Indexing a year of video locally on a 2021 MacBook with Gemma4-31B (50GB swap)

Original Article ↗ Hacker News Discussion ↗

A blog post about indexing a year’s worth of personal video on a 2021 M1 MacBook Pro using the Gemma 4 31B model sparks interest in how far local LLMs and vision tools can go for archival search, face detection, and creative workflows without relying on cloud services. Commenters trade practical details on running large models on Apple Silicon—memory limits, swap usage, quantization, and alternative runtimes and apps—while also exploring adjacent use cases like photo curation and semantic video search. A recurring theme is unease with AI‑assisted writing: many value the technical substance but find the LLM-influenced prose grating, raising broader questions about standards for AI-generated content in public forums.

Overall reaction & use case

  • Many commenters find the project impressive: indexing ~1TB / a year of video locally with a big vision-capable LLM on a 2021 MacBook is seen as a strong proof of what consumer hardware can do.
  • Several readers say they now have “a weekend project” and have similar piles of unorganized photos/videos they’d love to index.
  • There’s interest in extending this to still photos and to assist with editing in DaVinci Resolve or similar tools.

Hardware, performance & swap

  • Multiple reports of good local LLM performance on Apple Silicon (M1–M5) and even older x86 laptops, though with loud fans and heavy CPU usage.
  • Some question why so much swap is used, noting SSD wear and that a 4-bit 31B model should fit in far less RAM if the system is cleaned up.
  • Detailed benchmarks are shared for Qwen 3.6 models on M5 Pro (generation and prefill speeds), plus notes that MLX/oMLX are currently faster than llama.cpp on Apple Silicon.
  • Discussion of prefill vs generation speeds and how Apple Silicon currently lags GPUs for prefill-heavy workloads.

Implementation details & tooling

  • The main pipeline uses:
    • Sidecar .description.md files per clip for a plain-text “index” that travels with the media.
    • Face embeddings via insightface/ArcFace into a local SQLite DB.
    • EXIF GPS + Nominatim/OSM for location, keeping faces/locations out of the LLM’s remit.
  • Others describe similar setups using Whisper, ffmpeg, semantic search, vector DBs, and external LLMs for tagging and video chat.
  • Scene detection is flagged as a key next step. One implementation uses per-frame histograms to detect changes and select representative frames; ffmpeg’s built-in scene detection is also mentioned.

Local vs cloud, privacy & cost

  • Local models are praised for:
    • Cost control when processing large archives.
    • Privacy of personal/family media.
    • Freedom from API limits, uptime issues, and safety-policy false positives.
  • Some argue equivalent setups could work in the cloud if data already lived there; others emphasize the psychological benefit of “no limits” with local models.

AI-assisted writing & HN norms

  • Many readers find the article’s style “AI-sloppy”: staccato, trope-heavy, and distracting despite solid technical content.
  • Others say they enjoyed it and see it as one of the better LLM-assisted posts, with real substance underneath.
  • Several suggest tools and strategies to “de-AI” prose, remove common tropes, and demand higher editorial standards.
  • A moderator reiterates that AI-generated comments are against HN rules; AI-assisted articles are described as a gray area under active consideration.