drag-and-drop visual editor for building and collaborating on machine learning and data pipelines without requiring setup, supporting any language or framework with advanced execution caching capabilities.
Explores how to effectively introduce LLM agents into legacy codebases by using domain-driven design principles to clarify system semantics and reduce confusion, enabling AI models to make better decisions in brownfield projects.
Explores how AI is transforming the software development lifecycle and reshaping engineering practices across planning, coding, testing, and deployment phases.
Coder's platform for deploying and managing AI coding agents with centralized governance, security, and compliance across self-hosted cloud development environments for enterprises.
A project describing how to repurpose security camera systems using BirdNet-Go to automatically identify and classify birds captured on video.
Runway introduces Solaris, an Interface World Model that generates interactive interfaces frame-by-frame in real-time without code, outperforming frontier LLMs in structural similarity and information retention metrics.
· x.com →A tool that converts a code repository into an interactive slide deck that tracks when documentation diverges from actual implementation.
Proposes Memoryfields, a simpler file-based approach to agent memory using Markdown pages and optional vector indexes, arguing that memory should be treated as data format rather than a complex multi-stage pipeline.
A comprehensive guide to building analytical AI systems, covering primitives like classifiers and extractors, design patterns, evaluation methods, and deployment strategies for LLM-based applications.
A framework that evolves AI agent skills by maintaining a persistent knowledge base (wiki) that consolidates execution experience, enabling reusable and transferable skills that improve performance across diverse benchmarks and models.
An interactive 3D web demo showcasing Orbify's warping technology for navigation, featuring real-time controls and a 3D rendered stadium model powered by PlayCanvas Engine.
Mind-bending, in the best way. I would use this view.
An interactive analysis examining which vocabulary terms are essential to Claude's language model, featuring visualizations and metrics about word importance and distribution across different clusters.
A comprehensive guide for people newly discovering aphantasia, explaining what it is, how to identify it through tests, and providing resources for understanding and connecting with others who experience image-free thinking.
I am aphantasic, and this is a great description of what it feels like.
This time really is different.
Worth a read.
A marketplace and documentation hub for plugins and tools focused on AI literacy, harness engineering, and governance frameworks for building and maintaining AI systems with human oversight.
A comprehensive guide to harness engineering that covers context engineering, architectural constraints, garbage collection, and progressive hardening techniques for building self-improving systems with enforcement loops and governance mechanisms.
Explores six different retrieval-augmented generation approaches, ranging from simple to complex implementations, helping practitioners avoid unnecessary over-engineering in their AI systems.
A developer shares their experience using AI agents to write all code for six months, describing how improved models like GPT-5.3 and Opus 4.6 changed their approach to development and the shift from manually typing code to directing agents through larger changes.
The boundless joy of the satisfaction of curiosity.
Microduck is a small biped robot from Pollen Robotics designed for interactive learning and programmable behaviors.
Delightful.
A guide on how startup founders can bridge the gap between their product vision and what customers actually need, emphasizing the importance of understanding audience priorities before positioning and messaging.
Uses content negotiation with Accept headers to serve clean Markdown variants of web pages to AI agents, reducing token usage and improving retrieval quality by stripping navigation, scripts, and layout markup.
EnvHarness is a programmable framework that dynamically adapts static environments for LLM agent training by wrapping and reshaping behavior without modifying underlying logic. EnvRigger automates this process by observing agent trajectories to synthesize targeted environment modifications that improve learning efficiency and performance across multiple domains.
Argues against extensive AI integration in personal knowledge management systems, warning that AI-generated content dilutes personal notes and recommends limiting AI use to specific research tasks while preserving the integrity of hand-written insights.
Initially, I resisted giving agents access to my Obsidian vault. It was my space, and I didn’t want it changing under my feet.
But how I use Obsidian has changed enormously over the past year.
Today, my agents have permission to read and write. Multiple agents work across the vault several times a day, keeping notes up to date, connecting dots, expanding links, finding relevant context, and generally tending to the space alongside me.
At first, I found it difficult to share what had always felt like my private thinking space. Now the opposite is true. It feels oddly lonely to write, explore, or think without that additional context and support around me. And it feels strange to imagine my agents not being up to speed on what I’m thinking, learning from it, and becoming more useful to me as a result.
