SemiAnalysis Founder: The real determinant of large model competition is the internal feedback loop. Dylan Patel mentioned that Anthropic's Mythos 2 is reportedly fully trained but might not be released publicly; OpenAI's much-hyped Astra could face similar restrictions on public release. In his view, this signals a shift in focus for cutting-edge model companies—from 'who gets their model to market first' to 'who can continuously use the latest models internally to fuel next-gen development.' If Mythos 2 can continue training Mythos 3, or Astra can assist in iterating future models, then the improvement in model capabilities won't stop just because public release is delayed. On the contrary, what’s truly worth watching is: as closed-source models slow down public accessibility, will the visible capability gap shrink externally, while leading companies continue to widen the gap internally through faster iteration?