Alexander Doria on the technique Chinese labs are using to train 3 trillion-parameter models without access to Blackwells: "The most interesting part about the models that have been released over the last few weeks is this is an unlock, because until now we never had models since two trillion, three trillion parameters range that are very large and in the open. It never happened before." "For a very simple reason. Normally everyone thought that to train that you need to have the latest…