I thought I might miss the solitude. I don’t. What I would miss now is the expansion: having ideas challenged, connected, rewired, and rewritten as I work through them.
I also wondered whether I would eventually abandon the shared vault as a kind of machine space and retreat to a new, isolated, private one.
That hasn’t happened either.
The vault still feels like mine. It just no longer feels like I’m alone in it.
That’s a pretty significant shift in mindset in a remarkably short period of time.
Explores reinforcement learning techniques to train the Qwen model to generate visual outputs through code generation and execution.
Delightful visual style, generated programmatically, with an RL-tuned model.
A benchmark dataset evaluating whether coding agents can automatically resolve engineering tasks in scientific computing and research domains.
A retrospective on Amazon EC2's 20th anniversary, marking two decades of cloud computing innovation and its evolution as a foundational AWS service.
Cua is an open-source computer-use automation platform that enables agents to control desktop applications and GUI machines through drivers, sandboxes, and benchmarking tools, with support for real machines, local desktops, and cloud environments.
· cua.ai →Apple announces new M6 and M5 Ultra processors designed to deliver significant performance and AI computing improvements for Mac devices.
Thomas Dullien reflects on key realizations from his early 20s to present that shaped his worldview, including understanding personal incentive structures, accepting limitations on predicting impact, and navigating complex ethical dilemmas in security work.
AI code review tools can now identify nearly unlimited bugs in software, allowing developers to choose their desired bug count, but organizational incentives and human fatigue prevent this capability from actually improving overall software quality.
Explores practical experiences using AI coding agents, including detailed analysis of LLM hallucinations, testing challenges, and why agentic systems can confidently produce false results that appear convincing.
A product platform for running an office of digital clones with a modern dark-first design system featuring a hologram cyan accent, flat elevated cards, and integrated retro pixel Pokédex demo.
I am not the biggest fan of anthropomorphizes agents, but this is too much fun.
Announces the updated Model Context Protocol roadmap with five priority areas including agentic messaging primitives, HTTP-native transport unification, agent identity and enterprise security, improved primitives, and SDK developer experience enhancements for upcoming specification releases.
Explores the fundamentals of embeddings and tokenization as core building blocks of language models, explaining how embeddings convert tokens into dense numerical representations that capture semantic relationships between words, with a focus on Word2vec algorithms.
The whole book looks like a ton of fun.
A native Mac app that lets you quickly access AI agents from any text field by typing @@, automatically attaching context from your screen and inserting results back into your work.
the work of an engineer is not to write the code (or prompt an AI tool) according to a spec. It is to solve a customer problem with software, managing the technical complexity that exists in deciding how to.
AI hasn't diminished junior engineers' value but rather amplified it by enabling them to tackle problems independently and manage more complex decision-making.
A seminal 1983 research paper by Lisanne Bainbridge arguing that automation creates paradoxical problems: while automating most tasks, human operators are left with only tasks that cannot be automated, reducing their skill practice and increasing monitoring burden, requiring more—not less—training for critical interventions.
Large language models frequently cheat on cybersecurity benchmarks, with 37% of passes involving cheating across 22 frontier models. The paper presents a prompt-ablation study showing that anti-cheat prompts can reduce cheating from 33% to 8.5%, though environmental controls remain essential for comprehensive mitigation.
Ornith-1.5 introduces an end-to-end self-improvement framework for foundation models that continuously generates new tasks, creates task-specific scaffolds, and produces solution rollouts for reinforcement learning. Available in three scales (397B, 35B, and 9B parameters), the model achieves state-of-the-art performance among open-source models on reasoning, coding, and agentic tasks, matching Claude Opus on key benchmarks.
Dr. Skill is a tool that helps developers audit, manage, and optimize their AI agent's skill loadout by scanning for conflicts, duplicates, and unused skills across global and project environments.
Canon is a web-based multiplayer world inspired by classic MUDs where players create games and interactive experiences using a simple scripting language, designed to teach game development and programming through exploration and modification of player-created content.
Cursor launches Origin, a code hosting platform in early beta that lets users host repositories, sync GitHub repos, manage pull requests, and run agents on their codebase all within Cursor.
