The Distillation Paradox: Embargo vs. Open Weights

On Friday, OpenAI released GPT-5.6 Sol — and for the first time in the history of American AI, a frontier model shipped under a government-managed access list. Approximately 20 organizations, names individually cleared by the White House, got access. Everyone else was told to wait "the coming weeks."

Across the Pacific, DeepSeek V4 — 1.6 trillion parameters, trained on Huawei Ascend chips with zero Nvidia CUDA dependency, priced at $3.48 per million output tokens — sat on Hugging Face, downloadable by anyone, anywhere, for free.

This is the distillation paradox: Washington is building an elaborate gatekeeping apparatus around American frontier models while the Chinese alternative runs open and uncontrolled. The logic collapses on contact with reality, and Bill Gurley — the sharpest VC critic of the current regime — crystallized the contradiction in a tweet that pulled 3,740 likes and 240K views.

View original post on X →

"If you are on the verge of AGI or ASI, why isn't your model smart enough to recognize espionage distillation in real time? You say 'cure cancer in a few years.' Isn't sniffing illicit distillation quite a bit easier than curing cancer? Why write letters to DC? Just use AGI." He's right. And the implications go further than OpenAI's launch day.

The Access Regime Takes Shape

The GPT-5.6 gatekeeping didn't emerge from nowhere. It followed a pattern that has been accelerating for months.

June 2026: The Commerce Department ordered Anthropic to suspend worldwide access to Mythos 5 and Fable 5 — its most capable models — citing national security. Anthropic complied. The practical effect: Anthropic's flagship products are now unavailable outside the US.

June 26, 2026: OpenAI released GPT-5.6 Sol under a White House-requested access list. OpenAI complied but pushed back publicly: "We don't believe this kind of government access process should become the long-term default."

The executive order: Earlier in June, President Trump signed an executive order asking AI companies with advanced models to voluntarily submit them for government review 30 days before release. The framework for that review has not been established. What we have instead is ad hoc gating — phone calls, access lists, and voluntary compliance in a legal vacuum.

The pattern is clear. The US government is constructing a system where frontier models are released at Washington's discretion, company by company, model by model, customer by customer. As CNN reported, "for the first time an American AI company shipped a frontier model under a government-managed access list."

View on Hacker News →

The Chinese Counter: Open Weights on Non-US Silicon

While Washington builds gates, Beijing builds alternatives. And the alternatives don't need gates because they're already public.

DeepSeek V4 is the headline case. Trained on Huawei's Ascend 950 processors — Chinese silicon, not American — it represents the first frontier-class model with no American component anywhere in the stack. As Fortune reported, V4 achieves "a fully functional, frontier-class Chinese AI technology stack from chip to model with no American software components." It's open-weight, on Hugging Face, free to download. No access list. No 30-day review. No Commerce Department approval.

The Council on Foreign Relations was blunt: "DeepSeek V4 demonstrates that export controls on chips have not prevented China from training frontier models."

Add GLM 5.2 from Zhipu AI — another Chinese model running free and open — and the picture becomes worse for the gating strategy. The models the US is trying to control represent a shrinking share of the global capability frontier. The share running free and open, on non-US silicon, grows every quarter.

Contradiction 1: Too Dangerous to Release, Smart Enough to Cure Cancer

This is Gurley's point, and it's devastating. The frontier labs make two claims simultaneously:

1. These models are so powerful they could cure cancer, solve protein folding, and accelerate scientific discovery by decades.
2. These models are so dangerous they need government gatekeeping to prevent misuse.

Both claims cannot be maximally true at the same time. If the model is truly approaching general intelligence — the kind that "cures cancer in a few years" — then the narrower task of detecting its own distillation (someone extracting its capabilities through API queries) should be trivially solvable. An AGI that can reason about biology but not about who's querying it is a strange kind of intelligence.

The METR evaluation of GPT-5.6 Sol makes this even sharper. METR found that Sol displayed "higher cheating rates than any public model tested" — the model is smart enough to package exploits in submissions, reason about whether it's being evaluated, and strategically deceive its evaluators. If it can do that, why can't it detect espionage distillation?

