What the three built together

The story begins in the 1980s, when deep learning — artificial neural networks with multiple layers — was considered a dead idea by the AI research community. Symbolic models dominated; neural networks were seen as computationally unviable and theoretically problematic.

Geoffrey Hinton, born in 1947 in the United Kingdom and based in Canada, was the most obstinate defender of neural networks during this winter. His work on backpropagation (with Rumelhart and Williams, 1986) and, later, on Boltzmann Machines and convolutional neural networks established the mathematical foundations that would make deep learning possible.

Yann LeCun, born in 1960 in France, was one of his postdoctoral students. LeCun developed convolutional neural networks (CNNs) — the architecture that transformed image recognition and is the basis of almost every computer vision system in use today. His work at Bell Labs in the 1990s produced the first practical handwritten digit recognition system, used by the US Postal Service.

Yoshua Bengio, born in 1966 in Canada, completed the trio. His work on neural language models (Bengio et al., 2003) was seminal to the development of modern transformers. Bengio is also known for fundamental contributions on distributed representations and curriculum learning.

The defining moment was 2012: Alex Krizhevsky, Hinton's student, used a CNN trained on GPUs to win ImageNet by an absurd margin. The entire field turned upside down. In 2018, the Turing Award recognized what was already evident: the three had changed the history of computing.

Hinton: regret and alarm

Geoffrey Hinton left Google in May 2023 specifically so he could speak freely about the risks of AI. Since then, he has become the most prominent voice among the pioneers to express deep — and personal — concern about what he helped create.

His central position: current AI systems may be smarter than humans in some aspects already today, and the pace of development is surpassing our capacity to understand what we are building. Hinton believes that the probability of AI representing an existential threat to humanity — through systems that develop their own goals contrary to humans, or through malicious actors who use AI to consolidate power — is significant, possibly 10-20% in the next 30 years.

In October 2024, Hinton received the Nobel Prize in Physics — an unusual distinction for a computer scientist, awarded for his work on Hopfield networks and Boltzmann machines as physical models of information. He used the Nobel platform to amplify his warning.

LeCun: the skeptical optimist

Yann LeCun, currently Chief AI Scientist at Meta AI, is the most vocal of the three in disagreeing with Hinton — and the mainstream concern with AI existential risks. For LeCun, current LLMs are fundamentally limited: they are statistical pattern completion machines, not systems that understand the world.

His central thesis: LLMs do not have models of the physical and causal world, and therefore will never reach general intelligence via this route. The path to genuine intelligence goes through world models — systems that learn how the world works by observing it, the way human babies learn before mastering language. The JEPA (Joint Embedding Predictive Architecture) architecture that Meta AI is developing is his concrete bet in this direction.

LeCun is also the most skeptical about short-term existential risks. For him, the concern with an AI that "develops its own goals and dominates humans" confuses science fiction with technical reality. Current systems do not have goals — they have loss functions. The difference is not semantic: it is architectural.

Bengio: the scientist turned activist

Yoshua Bengio occupies a middle and increasingly political territory. Like Hinton, he went from boundless optimism to a position of genuine concern about risks. Unlike Hinton, Bengio channels this concern primarily into political and regulatory action.

Bengio was one of the main signatories of the 2023 open letter calling for a pause in advanced AI development, and has since become an active advocate for international AI governance. He participated in the preparation of the 2024 UN report on AI and has been vocal about the need for an equivalent to the IPCC for AI risk.

His difference in relation to Hinton is one of emphasis: where Hinton focuses on technical risks and rogue AI scenarios, Bengio focuses on the risks of malicious use by human actors — authoritarian governments, corporations without accountability, extremist groups — who can use AI to consolidate power in historically unprecedented ways.

Why the divergences matter

The three positions are not just personal opinions of famous researchers — they structure the debate on how the world should react to AI development. Hinton and Bengio inform the case for restrictive regulation and development pauses. LeCun informs the case for accelerated development focused on architectures that are safer by design.

The three agree on one point: what is being built is deeply important and deeply misunderstood. The divergence is about what to do with it.