Shazeer to openai, then John Jumper (the alphaFold nobel guy) to anthropic, plus Adler and Pritzler out the same door within a week. Every time one of these drops the framing is google is bleeding. I think people are reading it backwards. If the people who actually trained the thing can leave and instantly matter at a competitor, the weights were never the asset. The judgment about how to steer a model, what to eval it on, where it breaks, that stuff lives in heads not in checkpoints. Hardware...