Who's Afraid of Chinese Models?

Most of the reaction accepted the broad pressure on U.S. labs, especially on valuation. People kept coming back to the same point: if token pricing keeps falling and model quality keeps converging, the giant private-market marks on OpenAI and Anthropic look fragile. But the strongest pushback was on the essay’s cost logic. Several commenters said it is sloppy to treat open models as “free” just because the weights can be downloaded. Inference still has real manufacturing-style costs, and listed API prices do not tell you marginal cost. Others answered that this misses where open models are already winning. Many business workflows do not need frontier-level reasoning or long agent loops. For internal tooling, domain apps, and backend automation, open models can already be good enough, avoid lock-in, and keep proprietary data inside your own stack. The other live question was stickiness. A lot of individual users said switching among Claude Code, Codex, Cursor, and similar tools is easy because the products are converging. The more persuasive version of the lock-in argument was organizational, not technical. Once a company wires in connectors, permissions, approvals, and habits, the chosen tool tends to linger even if the underlying models are interchangeable. That leaves a narrower moat than the essay implies. It is less about magical agent harnesses and more about ordinary enterprise software inertia. On distillation, the center of gravity was that frontier labs are asking for one set of rules for themselves and another for everyone else. If training on scraped internet data is acceptable, many saw it as hard to morally distinguish model-on-model distillation. A few commenters drew a practical distinction, saying distillation extracts value from someone else’s expensive refinement rather than building from raw data, but even there the larger point held: legal barriers are doing more to protect incumbent labs than to prove Chinese labs are incapable of training strong models on their own. Multiple commenters flatly rejected the idea that Chinese progress can be explained mainly by copying. They pointed to the scale of Chinese research talent, data collection, compute buildout, and the weakness of “the model called itself Claude” style evidence. Security worries showed up, but mostly in a different form than “Chinese model bad.” Several people argued the key line is open versus closed, not Chinese versus American. Open weights are at least inspectable and self-hostable. Closed APIs require blind trust. Others pushed back that open weights are not automatically safe. A model can still be backdoored during training, and full reproducibility remains hard even with code and data. That left a pragmatic conclusion: open models reduce one kind of dependency risk, but they do not erase security review. The overall landing was blunt. Chinese competition looks real, distillation is probably part of the story but not the whole story, and the strongest moats are looking less like pure model superiority and more like distribution, enterprise process, and who can deliver the lowest cost per useful task.