Ion Stoica

Stoica's name is on the founding papers of Spark (distributed data processing), Ray (distributed Python for ML workloads), and the broader stack vLLM uses to serve language models efficiently — artifacts that have become the de facto substrate for ML in industry. The pattern is consistent: design a system around the actual computational bottlenecks researchers and engineers face, ship it open-source with a commercial company alongside, and let the ecosystem prove the design choices. For ai100, which runs hundreds of audits against multiple LLM engines, the infrastructure side Stoica's group built is not background — it's the layer that determines how feasibly you can run any large-scale model comparison at all.

Worth following when

you want to understand the infrastructure constraints that shape what large-scale LLM evaluation can actually do, beyond the methodological framing.

Topics

distributed ML systems and infrastructure (Spark, Ray, vLLM); the economics and engineering of large-scale model serving; the open-source-plus-commercial-company pattern in ML infrastructure.

Key works

Apache Spark (2010, co-creator); Ray distributed framework (2018, co-lead); contributions to the vLLM efficient-serving stack (2023, ongoing).