Jeff Dean

Jeff Dean is an AI researcher and systems engineer who coauthored work on model distillation, sparse mixture-of-experts networks, TensorFlow and Tensor Processing Units (TPUs). At Google, he advocated combining separate research efforts into a unified multimodal model—the effort he says became Gemini. In August 2026, he left Google to cofound Discovery Loop, pursuing automated machine-learning research and scientific experimentation.[3][4][6][7][1][10][9]

On August 5, 2026, Google announced that Dean and longtime collaborator Sanjay Ghemawat were launching an independent public benefit corporation after Dean's approximately 27 years at Google. Discovery Loop names Dean, Ghemawat, Quoc Le and Oriol Vinyals as its founding team. Google said it would participate as a founding investor and Cloud partner and collaborate on a research framework for ML systems and related infrastructure.[8][9]

In Steven Levy's prelaunch interview for WIRED, Dean said the idea had emerged only weeks earlier: “We were working on different things, but we were all starting to see the possibility of AI being able to automate scientific and engineering loops.” Dean said Sundar Pichai had tried over several meetings to persuade the founders to stay; Levy reported that the team ultimately chose the fun and freedom of a startup. Asked who would be CEO, Dean replied, “I think I’m the CEO,” adding, “Everyone pointed at me.”[10]

The venture aims to automate the experimental loop: propose an experiment, implement and run it, evaluate the results, then refine the approach. Its initial plan is to automate machine-learning research and engineering, using those capabilities to optimize its own technology stack before expanding into other scientific and engineering domains.[9][10]

Earlier that year, in his February 12, 2026 Latent Space interview, Dean recalled writing a one-page memo arguing that separate language and multimodal projects across Google Research, Brain and legacy DeepMind were fragmenting compute and talent. He proposed one effort to build a model that was multimodal from the start, calling the memo the origin of Gemini. He also said he supplied the name, invoking both the two organizations coming together as twins and NASA's Gemini program preceding Apollo.[1]

Dean argued for developing both frontier models and affordable, low-latency models. The larger models expose new capabilities and tackle difficult reasoning tasks; they also teach smaller models through distillation. As he put it, “you have to have the frontier model in order to then distill it into your smaller model.” He said successive Gemini Flash generations had matched or exceeded the preceding generation's Pro model.[1]

In the February 2026 interview, Dean described TPU development as a collaboration between chip architects and model researchers, planning for workloads two to six years ahead. Researchers can help identify useful future hardware features, while model architectures must also adapt to the chips available for training and inference. Small speculative features may justify their chip-area cost; larger changes require more extensive ML experimentation.[1]

His serving-efficiency analysis emphasized the energy required to move data into arithmetic units. Batching reuses weights after that movement; lower precision reduces the number of bits transferred. He also explained speculative decoding as a way to amortize weight movement across several candidate tokens. These choices must be considered alongside latency and throughput: a batch of one can offer attractive latency while using hardware inefficiently.[1]

Dean's work on parallel neural-network training and Google's distributed infrastructure predates Gemini. That earlier work provides context for his later contributions to systems that distribute computation across machines and accelerators.[1][5][6]

Dean recalls writing a 1990 undergraduate thesis comparing forms of parallel neural-network training. When he began working on neural networks at Google in late 2011, he revisited model-parallel and data-parallel ideas from that thesis to train larger networks. He had joined Google in 1999.[1][2]

One earlier engineering decision concerned Google's search index. Dean recalls that in 2001, growing traffic had required enough replicas that the machines collectively had sufficient memory to hold an entire copy of the index. Moving the index into memory made broader query expansion practical: synonyms such as “cafe” and “bistro” could accompany a restaurant query without incurring a disk seek for every added term on every shard.[1]

Dean and Sanjay Ghemawat coauthored the 2004 MapReduce paper. Programmers specify a map function that generates intermediate key/value pairs and a reduce function that merges values sharing a key. The runtime handles input partitioning, scheduling, machine failures and communication, allowing programmers to use a commodity cluster without implementing those distributed-system details themselves.[5]

With Geoffrey Hinton and Oriol Vinyals, Dean coauthored Distilling the Knowledge in a Neural Network, associated with the NIPS 2014 Deep Learning Workshop and submitted to arXiv in March 2015. Building on earlier ensemble-compression work by Caruana and collaborators, the paper developed a technique for transferring an ensemble's knowledge into a single deployable model. Dean later recalled the practical motivation: specialist image models improved performance, but an ensemble of perhaps 50 models was cumbersome to serve. He explained that larger-model logits supplied richer supervision for the smaller model than hard labels alone.[3][1]

Dean was part of the broad team behind the 2016 TensorFlow systems paper, including Martín Abadi, Paul Barham and Ghemawat. TensorFlow represents computation and mutable shared state as dataflow graphs, mapping their nodes across machines and CPUs, GPUs and TPUs. Making state management expressible in those graphs gave developers flexibility to experiment with optimizations and training algorithms.[6]

In 2017, Dean coauthored Outrageously Large Neural Networks with Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le and Hinton. Their sparsely gated mixture-of-experts layer uses a trainable gate to select a sparse combination of expert subnetworks for each example, increasing capacity without a proportional increase in computation. The paper described architectures with up to 137 billion parameters between stacked LSTM layers, evaluated on language modeling and machine translation.[4]

Dean also coauthored the 2017 TPU performance study, whose author list begins with Norman P. Jouppi, Cliff Young, Nishant Patil and David Patterson. It evaluated inference TPUs deployed since 2015, emphasizing deterministic execution for response-time requirements. On the production inference workloads studied, the authors reported average speed advantages of about 15–30 times and TOPS-per-watt advantages of about 30–80 times over contemporary Intel Haswell CPUs or Nvidia K80 GPUs.[7]

In February 2026, Dean proposed systems that give the effect of attending to trillions of tokens through staged retrieval. Lightweight processing would identify candidates, increasingly sophisticated models would narrow them, and a highly capable model would reason over the final selection. For personal assistants, he favored a shared model retrieving emails, photos and documents with the user's permission and opt-in.[1]

Dean expected general models to prevail over specialized models in most cases, while still seeing value in domain adaptation. A strong base model could receive additional robotics or healthcare training; specialized modules might also work alongside it. He treated the training-data mix as a trade-off among capabilities, rather than assuming a single model should ingest all available data from every domain.[1]

For coding agents, Dean argued that interaction should vary with the task: some work benefits from frequent guidance, while well-specified tasks can be delegated more independently. Clear specifications, including corner cases and performance requirements, shape the quality of the result. He also advocated lower latency to support longer reasoning and more work between human interactions; 10,000 tokens per second was an aspiration he considered useful for reasoning, parallel rollouts and code checking.[1]