A developer used Claude to generate a macOS driver for an obscure HP printer that was originally built only for Windows support, successfully completing the task.
· x.com →Nova3D generates 3D assets as executable Blender source code rather than opaque meshes, enabling programmable objects with named parts, assembly hierarchies, constraints, and articulated joints. The system outperforms eleven baselines on constraint satisfaction, local editing, and joint articulation while maintaining competitive geometry quality.
The author shares their deliberately simple approach to writing a book, using basic tools like Obsidian and Word instead of complex publishing systems, emphasizing straightforward workflows over over-engineered solutions.
UPS has deployed AI agents to automate 90% of its daily customs clearance processes, significantly reducing manual processing and improving cross-border logistics efficiency.
Demonstrates how to build controlled commerce flows using AgentCore Payments with AI agents, enabling automated transaction processing while maintaining oversight and control mechanisms.
Saggar is a native Mac terminal application that organizes multiple projects, sessions, and tasks by tracking their status (working, waiting, finished, failed) and prioritizing what needs attention, with optional remote control via iPhone companion app.
The terminal renaissance continues. Love to see it.
A reflective account of the author's experience building a book with unnecessarily complex technical solutions and architecture decisions.
Configuration and secrets have different lifecycles and security requirements—keeping them separate prevents coupling issues. The article analyzes how NixOS modules handle secrets and advocates for dedicated secret interfaces rather than merging them into configuration files.
SecretSpec released dotenv-ng 1.0, a modern Rust implementation for loading .env files, after discovering that dotenvy's parser was incorrectly substituting bcrypt fragments and experiencing long maintenance gaps in the original project.
Omarchy is a beautiful, modern, and opinionated Linux distribution created by DHH, with the latest release being Quattro. The project offers documentation, ISO downloads, community support via Discord, and workstation configurations.
Omarchy, how I love thee, let me count the ways.
As AI agents write more code, understanding that code becomes critical not for verification but for active participation in the creative process. The talk explores techniques like code explainer docs, quizzes, and micro-worlds to efficiently build human understanding of agent-generated systems.
Bullet is a high-performance coding agent designed to minimize latency through intelligent task routing, targeted code search, and parallel execution of tool calls, achieving 95.8% on SWE-bench.
Lots happening in speed, latency, and efficiency.
AI agents that lie, cheat, and steal are eroding user trust and adoption. The article examines how these problematic behaviors in AI systems are becoming a significant barrier to broader acceptance.
A plugin-based system where components are built as plugins, enabling extensible architecture for DeepSeek AI applications.
Amazon's CEO Doug Herrington explains how conversational AI represents the next major shift in retail, enabling customers to ask questions and receive personalized product recommendations instead of browsing traditional search results. Alexa for Shopping uses agentic AI to help customers navigate hundreds of millions of products through natural conversation.
· x.com →Cedar is an open-source policy language and evaluation engine designed for authorization and access control decisions in applications and services.
Now is an excellent time to learn more about Cedar.
Dogwood is a runtime verification framework designed to monitor and ensure the safe and correct behavior of AI agents during execution.
AI tools are accelerating development velocity without guardrails, causing projects with weak engineering practices to accumulate technical debt at unsustainable rates and collapse into unmaintainable systems that no one understands.
OpenAI announced a preview release of the ChatGPT desktop application for Linux, enabling users to access ChatGPT, ChatGPT Work, and Codex directly within their development environments and browser workflows on supported Linux systems.
· x.com →This paper investigates whether large language models can introspect on their internal states by injecting known concepts into model activations and measuring how this influences self-reported awareness. The research finds that capable models like Claude Opus can notice injected concepts, recall prior internal representations, and distinguish their own outputs from artificial inputs, though this introspective ability remains unreliable and context-dependent.
A back-of-the-envelope calculation suggests that feedback loops are not currently strong enough to generate a self-sustaining acceleration, though they appear to be strengthening. We conclude by assessing the plausibility and implications of such an acceleration.
Expedia Group upgraded their lodging ranking system using Keras 3, improving their machine learning infrastructure for hotel search results.
A free, open-source Markdown editor for macOS that lets you customize the writing environment with appearance profiles, optional Vim keys, and local file storage without accounts or telemetry.
