Who Are the Godfathers of AI? Hinton, Bengio and LeCun Explained

The godfathers of AI are Geoffrey Hinton, Yoshua Bengio and Yann LeCun, the three computer scientists who shared the 2018 Turing Award for deep learning. Hinton popularised the training method behind almost every large AI model, Bengio’s language research fed into the transformer behind ChatGPT, and LeCun developed the convolutional network that taught computers to see.

In March 2019 the Association for Computing Machinery announced that its 2018 A.M. Turing Award, often called the Nobel Prize of computing, would go jointly to the three men for their work on deep learning. Deep learning trains artificial neural networks, software loosely modelled on the brain and built from many stacked layers of simple connected units, to find patterns in data. ACM’s announcement noted that their efforts to revive interest in neural networks were at first met with scepticism. The press often calls them the godfathers of AI.

The three were honoured together again in 2025, when the Queen Elizabeth Prize for Engineering went to seven pioneers of modern machine learning. That group also included John Hopfield, Fei-Fei Li, Bill Dally and Jensen Huang. Quantum Zeitgeist has profiled each of the three separately over the years, and this guide brings their stories together, with links to our fuller profiles and news coverage throughout.

The godfathers of AI at a glance

The three men share a prize and a field, but their careers, honours and views have diverged. The table sets them side by side.

Geoffrey Hinton popularised backpropagation and won the 2024 Nobel Prize in Physics

Hinton came to AI from psychology. He was born in Wimbledon, London, on 6 December 1947, and studied experimental psychology at King’s College, Cambridge, before completing a PhD in artificial intelligence at the University of Edinburgh in 1978. In 1986 he co-authored, with David Rumelhart and Ronald Williams, the paper that popularised backpropagation. The method works out how much each connection inside a neural network contributed to a wrong answer. It then adjusts every connection slightly to reduce the error. Almost every large AI model is still trained this way.

Earlier, in 1983, he and Terry Sejnowski invented the Boltzmann machine. This network borrows from statistical physics, the branch of physics that describes systems made of many interacting parts, to learn the hidden structure in its data.

Science runs through Hinton’s family. His Turing Award biography counts George Boole, the Victorian logician whose algebra of true and false underpins computer science, among his ancestors. His great-grandfather Charles Hinton coined the word “tesseract”, and his father, Howard Everest Hinton, was an entomologist. Hinton himself took a detour after Cambridge, working briefly as a carpenter before returning to academia in 1972 to study artificial intelligence at Edinburgh.

He kept inventing long after backpropagation. In July 2012 he and four Toronto co-authors described the technique now called dropout, which randomly switches off half of a network’s feature detectors on each training example. That sharply reduces overfitting, the habit of memorising training data instead of learning general rules, much as a football squad in which any player might miss match day ends up with nobody indispensable. In December 2022, aged 75, he posted the forward-forward algorithm, which replaces backpropagation’s backward pass with a second forward pass on negative data.

Hinton built his career at the University of Toronto. When Google bought his small start-up DNNresearch in 2013, he began dividing his time between the university and Google, where he became a vice president and engineering fellow. He resigned from Google in May 2023. He wanted to speak freely about the risks of the technology he had helped create.

He explained the decision in an interview with the BBC. He said some of the dangers of AI chatbots were “quite scary”, adding “Right now, they’re not more intelligent than us, as far as I can tell. But I think they soon may be.” He accepted that age was part of it too, saying “I’m 75, so it’s time to retire.” He was careful not to blame his employer, saying Google had been “very responsible”.

In October 2024 he shared the Nobel Prize in Physics with John Hopfield “for foundational discoveries and inventions that enable machine learning with artificial neural networks”. He has also warned of large-scale job losses and, in a 2024 BBC interview, called for treaties similar to the Geneva Conventions to regulate military AI. On 16 September 2026 he spoke at a closed briefing on AI risks for US senators, convened by Bernie Sanders, and said Congress had “maybe a year, but not much more than a year” to put guardrails around AI. Our full profile of Geoffrey Hinton covers his research and awards in more depth.

Yoshua Bengio’s language research fed the transformer, and he now runs the LawZero safety lab

Bengio was born in Paris on 5 March 1964 and has spent his academic career at the Université de Montréal, where he founded Mila, the Quebec AI institute. His group taught neural networks to model language. Its work on the attention mechanism fed directly into the transformer architecture behind today’s large language models. Attention lets a model decide which earlier words matter most when it interprets the next one. These are the systems, such as ChatGPT, that are trained on vast quantities of text to predict the next word.

His paper “A Neural Probabilistic Language Model” dates from 2000, and its full version appeared in the Journal of Machine Learning Research in 2003. It gave every word a learned list of numbers, so that a sentence the model had never seen could still be judged likely if it was built from words similar to familiar ones. The paper’s own example is that having seen “The cat is walking in the bedroom” should help a model treat “A dog was running in a room” as almost as likely. On two text corpora, the authors report, the approach significantly improved on state-of-the-art n-gram models, which generalise by stitching together very short sequences of words seen in training.

