Burla demo processing 1.7M Airbnb photos via CLIP plus Claude Haiku Vision and 50.7M reviews on a 1.7K-CPU/20-A100 dynamic cluster to surface weird listings.
Key Takeaways
Pipeline: CLIP embeds 1.7M photos against text prompts, top suspects sent to Claude Haiku Vision for category confirmation (pets, drug-den vibes, bad TV mounts, chaotic kitchens).
Reviews use a 3-tier funnel: regex on all 50M, SBERT embedding cluster on top 200K, Claude Haiku reranking on top 12K.
Burla’s remote_parallel_map scaled to 1.7K CPU workers and 20 A100s on one dynamic cluster; no Docker or Kubernetes required.
Demand validation uses bootstrap 95% CI on 365-night calendar occupancy per listing; results labeled accepted only when group bars don’t overlap.
Data source is Inside Airbnb’s public dump across 119 cities and 4 quarterly snapshots.
Hacker News Comment Review
Commenters largely treated this as content marketing for Burla’s managed cloud service, noting the prominent Burla branding and that the author works there.
Classification quality drew skepticism: “drug-den vibes” flags appeared to catch poorly lit small rooms rather than genuinely suspicious listings, suggesting CLIP prompt engineering was too coarse.
Inside Airbnb’s community guidelines explicitly request no scraping, raising legal and ethical flags several commenters noted alongside general concerns about resource waste for novelty output.
Notable Comments
@devmor: classification logic is “insane leaps” – dark or obscured photos flagged as drug dens, mostly just poor photography.
@danhon: Inside Airbnb guidelines explicitly prohibit scraping; direct data access requires emailing data@insideairbnb.com.