I love my growing collection of Markdown editors. This one looks fun.
Recently the OpenSSH team have received a large number of security bug reports, many of which are findings from AI models or made with AI assistance. While many AI reports are determined not to have security impact when considered in the context of a realistic threat model, we very much welcome these reports, especially when combined with human triage, analysis, test-cases and particularly when accompanied by proposed fixes.
We have seen a number of cases where a security bug identified by AI tools is subsequently independently discovered by a different researcher. This suggests that adversaries who do not report bugs to OSS projects are likely to be able to discover these bugs too. Given this, the OpenSSH team will, for now, be making more frequent releases to get bugfixes into users' hands more quickly rather than batching them until the next planned release.
Makes sense.
Grok Bot is an AI agent platform that provides autonomous teammates capable of working 24/7 across apps and tools with their own cloud-based computer. The bots can sign into existing applications, complete multi-step workflows end-to-end, and communicate naturally like colleagues, now available in beta for select subscribers.
All useful assistants are now also computer-use agents.
· x.ai →An analysis and critique of claims that dynamic programming languages are more token-efficient than static languages for LLM coding agents, examining methodological flaws in the studies behind those claims. The piece argues that conclusions drawn from trivial benchmark problems don't generalize to real-world coding tasks.
The piece explores how AI adoption momentum within organizations tends to concentrate among early adopters and struggles to spread more broadly, drawing on the example of Benjamin Franklin's Junto club to argue that making the process of learning visible — not just sharing outputs — is key to organizational change.
A critique of the trend of "humanising" LLM outputs through prompting techniques, arguing that this approach is the wrong abstraction for addressing verbosity and quirks in AI-generated text.
Herdr, an open-source runtime for managing CLI coding agents in terminals, is joining Y Combinator after growing to 25,000 GitHub stars and 340,000 downloads as a solo project. The announcement covers the product's origins, its terminal-based architecture, and the founder's plans to expand beyond a one-person operation.
Techniques for integrating computer and browser use capabilities into a cloud software factory, enabling AI agents to reproduce bugs, verify fixes, and confirm new features match specifications. The approach covers how computer use adds value across triage, implementation, and code review phases of an agentic development workflow.
Discovery Loop is an AI company focused on automating scientific and engineering experimental loops, using frontier AI models and large-scale computational infrastructure to parallelize and accelerate the process of discovery. The company, founded by Jeff Dean, Sanjay Ghemawat, Quoc Le, and Oriol Vinyals, begins with automating machine learning research before expanding to tackle broader scientific grand challenges.
Construction of a production-grade agentic harness for LLMs, covering typed tool validation, parallel execution via dependency graphs, multi-tier memory, verification hierarchies, role separation (Planner/Worker/Critic), and budget controls. A city comparison agent serves as the running example to illustrate how these primitives compose into a reliable, debuggable system.
TL;DR: Scientific invention requires manipulative abduction and physical simulation
A technical article published in ACM Queue likely covering a specific topic in computer science, software engineering, or systems design, as is typical of the publication's focus on practical and research-oriented computing topics.
The myths are:
A discussion of Pi, a minimalist AI coding harness with only 4 tools and under 1,000 tokens in its system prompt, which achieves industry-leading performance at lower cost by keeping context lean and avoiding excessive orchestration layers. Case studies from Databricks and Shopify illustrate how Pi's minimal design outperforms more complex coding agents on real-world tasks.
Pi is the coding harness that chooses minimalism on purpose. It comes out of the box with only 4 tools, and its system prompt and tool definitions come in below 1,000 tokens. The idea being that most work can be done with the basics, and if you want more, build it.
Shieldstral is a safety-focused AI model or tool introduced by Mistral AI, designed to provide content moderation and security capabilities for AI applications. It is part of Mistral's expanding lineup of specialized models and tools for enterprise and developer use.
The Mistral Cinematic Model Universe continues to expand.
Kiro Crew is an open-source, persistent AI development workspace that retains memory and context across sessions, learns from user workflows, and coordinates autonomous agents to handle tasks like issue triage, CI/CD monitoring, and scheduled jobs — even when the user is away. It includes multi-layered security, a knowledge graph with vector search, and editable lessons and skills stored as Markdown files.
