DeepMind's Andrew Trask: Scaling Laws Are Pushing AI From Monolithic Models to Protocols

iamtrask · x · 2026-07-30

DeepMind's Andrew Trask discusses how scaling laws are reshaping the future architecture of AI. He notes that as single companies hit bottlenecks in data, compute, and talent, AI is evolving from a massive monolithic "program" into a "protocol."

By combining models from multiple providers, we implicitly integrate the resources they were trained on, yielding faster and better models at lower costs. Trask emphasizes that ensembling top-tier models will always win for absolute accuracy, while mixing open and closed models will quietly dominate the Pareto frontier for price-to-performance ratio.

Original post →

More from AGI Musings

AGI Musings channel →