US holds off blacklisting DeepSeek, more than 100 firms deemed security risks
U.S. export controls on Chinese AI firms, including a decision to delay blacklisting DeepSeek, are being read as part of a broader struggle over who will control the economics and infrastructure of advanced AI. Commenters weigh claims that Chinese labs illicitly trained on U.S. models against the fact that Western companies themselves built their systems on massive, often contested data scraping, highlighting what many see as hypocrisy and a de facto global IP free‑for‑all. The thread also probes whether blacklisting and chip bans are genuinely about national security or primarily about protecting U.S. incumbents from cheaper open‑weight Chinese models that undercut them on price and could weaken their market dominance.
IP, Scraping, and “Illicit” Training
- Many see Anthropic’s complaint about DeepSeek scraping Claude as hypocritical, given US labs trained on massive web and book corpora without consent.
- Some argue outputs from LLMs should be fully legal to scrape and reuse, and ToS bans on this should be unenforceable to avoid entrenched monopolies.
- Others stress that ToS breaches and pirated datasets still matter legally and morally, even if everyone is “stealing from each other.”
- Disagreement over evidence: some say claims of OpenAI/Anthropic “IP theft” are overblown; others cite lawsuits, settlements, and verbatim memorization studies as proof.
US Policy, Entity Lists, and Protectionism
- Strong view that US actions (entity lists, export controls, potential DeepSeek blacklisting) are about protecting domestic AI incumbents and capital, not “safety.”
- Counterview: measures are framed as national security and industrial policy, especially in a broader “America vs China” tech rivalry.
- Some note Chinese AI labs already on the Entity List still ship top open-weight models, implying limited practical impact once weights are public.
- Debate over whether export controls on GPUs are meaningful or easily circumvented via third countries and gray markets.
Security, Sovereignty, and Whom to Trust
- Non‑US commenters often say they now distrust US clouds more than Chinese ones, citing the CLOUD Act and NSA-style collection.
- Others counter that Chinese firms are tightly tied to the state and should be assumed to serve CCP interests; US still seen as relatively more constrained by law.
- Several point out that open‑weights models can be self‑hosted in EU/US clouds, separating “Chinese model” from “Chinese server.”
Economics, Competition, and Industrial Strategy
- DeepSeek’s extremely low pricing is viewed as a potential “commoditize the complement” move that threatens US hyperscaler business models.
- Some suspect state support and deliberate price undercutting; others frame it as compute‑efficient engineering driven by China’s constrained access to high‑end chips.
- Large argument over protectionism: some want to block Chinese AI (and cars like BYD) to preserve domestic industry; others argue this entrenches uncompetitive incumbents.
Developer Experiences and Practical Use
- Multiple developers report using DeepSeek (especially Flash and v4 Pro) heavily for coding and documentation at a tiny fraction of Claude/OpenAI cost.
- Perceived quality: slightly behind top US frontier models on very hard tasks but “good enough” for everyday coding, making price/performance compelling.
- Many expect cheap, open‑weight models to erode the moat of closed US providers and shift value to hardware and self‑hosted deployments.