Zuckerberg's Open Letter: The model has to learn everything

Zuckerberg published a long paper today, publicly defending the distillation of the model. In his view, model learning from other models was already part of the open source ecology. If the United States is to continue running open source AI, it will have to ease the restrictions on distillation and training data. He also directly refutes the claim that distillation is harmful: what is observed should be allowed to learn. Zuckerberg noted that if the United States tied itself up first, it would only make the open source models of other countries follow faster. Contrary to recent moves by OpenAI and Anthropic, both companies are protecting their competitors from circumventing restrictions, while Zuckerberg is of the view that learning between models cannot be limited together. Meta just released Muse Glimmer today, which itself was distilled from Muse Spark。