That was an interesting read, I don't understand the jump to Anthropic 'compressing' its soul document into the models weights though. Seems like progressive disclosure to inject this huge doc into the models context during its turn only when its potentially relevant. To my understanding that was the core innovation of Claude 'Skills' and likely what they're leaning on here. The lossy quality you observed could be the model 'cleaning up' any formatting and metadata present in that context injection.
That was an interesting read, I don't understand the jump to Anthropic 'compressing' its soul document into the models weights though. Seems like progressive disclosure to inject this huge doc into the models context during its turn only when its potentially relevant. To my understanding that was the core innovation of Claude 'Skills' and likely what they're leaning on here. The lossy quality you observed could be the model 'cleaning up' any formatting and metadata present in that context injection.