The attention work came in September 2014, when Bengio, Dzmitry Bahdanau and Kyunghyun Cho posted a translation paper that let the network search the source sentence for the words most relevant to each word it was about to write. Earlier systems squeezed a whole sentence into one fixed-length summary, as if a translator had to memorise a paragraph before writing a word. That search is the mechanism now called attention. Earlier that year, in June 2014, Bengio had co-authored, with his PhD student Ian Goodfellow as lead author, the paper that introduced generative adversarial networks, in which one network forges data and a second learns to spot the fakes.

In October 2025 he became the first living scientist to pass one million citations on Google Scholar. His more recent research has pushed toward causal reasoning, the ability of a model to understand why something happens as well as recognising that it does.

Bengio’s interest in the social effects of AI predates his 2023 change of course. He sat on the scientific team behind the Montréal Declaration for a Responsible Development of Artificial Intelligence, whose co-construction process the Université de Montréal launched on 3 November 2017. Over three months, 15 deliberation workshops involved more than 500 citizens, experts and stakeholders. The finished declaration sets out 10 principles, built on values that include well-being, autonomy, privacy, solidarity, democracy, equity and environmental sustainability.

In 2023 Bengio changed course. He decided to devote the rest of his career to reducing catastrophic risks from AI. In June 2025 he launched LawZero, a Canadian nonprofit that raised around $30 million in philanthropic funding. LawZero is building what Bengio calls Scientist AI. The system is designed to explain and predict the world without pursuing goals of its own. It could sit alongside AI agents, programs that take actions on a user’s behalf, and block any action it judges likely to cause harm.

He chairs the International AI Safety Report. He also co-chairs the UN’s new Independent International Scientific Panel on AI, whose 40 experts released their preliminary report on 1 July 2026. It found that current safeguards cannot keep pace with the growth of AI capabilities. Our profile of Yoshua Bengio tells the longer story.

The safety report Bengio chairs traces back to the first AI Safety Summit at Bletchley Park in November 2023, where the nations present agreed to support its development. The first full report was published on 29 January 2025, following an interim version in May 2024. The report itself says 96 AI experts contributed, including an expert advisory panel nominated by 30 countries, the OECD, the European Union and the United Nations. By design it summarises the evidence and does not recommend specific policies. A second edition followed on 3 February 2026.

Yann LeCun developed the convolutional network and has raised $1.03 billion for world models

LeCun was born near Paris in 1960 and joined Bell Labs in 1988. There he developed the convolutional neural network, a design that slides small learned filters across an image so that the network can recognise an edge or a shape wherever it appears in the picture. The approach became the standard tool of computer vision, as our profile of Yann LeCun explains in detail.

At Bell Labs the convolutional network was put to work reading the numbers written by hand on bank cheques. The system was licensed to suppliers of banking equipment such as National Cash Register, and cheque clearing suited it because the work was done in central offices with powerful computers. By LeCun’s own estimate, the technology was reading between 10 and 20 per cent of all cheques written in the United States in the early 2000s.

He joined New York University in 2003. In December 2013 he became the founding director of Facebook AI Research, later serving as Meta’s chief AI scientist, while keeping his professorship at NYU.

LeCun has argued for years that large language models will never reach human-level intelligence. Fluency is not enough. In his view text captures only a small and lossy slice of how the physical world works. He announced on 19 November 2025 that he would leave Meta after twelve years. In March 2026 his Paris-based start-up, Advanced Machine Intelligence Labs, closed a $1.03 billion seed round at a reported $3.5 billion pre-money valuation.

AMI Labs is building world models. These systems learn how the physical world behaves from video and other sensory data, so that they can predict and plan. They use an architecture called JEPA, short for Joint Embedding Predictive Architecture, which makes its predictions inside an abstract internal representation of a scene instead of generating pixels or words. Our complete guide to world models and LeCun’s bet against large language models sets out the technical case.

LeCun’s colleagues at Meta tested the approach before he left. I-JEPA, described in a January 2023 paper by Meta AI researchers including LeCun, learned about images by predicting the internal representation of hidden blocks of a picture, and its largest version trained in under 72 hours on 16 graphics processors. V-JEPA followed in February 2024 and applied the idea to 2 million videos. Predicting the gist of a hidden patch instead of every pixel is like knowing a missing jigsaw piece shows sky without having to paint the clouds.

2012 was the year deep learning won

The turning point came in the autumn of 2012. Hinton and his students Alex Krizhevsky and Ilya Sutskever entered a deep convolutional network, now known as AlexNet, in the ImageNet challenge, which asks software to sort photographs into 1,000 categories. Their paper reports a winning top-5 error rate of 15.3 per cent, against 26.2 per cent for the second-best entry. Put simply, the right answer was missing from the Toronto network’s five best guesses for about 15 pictures in every 100, against about 26 for the runner-up.

The network combined LeCun’s convolutional layers with Toronto’s own techniques, including dropout. It learned from 1.2 million labelled images in five to six days on two NVIDIA GTX 580 graphics processors, with the network split across both because a single card’s 3GB of memory was too small. The authors credit current GPUs, paired with a highly optimised implementation of convolution, with making networks of this size trainable at all.

