The dominant work mode at DeepMind is for researchers to ~self organize into project teams and work on the research they think is important (with some filtration by management.) But the ideas are very much coming from the bottom-up. (Where do ideas come from?) I get the impression that it’s fairly collaboratively interconnected - people are working on several projects at once with several groups of people. Here the goals are perhaps fairly fuzzy. This work mode is reminiscent of how Bell Labs enabled free radicals.
However, occasionally top management (CEO or people close to that level) decide to create a ‘strike team’ to go after a specific problem. My impression is that in addition to the impetus coming from a different place these strike teams also operate in a very different mode from people in the rest of the org - they silo off to some extent and focus exclusively on a very well defined goal. These strike teams were how both AlphaFold and AlphaGo happened.
One could draw a parallel between this ‘active reserves’ approach and the §ARPA model, where US Academia is the more diffuse self-directed ‘normal work’ with DARPA and corporate R&D filled overlapping but different niches jumping in to create programmatic strike teams.
The clutch question that I have no answer to is ‘how do Demis Hassabis or other executives decide the moment is right to initiate a strike team?’ Which ties into the bigger question of When is an intense deliverable-focused push helpful and when is it harmful?