Quoc

Quoc V. Le is an AI researcher whose contributions include sequence-to-sequence learning, pretraining recurrent networks with unlabeled data, and automated neural-network design. In August 2026, he cofounded Discovery Loop with Jeff Dean, Sanjay Ghemawat, and Oriol Vinyals, pursuing systems to automate machine-learning research and engineering.[6][7][8][14]

Dean announced the venture on August 5, 2026. Latent.Space reported the following day that Le and the three other founders had departed DeepMind to establish it. Discovery Loop was founded as a Public Benefit Corporation with a mission to automate machine learning, science, and engineering.[3][14]

The founders’ stated approach is to automate the experimental loop: proposing experiments, implementing and running them, examining results, and refining the next attempt. Their initial focus is machine-learning research and engineering. In Dean’s launch announcement, the team planned to become its own first customer, using automated experimentation to improve its technology before expanding into other domains.[12][14]

Le’s work on automated experimentation followed earlier collaborations on learning representations, translating sequences, and searching for network architectures. Those projects provide technical context for his career before Discovery Loop.[9][6][8][14]

Le undertook undergraduate work at ANU and NICTA in Canberra under Alex Smola, visited the Max Planck Institute for Biological Cybernetics in Tübingen, and became a Stanford PhD student advised by Andrew Ng. His personal Stanford page dates the start of his Google research-scientist role to 2013.[5]

Tom Simonite’s Innovators Under 35 profile, which records Le’s 2014 honor, describes his childhood in rural Vietnam without electricity at home but near a library. According to Simonite, Le imagined around age 14 a machine capable of inventing. Discussing his frustration with machine learning’s reliance on manual labeling and feature specification, Le said, “I’m a guy without a lot of patience.”[15]

Le was the first-listed author of Building high-level features using large scale unsupervised learning, first submitted in December 2011 and revised in 2012. His collaborators were Marc’Aurelio Ranzato, Rajat Monga, Matthieu Devin, Kai Chen, Greg S. Corrado, Jeff Dean, and Andrew Y. Ng. They trained a nine-layer locally connected sparse autoencoder on ten million images, using model parallelism and asynchronous stochastic gradient descent across 1,000 machines with 16,000 cores for three days.[9]

The team reported learning a face detector without labels identifying which images contained faces; the network was also sensitive to cat faces and human bodies. In a subsequent classification experiment using the learned features, they reported 15.8% accuracy across 20,000 ImageNet object categories—a 70% relative improvement over the previous state of the art.[9]

With Tomas Mikolov, Le introduced Paragraph Vector in Distributed Representations of Sentences and Documents, submitted in May 2014. The unsupervised method learns fixed-length representations of variable-length sentences, paragraphs, and documents by training a dense document vector to predict words in that document. It addressed limitations of bag-of-words representations, which discard word order and ignore word semantics.[10]

In September 2014, Le, Ilya Sutskever, and Oriol Vinyals submitted Sequence to Sequence Learning with Neural Networks. Their encoder–decoder method uses a multilayer Long Short-Term Memory network to map an input sequence into a fixed-dimensional vector, then another deep LSTM to generate the target sequence. A consequential training choice was to reverse the words in source sentences while preserving target order. The authors reported that this introduced shorter dependencies between source and target, making optimization easier.[6]

For English-to-French translation on the entire WMT’14 test set, the authors reported BLEU 34.8 for the LSTM, with penalties for out-of-vocabulary words, compared with 33.3 for a phrase-based statistical translation system. A separate configuration used the LSTM to rerank that system’s 1,000 hypotheses and reached BLEU 36.5.[6]

Le and Andrew M. Dai’s Semi-supervised Sequence Learning, submitted in November 2015, explored two ways to pretrain recurrent networks on unlabeled data: predicting the next item in a sequence and reconstructing an input sequence through a sequence autoencoder. The learned parameters then initialized supervised models. The authors reported more stable, better-generalizing LSTMs in text-classification experiments including IMDB, DBpedia, and 20 Newsgroups.[7]

In an October 2023 interview, Jeremy Howard acknowledged that Dai and Le had already trained and fine-tuned a language model before his ULMFiT work. He distinguished their domain-specific corpus testing from his approach using a general-purpose corpus.[4]

With Barret Zoph, Le coauthored Neural Architecture Search with Reinforcement Learning, first submitted in November 2016. Their recurrent controller generated descriptions of neural networks and learned through reinforcement learning to maximize the generated architectures’ expected validation accuracy. The approach searched for both an image-classification architecture on CIFAR-10 and a recurrent cell for language modeling on Penn Treebank.[8]

Le also coauthored Chain-of-Thought Prompting Elicits Reasoning in Large Language Models, first submitted in January 2022, with Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, and Denny Zhou. The method supplies demonstrations containing intermediate reasoning steps as prompt exemplars. Across three evaluated large language models, the authors reported improvements on arithmetic, commonsense, and symbolic reasoning tasks; their concrete GSM8K example used a 540-billion-parameter model with eight exemplars.[11]

Le was one of Yi Tay’s two managers at Google. In a July 2024 interview, Tay recalled a friendship-like relationship built around brainstorming, including discussions of the U-PaLM paper. He credited Le’s research intuition and taste with influencing his own judgment about promising work, describing him as “very much a researcher more than like a corporate, you know, manager kind of thing.”[1]