How Did DeepMind Decide What Code to Keep?

David Senra, reading Sebastian Mallaby, says DeepMind decided what stayed in the master code by measurement alone. The lab embraced a strict unity, and all team members poured their energies into improving one single model. Next came meritocracy. Any team member was welcome to propose an improvement to the model and test it. If the upgrade boosted performance, it was added to the master code on which everyone was building. Seniority, force of personality, and dazzling theoretical claims as to why something should work did not affect what went into the program. Only measurement mattered.

Senra says Demis Hassabis is running all of Google's AI, and the project is so big that Hassabis does not code anymore and does not design things directly. The skills are more about holding a hundred different projects in mind, context switching between complicated things with negative minutes of time between, laying out division, picking the right intermediate targets on the way to the big goal, and nurturing people to take things on. The word used most is relentless: relentless progress, relentless shipping, a relentless production machine for innovation.

Less than two years after the messy shotgun marriage that created Google DeepMind, the team had closed the technical gap. It was a considerable achievement.