Theory Ties ML Impact to Infra Team Pain
Horace He proposes that effective ML techniques increase operational demands on infrastructure teams.
TLDR
Horace He developed a theory with collaborators arguing that ML researcher impact scales directly with infrastructure strain created by their methods. Replies cite examples including mixture-of-experts models, complex architectures such as GDN, Muon optimizer, and RL scaling. Participants from research and engineering roles respond with agreement, jokes about added workloads, and observations that simpler ideas see wider adoption while complex ones require justification through results. The discussion generalizes the pattern beyond ML to other technical fields.
