How Anthropic actually builds Claude

We treat the model as if it's a product to some degree. With every new model, we are specking out exactly what are the requirements for this model? What do we want this model to be good at? What do we think it's going to be good at? Because that's actually part of the interesting part of model development compared to product development is that in many ways we're growing the model and it's not exactly like we're specking it out to the fullest extent that you can when you're designing like a software layout. There's like buckets of capabilities that we care a lot about. Of course, coding has been a big one, right? Knowledge work recently one as well. So, with some of our recent models, we've tried to make them really good at like working within our products like Cloud for Excel. But then there's also the side of on every model, we want to fix and improve on the things the last model didn't do so well. If we notice some like really interesting behavior, are there tweaks and interventions that we can take in training on the next one? One of the cool parts about working on these models is that they just touch so many different surface areas. So, as a research PM, you need to think about how this model is going to be exposed through all our surfaces, whether it's API or it's cloud code or it's co-work.