Sanjay

Sanjay Ghemawat is a computer scientist and cofounder of Discovery Loop, an effort to automate machine learning, science, and engineering. His earlier contributions include developing MapReduce with Jeff Dean and coauthoring the 2016 TensorFlow systems paper with Martin Abadi, Paul Barham, Dean, Derek G. Murray, and other collaborators.[6][5][4]

On August 6, 2026, Latent.Space reported that Ghemawat, Dean, Oriol Vinyals, and Quoc Le were departing DeepMind to cofound Discovery Loop. Dean’s announcement, dated August 5, named the four founders and described the company as a public benefit corporation.[1][6]

In his founding announcement, Dean described the team’s approach as automating the experimental loop, initially concentrating on large-scale experimentation for ML research and engineering. The founders planned to use their own systems as their first customers, gaining rapid feedback from that work.[6]

Ghemawat’s earlier work at Google provides context for the large-scale systems expertise behind that founding team: he and Dean had already built software that let programmers distribute work across many unreliable machines.[6][5][3]

Ghemawat studied at Cornell and pursued graduate work at MIT, where Barbara Liskov was his adviser. He worked at Digital Equipment Corporation before following his colleague Dean to Google in December 1999. Their practice of writing code together at one computer began before Google.[3]

During Google’s March 2000 indexing emergency, Ghemawat and Dean examined missing and misordered words in the stalled index. After days of checking software logic, they inspected the binary representation. Ghemawat pointed out a bit that should have been zero but was one; the pair recognized a pattern of memory corruption and wrote code to compensate for the offending machines. They also added crawling checkpoints so interrupted work could resume midstream, and in 2001 demonstrated that Google’s index could be stored in memory rather than on hard drives.[3]

In 2003, repeated rewrites of Google’s crawler and indexer led Ghemawat and Dean to generalize the coordination work they kept doing. They developed MapReduce over four months, then rewrote the crawling and indexing system in terms of MapReduce tasks the following year. Their jointly authored 2004 paper made the model public.[3][5]

MapReduce separates a computation into a map function that produces intermediate key/value pairs and a reduce function that merges values sharing a key. For word counting, mapping can count occurrences on individual pages, while reducing adds those counts together. The runtime handles input partitioning, scheduling, machine failures, and communication, relieving each programmer of implementing those distributed-system details.[5][3]

Ghemawat was one of 22 coauthors of TensorFlow: A system for large-scale machine learning, published at OSDI in 2016. TensorFlow represents computation, shared state, and state-changing operations as dataflow graphs. It places graph nodes across machines and devices—including CPUs, GPUs, and TPUs—and exposes state management so developers can experiment with training algorithms and optimizations.[4]

In a TensorFlow Lite session observed by James Somers in 2018, Ghemawat and Dean analyzed code size and optimized a tool they had built for that analysis. When Dean proposed refactoring a section, Ghemawat suggested first understanding the patterns and recording notes. The same profile reported that they were planning an unnamed AI prototype intended, in Dean’s description, to train a large model for thousands or millions of tasks.[3]

Ghemawat expressed enthusiasm for pair programming: “I don’t know why more people don’t do it,” he told Somers. Their longtime colleague Wilson Hsieh contrasted Dean’s rapid ideas and prototypes with Ghemawat’s durability: “Sanjay was the one who built things to last.” Craig Silverstein praised the balance of readability and density in Ghemawat’s code, and how readily it accommodated new functionality.[3]

In 2018, Ghemawat was an individual contributor who managed no one, while participating in weekly meetings of Google’s Area Tech Leads to make company-wide technical decisions. His preference for that role was explicit: “I would not want Jeff’s job,” he said of Dean’s management responsibilities.[3]