DeepSeek Spared the Blacklist, Drug Repurposing Slashes Costs, and Madrid Builds Cheap Metros

The US government's decision to hold off on blacklisting DeepSeek — while simultaneously designating more than 100 other Chinese technology firms as security risks — drew 467 points and 516 comments. Reuters reporting suggests the decision is at least partly pragmatic: DeepSeek's models are open-weight, meaning the weights are publicly available and a blacklist of the company does not affect availability of the underlying capability. The parallel drawn in the thread is to the NSA's largely unsuccessful attempt to classify public-key cryptography in the 1970s — restrictions on mathematical knowledge have a poor historical track record. The more than 100 firms that were designated reportedly span surveillance technology, telecommunications infrastructure, and AI-assisted military systems, suggesting capability-specific assessments rather than broad sector blacklisting.

Against that policy backdrop, the episode examines a structural assumption running through coverage of open-weight AI: that distributing model weights is equivalent to distributing the capability. The counterargument is hardware-gated access. GPT-4-class models require multiple high-end GPUs to run at useful speeds; even smaller open models require hardware costing several thousand dollars. The population currently running local models is described as small, technically sophisticated, and relatively affluent or institutionally affiliated. A second complicating factor is training concentration: the organizations capable of training frontier-class models may be narrowing rather than widening as training costs increase faster than hardware cheapens. Two falsifiable signals are proposed for tracking the thesis: whether hardware costs for capable model inference fail to decrease at the rate of 1990s-era PC hardware, and whether the number of organizations capable of training competitive frontier models contracts rather than expands by 2028.

A lower-scoring story — 19 points — on drug repurposing from researchers at King's College London carries outsized substantive weight. The research documents hospitals and universities successfully repurposing existing drugs for new indications at costs reportedly 90 percent lower than developing new drugs. The mechanism is that off-patent generics can be tested for new uses through academic clinical trials without a full FDA new drug application, provided the indication covers a population already eligible for the drug. The HN thread, small but substantive, identifies a structural barrier: academic institutions lack the commercial infrastructure to bring repurposed drugs through regulatory approval and to market at scale, and pharmaceutical companies have little commercial interest in running expensive trials for drugs they cannot patent — a market failure where clear social value exists but no private actor is positioned to capture enough of it to justify the investment.

The Madrid metro construction piece from Works in Progress — 157 points, 101 comments — examines how Madrid built its metro system at dramatically lower per-kilometer costs than comparable systems in New York, London, or Paris. The attributed factors are continuous tunneling rather than cut-and-cover excavation, a national workforce that specialized and scaled through repetition, and a procurement model that kept contracts simple and made changes expensive. The HN thread extends this into a broader discussion of why US infrastructure costs are high, with the Madrid case providing what commenters describe as unusually clear comparative data. Australia's new SMS sender ID registration requirement from the ACMA — 106 points, 61 comments — rounds out the policy segment, with Australia implementing a mandatory registry for alphanumeric SMS sender IDs to prevent phishing and impersonation, a system the UK has adopted but the US has not, partly due to First Amendment complexity around compelled registration for communication.