Open3DLab’s Salaryman reports a crawler downloading 145.5 GB of community-made 3D assets across three sites. Its user agent identified it as doing 3D foundation-model research at Roblox. The operators say it congested free download servers and ignored restrictions on automated collection.

Their account distinguishes page visits from completed downloads:

  • The crawler visited 18,926 project pages in 26,429 requests over 34 hours, from one Amazon cloud address.
  • It downloaded 1,027 files from 712 projects. Of those projects, 423 carried noncommercial licenses and 345 contained adult material.
  • Download-speed and concurrency caps constrained collection. The operators blocked later attempts and report receiving no answer to two emails.

The author suspects collection for a generative 3D model, but neither that purpose nor actual use in training is confirmed. The Roblox attribution rests on the crawler’s self-description, which can be forged; its network address does not independently establish ownership.

The account connects AI data collection to the costs borne by communities hosting the material. Its argument is about permission, attribution and unequal bargaining power. The operators openly acknowledge their own catalogue’s fan ports and copyright grey areas.

The 132-comment thread on Hacker News adds a practical defense, its collateral damage, and competing readings of the scrape.

What the thread adds

  • hypfer — proposes refusing traffic from major cloud providers as a partial bot defense: “you can get rid of a lot of bots by just refusing to talk to where they tend to live.”
  • cube00 — warns that the same block can exclude real visitors: “they connect through VPN gateways running on those cloud services (especially Azure)”. jeremyjh also challenges the idea that residential-proxy evasion is merely a future possibility.
  • echoangle and philipwhiuk — both suggest the adult files could be intended for moderation training. Echoangle asks whether they could “train a model that rejects adult uploads to roblox”. That is an alternative hypothesis, with no evidence offered for this particular scrape.
  • phn — challenges the complaint’s moral consistency because the site hosts apparent third-party game assets. amiga386 replies that the article already acknowledges this and argues against unequal enforcement. The disagreement leaves the actual collection purpose unresolved.

HN handles are pseudonymous and expose no per-comment scores; ordering is HN’s ranking. This is a slice of the thread, not a consensus.