AI Thinkers

Hinton co-authored the 1986 Nature paper that made backpropagation widely known, and in 2006 showed how networks with many layers could be trained. His Toronto group built AlexNet in 2012, the result that turned computer vision towards deep learning.

He left Google in May 2023 after about a decade there, saying he wanted to talk about the risks of AI without worrying about how it affected Google's business.

Honours Turing Award 2018 (with Bengio and LeCun) · Nobel Prize in Physics 2024 (with John Hopfield)

Also sourced from Nobel Prize press release, 2024 · AP on leaving Google, May 2023

In the History of AI Backpropagation, 1986 · Deep belief networks, 2006 · AlexNet, 2012

Bengio's 2003 paper on a neural probabilistic language model introduced the idea of learning word representations inside a network that predicts text, an ancestor of today's language models. He founded the Mila institute in Montréal.

In October 2025 Mila reported that he had become the first living scientist to pass one million citations on Google Scholar. In June 2025 he launched LawZero, a nonprofit lab building AI systems designed to be safe.

Honours Turing Award 2018 (with Hinton and LeCun)

Also sourced from Mila, one million citations, October 2025 · SiliconANGLE on LawZero, June 2025

In the History of AI Deep belief networks, 2006

LeCun developed convolutional neural networks, which learn to recognise patterns in images, and used them to read handwritten digits on cheques. The design underlies most computer vision systems built since.

He was the founding director of Meta's FAIR research lab and later its chief AI scientist. In November 2025 he confirmed he was leaving Meta to start a company built on his “advanced machine intelligence” research agenda: systems that understand the physical world, remember and plan.

Honours Turing Award 2018 (with Hinton and Bengio)

Also sourced from The Decoder on leaving Meta, November 2025

In the History of AI AlexNet, 2012

With Sepp Hochreiter, Schmidhuber co-authored the 1997 paper introducing long short-term memory, a kind of recurrent network that can carry information across long sequences. LSTMs powered speech recognition and machine translation for much of the 2010s.

He has argued publicly that the field under-credits earlier work, including his own lab's, which makes him one of its most outspoken voices on the history of AI.

As a student in 1991, Hochreiter analysed why deep and recurrent networks struggled to learn: error signals shrink as they pass back through many steps, the vanishing gradient problem. His answer, developed with Jürgen Schmidhuber, was long short-term memory, published in 1997.

LSTM became one of the most widely used neural network designs before Transformers took over.

Sutskever co-authored AlexNet in 2012 and, in 2014, sequence-to-sequence learning, which showed that a neural network could translate whole sentences. He co-founded OpenAI and was its chief scientist through the GPT era.

He left OpenAI in 2024 and launched Safe Superintelligence (SSI). In July 2025 he became its chief executive after his co-founder Daniel Gross left.

Also sourced from TechCrunch on leading SSI, July 2025

In the History of AI AlexNet, 2012 · GPT-3, 2020

He was first author of the Microsoft Research paper on deep residual learning, which added shortcut connections so that networks with a hundred or more layers could be trained. ResNet won the 2015 ImageNet competition, and residual connections are now standard inside Transformers too.

In April 2025 a Nature analysis found the ResNet paper to be the most-cited paper of the twenty-first century.

Also sourced from Nature, most-cited papers of the 21st century, April 2025

Vaswani is the first author of “Attention Is All You Need”, the 2017 paper by eight researchers that introduced the Transformer. Almost every large language model since, from GPT to Claude and Gemini, is built on that design.

The same Nature analysis in April 2025 placed the paper among the ten most-cited papers of the twenty-first century.

Also sourced from Nature, most-cited papers of the 21st century, April 2025

In the History of AI Attention Is All You Need, 2017

In 2014 Goodfellow and colleagues introduced generative adversarial networks, in which one network generates samples and another tries to spot the fakes. GANs produced the first convincingly realistic synthetic faces.

He also co-wrote, with Yoshua Bengio and Aaron Courville, the textbook Deep Learning (2016), which is free to read online.

