It's kinda weird to think of brain as an X project. DistBelief was developed out in the open in the main source code repo by people you think of as googlers, like Jeff. So what makes it "X"?
My own personal image of X was that if you had to drive to some old Air Force base to see it, it was X. Since brain was software, it doesn’t fit the model.
Waymo may / will earn them some money, either they sell it to a crazy big investor, or they license the technology out to a company backed by a crazy big investor.
alright I'm gonna ask the naive question: how exactly does Brain make money? is it offered as an independent product? or is it's incremental contribution to the rest of Google that high? what even is Brain, to Google devs?
To me it's an umbrella for all of the projects that fell out of an initial realization (by Jeff Dean and his circle) that Google's infrastructure was large enough to put into practice a class of older ML ideas that had spent years being thought infeasible. They developed the means of training huge models, then useful models that resulted from that, and follow-on project like TPUs.
Is Brain doing great? From a few things people have said, it sounded like Brain was struggling so Google bought Deep Mind instead. All of the really cutting edge AI stuff has come out of Deep Mind, not Brain.
Just to clarify here, Deep Mind in no way replaces Brain. The AI space, like any other space, isn't a single unified landscape. There are areas where Deep Mind is a true standard setter--particularly deep reinforcement learning (AlphaGo, for example)--and areas where Brain has been one of the standard setters--natural language processing springs to mind, where Brain has historically pushed the field forward with projects like word2vec, BERT, and the Transformer architecture generally.
Brain is also a much more general organization. They do things like develop Tensorflow, which is one of two most popular ML frameworks in the world (and until recently was far and away the most popular), and TPUs, which are ML-dedicated ASICs that have a huge impact on training and inference.
Both orgs are world class and historically important, and they certainly overlap, but they aren't replacements for one another.
I strongly disagree. Without a shadow of a doubt, the most successful/important AI model of recent times is the Transformer (from Brain). As another example, the fastest accelerators, TPUs, are from Brain. If you only look at "amount of output" and use # publications at e.g. NeurIPS as proxy, Brain vastly outperforms Deepmind. Deepmind really exceeds at flashy and showy PR releases like AlphaStar, not so much at laying the groundwork for AI breakthroughs. (Though of course AlphaFold or AlphaGo were great).
Brain is not a pure research group. It builds infrastructures/platforms for machine learning related tasks and its research part is relatively small compared to the entire organization.
