Jeff Dean & Noam Shazeer — 25 years at Google: from PageRank to AGI

The interview presents Jeff Dean and Noam Shazeer as long-running contributors to Google's technical evolution: Dean worked on systems including MapReduce, BigTable, TensorFlow, and Gemini, while Shazeer helped develop Transformer and mixture-of-experts methods. Dean describes how organizational scale changes what any one employee can know, from recognizing every colleague in a 25-person company to encountering projects he has never heard of. He argues that broad networks inside a large organization help people find the right expertise.

Both guests describe careers that began before AI became a mainstream business. Shazeer joined Google after seeing its search growth and hoping eventually to fund long-term AI research; he was already interested in AI around 2000. Dean says Google's mission to organize the world's information was broad enough to require increasingly advanced systems. Their accounts connect early search infrastructure and neural-network experiments with later large-scale AI work.

Dean says the hardware landscape shifted from rapid gains in general-purpose CPUs toward specialized accelerators such as TPUs and machine-learning GPUs. Because moving data has become more expensive than arithmetic, deep learning benefits from matrix-heavy computation. Google designed TPUs around low-precision linear algebra, including eight-bit inference and later lower-precision formats. Dean frames this as hardware-algorithm co-design: quantization can increase throughput and lower cost, but algorithm researchers must account for the trade-offs rather than treating reduced precision only as a risk.

Dean recounts that in 1990 his undergraduate thesis compared model-parallel and data-parallel backpropagation on a 32-processor Hypercube. He was excited by neural networks' ability to solve toy problems, though the available compute was far short of what real applications required. Moore's Law eventually made neural networks useful at practical scale. In 2007, a machine-translation team at Google used a two-trillion-word five-gram model; Dean redesigned its storage and lookup system so that a sentence translation took roughly 100 milliseconds instead of a night. The episode illustrates how systems engineering can turn a promising model into a usable service.

Dean recounts improving Google's 2007 machine-translation system, which initially took 12 hours to translate a sentence because it made roughly 100,000 disk seeks for each word. He and the team built a compressed in-memory representation of a five-gram model trained on two trillion words, distributed across 200 machines, with batched lookups. The change reduced translation time to about 100 milliseconds, illustrating how systems design can make a promising model practical.