In the History of AI GANs, 2014

Kingma co-created two tools found in a vast amount of machine learning code: the variational autoencoder (2013, with Max Welling), a way to train generative models, and Adam (2014, with Jimmy Ba), the optimiser used to train a large share of neural networks.

He was part of OpenAI's founding team, later worked at Google, and joined Anthropic in October 2024.

Also sourced from TechCrunch on joining Anthropic, October 2024

Sutton, with Andrew Barto, laid the foundations of reinforcement learning, including temporal-difference learning, and co-wrote the field's standard textbook. ACM lists him as a professor at the University of Alberta and a research scientist at Keen Technologies.

His 2019 essay “The Bitter Lesson” argued that, over the long run, AI progress has come from general methods that scale with computing power rather than from building in human knowledge. It has become a touchstone for the scaling debate.

Honours Turing Award 2024 (with Barto)

Also sourced from ACM, 2024 Turing Award

Barto, a professor emeritus at the University of Massachusetts Amherst, developed the core ideas of reinforcement learning with Richard Sutton from the late 1970s onwards, drawing on psychology and neuroscience as well as computing.

Their textbook trained generations of researchers, and the methods behind it sit inside systems from AlphaGo to the fine-tuning of today's chatbots.

Honours Turing Award 2024 (with Sutton)

Also sourced from ACM, 2024 Turing Award

Silver led the research behind AlphaGo, which beat Lee Sedol at Go in 2016, and AlphaGo Zero and AlphaZero, which learned Go, chess and shogi entirely by playing against themselves.

The work showed that reinforcement learning combined with deep networks and search could reach superhuman play without human examples.

In the History of AI AlphaGo, 2016

Li led the creation of ImageNet, a database of millions of labelled images presented in 2009. The annual ImageNet challenge became the benchmark on which deep learning proved itself in 2012.

She is a co-founder of World Labs, which describes itself as a spatial intelligence company building models that perceive, generate and interact with 3D worlds.

Also sourced from World Labs, about

In the History of AI AlexNet, 2012

Ng's group at Stanford showed in 2009 that graphics processors could speed up the training of large neural networks dramatically, an idea that became central to modern AI.

He co-founded Coursera with Daphne Koller in 2012, and his online machine learning courses introduced the field to millions of learners.

Also sourced from Coursera, about

Koller's work on probabilistic graphical models, set out in the textbook she wrote with Nir Friedman, gave machine learning a rigorous way to represent uncertainty across many related variables.

She co-founded Coursera with Andrew Ng in 2012, and has since applied machine learning to biology and drug discovery.

Also sourced from Coursera, about

Pearl developed Bayesian networks in the 1980s, giving computers a practical way to reason with probabilities, and then built a mathematical framework for causation: how to tell whether one thing causes another, not just whether they occur together.

His ideas are now used across medicine, economics and the social sciences, and he wrote The Book of Why (2018) for general readers.

Honours Turing Award 2011

Also sourced from Wikipedia, Turing Award recipients

Jordan helped bring statistics and machine learning together. His work spans graphical models, variational inference and Bayesian methods, and his 2003 paper with David Blei and Andrew Ng introduced latent Dirichlet allocation, a widely used method for finding topics in large collections of text.

Many of today's leading machine learning researchers trained in his group at the University of California, Berkeley.

Russell co-wrote, with Peter Norvig, Artificial Intelligence: A Modern Approach, the textbook used in AI courses around the world. He is a professor at the University of California, Berkeley.

His 2019 book Human Compatible argues that AI systems should be built to be uncertain about what humans want, so that they defer to us. That approach is the focus of the Center for Human-Compatible AI at Berkeley.

Also sourced from Center for Human-Compatible AI

In 2018 Gebru and Joy Buolamwini published Gender Shades, which found that commercial face-analysis systems were far less accurate for darker-skinned women than for lighter-skinned men. She co-authored the 2021 paper “On the Dangers of Stochastic Parrots”, an early critique of ever-larger language models.

She left Google in December 2020 and founded the Distributed AI Research Institute (DAIR) in December 2021, which studies the harms of AI systems and how to reduce them.

Also sourced from DAIR, Distributed AI Research Institute