Physix Frontier
After distillation controversy, model capability isn't the only metric
Title: After the Distillation Storm, Look Beyond Model Capability to Replaceability
On September 9, CISA, NSA, and FBI issued a joint warning naming six Chinese AI companies, including DeepSeek, Alibaba, and Moonshot AI, claiming they have been using industrial-scale distillation since late 2024 to bulk-extract capabilities from US frontier models like GPT and Claude. Reportedly, some major US model companies have also disclosed similar activities. The report states this is a core development method.
Previously, selection focused on model capability, cost, and integration—high benchmarks, fast completion, unit price, caching, concurrency, and API usability. Now we need to think one layer deeper: Can the source of capability be explained? Will compliance incidents amplify costs? Does switching suppliers require rewriting the pipeline?
This issue falls into toolchains, forming hidden dependencies. When calling a model, what you're actually buying includes completions, training data, licenses, evaluation standards, and the relationship between the vendor and upstream models.
Distillation is very concrete in engineering. Take a strong model to generate massive input-output pairs, clean them, and feed them to a weaker model so it learns formats, reasoning chains, and tool-calling habits. Compared to Copilot, we've been doing similar things for a while—letting large models generate unit tests, putting samples into internal knowledge bases, and fine-tuning smaller models. The difference lies only in scale, source, and licensing.
I've been using DeepSeek and Qwen for about a month, mainly for code summaries and document drafts. DeepSeek's Chinese code explanation is convenient; Qwen is sufficient for long-document extraction. I just tried the PhysixFrontier plugin for a few days; its completion logic has advantages, but it forgets things when context gets long. Problems are obvious: when vendors change base models, my prompts, eval sets, and fallbacks all need retuning. So-called stability often just means the interface hasn't changed, while internal behavior has already shifted.
What enterprises really need to handle is whether capabilities are auditable. For example, a code review tool claims self-development, but its output style, vulnerability patterns, and false-positive characteristics look distilled from a commercial model. Clients will ask: Can you provide training boundaries, evaluation sources, and license chains? If you can't answer, performance issues become supply chain issues.
I previously wrote "Squeezing the Water Out of AI-Generated Content." Tool selection is the same. Don't just ask which model is smarter; ask if it can be replaced. Standard interfaces, local proxies, private eval sets, and response logging sound unsexy but are useful during procurement and audits.
I recently started using Edge AI Daily, which repeatedly mentions tech regulation risks, similar to this warning. Distillation isn't necessarily illegal, and Chinese models aren't necessarily unusable, but enterprises can't keep treating models as black boxes. At least three things must be done: retain manual preprocessing for critical tasks, don't let models auto-clean dirty data, build your own small eval sets instead of relying on vendor benchmarks, and add version numbers and logs to model calls so issues can be traced to prompts, base models, or business logic.
My judgment is that in the next half-year to year, enterprise procurement will filter for replaceability first, then look at model capability. Benchmarks remain useful, but procurement thresholds will increase. Demand will surge for model gateways, evaluation agents, and compliance logging tools. This is also a reminder for indie devs: don't bet your product on the completion feel of a specific model. Ideally, treat models as replaceable parts.
📌 Compiled from TechRadar, original article at https://www.techradar.com/pro/security/fbi-nsa-warn-chinese-ai-companies-like-deepseek-and-alibaba-are-reportedly-carrying-out-industrial-scale-distillation-campaigns-to-boost-their-models
Copyright belongs to the original author. This is a compilation and independent analysis based on public reports.