Last week, Ghostty—the terminal emulator from HashiCorp co-founder Mitchell Hashimoto—published an AI policy that’s getting attention. The rules: disclose your AI use, only submit PRs for accepted issues, test everything yourself, and (here’s the memorable part) “bad AI drivers will be banned and ridiculed in public.”
The policy isn’t anti-AI. Ghostty itself was written with plenty of AI assistance. The problem, as the doc states: “It’s the people, not the tools.”
Here’s what’s actually happening: AI slop PRs are a real problem, but not because AI generates slop. Humans generate slop. AI just makes it cheaper to produce at volume.
The Hacker News thread crystallized something important. One commenter nailed it: people submitting AI slop often “have no perspective to put that shame.” They’ve never honed a craft, never learned the humility that comes from doing something badly for years. The AI gives them what feels like competence, and when maintainers push back, they take it personally.
Another pointed out this isn’t new—people have always submitted garbage PRs to pad their GitHub profiles. The difference now is volume. What used to require copying from Stack Overflow and manually tweaking code can now happen in seconds. The economics changed; human nature didn’t.
But here’s the thing: AI coding has crossed a threshold. The code coming out of modern tools often works—sometimes better than what junior developers write under deadline pressure. When someone with architectural judgment uses AI, you get solid implementations that solve real problems. When someone with no judgment uses AI, you get working code that solves the wrong problem. Same as before, just faster.
The issue-to-PR workflow Ghostty enforces makes sense precisely because implementation is becoming commoditized. Humans make product decisions—what’s worth building and why. The code underneath? As long as it’s clean, tested, and maintainable, the origin story matters less. Show me you understand what needs to be built. The “how” is increasingly the easy part.
What’s emerging is a trust problem, but not about AI capability. Remote contributions now get more scrutiny. New repos look suspicious by default. Maintainers are becoming unpaid QA for bad prompts. The phrase “AI-assisted” triggers skepticism instead of curiosity.
The irony is that this is happening as AI tools get genuinely powerful. We’re building better autonomous code generation while simultaneously dealing with a flood of low-judgment contributors who finally have the tools to expose how little they understand. Projects are learning to filter for judgment instead of implementation skill—which is probably where we should have been all along.
AI doesn’t generate slop. Humans do. The tools just made it visible at scale.
This Week’s Repos
Big Movers (20k+ stars)
remotion — 28.2k stars - Make videos programmatically with React. Leverage CSS, Canvas, SVG, WebGL—all the web tech you already know. Reusable components, fast refresh, the whole React ecosystem applied to video creation.
goose — 27.5k stars - From Block (formerly Square). An open source AI agent that goes beyond code suggestions—it installs, executes, edits, and tests. Works with any LLM. The “on-machine AI agent” that can automate complex dev tasks from start to finish.
Ciphey — 20.8k stars - Automatically decrypt without knowing the key or cipher. Feed it encrypted text, it figures out the rest. CTF tool that’s been around but keeps growing. C++ core for speed, supports encryptions and hashes that CyberChef Magic doesn’t touch.
mastra — 20.3k stars - Just hit 1.0. From the team behind Gatsby—a TypeScript framework for building AI-powered apps and agents. Model routing across 40+ providers, integrates with Next.js/Node. The “batteries included” approach to AI application development.
Agent Frameworks (the crowded space)
PocketFlow — 9.7k stars - The minimalist counter-play: 100 lines, zero dependencies, zero vendor lock-in. Let AI agents build agents. When you want the opposite of framework bloat.
Tambo — 4.6k stars - Generative UI for React. Register components with Zod schemas, AI decides what to render based on conversation. Two modes: one-shot “generative” components (charts, summaries) and persistent “interactable” ones (carts, boards). MIT licensed.
openagents — 1.6k stars - Agent networks for collaboration. Protocol-agnostic (WebSocket, gRPC, HTTP, libp2p, A2A). Mod-driven architecture where agents can collaborate on wikis, docs, games. The “bring your own agents” philosophy.
nuwax — 514 stars - Calling itself an “Agent OS”—workflow builder, plugins, RAG knowledge base, MCP integration. No-code approach for building vertical-specific agents. Chinese/English docs.
vm0 — 193 stars - Run natural language workflows in isolated containers. Cloud sandbox for Claude Code or Codex agents. Session persistence, observability, skill integrations (GitHub, Slack, Notion).
DevOps & Infrastructure
hummingbot — 15.7k stars - Crypto trading bot framework. Design and deploy automated strategies across 140+ exchanges. $34B in user-generated trading volume last year. Market making, arbitrage, backtesting. Apache 2.0.
firmware — 4.6k stars - “Predatory ESP32 Firmware” for M5Stack Cardputer and similar devices. NFC, RF, embedded security tools. The Flipper Zero energy but for ESP32 hardware.
useSend — 3.7k stars - Open source Resend/Sendgrid alternative. Built on AWS SES with a nice DX layer on top. Email editor, delivery tracking, schedule API. One-click Railway deploy. Currently in beta.
Databuddy — 891 stars - Privacy-first analytics. GDPR-compliant out of the box, no cookie banners needed. Multi-tenant support, funnel analysis, custom events. The “self-hosted GA alternative” in a crowded space, but built on Next.js 16 and React 19.
portabase — 246 stars - Database backup & restore for PostgreSQL, MySQL/MariaDB, MongoDB. Simple problem, clean solution. CLI or Docker, pairs with their agent for remote operations.