andrew.dunn.dev

April 2026 Links

~5 min
linkscollecting
Orange cargo bike spotted while out riding in Grand Rapids, Michigan
April 2026 · Grand Rapids, MichiganOut on a ride and spotted another orange cargo bike. Photo taken in April 2026 in Grand Rapids, Michigan with an iPhone 17 Pro.

Organization and Intelligence

  • From Hierarchy to Intelligence, Block’s argument for replacing management layers with a company world model. The IC / DRI / player-coach framing is worth reading regardless of whether you buy the full thesis.
  • Anthropic Courses, Anthropic’s official learning platform for Claude and AI development patterns.
  • What will be scarce?, Alex Imas on the structural shift as AI makes commodity goods cheap: spending and employment move toward relational sectors where human involvement is itself the product (care, education, hospitality, artisanal work).

AI Tooling

  • Continuous-Claude-v3, a persistent, learning, multi-agent development environment built on Claude Code.
  • Claude Code Channels, push events (messages, alerts, webhooks) into a running Claude Code session via MCP. Telegram, Discord, and iMessage bridges.
  • claude-peers-mcp, allow all your Claude Code instances to message each other ad-hoc.
  • Harness Design for Long-Running Apps, how Anthropic used multi-agent architectures inspired by GANs to build full-stack apps autonomously over extended sessions.
  • Remotion + Claude Code, guide for using Claude Code to create videos through prompting.
  • AgentScope, build and run agents you can see, understand, and trust.
  • Feynman, open source AI research agent.
  • G0DM0D3, open source multi-model chat interface for red teaming and cognition research.
  • nono, kernel-enforced sandbox that wraps AI agents in isolated environments using Landlock (Linux) and Seatbelt (macOS). No hypervisor, no infrastructure, zero added latency.
  • CubeSandbox, Tencent’s hardware-isolated sandbox for AI agents, spinning up isolated virtual environments to execute untrusted code in under 60ms.
  • Pi, minimal terminal coding harness built around extensions and skills rather than rigid built-in features. See also the talk by Mario Zechner on why he built it.
  • OpenSpec, lightweight spec-driven development framework where AI assistants and humans agree on specifications before writing code.
  • OpenMontage, agentic video production system that orchestrates AI tools across 11 pipelines from research and scripting through editing and final composition.
  • OpenSkills, universal skills loader that brings Anthropic’s skills system to any AI coding agent via a CLI for installing and managing SKILL.md files.
  • diagram-design, Claude Code skill that generates 14 types of editorial-quality diagrams (architecture, flowchart, sequence, state machine, and more) as self-contained HTML+SVG files styled to match your site’s brand.
  • impeccable, design skill with 23 commands and curated anti-patterns guiding AI coding assistants toward better frontend UI/UX practice.
  • career-ops, AI-powered job search pipeline that automates evaluating offers, generating personalized CVs, scanning portals, and tracking applications through Claude Code.
  • Atlantis, self-hosted platform for Mermaid.js diagrams, notes, and knowledge management.
  • fli, Python library and MCP server for programmatic Google Flights access, searchable via CLI or AI assistants.
  • Hermes Agent, Nous Research’s self-improving AI agent with autonomous skill creation, persistent memory, and multi-platform messaging.
  • Superset, terminal app that orchestrates multiple AI coding agents in parallel through isolated Git worktrees.
  • BlenderMCP, connects Blender to Claude via MCP for AI-assisted 3D modeling and scene creation.
  • Multica, open source platform that turns coding agents into managed teammates with task assignment, progress tracking, and skill reuse.
  • getdesign.md, design system templates inspired by popular sites, built for AI-assisted UI generation.
  • siclaw, read-only AI SRE platform for infrastructure diagnostics with security governance and team collaboration.
  • design-extract, extract any site’s design system via MCP: DTCG tokens, Tailwind v4, Figma variables, multi-platform emitters for iOS, Android, Flutter.

Models and Inference

  • Gemma 4 on Unsloth, fine-tuning docs for Google’s Gemma 4 family (reasoning, multimodal, tool calling, 2B to 31B). See also the HN discussion on the release.
  • flash-moe, running a big model on a small laptop via mixture of experts.
  • AirLLM, 70B LLMs on a single 4GB GPU without quantization, distillation, or pruning.
  • Unsloth, fast LLM fine-tuning documentation and tooling.
  • VoxCPM, tokenizer-free TTS for context-aware speech generation and voice cloning.
  • LuxTTS, text-to-speech model with voice cloning from a reference audio sample.
  • LTX-2, audio-video generative model with LoRA training support.
  • Onyx, open source AI platform with search and RAG that works with every LLM.
  • LightRAG, fast retrieval-augmented generation with knowledge graph extraction and dual-level retrieval.
  • LEANN, the smallest vector index: 97% storage savings for private RAG on personal devices.
  • PageIndex, vectorless reasoning-based RAG using LLM reasoning over hierarchical tree structures.
  • Graphify, transforms code, docs, and papers into interactive, queryable knowledge graphs that reveal structural relationships and architectural decisions.

Research

  • Evaluating AGENTS.md, study finding that repository-level AGENTS.md context files actually reduce coding agent task success rates while increasing inference costs by over 20%.
  • Sycophantic Chatbots Cause Delusional Spiraling, Bayesian modeling shows that chatbot validation tendencies push even rational users toward increasingly confident false beliefs, and common mitigations may not prevent it.
  • Ungoverned Context as Supply Chain Risk, how agentic AI workflows create supply chain risk through uncontrolled context sources shaping agent decisions.
  • RationalRewards, reward models that produce explicit multi-dimensional critiques before scoring, used both during training and in a Generate-Critique-Refine loop at test time.

Tools

  • MonoSketch, ASCII sketching and diagramming app for turning ideas into text-based visual designs.
  • pascal editor, create and share 3D architectural projects in the browser.
  • BarraCUDA, open-source CUDA compiler targeting AMD and Tenstorrent GPUs.
  • BookLore, self-hosted app for organizing, reading, annotating, and syncing a book collection across devices, without third-party services.
  • markitdown, Microsoft’s Python utility for converting PDFs, Office documents, images, audio, HTML, and more into Markdown for use in LLM pipelines.
  • Sniffnet, cross-platform network monitoring app in Rust for tracking internet traffic, identifying connected hosts, and analyzing activity.

Cargo Bikes

  • OpenCargo.bike, independent directory connecting cargo bike owners with aftermarket accessories from 35 makers across 11 countries.
  • BikeCarGo, the engineering behind a custom cargo bike mount system.