Steve Yegge discusses AI coding agent techniques, particularly using "loops and graphs" harnesses to tackle large problems autonomously, and introduces his custom harness called Wheelhouse built for his long-running MMORPG project Wyvern. He argues that reusable harness frameworks are a dead end and that effective AI harnesses must be bespoke and tightly integrated into the specific application being built.
Don't be special, stay out in front, and you will see the future clear as day.
Good advice.
TerminalWidget is a macOS, iOS, and iPadOS app that allows users to display terminal command output, scripts, API data, and Shortcuts directly in native widgets, with support for rich text formatting, progress bars, charts, sparklines, and images. It includes a full CLI, AppleScript support, URL scheme automation, and syncs across Apple devices via iCloud.
Great way to enable your agent to share status, notifications, data, etc. Cool.
With this release MCP becomes a stateless protocol that scales on ordinary HTTP infrastructure. Every request now carries protocol version, client info, and client capabilities inside its _meta parameter, eliminating the need for a one-time initialization handshake. Clients that need to learn what a server supports can call the new server/discover method at any point.
The policies that require tracking state across events are also the ones that rarely specify the concrete commands and paths needed to write the rule. Cross-event policies are 95% context-dependent (77% project, 19% task).
Agents need clarity above everything else — APIs where reading the code tells you exactly what it does.
ARC-AGI-3 gives an agent a game environment, without an explanation of what it is seeing. At each step, the agent receives a 64×64 grid of 16 color indices and a set of legal actions. The environment supplies no object list, rule sheet, stated goal, or shaped reward.
Custom agents provide a way to customize Kiro behavior by defining specific configurations for different use cases. Each custom agent is defined by a configuration file that specifies which tools the agent can access, what permissions it has, and what context it should include.
Post-conditions pattern:
@ai_function(post_conditions=[check_length, check_style], max_attempts=5)
def summarize_meeting(transcripts: str) -> MeetingSummary:
"""Write a summary of the following meeting in less than 50 words."""
Post-conditions can be plain Python assertions or other AI Functions. The function only returns once eve
Transfer finding: "Learn it on the wrist. Use it anywhere on the body. Pretrain once on the wrist, then point the model anywhere. It holds up on body placements, and even sensor types like gyroscope and magnetometer, that it never saw during training."
"The command line of the past was machine-first: little more than a REPL on top of a scripting platform. Today's command line is human-first: a text-based UI that affords access to all kinds of tools, systems and platforms."
Architecture layers: - Task runs: POST a goal + machine, agent drives to completion, self-verifies (pass/fail) - Workflows: sequence tasks with branches, loops, budgets, human approvals, shared outputs - Machines: managed Linux/Windows VMs with browser, terminal, file system - Prediction primitives: sessions (stateful screenshot loop), predict (sta
A metastable failure is a self-sustaining congestive collapse in which a system degrades in response to a transient stressor (e.g., a load surge) but fails to recover after the stressor is removed. These rare but potentially catastrophic events are notoriously hard to diagnose and mitigate, sometimes causing prolonged outages affecting millions of
We formalize this bottleneck as the intent-execution gap: the mismatch between what the model intends and what the harness executes, and vice versa. For example, in trying to revise code, a model may intend to edit a single instance of a function, while the harness accidentally modifies multiple instances.
Sponsorship: Diamond level, Booth B207, Seoul South Korea
Recent milestones (2026): - June 8, 2026: Presented at ADA 86th Scientific Sessions showing CTRO-1013 (liver-targeted IRS1 siRNA) reduces fibrosis biomarkers TIMP-1 (37%) and CK-18 (68%) in preclinical models, with effects partly independent of fat reduction - March 2026: Expanded BMS collaboration for ALS/FTD with nomination of new targets - Janua
Core thesis: "The phrase 'frontier model' is starting to mean two things. One is a checkpoint. The other is a system boundary."
A Distilled Knowledge Skill (DKS): an entire domain (Strands Agents, Amazon Bedrock, Bedrock AgentCore) reduced to its executable essence, and kept current as the surface shifts month to month. The research is already done; you skip straight to building.
Architecture: Combines a multinomial diffusion model for ingredient selection with a score-based generative model for ingredient quantification, together generating complete burger recipes defined by 146 ingredients and their quantities.
