Anthropic says Alibaba illicitly extracted Claude AI model capabilities

Perceived hypocrisy & IP disputes

• Many see Anthropic’s complaint as ironic: they trained on massive scraped and even pirated corpora (including books, leading to a large settlement) and now object to others “extracting capabilities” from their models.
• Some argue Anthropic forfeits moral high ground; others counter that piracy of books and ToS‑violating distillation are legally distinct (copyright vs contract/trade secrets).
• Debate over whether any party has sympathy: some see it as “thieves stealing from thieves,” others emphasize that serious R&D and late‑stage training still cost real money.

What “distillation attacks” are

• Commenters distinguish:
  • Full/logit distillation (internal, needs access to weights/logits).
  • Black‑box distillation / supervised fine‑tuning on API outputs.
  • RL/RLAIF using a stronger model as reward model or teacher.
• Some say calling distillation an “attack” is marketing and protectionism; they liken it to web scraping, API re‑implementation, or fast‑follower reverse engineering.
• Others stress it’s a clear ToS violation and an attempt to free‑ride on expensive post‑training (especially RL for coding and reasoning), which can materially close capability gaps.

Chinese token resellers & fraud economy

• Detailed reports describe a gray market in China reselling Claude access at 70–90% below official API prices.
• Mechanisms cited: pooled consumer subscriptions, payment fraud, residential proxies, and industrial bot farms creating tens of thousands of accounts; conversations and traces are logged and resold as training data.
• Identity verification and geoblocking are seen as partial, easily bypassed mitigations; some say Anthropic’s own loss‑leader pricing made arbitrage inevitable.

Geopolitics & regulatory capture

• Strong suspicion that Anthropic is framing distillation as a China‑linked “national security” issue to woo US regulators, justify export controls, and kneecap Chinese labs and possibly future non‑US competitors.
• Others say Chinese labs also benefit from US‑side openness while blocking Western services at home, and that US would likely behave similarly if roles were reversed.

Open vs closed models and commoditization

• Many cheer Chinese open‑weight models and distillation as “Robin Hooding” proprietary capabilities into the commons and accelerating cheap competition.
• Counterpoint: heavy reliance on distillation may under‑incentivize original frontier R&D and could eventually leave the world dependent on a few state‑aligned suppliers.
• Broad expectation that inference will become low‑margin and commoditized; value may shift to proprietary data, agents, and application layers rather than the base models themselves.

Safety, censorship & access

• Commenters criticize Anthropic’s safety posture and export‑ban narratives as inconsistent: models are hyped as dangerously powerful yet often heavily censored and still jailbreakable.
• Widespread fear that true frontier models may end up available only to governments and a handful of megacorps, with everyone else relying on slightly‑behind open or distilled models.