📦The Greats

Google's 2016 release of TensorFlow as open source gave researchers and production engineers a common framework for describing and executing many machine-learning workloads. Its spread highlighted how frameworks, tooling and deployment environments influence which ideas can move from research into products.

At Stanford, Koller co-authored the textbook Probabilistic Graphical Models, which gave students and practitioners a systematic vocabulary for uncertainty and structured prediction. Her later work building organizations around online learning and drug discovery illustrates the systems needed around models.

Rudin has argued that for high-stakes decisions, an opaque model paired with a post-hoc explanation is often a poor substitute for a model designed to be interpretable from the start. The argument directly affects MLOps evaluation and model-approval practices.

His 2017 book Designing Data-Intensive Applications translated hard-won distributed-systems lessons into a widely read engineering guide. MLOps engineers use the same questions about lineage, latency, correctness and failure when building training and inference data flows.

Through fast.ai courses, Thomas and collaborators made practical deep-learning instruction accessible to many learners outside elite research labs. Her later advocacy on accountability has reinforced the idea that deployment teams must measure harms, not only accuracy.

Her 2022 book Designing Machine Learning Systems collected operational lessons—data distribution shifts, feedback loops, evaluation and infrastructure—that had often been scattered across engineering teams. It became a common reference for people trying to make MLOps concrete.

Gebru co-authored the 2018 “Datasheets for Datasets” paper, proposing standardized documentation for datasets' motivation, composition and collection. The idea gives operational teams a way to ask what a model's training material represents before deploying it in a new setting.