Yann LeCun

Developed CNNs and demonstrated their effectiveness on handwritten digit recognition (MNIST). For 20 years this was a niche result; after AlexNet (2012), CNNs became the dominant paradigm for all computer vision tasks. LeCun's patient persistence during the 'AI winter' is one of the great stories in AI history.

LeCun's most sustained public position (2022–present): autoregressive LLMs are fundamentally limited because they predict tokens rather than building internal models of how the world works. His proposed alternative — 'Joint Embedding Predictive Architecture' (JEPA) — learns world models in a self-supervised way without generating tokens. He argues this is closer to how animal intelligence actually works.

Argues forcefully (especially since Meta's LLaMA releases) that open AI is inherently more transparent, more auditable, and more democratically accountable than proprietary closed models. This directly contradicts the OpenAI/Anthropic position. His view: monocultures of AI are dangerous; diversity of open models creates resilience.

Maintains that LLMs are fundamentally 'autocomplete on steroids' — impressive interpolation machines that lack genuine understanding, causality, planning, or common sense. Regularly clashes with Sam Altman and others who claim near-AGI. His view: we need a completely new architecture before real machine intelligence is possible.

Publicly dismisses AI existential risk (as articulated by Hinton, Bengio, Yudkowsky) as 'science fiction.' His argument: AGI is further away than the doomers claim, and fear-mongering distracts from actual present-day harms like bias, misinformation, and labor displacement. This stance is one of the most controversial in the field.