Introduction

AI-assisted coding is incredibly convenient, but those lazy type definitions that "seem to work" but lack actual type safety can drive you crazy. This week, anti-slop, which rocketed to over 4,100 GitHub stars in just 30 days, was born to solve exactly this! I'm Muyan. Today, besides sharing this TS gatekeeper, I'll also introduce a piece of black magic that lets a 16GB Mac run a 10-billion-parameter LLM, and a massive update to a classic video editor. Stay tuned to find out how they did it!

---

Open Source Picks of the Week

1. anti-slop: The Oxlint Gatekeeper Curing Lazy AI-Generated Code

• What it is: Everyone is heavily using AI to write TypeScript code these days, resulting in a bunch of lazy types that "look like they work but lack evidence." This tool is an Oxlint ruleset specifically designed to catch these lazy AI-generated types ("slop") for your team. It's completely free, requires no API keys, and can be set up as a gatekeeper in your team's project today.
• Buzz this week: A viral new project that has exploded to 4,115 GitHub Stars within 30 days, licensed under MIT.
• Key Features:
  • Built-in 16 strict rules, such as no-chained-type-assertions (forbidding consecutive as assertions), no-unsafe-dictionary-type (disallowing vague definitions like Record<string, unknown>), and require-safety-comment-for-type-assertion (requiring every type assertion to have a comment explaining "why it is safe").
  • Powered by Oxlint written in Rust, it is incredibly fast. It is also designed in a "vendored" mode, where you copy src/ directly into your project. The rules are entirely yours to modify, rather than being locked to a versioned npm dependency.
• Quick Start: The fastest way is to install it automatically via the agent:

  npx skills add dmmulroy/anti-slop --skill install-anti-slop
  

  Or manually copy src/ into your repo, register the entry point in oxlint.config.ts, and select your rules.
• Comparison: Compared to the commonly used typescript-eslint, it runs blazingly fast on Rust-based Oxlint and takes a much stricter stance: it doesn't catch syntax errors, but rather "insufficiently evidenced patterns." Additionally, instead of using an npm dependency, it forces you to vendor it into your repo, defaulting to the idea that every team should customize their own rules—which is highly unique.
• Who it's for: Development teams who want to eliminate sloppy AI-generated types and elevate their TypeScript code quality.
• Link: https://github.com/dmmulroy/anti-slop

---

2. slotstream: The Black Magic That Lets a 16GB Mac Smoothly Run 10B+ Parameter LLMs

• What it is: Designed to solve the pain point of developers who lack local memory but don't want to send sensitive code to the cloud. Through specialized memory management techniques, it allows you to run the 104GB Qwen3.8-Flash-Next model—which normally requires massive memory—on a 16GB Mac.
• Buzz this week: Garnered 238 points (with 116 comments) on Hacker News, currently has 295 stars on GitHub, and is licensed under MIT.
• Key Features:
  • Leveraging the characteristic of MoE (Mixture of Experts) models where only a few experts are used per inference, it avoids loading all 68GB of routed experts into memory. Instead, it maintains a fixed-size cache pool and streams missing experts from the SSD in real-time.
  • The cache pool is shared across layers; layers requiring more slots can "borrow" from other layers. In the author's benchmarks, a 48GB M5 Pro achieved 12 tok/s, a 24GB machine got around 8 tok/s, and even a 16GB machine maintained a usable speed of 4 tok/s.
  • Honest Caveat: Currently, it only supports a single model, Qwen3.8-Flash-Next:4bit. Llama, DeepSeek, and other versions of Qwen are not supported at all. It also requires macOS 14 or above and about 110GB of free disk space.
• Quick Start: One-click installation command:

  curl -fsSL https://raw.githubusercontent.com/carloslfu/slotstream/main/install.sh | sh
  

  Run directly after installation:

  slotstream run --prompt "Why is the sky blue?"
  

  (Note: The first run will automatically download about 110GB of model weights. Please ensure you have enough disk space.)
• Comparison: While tools like Ollama or LM Studio operate on the logic of "if it doesn't fit in memory, it won't run," slotstream treats the SSD as a second-tier memory layer to trade speed for space. The author honestly admits that they haven't done a full side-by-side comparison with llama.cpp or MLX on the same Mac yet; the data are benchmarks from their respective environments.
• Who it's for: Developers who want to run large models offline locally on a Mac, with plenty of disk space but limited RAM.
• Link: https://github.com/carloslfu/slotstream

---

3. OpenShot 4.0: A Completely Free Video Editing Powerhouse with Built-in Local AI Masking and Professional Color Grading

• What it is: Solves the pain point for developers who need to make technical tutorials or product demos but don't want to pay hefty subscriptions for commercial video editors or worry about cloud AI privacy leaks. This veteran GPL video editor has received an epic overhaul in version 4.0, bringing screen recording, color grading, and local AI features all in one go.
• Buzz this week: Highly discussed on Hacker News after the release of the new version, gaining 602 points and 139 comments.
• Key Features:
  • Record microphone, screen, webcam, and system audio directly within your project. Each source gets its own independent track on the timeline, making post-editing a breeze.
  • Built-in brand new Color View, offering color wheels, curves, LUTs, and four professional scopes including Luma Waveform and Histogram. It also features AI object masking, which automatically detects and tracks subjects. The machine learning models run entirely locally—no internet, no accounts, and no subscriptions required.
  • The timeline has been rewritten in native Qt for smoother zooming, and ten new effects have been added, including audio visualization and beat sync.
• Quick Start: Go directly to the official website to download the installer for Windows, macOS, or Linux (supporting Wayland and PipeWire). When using the AI masking feature, the software will automatically download the model locally.
• Comparison: Compared to DaVinci Resolve or Premiere, OpenShot 4.0's greatest strength is that it is completely free GPL v3 software that packs screen recording, color grading, and AI background removal all into a single app. Best of all, the AI features run entirely locally, with absolutely no pressure to tie you to a cloud subscription.
• Who it's for: Developers and creators who need to make tutorial or demo videos, value privacy, and want to avoid subscription fees.
• Link: https://www.openshot.org

---

Conclusion

As we can see from this week's open-source picks, whether it's preventing AI from writing sloppy code, squeezing hardware to run large models on the edge, or bringing AI locally for video editing, "local-first and on-device AI" has become an irreversible trend. I highly recommend installing anti-slop today to safeguard your project's code quality. See you next week!