Industry moved fast. In December 2012, during the NIPS conference at Lake Tahoe, Hinton ran an email auction from his hotel room for DNNresearch, the company he had formed with Krizhevsky and Sutskever. Four bidders took part, Baidu, Google, Microsoft and DeepMind. He stopped the bidding at $44 million and sold to Google, and a year later Facebook hired LeCun. Bengio stayed a full professor in Montreal and built up Mila, and by 2020 Google, Facebook, Microsoft and Samsung had all opened labs in the city.

A Toronto postdoc in 1987 linked all three careers

The three godfathers are usually presented as parallel pioneers who happened to share a prize. In fact their careers form a single chain of mentoring and collaboration that began in the mid-1980s. As a PhD student in Paris, LeCun had independently worked out a learning method close to backpropagation, and in 1985 Hinton introduced himself to LeCun during a visit to Paris. LeCun went on to spend 1987 to 1988 as a postdoctoral researcher in Hinton’s group at the University of Toronto.

Bengio, then a graduate student at McGill University in Montreal, met LeCun at neural network conferences and workshops. He finished his PhD at McGill in 1991 and a postdoc at MIT, then joined LeCun at Bell Labs as a postdoctoral fellow in 1992.

At Bell Labs Bengio worked with LeCun on handwriting recognition, combining neural networks with probabilistic models of sequences in a system AT&T used to read cheques. His neural language model, the line of research that later led to his work on attention, followed in 2000 after he moved to Montreal. In 1998 he and LeCun were co-authors, with Léon Bottou and Patrick Haffner, of the paper on gradient-based learning for document recognition that described the LeNet convolutional network.

Neural networks fell out of fashion in the 2000s. Through those lean years the Canadian Institute for Advanced Research kept the group together. Its programme on neural computation was directed by Hinton from 2004 to 2013, and LeCun and Bengio have co-directed it since 2014 under the name Learning in Machines and Brains. In 2013 LeCun and Bengio also co-founded the International Conference on Learning Representations, now one of the main venues for deep learning research. The current argument over AI risk is a disagreement among colleagues who have known each other for close to forty years.

Hinton’s Boltzmann machine is the godfathers’ one direct line into quantum computing

A Boltzmann machine gives every possible arrangement of its units an energy, in the same way physicists describe a magnet made of many tiny interacting spins. It learns by adjusting its connections until the patterns in its training data become the low-energy, most probable states.

In January 2016 researchers at D-Wave Systems and the University of Waterloo, including Roger Melko of the Perimeter Institute, proposed a quantum Boltzmann machine. It replaces the units with qubits. A qubit is the basic unit of quantum information. The authors suggested that quantum annealers, machines that settle a set of qubits into low-energy states, could be used to train it.

None of the three has made quantum computing a research focus, although Bengio and LeCun have both crossed paths with it. Bengio gave the keynote at a November 2023 workshop at the Perimeter Institute that launched a quantum and AI collaboration with IVADO, Mila and Institut Courtois. Melko leads Perimeter’s Quantum Intelligence Lab. LeCun has been the sceptic of the group. He questioned the future of quantum computing at the tenth anniversary of Meta’s research lab, arguing that classical computers can solve most problems more efficiently.

Hinton and Bengio want AI regulated, while LeCun calls the fear premature

A shared Turing Award has not produced a shared view of where AI is heading, and the argument between the three has become one of the most public in the field. Hinton and Bengio both believe that AI systems could overtake human abilities. They judge the risk of losing control serious enough to justify new laws. In 2024 both backed California’s SB 1047 AI safety bill, and Hinton said the state needed legislation with real teeth.

On 30 May 2023 Bengio signed, and Hinton backed, a one-sentence statement published by the Center for AI Safety, which reads “Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.” Reuters reported more than 350 signatories at launch, including the chief executives of OpenAI, DeepMind and Anthropic. LeCun did not sign, and Reuters reported that the Center for AI Safety singled out Meta, where he worked, for not signing.

Hinton has since attached a number to his fear. In December 2024 he told BBC Radio 4’s Today programme that there was a “10% to 20%” chance of AI leading to human extinction within the next three decades, up from the one in ten he had given before. “You see, we’ve never had to deal with things more intelligent than ourselves before,” he said.

LeCun regards those fears as premature. In a 2024 lecture setting out his vision for human-level AI he argued that open research and open-source AI must not be regulated out of existence.

LeCun has answered his old colleagues in public. When the extinction statement appeared, the BBC noted that he had dismissed such warnings on Twitter, now X, writing that “the most common reaction by AI researchers to these prophecies of doom is face palming”. Leaving Meta has not softened his view. In an interview with Fortune published on 1 October 2026 he said he was not worried “at all” about AI wiping out humanity, and called AI chiefs’ warnings that the technology could kill everyone “the worst marketing campaign you can possibly imagine.”

Much of the split comes down to a technical judgement about today’s language models. LeCun sees them as a dead end. So he finds little in them to fear. Hinton and Bengio watch the same models keep gaining abilities, and conclude that the time to put safeguards in place is now, before more capable systems arrive.