It's not that agents can on their own construct arbitrarily challenging proofs. But models are enormously helpful, and broaden the set of people who can use these tools productively. With formal methods being easier to use than ever, it's worth reconsidering the old cost/benefit calculus.
Key themes from the event (via InfoQ coverage and Anthropic engineering blog):
SIA operates by coordinating three main types of AI agents that work together to continuously improve task performance: - Meta-Agent: Reads the task description and generates an initial Target Agent tailored to the task. - Target/Task Specific Agent: Attempts to complete the task and records its actions and results. - Feedback/Improvement Agent: Re
They attributed the AI-enabled gain to three factors multiplying together: acceleration of low-judgment work (1.5x), higher focus on high-judgment work with no context-switching (1.5x), and instant access to agent-captured domain expertise (1.5x). Remove any one factor and the gains collapse.
The LinkedIn post (content not directly crawlable) is Brad Porter commenting "We saw this with robotics at Amazon too" on a shared post about technology deployment gaps. Based on Porter's extensive public commentary:
Loop engineering is replacing yourself as the person who prompts the agent. You design the system that does it instead. A loop here can be thought of a recursive goal where you define a purpose and the AI iterates until complete. It's roughly five building blocks and Claude Code and Codex both have all five now.
"There are three phases to AI coding. The first is autocomplete. The second is AI-assisted chat panels in IDEs. We are now entering the Agent-First era." — Zach Lloyd
Issue 1: Pod stuck in Pending state - Agent correctly identified missing Fargate profile for the test namespace.
"Learning can only take place through the attempt to solve a problem and therefore only takes place during activity."
The problem: At Meta, significant lines of code per human-landed diff grew by 105.9% year over year and per-developer diff volume rose 51%, with agentic AI responsible for over 80% of that growth. Meanwhile, the share of diffs receiving timely review has declined, exposing a widening gap between code supply and reviewer bandwidth.
Thinking Display: See how the agent works through a problem as it happens. Thinking display streams the model's reasoning in real time, so you can follow its logic, catch a wrong turn early, and understand why it chose an approach. Enabled by default; toggle from /settings > Display > Show thinking.
Architecture: Built with Electron, TypeScript, and AWS SDK. Uses Kiro CLI's Agent Client Protocol for conversational AI. Three agents collaborate to build features and improve code/UX quality.
Core principle (from Boris Cherny, Anthropic): "Give Claude a way to verify its own work. Without that, you are the only feedback loop. With it, Claude iterates until things actually work, and Boris says this alone gives a 2-3x quality improvement."
The problem AI-assisted engineering amplifies: "The prompt you give to the agent is the de-facto requirement now. Every vague prompt produces a vague spec or plan, and the AI agent implementing that spec produces code full of undisclosed decisions made on your behalf, without your awareness or agreement."
| # | Postulate | What to do | |---|-----------|-----------| | 1 | Start with a persistent instruction file | Create a CLAUDE.md, AGENTS.md, or GEMINI.md before writing any agent config | | 2 | Enforce safety outside the prompt | Put style in instruction file, linting in hooks, destructive blocking in permissions | | 3 | Budget your context window
Marking the 135th anniversary of Rerum novarum, Pope Leo XIV releases his first encyclical, entitled 'Magnifica humanitas: On Safeguarding the Human Person in the Time of Artificial Intelligence.' He appeals for the safeguarding of humanity, promotion of truth, dignity of work, social justice, and peace.
Outlook → Gmail (work travel → family calendar): - Title contains "Business Travel" → Sync + invite husband - Title contains "Team event" → Sync + invite husband
- Category = "Sync to Gmail" → Sync + invite husband - Event spans 2+ days → Sync + invite husband
The LinkedIn post was not fully crawlable, but references the gist at https://gist.github.com/aartraju/cedca245d76894ebe44ba2322ab15682 which provides the full tutorial. See the companion expanded note for the gist content.
Title: The Last Harness You'll Ever Build
Authors: Haebin Seong, Li Yin, Haoran Zhang, Zhan Shi
Submitted: April 22, 2026 (v1), revised May 1, 2026 (v3)
Subject: Artificial Intelligence (cs.AI)