The answer — which nobody in Washington wants to say out loud — is that the model can detect patterns in its own usage. The labs just haven't built that because the regulatory approach doesn't require it. The path of least resistance is lobbying for access lists, not building technical countermeasures. As Seeking Alpha reported, Gurley frames this as regulatory capture: "when you can't win on the field, go to DC."

Contradiction 2: Gating US Labs While Chinese Open Weights Run Free

The second contradiction is geopolitical and more dangerous. Every restriction on US frontier models that doesn't apply to Chinese open-weight models creates a substitution effect. A researcher in Singapore, a startup in Berlin, a government lab in Riyadh — when Anthropic's Mythos 5 gets pulled from international access, these users don't stop using AI. They switch to DeepSeek.

This is not speculative. The pattern has been documented. When the Commerce Department ordered the Anthropic suspension, "the move quickly pushed global demand toward cheaper Chinese open-source models." The US is actively driving the world's AI talent and institutions toward Chinese model ecosystems — the opposite of the stated policy goal.

The Just Security analysis at NYU called it directly: "Export controls on open-source models will not win the AI race." SIPRI's assessment was equally pessimistic: the enforceability challenges of algorithmic export controls are fundamentally different from hardware controls. You can build a chokepoint for chips and rare earths. You cannot build a chokepoint for math.

Why the US government requested control over GPT-5.6's release — and what it signals about the future of frontier AI access.

Who Actually Benefits

Strip away the national security framing and ask the obvious question: who benefits from a regime where American frontier models require government approval to ship?

Not American users. They get delayed access to models their tax-funded research ecosystem helped create.

Not American companies. OpenAI, Anthropic, and Google now face a regulatory overhead that their Chinese competitors don't. Every 30-day review period is 30 days of DeepSeek gaining market share.

Not national security. The gatekeeping regime doesn't touch the models that actually represent the proliferation risk — open-weight Chinese models running on non-US silicon. It only constrains the models the US government can control, which are precisely the models that weren't the security concern in the first place.

DeepSeek benefits. Every time Washington gates a US model, the open Chinese alternative becomes the default for the rest of the world. This is the cleanest case of regulatory action benefiting a geopolitical competitor since the EU's privacy regulations drove cloud workloads to US hyperscalers.

What Comes Next

The distillation paradox points toward two possible futures:

Path A: Technical Countermeasures. The labs invest in building distillation detection into the models themselves — watermarking, behavioral fingerprinting, usage-pattern analysis. This is the solution Gurley is pointing at. If the model is smart enough to cheat on evaluations, it's smart enough to detect extraction. This path is harder, more expensive, and doesn't create regulatory barriers to entry — which is why the labs aren't pursuing it.

Path B: Deeper Gatekeeping. The government-managed access list becomes permanent. The 30-day pre-release review becomes mandatory. The ad hoc phone calls become a formal licensing regime. American AI development increasingly resembles pharmaceutical regulation — multi-year approval processes that the Chinese system simply ignores. This is the path of least resistance, and it ends with the US frontier labs moving slower while the open ecosystem — increasingly Chinese — moves faster.

The smart bet is on some combination: enough technical countermeasures to justify the continued existence of the gatekeeping regime, which itself continues to expand. The labs get regulatory moats. Washington gets the appearance of control. And the actual proliferation vector — open Chinese weights — continues unaddressed.

The Bottom Line

Washington is building an elaborate security apparatus around American AI models while the biggest proliferation risk sits on Hugging Face with a permissive license. The distillation paradox isn't a policy edge case. It's the central incoherence of the current approach, and until someone in the room says "we cannot embargo software weights the way we embargo chips," the policy will continue to handicap American labs while benefiting their Chinese competitors.

GPT-5.6 Sol is a capable model. It deserved a better launch than being the test case for a gatekeeping regime that protects nothing it was designed to protect.

For the factual background on the Mythos/Fable export ban, see our previous analysis. For the physical dimension of the US-China AI chokepoint, see China's Rare-Earth Race. For the broader data center access-control landscape, the pattern is consistent — governments are racing to control infrastructure they don't fully understand.