The Pentagon's AI advantage is being threatened by a Chinese distillation technique that enables the rapid replication of AI capabilities. This technique, which involves training a less capable model to replicate the behavior of a more powerful one, has allowed Chinese firms to narrow the gap with US frontier models. The US government has attempted to restrict the export of high-end chips, but this has not been enough to maintain a strategic lead.
The distillation technique has been used by Chinese labs to extract the capabilities of US frontier models, such as those developed by Anthropic and Google. This has enabled them to deploy frontier-class models at a significantly lower cost and with a nearly 90 percent operational discount. The US Department of Defense is now looking to adopt a two-pronged strategy to counter this threat, including partnering with frontier companies to gain early access to new models and establishing a high-velocity refinement pipeline to turn these models into operational assets.
The Pentagon is also exploring ways to secure early access to frontier models, including embedding liaisons within tech labs and negotiating exclusive windows of access prior to public release. This would provide a head start in integrating these capabilities into critical systems before they can be distilled by adversaries.
